drivers/cpufreq/powernow-k8.c

   1 /*
   2  *   (c) 2003-2010 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 on behalf of SuSE Labs
  11  *  (C) 2004 Dominik Brodowski <linux@brodo.de>
  12  *  (C) 2004 Pavel Machek <pavel@ucw.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, Jacob Shin, and others.
  18  *  Originally developed by Paul Devriendt.
  19  *  Processor information obtained from Chapter 9 (Power and Thermal Management)
  20  *  of the "BIOS and Kernel Developer's Guide for the AMD Athlon 64 and AMD
  21  *  Opteron Processors" available for download from www.amd.com
  22  *
  23  *  Tables for specific CPUs can be inferred from
  24  *     http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/30430.pdf
  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 #include <linux/sched.h>        /* for current / set_cpus_allowed() */
  36 #include <linux/io.h>
  37 #include <linux/delay.h>
  38
  39 #include <asm/msr.h>
  40
  41 #include <linux/acpi.h>
  42 #include <linux/mutex.h>
  43 #include <acpi/processor.h>
  44
  45 #define PFX "powernow-k8: "
  46 #define VERSION "version 2.20.00"
  47 #include "powernow-k8.h"
  48 #include "mperf.h"
  49
  50 /* serialize freq changes  */
  51 static DEFINE_MUTEX(fidvid_mutex);
  52
  53 static DEFINE_PER_CPU(struct powernow_k8_data *, powernow_data);
  54
  55 static int cpu_family = CPU_OPTERON;
  56
  57 /* core performance boost */
  58 static bool cpb_capable, cpb_enabled;
  59 static struct msr __percpu *msrs;
  60
  61 static struct cpufreq_driver cpufreq_amd64_driver;
  62
  63 #ifndef CONFIG_SMP
  64 static inline const struct cpumask *cpu_core_mask(int cpu)
  65 {
  66         return cpumask_of(0);
  67 }
  68 #endif
  69
  70 /* Return a frequency in MHz, given an input fid */
  71 static u32 find_freq_from_fid(u32 fid)
  72 {
  73         return 800 + (fid * 100);
  74 }
  75
  76 /* Return a frequency in KHz, given an input fid */
  77 static u32 find_khz_freq_from_fid(u32 fid)
  78 {
  79         return 1000 * find_freq_from_fid(fid);
  80 }
  81
  82 static u32 find_khz_freq_from_pstate(struct cpufreq_frequency_table *data,
  83                 u32 pstate)
  84 {
  85         return data[pstate].frequency;
  86 }
  87
  88 /* Return the vco fid for an input fid
  89  *
  90  * Each "low" fid has corresponding "high" fid, and you can get to "low" fids
  91  * only from corresponding high fids. This returns "high" fid corresponding to
  92  * "low" one.
  93  */
  94 static u32 convert_fid_to_vco_fid(u32 fid)
  95 {
  96         if (fid < HI_FID_TABLE_BOTTOM)
  97                 return 8 + (2 * fid);
  98         else
  99                 return fid;
 100 }
 101
 102 /*
 103  * Return 1 if the pending bit is set. Unless we just instructed the processor
 104  * to transition to a new state, seeing this bit set is really bad news.
 105  */
 106 static int pending_bit_stuck(void)
 107 {
 108         u32 lo, hi;
 109
 110         if (cpu_family == CPU_HW_PSTATE)
 111                 return 0;
 112
 113         rdmsr(MSR_FIDVID_STATUS, lo, hi);
 114         return lo & MSR_S_LO_CHANGE_PENDING ? 1 : 0;
 115 }
 116
 117 /*
 118  * Update the global current fid / vid values from the status msr.
 119  * Returns 1 on error.
 120  */
 121 static int query_current_values_with_pending_wait(struct powernow_k8_data *data)
 122 {
 123         u32 lo, hi;
 124         u32 i = 0;
 125
 126         if (cpu_family == CPU_HW_PSTATE) {
 127                 rdmsr(MSR_PSTATE_STATUS, lo, hi);
 128                 i = lo & HW_PSTATE_MASK;
 129                 data->currpstate = i;
 130
 131                 /*
 132                  * a workaround for family 11h erratum 311 might cause
 133                  * an "out-of-range Pstate if the core is in Pstate-0
 134                  */
 135                 if ((boot_cpu_data.x86 == 0x11) && (i >= data->numps))
 136                         data->currpstate = HW_PSTATE_0;
 137
 138                 return 0;
 139         }
 140         do {
 141                 if (i++ > 10000) {
 142                         pr_debug("detected change pending stuck\n");
 143                         return 1;
 144                 }
 145                 rdmsr(MSR_FIDVID_STATUS, lo, hi);
 146         } while (lo & MSR_S_LO_CHANGE_PENDING);
 147
 148         data->currvid = hi & MSR_S_HI_CURRENT_VID;
 149         data->currfid = lo & MSR_S_LO_CURRENT_FID;
 150
 151         return 0;
 152 }
 153
 154 /* the isochronous relief time */
 155 static void count_off_irt(struct powernow_k8_data *data)
 156 {
 157         udelay((1 << data->irt) * 10);
 158         return;
 159 }
 160
 161 /* the voltage stabilization time */
 162 static void count_off_vst(struct powernow_k8_data *data)
 163 {
 164         udelay(data->vstable * VST_UNITS_20US);
 165         return;
 166 }
 167
 168 /* need to init the control msr to a safe value (for each cpu) */
 169 static void fidvid_msr_init(void)
 170 {
 171         u32 lo, hi;
 172         u8 fid, vid;
 173
 174         rdmsr(MSR_FIDVID_STATUS, lo, hi);
 175         vid = hi & MSR_S_HI_CURRENT_VID;
 176         fid = lo & MSR_S_LO_CURRENT_FID;
 177         lo = fid | (vid << MSR_C_LO_VID_SHIFT);
 178         hi = MSR_C_HI_STP_GNT_BENIGN;
 179         pr_debug("cpu%d, init lo 0x%x, hi 0x%x\n", smp_processor_id(), lo, hi);
 180         wrmsr(MSR_FIDVID_CTL, lo, hi);
 181 }
 182
 183 /* write the new fid value along with the other control fields to the msr */
 184 static int write_new_fid(struct powernow_k8_data *data, u32 fid)
 185 {
 186         u32 lo;
 187         u32 savevid = data->currvid;
 188         u32 i = 0;
 189
 190         if ((fid & INVALID_FID_MASK) || (data->currvid & INVALID_VID_MASK)) {
 191                 printk(KERN_ERR PFX "internal error - overflow on fid write\n");
 192                 return 1;
 193         }
 194
 195         lo = fid;
 196         lo |= (data->currvid << MSR_C_LO_VID_SHIFT);
 197         lo |= MSR_C_LO_INIT_FID_VID;
 198
 199         pr_debug("writing fid 0x%x, lo 0x%x, hi 0x%x\n",
 200                 fid, lo, data->plllock * PLL_LOCK_CONVERSION);
 201
 202         do {
 203                 wrmsr(MSR_FIDVID_CTL, lo, data->plllock * PLL_LOCK_CONVERSION);
 204                 if (i++ > 100) {
 205                         printk(KERN_ERR PFX
 206                                 "Hardware error - pending bit very stuck - "
 207                                 "no further pstate changes possible\n");
 208                         return 1;
 209                 }
 210         } while (query_current_values_with_pending_wait(data));
 211
 212         count_off_irt(data);
 213
 214         if (savevid != data->currvid) {
 215                 printk(KERN_ERR PFX
 216                         "vid change on fid trans, old 0x%x, new 0x%x\n",
 217                         savevid, data->currvid);
 218                 return 1;
 219         }
 220
 221         if (fid != data->currfid) {
 222                 printk(KERN_ERR PFX
 223                         "fid trans failed, fid 0x%x, curr 0x%x\n", fid,
 224                         data->currfid);
 225                 return 1;
 226         }
 227
 228         return 0;
 229 }
 230
 231 /* Write a new vid to the hardware */
 232 static int write_new_vid(struct powernow_k8_data *data, u32 vid)
 233 {
 234         u32 lo;
 235         u32 savefid = data->currfid;
 236         int i = 0;
 237
 238         if ((data->currfid & INVALID_FID_MASK) || (vid & INVALID_VID_MASK)) {
 239                 printk(KERN_ERR PFX "internal error - overflow on vid write\n");
 240                 return 1;
 241         }
 242
 243         lo = data->currfid;
 244         lo |= (vid << MSR_C_LO_VID_SHIFT);
 245         lo |= MSR_C_LO_INIT_FID_VID;
 246
 247         pr_debug("writing vid 0x%x, lo 0x%x, hi 0x%x\n",
 248                 vid, lo, STOP_GRANT_5NS);
 249
 250         do {
 251                 wrmsr(MSR_FIDVID_CTL, lo, STOP_GRANT_5NS);
 252                 if (i++ > 100) {
 253                         printk(KERN_ERR PFX "internal error - pending bit "
 254                                         "very stuck - no further pstate "
 255                                         "changes possible\n");
 256                         return 1;
 257                 }
 258         } while (query_current_values_with_pending_wait(data));
 259
 260         if (savefid != data->currfid) {
 261                 printk(KERN_ERR PFX "fid changed on vid trans, old "
 262                         "0x%x new 0x%x\n",
 263                        savefid, data->currfid);
 264                 return 1;
 265         }
 266
 267         if (vid != data->currvid) {
 268                 printk(KERN_ERR PFX "vid trans failed, vid 0x%x, "
 269                                 "curr 0x%x\n",
 270                                 vid, data->currvid);
 271                 return 1;
 272         }
 273
 274         return 0;
 275 }
 276
 277 /*
 278  * Reduce the vid by the max of step or reqvid.
 279  * Decreasing vid codes represent increasing voltages:
 280  * vid of 0 is 1.550V, vid of 0x1e is 0.800V, vid of VID_OFF is off.
 281  */
 282 static int decrease_vid_code_by_step(struct powernow_k8_data *data,
 283                 u32 reqvid, u32 step)
 284 {
 285         if ((data->currvid - reqvid) > step)
 286                 reqvid = data->currvid - step;
 287
 288         if (write_new_vid(data, reqvid))
 289                 return 1;
 290
 291         count_off_vst(data);
 292
 293         return 0;
 294 }
 295
 296 /* Change hardware pstate by single MSR write */
 297 static int transition_pstate(struct powernow_k8_data *data, u32 pstate)
 298 {
 299         wrmsr(MSR_PSTATE_CTRL, pstate, 0);
 300         data->currpstate = pstate;
 301         return 0;
 302 }
 303
 304 /* Change Opteron/Athlon64 fid and vid, by the 3 phases. */
 305 static int transition_fid_vid(struct powernow_k8_data *data,
 306                 u32 reqfid, u32 reqvid)
 307 {
 308         if (core_voltage_pre_transition(data, reqvid, reqfid))
 309                 return 1;
 310
 311         if (core_frequency_transition(data, reqfid))
 312                 return 1;
 313
 314         if (core_voltage_post_transition(data, reqvid))
 315                 return 1;
 316
 317         if (query_current_values_with_pending_wait(data))
 318                 return 1;
 319
 320         if ((reqfid != data->currfid) || (reqvid != data->currvid)) {
 321                 printk(KERN_ERR PFX "failed (cpu%d): req 0x%x 0x%x, "
 322                                 "curr 0x%x 0x%x\n",
 323                                 smp_processor_id(),
 324                                 reqfid, reqvid, data->currfid, data->currvid);
 325                 return 1;
 326         }
 327
 328         pr_debug("transitioned (cpu%d): new fid 0x%x, vid 0x%x\n",
 329                 smp_processor_id(), data->currfid, data->currvid);
 330
 331         return 0;
 332 }
 333
 334 /* Phase 1 - core voltage transition ... setup voltage */
 335 static int core_voltage_pre_transition(struct powernow_k8_data *data,
 336                 u32 reqvid, u32 reqfid)
 337 {
 338         u32 rvosteps = data->rvo;
 339         u32 savefid = data->currfid;
 340         u32 maxvid, lo, rvomult = 1;
 341
 342         pr_debug("ph1 (cpu%d): start, currfid 0x%x, currvid 0x%x, "
 343                 "reqvid 0x%x, rvo 0x%x\n",
 344                 smp_processor_id(),
 345                 data->currfid, data->currvid, reqvid, data->rvo);
 346
 347         if ((savefid < LO_FID_TABLE_TOP) && (reqfid < LO_FID_TABLE_TOP))
 348                 rvomult = 2;
 349         rvosteps *= rvomult;
 350         rdmsr(MSR_FIDVID_STATUS, lo, maxvid);
 351         maxvid = 0x1f & (maxvid >> 16);
 352         pr_debug("ph1 maxvid=0x%x\n", maxvid);
 353         if (reqvid < maxvid) /* lower numbers are higher voltages */
 354                 reqvid = maxvid;
 355
 356         while (data->currvid > reqvid) {
 357                 pr_debug("ph1: curr 0x%x, req vid 0x%x\n",
 358                         data->currvid, reqvid);
 359                 if (decrease_vid_code_by_step(data, reqvid, data->vidmvs))
 360                         return 1;
 361         }
 362
 363         while ((rvosteps > 0) &&
 364                         ((rvomult * data->rvo + data->currvid) > reqvid)) {
 365                 if (data->currvid == maxvid) {
 366                         rvosteps = 0;
 367                 } else {
 368                         pr_debug("ph1: changing vid for rvo, req 0x%x\n",
 369                                 data->currvid - 1);
 370                         if (decrease_vid_code_by_step(data, data->currvid-1, 1))
 371                                 return 1;
 372                         rvosteps--;
 373                 }
 374         }
 375
 376         if (query_current_values_with_pending_wait(data))
 377                 return 1;
 378
 379         if (savefid != data->currfid) {
 380                 printk(KERN_ERR PFX "ph1 err, currfid changed 0x%x\n",
 381                                 data->currfid);
 382                 return 1;
 383         }
 384
 385         pr_debug("ph1 complete, currfid 0x%x, currvid 0x%x\n",
 386                 data->currfid, data->currvid);
 387
 388         return 0;
 389 }
 390
 391 /* Phase 2 - core frequency transition */
 392 static int core_frequency_transition(struct powernow_k8_data *data, u32 reqfid)
 393 {
 394         u32 vcoreqfid, vcocurrfid, vcofiddiff;
 395         u32 fid_interval, savevid = data->currvid;
 396
 397         if (data->currfid == reqfid) {
 398                 printk(KERN_ERR PFX "ph2 null fid transition 0x%x\n",
 399                                 data->currfid);
 400                 return 0;
 401         }
 402
 403         pr_debug("ph2 (cpu%d): starting, currfid 0x%x, currvid 0x%x, "
 404                 "reqfid 0x%x\n",
 405                 smp_processor_id(),
 406                 data->currfid, data->currvid, reqfid);
 407
 408         vcoreqfid = convert_fid_to_vco_fid(reqfid);
 409         vcocurrfid = convert_fid_to_vco_fid(data->currfid);
 410         vcofiddiff = vcocurrfid > vcoreqfid ? vcocurrfid - vcoreqfid
 411             : vcoreqfid - vcocurrfid;
 412
 413         if ((reqfid <= LO_FID_TABLE_TOP) && (data->currfid <= LO_FID_TABLE_TOP))
 414                 vcofiddiff = 0;
 415
 416         while (vcofiddiff > 2) {
 417                 (data->currfid & 1) ? (fid_interval = 1) : (fid_interval = 2);
 418
 419                 if (reqfid > data->currfid) {
 420                         if (data->currfid > LO_FID_TABLE_TOP) {
 421                                 if (write_new_fid(data,
 422                                                 data->currfid + fid_interval))
 423                                         return 1;
 424                         } else {
 425                                 if (write_new_fid
 426                                     (data,
 427                                      2 + convert_fid_to_vco_fid(data->currfid)))
 428                                         return 1;
 429                         }
 430                 } else {
 431                         if (write_new_fid(data, data->currfid - fid_interval))
 432                                 return 1;
 433                 }
 434
 435                 vcocurrfid = convert_fid_to_vco_fid(data->currfid);
 436                 vcofiddiff = vcocurrfid > vcoreqfid ? vcocurrfid - vcoreqfid
 437                     : vcoreqfid - vcocurrfid;
 438         }
 439
 440         if (write_new_fid(data, reqfid))
 441                 return 1;
 442
 443         if (query_current_values_with_pending_wait(data))
 444                 return 1;
 445
 446         if (data->currfid != reqfid) {
 447                 printk(KERN_ERR PFX
 448                         "ph2: mismatch, failed fid transition, "
 449                         "curr 0x%x, req 0x%x\n",
 450                         data->currfid, reqfid);
 451                 return 1;
 452         }
 453
 454         if (savevid != data->currvid) {
 455                 printk(KERN_ERR PFX "ph2: vid changed, save 0x%x, curr 0x%x\n",
 456                         savevid, data->currvid);
 457                 return 1;
 458         }
 459
 460         pr_debug("ph2 complete, currfid 0x%x, currvid 0x%x\n",
 461                 data->currfid, data->currvid);
 462
 463         return 0;
 464 }
 465
 466 /* Phase 3 - core voltage transition flow ... jump to the final vid. */
 467 static int core_voltage_post_transition(struct powernow_k8_data *data,
 468                 u32 reqvid)
 469 {
 470         u32 savefid = data->currfid;
 471         u32 savereqvid = reqvid;
 472
 473         pr_debug("ph3 (cpu%d): starting, currfid 0x%x, currvid 0x%x\n",
 474                 smp_processor_id(),
 475                 data->currfid, data->currvid);
 476
 477         if (reqvid != data->currvid) {
 478                 if (write_new_vid(data, reqvid))
 479                         return 1;
 480
 481                 if (savefid != data->currfid) {
 482                         printk(KERN_ERR PFX
 483                                "ph3: bad fid change, save 0x%x, curr 0x%x\n",
 484                                savefid, data->currfid);
 485                         return 1;
 486                 }
 487
 488                 if (data->currvid != reqvid) {
 489                         printk(KERN_ERR PFX
 490                                "ph3: failed vid transition\n, "
 491                                "req 0x%x, curr 0x%x",
 492                                reqvid, data->currvid);
 493                         return 1;
 494                 }
 495         }
 496
 497         if (query_current_values_with_pending_wait(data))
 498                 return 1;
 499
 500         if (savereqvid != data->currvid) {
 501                 pr_debug("ph3 failed, currvid 0x%x\n", data->currvid);
 502                 return 1;
 503         }
 504
 505         if (savefid != data->currfid) {
 506                 pr_debug("ph3 failed, currfid changed 0x%x\n",
 507                         data->currfid);
 508                 return 1;
 509         }
 510
 511         pr_debug("ph3 complete, currfid 0x%x, currvid 0x%x\n",
 512                 data->currfid, data->currvid);
 513
 514         return 0;
 515 }
 516
 517 static void check_supported_cpu(void *_rc)
 518 {
 519         u32 eax, ebx, ecx, edx;
 520         int *rc = _rc;
 521
 522         *rc = -ENODEV;
 523
 524         if (__this_cpu_read(cpu_info.x86_vendor) != X86_VENDOR_AMD)
 525                 return;
 526
 527         eax = cpuid_eax(CPUID_PROCESSOR_SIGNATURE);
 528         if (((eax & CPUID_XFAM) != CPUID_XFAM_K8) &&
 529             ((eax & CPUID_XFAM) < CPUID_XFAM_10H))
 530                 return;
 531
 532         if ((eax & CPUID_XFAM) == CPUID_XFAM_K8) {
 533                 if (((eax & CPUID_USE_XFAM_XMOD) != CPUID_USE_XFAM_XMOD) ||
 534                     ((eax & CPUID_XMOD) > CPUID_XMOD_REV_MASK)) {
 535                         printk(KERN_INFO PFX
 536                                 "Processor cpuid %x not supported\n", eax);
 537                         return;
 538                 }
 539
 540                 eax = cpuid_eax(CPUID_GET_MAX_CAPABILITIES);
 541                 if (eax < CPUID_FREQ_VOLT_CAPABILITIES) {
 542                         printk(KERN_INFO PFX
 543                                "No frequency change capabilities detected\n");
 544                         return;
 545                 }
 546
 547                 cpuid(CPUID_FREQ_VOLT_CAPABILITIES, &eax, &ebx, &ecx, &edx);
 548                 if ((edx & P_STATE_TRANSITION_CAPABLE)
 549                         != P_STATE_TRANSITION_CAPABLE) {
 550                         printk(KERN_INFO PFX
 551                                 "Power state transitions not supported\n");
 552                         return;
 553                 }
 554         } else { /* must be a HW Pstate capable processor */
 555                 cpuid(CPUID_FREQ_VOLT_CAPABILITIES, &eax, &ebx, &ecx, &edx);
 556                 if ((edx & USE_HW_PSTATE) == USE_HW_PSTATE)
 557                         cpu_family = CPU_HW_PSTATE;
 558                 else
 559                         return;
 560         }
 561
 562         *rc = 0;
 563 }
 564
 565 static int check_pst_table(struct powernow_k8_data *data, struct pst_s *pst,
 566                 u8 maxvid)
 567 {
 568         unsigned int j;
 569         u8 lastfid = 0xff;
 570
 571         for (j = 0; j < data->numps; j++) {
 572                 if (pst[j].vid > LEAST_VID) {
 573                         printk(KERN_ERR FW_BUG PFX "vid %d invalid : 0x%x\n",
 574                                j, pst[j].vid);
 575                         return -EINVAL;
 576                 }
 577                 if (pst[j].vid < data->rvo) {
 578                         /* vid + rvo >= 0 */
 579                         printk(KERN_ERR FW_BUG PFX "0 vid exceeded with pstate"
 580                                " %d\n", j);
 581                         return -ENODEV;
 582                 }
 583                 if (pst[j].vid < maxvid + data->rvo) {
 584                         /* vid + rvo >= maxvid */
 585                         printk(KERN_ERR FW_BUG PFX "maxvid exceeded with pstate"
 586                                " %d\n", j);
 587                         return -ENODEV;
 588                 }
 589                 if (pst[j].fid > MAX_FID) {
 590                         printk(KERN_ERR FW_BUG PFX "maxfid exceeded with pstate"
 591                                " %d\n", j);
 592                         return -ENODEV;
 593                 }
 594                 if (j && (pst[j].fid < HI_FID_TABLE_BOTTOM)) {
 595                         /* Only first fid is allowed to be in "low" range */
 596                         printk(KERN_ERR FW_BUG PFX "two low fids - %d : "
 597                                "0x%x\n", j, pst[j].fid);
 598                         return -EINVAL;
 599                 }
 600                 if (pst[j].fid < lastfid)
 601                         lastfid = pst[j].fid;
 602         }
 603         if (lastfid & 1) {
 604                 printk(KERN_ERR FW_BUG PFX "lastfid invalid\n");
 605                 return -EINVAL;
 606         }
 607         if (lastfid > LO_FID_TABLE_TOP)
 608                 printk(KERN_INFO FW_BUG PFX
 609                         "first fid not from lo freq table\n");
 610
 611         return 0;
 612 }
 613
 614 static void invalidate_entry(struct cpufreq_frequency_table *powernow_table,
 615                 unsigned int entry)
 616 {
 617         powernow_table[entry].frequency = CPUFREQ_ENTRY_INVALID;
 618 }
 619
 620 static void print_basics(struct powernow_k8_data *data)
 621 {
 622         int j;
 623         for (j = 0; j < data->numps; j++) {
 624                 if (data->powernow_table[j].frequency !=
 625                                 CPUFREQ_ENTRY_INVALID) {
 626                         if (cpu_family == CPU_HW_PSTATE) {
 627                                 printk(KERN_INFO PFX
 628                                         "   %d : pstate %d (%d MHz)\n", j,
 629                                         data->powernow_table[j].index,
 630                                         data->powernow_table[j].frequency/1000);
 631                         } else {
 632                                 printk(KERN_INFO PFX
 633                                         "fid 0x%x (%d MHz), vid 0x%x\n",
 634                                         data->powernow_table[j].index & 0xff,
 635                                         data->powernow_table[j].frequency/1000,
 636                                         data->powernow_table[j].index >> 8);
 637                         }
 638                 }
 639         }
 640         if (data->batps)
 641                 printk(KERN_INFO PFX "Only %d pstates on battery\n",
 642                                 data->batps);
 643 }
 644
 645 static u32 freq_from_fid_did(u32 fid, u32 did)
 646 {
 647         u32 mhz = 0;
 648
 649         if (boot_cpu_data.x86 == 0x10)
 650                 mhz = (100 * (fid + 0x10)) >> did;
 651         else if (boot_cpu_data.x86 == 0x11)
 652                 mhz = (100 * (fid + 8)) >> did;
 653         else
 654                 BUG();
 655
 656         return mhz * 1000;
 657 }
 658
 659 static int fill_powernow_table(struct powernow_k8_data *data,
 660                 struct pst_s *pst, u8 maxvid)
 661 {
 662         struct cpufreq_frequency_table *powernow_table;
 663         unsigned int j;
 664
 665         if (data->batps) {
 666                 /* use ACPI support to get full speed on mains power */
 667                 printk(KERN_WARNING PFX
 668                         "Only %d pstates usable (use ACPI driver for full "
 669                         "range\n", data->batps);
 670                 data->numps = data->batps;
 671         }
 672
 673         for (j = 1; j < data->numps; j++) {
 674                 if (pst[j-1].fid >= pst[j].fid) {
 675                         printk(KERN_ERR PFX "PST out of sequence\n");
 676                         return -EINVAL;
 677                 }
 678         }
 679
 680         if (data->numps < 2) {
 681                 printk(KERN_ERR PFX "no p states to transition\n");
 682                 return -ENODEV;
 683         }
 684
 685         if (check_pst_table(data, pst, maxvid))
 686                 return -EINVAL;
 687
 688         powernow_table = kmalloc((sizeof(struct cpufreq_frequency_table)
 689                 * (data->numps + 1)), GFP_KERNEL);
 690         if (!powernow_table) {
 691                 printk(KERN_ERR PFX "powernow_table memory alloc failure\n");
 692                 return -ENOMEM;
 693         }
 694
 695         for (j = 0; j < data->numps; j++) {
 696                 int freq;
 697                 powernow_table[j].index = pst[j].fid; /* lower 8 bits */
 698                 powernow_table[j].index |= (pst[j].vid << 8); /* upper 8 bits */
 699                 freq = find_khz_freq_from_fid(pst[j].fid);
 700                 powernow_table[j].frequency = freq;
 701         }
 702         powernow_table[data->numps].frequency = CPUFREQ_TABLE_END;
 703         powernow_table[data->numps].index = 0;
 704
 705         if (query_current_values_with_pending_wait(data)) {
 706                 kfree(powernow_table);
 707                 return -EIO;
 708         }
 709
 710         pr_debug("cfid 0x%x, cvid 0x%x\n", data->currfid, data->currvid);
 711         data->powernow_table = powernow_table;
 712         if (cpumask_first(cpu_core_mask(data->cpu)) == data->cpu)
 713                 print_basics(data);
 714
 715         for (j = 0; j < data->numps; j++)
 716                 if ((pst[j].fid == data->currfid) &&
 717                     (pst[j].vid == data->currvid))
 718                         return 0;
 719
 720         pr_debug("currfid/vid do not match PST, ignoring\n");
 721         return 0;
 722 }
 723
 724 /* Find and validate the PSB/PST table in BIOS. */
 725 static int find_psb_table(struct powernow_k8_data *data)
 726 {
 727         struct psb_s *psb;
 728         unsigned int i;
 729         u32 mvs;
 730         u8 maxvid;
 731         u32 cpst = 0;
 732         u32 thiscpuid;
 733
 734         for (i = 0xc0000; i < 0xffff0; i += 0x10) {
 735                 /* Scan BIOS looking for the signature. */
 736                 /* It can not be at ffff0 - it is too big. */
 737
 738                 psb = phys_to_virt(i);
 739                 if (memcmp(psb, PSB_ID_STRING, PSB_ID_STRING_LEN) != 0)
 740                         continue;
 741
 742                 pr_debug("found PSB header at 0x%p\n", psb);
 743
 744                 pr_debug("table vers: 0x%x\n", psb->tableversion);
 745                 if (psb->tableversion != PSB_VERSION_1_4) {
 746                         printk(KERN_ERR FW_BUG PFX "PSB table is not v1.4\n");
 747                         return -ENODEV;
 748                 }
 749
 750                 pr_debug("flags: 0x%x\n", psb->flags1);
 751                 if (psb->flags1) {
 752                         printk(KERN_ERR FW_BUG PFX "unknown flags\n");
 753                         return -ENODEV;
 754                 }
 755
 756                 data->vstable = psb->vstable;
 757                 pr_debug("voltage stabilization time: %d(*20us)\n",
 758                                 data->vstable);
 759
 760                 pr_debug("flags2: 0x%x\n", psb->flags2);
 761                 data->rvo = psb->flags2 & 3;
 762                 data->irt = ((psb->flags2) >> 2) & 3;
 763                 mvs = ((psb->flags2) >> 4) & 3;
 764                 data->vidmvs = 1 << mvs;
 765                 data->batps = ((psb->flags2) >> 6) & 3;
 766
 767                 pr_debug("ramp voltage offset: %d\n", data->rvo);
 768                 pr_debug("isochronous relief time: %d\n", data->irt);
 769                 pr_debug("maximum voltage step: %d - 0x%x\n", mvs, data->vidmvs);
 770
 771                 pr_debug("numpst: 0x%x\n", psb->num_tables);
 772                 cpst = psb->num_tables;
 773                 if ((psb->cpuid == 0x00000fc0) ||
 774                     (psb->cpuid == 0x00000fe0)) {
 775                         thiscpuid = cpuid_eax(CPUID_PROCESSOR_SIGNATURE);
 776                         if ((thiscpuid == 0x00000fc0) ||
 777                             (thiscpuid == 0x00000fe0))
 778                                 cpst = 1;
 779                 }
 780                 if (cpst != 1) {
 781                         printk(KERN_ERR FW_BUG PFX "numpst must be 1\n");
 782                         return -ENODEV;
 783                 }
 784
 785                 data->plllock = psb->plllocktime;
 786                 pr_debug("plllocktime: 0x%x (units 1us)\n", psb->plllocktime);
 787                 pr_debug("maxfid: 0x%x\n", psb->maxfid);
 788                 pr_debug("maxvid: 0x%x\n", psb->maxvid);
 789                 maxvid = psb->maxvid;
 790
 791                 data->numps = psb->numps;
 792                 pr_debug("numpstates: 0x%x\n", data->numps);
 793                 return fill_powernow_table(data,
 794                                 (struct pst_s *)(psb+1), maxvid);
 795         }
 796         /*
 797          * If you see this message, complain to BIOS manufacturer. If
 798          * he tells you "we do not support Linux" or some similar
 799          * nonsense, remember that Windows 2000 uses the same legacy
 800          * mechanism that the old Linux PSB driver uses. Tell them it
 801          * is broken with Windows 2000.
 802          *
 803          * The reference to the AMD documentation is chapter 9 in the
 804          * BIOS and Kernel Developer's Guide, which is available on
 805          * www.amd.com
 806          */
 807         printk(KERN_ERR FW_BUG PFX "No PSB or ACPI _PSS objects\n");
 808         printk(KERN_ERR PFX "Make sure that your BIOS is up to date"
 809                 " and Cool'N'Quiet support is enabled in BIOS setup\n");
 810         return -ENODEV;
 811 }
 812
 813 static void powernow_k8_acpi_pst_values(struct powernow_k8_data *data,
 814                 unsigned int index)
 815 {
 816         u64 control;
 817
 818         if (!data->acpi_data.state_count || (cpu_family == CPU_HW_PSTATE))
 819                 return;
 820
 821         control = data->acpi_data.states[index].control;
 822         data->irt = (control >> IRT_SHIFT) & IRT_MASK;
 823         data->rvo = (control >> RVO_SHIFT) & RVO_MASK;
 824         data->exttype = (control >> EXT_TYPE_SHIFT) & EXT_TYPE_MASK;
 825         data->plllock = (control >> PLL_L_SHIFT) & PLL_L_MASK;
 826         data->vidmvs = 1 << ((control >> MVS_SHIFT) & MVS_MASK);
 827         data->vstable = (control >> VST_SHIFT) & VST_MASK;
 828 }
 829
 830 static int powernow_k8_cpu_init_acpi(struct powernow_k8_data *data)
 831 {
 832         struct cpufreq_frequency_table *powernow_table;
 833         int ret_val = -ENODEV;
 834         u64 control, status;
 835
 836         if (acpi_processor_register_performance(&data->acpi_data, data->cpu)) {
 837                 pr_debug("register performance failed: bad ACPI data\n");
 838                 return -EIO;
 839         }
 840
 841         /* verify the data contained in the ACPI structures */
 842         if (data->acpi_data.state_count <= 1) {
 843                 pr_debug("No ACPI P-States\n");
 844                 goto err_out;
 845         }
 846
 847         control = data->acpi_data.control_register.space_id;
 848         status = data->acpi_data.status_register.space_id;
 849
 850         if ((control != ACPI_ADR_SPACE_FIXED_HARDWARE) ||
 851             (status != ACPI_ADR_SPACE_FIXED_HARDWARE)) {
 852                 pr_debug("Invalid control/status registers (%llx - %llx)\n",
 853                         control, status);
 854                 goto err_out;
 855         }
 856
 857         /* fill in data->powernow_table */
 858         powernow_table = kmalloc((sizeof(struct cpufreq_frequency_table)
 859                 * (data->acpi_data.state_count + 1)), GFP_KERNEL);
 860         if (!powernow_table) {
 861                 pr_debug("powernow_table memory alloc failure\n");
 862                 goto err_out;
 863         }
 864
 865         /* fill in data */
 866         data->numps = data->acpi_data.state_count;
 867         powernow_k8_acpi_pst_values(data, 0);
 868
 869         if (cpu_family == CPU_HW_PSTATE)
 870                 ret_val = fill_powernow_table_pstate(data, powernow_table);
 871         else
 872                 ret_val = fill_powernow_table_fidvid(data, powernow_table);
 873         if (ret_val)
 874                 goto err_out_mem;
 875
 876         powernow_table[data->acpi_data.state_count].frequency =
 877                 CPUFREQ_TABLE_END;
 878         powernow_table[data->acpi_data.state_count].index = 0;
 879         data->powernow_table = powernow_table;
 880
 881         if (cpumask_first(cpu_core_mask(data->cpu)) == data->cpu)
 882                 print_basics(data);
 883
 884         /* notify BIOS that we exist */
 885         acpi_processor_notify_smm(THIS_MODULE);
 886
 887         if (!zalloc_cpumask_var(&data->acpi_data.shared_cpu_map, GFP_KERNEL)) {
 888                 printk(KERN_ERR PFX
 889                                 "unable to alloc powernow_k8_data cpumask\n");
 890                 ret_val = -ENOMEM;
 891                 goto err_out_mem;
 892         }
 893
 894         return 0;
 895
 896 err_out_mem:
 897         kfree(powernow_table);
 898
 899 err_out:
 900         acpi_processor_unregister_performance(&data->acpi_data, data->cpu);
 901
 902         /* data->acpi_data.state_count informs us at ->exit()
 903          * whether ACPI was used */
 904         data->acpi_data.state_count = 0;
 905
 906         return ret_val;
 907 }
 908
 909 static int fill_powernow_table_pstate(struct powernow_k8_data *data,
 910                 struct cpufreq_frequency_table *powernow_table)
 911 {
 912         int i;
 913         u32 hi = 0, lo = 0;
 914         rdmsr(MSR_PSTATE_CUR_LIMIT, lo, hi);
 915         data->max_hw_pstate = (lo & HW_PSTATE_MAX_MASK) >> HW_PSTATE_MAX_SHIFT;
 916
 917         for (i = 0; i < data->acpi_data.state_count; i++) {
 918                 u32 index;
 919
 920                 index = data->acpi_data.states[i].control & HW_PSTATE_MASK;
 921                 if (index > data->max_hw_pstate) {
 922                         printk(KERN_ERR PFX "invalid pstate %d - "
 923                                         "bad value %d.\n", i, index);
 924                         printk(KERN_ERR PFX "Please report to BIOS "
 925                                         "manufacturer\n");
 926                         invalidate_entry(powernow_table, i);
 927                         continue;
 928                 }
 929                 /* Frequency may be rounded for these */
 930                 if ((boot_cpu_data.x86 == 0x10 && boot_cpu_data.x86_model < 10)
 931                                  || boot_cpu_data.x86 == 0x11) {
 932
 933                         rdmsr(MSR_PSTATE_DEF_BASE + index, lo, hi);
 934                         if (!(hi & HW_PSTATE_VALID_MASK)) {
 935                                 pr_debug("invalid pstate %d, ignoring\n", index);
 936                                 invalidate_entry(powernow_table, i);
 937                                 continue;
 938                         }
 939
 940                         powernow_table[i].frequency =
 941                                 freq_from_fid_did(lo & 0x3f, (lo >> 6) & 7);
 942                 } else
 943                         powernow_table[i].frequency =
 944                                 data->acpi_data.states[i].core_frequency * 1000;
 945
 946                 powernow_table[i].index = index;
 947         }
 948         return 0;
 949 }
 950
 951 static int fill_powernow_table_fidvid(struct powernow_k8_data *data,
 952                 struct cpufreq_frequency_table *powernow_table)
 953 {
 954         int i;
 955
 956         for (i = 0; i < data->acpi_data.state_count; i++) {
 957                 u32 fid;
 958                 u32 vid;
 959                 u32 freq, index;
 960                 u64 status, control;
 961
 962                 if (data->exttype) {
 963                         status =  data->acpi_data.states[i].status;
 964                         fid = status & EXT_FID_MASK;
 965                         vid = (status >> VID_SHIFT) & EXT_VID_MASK;
 966                 } else {
 967                         control =  data->acpi_data.states[i].control;
 968                         fid = control & FID_MASK;
 969                         vid = (control >> VID_SHIFT) & VID_MASK;
 970                 }
 971
 972                 pr_debug("   %d : fid 0x%x, vid 0x%x\n", i, fid, vid);
 973
 974                 index = fid | (vid<<8);
 975                 powernow_table[i].index = index;
 976
 977                 freq = find_khz_freq_from_fid(fid);
 978                 powernow_table[i].frequency = freq;
 979
 980                 /* verify frequency is OK */
 981                 if ((freq > (MAX_FREQ * 1000)) || (freq < (MIN_FREQ * 1000))) {
 982                         pr_debug("invalid freq %u kHz, ignoring\n", freq);
 983                         invalidate_entry(powernow_table, i);
 984                         continue;
 985                 }
 986
 987                 /* verify voltage is OK -
 988                  * BIOSs are using "off" to indicate invalid */
 989                 if (vid == VID_OFF) {
 990                         pr_debug("invalid vid %u, ignoring\n", vid);
 991                         invalidate_entry(powernow_table, i);
 992                         continue;
 993                 }
 994
 995                 if (freq != (data->acpi_data.states[i].core_frequency * 1000)) {
 996                         printk(KERN_INFO PFX "invalid freq entries "
 997                                 "%u kHz vs. %u kHz\n", freq,
 998                                 (unsigned int)
 999                                 (data->acpi_data.states[i].core_frequency
1000                                  * 1000));
1001                         invalidate_entry(powernow_table, i);
1002                         continue;
1003                 }
1004         }
1005         return 0;
1006 }
1007
1008 static void powernow_k8_cpu_exit_acpi(struct powernow_k8_data *data)
1009 {
1010         if (data->acpi_data.state_count)
1011                 acpi_processor_unregister_performance(&data->acpi_data,
1012                                 data->cpu);
1013         free_cpumask_var(data->acpi_data.shared_cpu_map);
1014 }
1015
1016 static int get_transition_latency(struct powernow_k8_data *data)
1017 {
1018         int max_latency = 0;
1019         int i;
1020         for (i = 0; i < data->acpi_data.state_count; i++) {
1021                 int cur_latency = data->acpi_data.states[i].transition_latency
1022                         + data->acpi_data.states[i].bus_master_latency;
1023                 if (cur_latency > max_latency)
1024                         max_latency = cur_latency;
1025         }
1026         if (max_latency == 0) {
1027                 /*
1028                  * Fam 11h and later may return 0 as transition latency. This
1029                  * is intended and means "very fast". While cpufreq core and
1030                  * governors currently can handle that gracefully, better set it
1031                  * to 1 to avoid problems in the future.
1032                  */
1033                 if (boot_cpu_data.x86 < 0x11)
1034                         printk(KERN_ERR FW_WARN PFX "Invalid zero transition "
1035                                 "latency\n");
1036                 max_latency = 1;
1037         }
1038         /* value in usecs, needs to be in nanoseconds */
1039         return 1000 * max_latency;
1040 }
1041
1042 /* Take a frequency, and issue the fid/vid transition command */
1043 static int transition_frequency_fidvid(struct powernow_k8_data *data,
1044                 unsigned int index)
1045 {
1046         u32 fid = 0;
1047         u32 vid = 0;
1048         int res, i;
1049         struct cpufreq_freqs freqs;
1050
1051         pr_debug("cpu %d transition to index %u\n", smp_processor_id(), index);
1052
1053         /* fid/vid correctness check for k8 */
1054         /* fid are the lower 8 bits of the index we stored into
1055          * the cpufreq frequency table in find_psb_table, vid
1056          * are the upper 8 bits.
1057          */
1058         fid = data->powernow_table[index].index & 0xFF;
1059         vid = (data->powernow_table[index].index & 0xFF00) >> 8;
1060
1061         pr_debug("table matched fid 0x%x, giving vid 0x%x\n", fid, vid);
1062
1063         if (query_current_values_with_pending_wait(data))
1064                 return 1;
1065
1066         if ((data->currvid == vid) && (data->currfid == fid)) {
1067                 pr_debug("target matches current values (fid 0x%x, vid 0x%x)\n",
1068                         fid, vid);
1069                 return 0;
1070         }
1071
1072         pr_debug("cpu %d, changing to fid 0x%x, vid 0x%x\n",
1073                 smp_processor_id(), fid, vid);
1074         freqs.old = find_khz_freq_from_fid(data->currfid);
1075         freqs.new = find_khz_freq_from_fid(fid);
1076
1077         for_each_cpu(i, data->available_cores) {
1078                 freqs.cpu = i;
1079                 cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
1080         }
1081
1082         res = transition_fid_vid(data, fid, vid);
1083         if (res)
1084                 return res;
1085
1086         freqs.new = find_khz_freq_from_fid(data->currfid);
1087
1088         for_each_cpu(i, data->available_cores) {
1089                 freqs.cpu = i;
1090                 cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
1091         }
1092         return res;
1093 }
1094
1095 /* Take a frequency, and issue the hardware pstate transition command */
1096 static int transition_frequency_pstate(struct powernow_k8_data *data,
1097                 unsigned int index)
1098 {
1099         u32 pstate = 0;
1100         int res, i;
1101         struct cpufreq_freqs freqs;
1102
1103         pr_debug("cpu %d transition to index %u\n", smp_processor_id(), index);
1104
1105         /* get MSR index for hardware pstate transition */
1106         pstate = index & HW_PSTATE_MASK;
1107         if (pstate > data->max_hw_pstate)
1108                 return -EINVAL;
1109
1110         freqs.old = find_khz_freq_from_pstate(data->powernow_table,
1111                         data->currpstate);
1112         freqs.new = find_khz_freq_from_pstate(data->powernow_table, pstate);
1113
1114         for_each_cpu(i, data->available_cores) {
1115                 freqs.cpu = i;
1116                 cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
1117         }
1118
1119         res = transition_pstate(data, pstate);
1120         freqs.new = find_khz_freq_from_pstate(data->powernow_table, pstate);
1121
1122         for_each_cpu(i, data->available_cores) {
1123                 freqs.cpu = i;
1124                 cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
1125         }
1126         return res;
1127 }
1128
1129 /* Driver entry point to switch to the target frequency */
1130 static int powernowk8_target(struct cpufreq_policy *pol,
1131                 unsigned targfreq, unsigned relation)
1132 {
1133         cpumask_var_t oldmask;
1134         struct powernow_k8_data *data = per_cpu(powernow_data, pol->cpu);
1135         u32 checkfid;
1136         u32 checkvid;
1137         unsigned int newstate;
1138         int ret = -EIO;
1139
1140         if (!data)
1141                 return -EINVAL;
1142
1143         checkfid = data->currfid;
1144         checkvid = data->currvid;
1145
1146         /* only run on specific CPU from here on. */
1147         /* This is poor form: use a workqueue or smp_call_function_single */
1148         if (!alloc_cpumask_var(&oldmask, GFP_KERNEL))
1149                 return -ENOMEM;
1150
1151         cpumask_copy(oldmask, tsk_cpus_allowed(current));
1152         set_cpus_allowed_ptr(current, cpumask_of(pol->cpu));
1153
1154         if (smp_processor_id() != pol->cpu) {
1155                 printk(KERN_ERR PFX "limiting to cpu %u failed\n", pol->cpu);
1156                 goto err_out;
1157         }
1158
1159         if (pending_bit_stuck()) {
1160                 printk(KERN_ERR PFX "failing targ, change pending bit set\n");
1161                 goto err_out;
1162         }
1163
1164         pr_debug("targ: cpu %d, %d kHz, min %d, max %d, relation %d\n",
1165                 pol->cpu, targfreq, pol->min, pol->max, relation);
1166
1167         if (query_current_values_with_pending_wait(data))
1168                 goto err_out;
1169
1170         if (cpu_family != CPU_HW_PSTATE) {
1171                 pr_debug("targ: curr fid 0x%x, vid 0x%x\n",
1172                 data->currfid, data->currvid);
1173
1174                 if ((checkvid != data->currvid) ||
1175                     (checkfid != data->currfid)) {
1176                         printk(KERN_INFO PFX
1177                                 "error - out of sync, fix 0x%x 0x%x, "
1178                                 "vid 0x%x 0x%x\n",
1179                                 checkfid, data->currfid,
1180                                 checkvid, data->currvid);
1181                 }
1182         }
1183
1184         if (cpufreq_frequency_table_target(pol, data->powernow_table,
1185                                 targfreq, relation, &newstate))
1186                 goto err_out;
1187
1188         mutex_lock(&fidvid_mutex);
1189
1190         powernow_k8_acpi_pst_values(data, newstate);
1191
1192         if (cpu_family == CPU_HW_PSTATE)
1193                 ret = transition_frequency_pstate(data, newstate);
1194         else
1195                 ret = transition_frequency_fidvid(data, newstate);
1196         if (ret) {
1197                 printk(KERN_ERR PFX "transition frequency failed\n");
1198                 ret = 1;
1199                 mutex_unlock(&fidvid_mutex);
1200                 goto err_out;
1201         }
1202         mutex_unlock(&fidvid_mutex);
1203
1204         if (cpu_family == CPU_HW_PSTATE)
1205                 pol->cur = find_khz_freq_from_pstate(data->powernow_table,
1206                                 newstate);
1207         else
1208                 pol->cur = find_khz_freq_from_fid(data->currfid);
1209         ret = 0;
1210
1211 err_out:
1212         set_cpus_allowed_ptr(current, oldmask);
1213         free_cpumask_var(oldmask);
1214         return ret;
1215 }
1216
1217 /* Driver entry point to verify the policy and range of frequencies */
1218 static int powernowk8_verify(struct cpufreq_policy *pol)
1219 {
1220         struct powernow_k8_data *data = per_cpu(powernow_data, pol->cpu);
1221
1222         if (!data)
1223                 return -EINVAL;
1224
1225         return cpufreq_frequency_table_verify(pol, data->powernow_table);
1226 }
1227
1228 struct init_on_cpu {
1229         struct powernow_k8_data *data;
1230         int rc;
1231 };
1232
1233 static void __cpuinit powernowk8_cpu_init_on_cpu(void *_init_on_cpu)
1234 {
1235         struct init_on_cpu *init_on_cpu = _init_on_cpu;
1236
1237         if (pending_bit_stuck()) {
1238                 printk(KERN_ERR PFX "failing init, change pending bit set\n");
1239                 init_on_cpu->rc = -ENODEV;
1240                 return;
1241         }
1242
1243         if (query_current_values_with_pending_wait(init_on_cpu->data)) {
1244                 init_on_cpu->rc = -ENODEV;
1245                 return;
1246         }
1247
1248         if (cpu_family == CPU_OPTERON)
1249                 fidvid_msr_init();
1250
1251         init_on_cpu->rc = 0;
1252 }
1253
1254 /* per CPU init entry point to the driver */
1255 static int __cpuinit powernowk8_cpu_init(struct cpufreq_policy *pol)
1256 {
1257         static const char ACPI_PSS_BIOS_BUG_MSG[] =
1258                 KERN_ERR FW_BUG PFX "No compatible ACPI _PSS objects found.\n"
1259                 FW_BUG PFX "Try again with latest BIOS.\n";
1260         struct powernow_k8_data *data;
1261         struct init_on_cpu init_on_cpu;
1262         int rc;
1263         struct cpuinfo_x86 *c = &cpu_data(pol->cpu);
1264
1265         if (!cpu_online(pol->cpu))
1266                 return -ENODEV;
1267
1268         smp_call_function_single(pol->cpu, check_supported_cpu, &rc, 1);
1269         if (rc)
1270                 return -ENODEV;
1271
1272         data = kzalloc(sizeof(struct powernow_k8_data), GFP_KERNEL);
1273         if (!data) {
1274                 printk(KERN_ERR PFX "unable to alloc powernow_k8_data");
1275                 return -ENOMEM;
1276         }
1277
1278         data->cpu = pol->cpu;
1279         data->currpstate = HW_PSTATE_INVALID;
1280
1281         if (powernow_k8_cpu_init_acpi(data)) {
1282                 /*
1283                  * Use the PSB BIOS structure. This is only available on
1284                  * an UP version, and is deprecated by AMD.
1285                  */
1286                 if (num_online_cpus() != 1) {
1287                         printk_once(ACPI_PSS_BIOS_BUG_MSG);
1288                         goto err_out;
1289                 }
1290                 if (pol->cpu != 0) {
1291                         printk(KERN_ERR FW_BUG PFX "No ACPI _PSS objects for "
1292                                "CPU other than CPU0. Complain to your BIOS "
1293                                "vendor.\n");
1294                         goto err_out;
1295                 }
1296                 rc = find_psb_table(data);
1297                 if (rc)
1298                         goto err_out;
1299
1300                 /* Take a crude guess here.
1301                  * That guess was in microseconds, so multiply with 1000 */
1302                 pol->cpuinfo.transition_latency = (
1303                          ((data->rvo + 8) * data->vstable * VST_UNITS_20US) +
1304                          ((1 << data->irt) * 30)) * 1000;
1305         } else /* ACPI _PSS objects available */
1306                 pol->cpuinfo.transition_latency = get_transition_latency(data);
1307
1308         /* only run on specific CPU from here on */
1309         init_on_cpu.data = data;
1310         smp_call_function_single(data->cpu, powernowk8_cpu_init_on_cpu,
1311                                  &init_on_cpu, 1);
1312         rc = init_on_cpu.rc;
1313         if (rc != 0)
1314                 goto err_out_exit_acpi;
1315
1316         if (cpu_family == CPU_HW_PSTATE)
1317                 cpumask_copy(pol->cpus, cpumask_of(pol->cpu));
1318         else
1319                 cpumask_copy(pol->cpus, cpu_core_mask(pol->cpu));
1320         data->available_cores = pol->cpus;
1321
1322         if (cpu_family == CPU_HW_PSTATE)
1323                 pol->cur = find_khz_freq_from_pstate(data->powernow_table,
1324                                 data->currpstate);
1325         else
1326                 pol->cur = find_khz_freq_from_fid(data->currfid);
1327         pr_debug("policy current frequency %d kHz\n", pol->cur);
1328
1329         /* min/max the cpu is capable of */
1330         if (cpufreq_frequency_table_cpuinfo(pol, data->powernow_table)) {
1331                 printk(KERN_ERR FW_BUG PFX "invalid powernow_table\n");
1332                 powernow_k8_cpu_exit_acpi(data);
1333                 kfree(data->powernow_table);
1334                 kfree(data);
1335                 return -EINVAL;
1336         }
1337
1338         /* Check for APERF/MPERF support in hardware */
1339         if (cpu_has(c, X86_FEATURE_APERFMPERF))
1340                 cpufreq_amd64_driver.getavg = cpufreq_get_measured_perf;
1341
1342         cpufreq_frequency_table_get_attr(data->powernow_table, pol->cpu);
1343
1344         if (cpu_family == CPU_HW_PSTATE)
1345                 pr_debug("cpu_init done, current pstate 0x%x\n",
1346                                 data->currpstate);
1347         else
1348                 pr_debug("cpu_init done, current fid 0x%x, vid 0x%x\n",
1349                         data->currfid, data->currvid);
1350
1351         per_cpu(powernow_data, pol->cpu) = data;
1352
1353         return 0;
1354
1355 err_out_exit_acpi:
1356         powernow_k8_cpu_exit_acpi(data);
1357
1358 err_out:
1359         kfree(data);
1360         return -ENODEV;
1361 }
1362
1363 static int __devexit powernowk8_cpu_exit(struct cpufreq_policy *pol)
1364 {
1365         struct powernow_k8_data *data = per_cpu(powernow_data, pol->cpu);
1366
1367         if (!data)
1368                 return -EINVAL;
1369
1370         powernow_k8_cpu_exit_acpi(data);
1371
1372         cpufreq_frequency_table_put_attr(pol->cpu);
1373
1374         kfree(data->powernow_table);
1375         kfree(data);
1376         per_cpu(powernow_data, pol->cpu) = NULL;
1377
1378         return 0;
1379 }
1380
1381 static void query_values_on_cpu(void *_err)
1382 {
1383         int *err = _err;
1384         struct powernow_k8_data *data = __this_cpu_read(powernow_data);
1385
1386         *err = query_current_values_with_pending_wait(data);
1387 }
1388
1389 static unsigned int powernowk8_get(unsigned int cpu)
1390 {
1391         struct powernow_k8_data *data = per_cpu(powernow_data, cpu);
1392         unsigned int khz = 0;
1393         int err;
1394
1395         if (!data)
1396                 return 0;
1397
1398         smp_call_function_single(cpu, query_values_on_cpu, &err, true);
1399         if (err)
1400                 goto out;
1401
1402         if (cpu_family == CPU_HW_PSTATE)
1403                 khz = find_khz_freq_from_pstate(data->powernow_table,
1404                                                 data->currpstate);
1405         else
1406                 khz = find_khz_freq_from_fid(data->currfid);
1407
1408
1409 out:
1410         return khz;
1411 }
1412
1413 static void _cpb_toggle_msrs(bool t)
1414 {
1415         int cpu;
1416
1417         get_online_cpus();
1418
1419         rdmsr_on_cpus(cpu_online_mask, MSR_K7_HWCR, msrs);
1420
1421         for_each_cpu(cpu, cpu_online_mask) {
1422                 struct msr *reg = per_cpu_ptr(msrs, cpu);
1423                 if (t)
1424                         reg->l &= ~BIT(25);
1425                 else
1426                         reg->l |= BIT(25);
1427         }
1428         wrmsr_on_cpus(cpu_online_mask, MSR_K7_HWCR, msrs);
1429
1430         put_online_cpus();
1431 }
1432
1433 /*
1434  * Switch on/off core performance boosting.
1435  *
1436  * 0=disable
1437  * 1=enable.
1438  */
1439 static void cpb_toggle(bool t)
1440 {
1441         if (!cpb_capable)
1442                 return;
1443
1444         if (t && !cpb_enabled) {
1445                 cpb_enabled = true;
1446                 _cpb_toggle_msrs(t);
1447                 printk(KERN_INFO PFX "Core Boosting enabled.\n");
1448         } else if (!t && cpb_enabled) {
1449                 cpb_enabled = false;
1450                 _cpb_toggle_msrs(t);
1451                 printk(KERN_INFO PFX "Core Boosting disabled.\n");
1452         }
1453 }
1454
1455 static ssize_t store_cpb(struct cpufreq_policy *policy, const char *buf,
1456                                  size_t count)
1457 {
1458         int ret = -EINVAL;
1459         unsigned long val = 0;
1460
1461         ret = strict_strtoul(buf, 10, &val);
1462         if (!ret && (val == 0 || val == 1) && cpb_capable)
1463                 cpb_toggle(val);
1464         else
1465                 return -EINVAL;
1466
1467         return count;
1468 }
1469
1470 static ssize_t show_cpb(struct cpufreq_policy *policy, char *buf)
1471 {
1472         return sprintf(buf, "%u\n", cpb_enabled);
1473 }
1474
1475 #define define_one_rw(_name) \
1476 static struct freq_attr _name = \
1477 __ATTR(_name, 0644, show_##_name, store_##_name)
1478
1479 define_one_rw(cpb);
1480
1481 static struct freq_attr *powernow_k8_attr[] = {
1482         &cpufreq_freq_attr_scaling_available_freqs,
1483         &cpb,
1484         NULL,
1485 };
1486
1487 static struct cpufreq_driver cpufreq_amd64_driver = {
1488         .verify         = powernowk8_verify,
1489         .target         = powernowk8_target,
1490         .bios_limit     = acpi_processor_get_bios_limit,
1491         .init           = powernowk8_cpu_init,
1492         .exit           = __devexit_p(powernowk8_cpu_exit),
1493         .get            = powernowk8_get,
1494         .name           = "powernow-k8",
1495         .owner          = THIS_MODULE,
1496         .attr           = powernow_k8_attr,
1497 };
1498
1499 /*
1500  * Clear the boost-disable flag on the CPU_DOWN path so that this cpu
1501  * cannot block the remaining ones from boosting. On the CPU_UP path we
1502  * simply keep the boost-disable flag in sync with the current global
1503  * state.
1504  */
1505 static int cpb_notify(struct notifier_block *nb, unsigned long action,
1506                       void *hcpu)
1507 {
1508         unsigned cpu = (long)hcpu;
1509         u32 lo, hi;
1510
1511         switch (action) {
1512         case CPU_UP_PREPARE:
1513         case CPU_UP_PREPARE_FROZEN:
1514
1515                 if (!cpb_enabled) {
1516                         rdmsr_on_cpu(cpu, MSR_K7_HWCR, &lo, &hi);
1517                         lo |= BIT(25);
1518                         wrmsr_on_cpu(cpu, MSR_K7_HWCR, lo, hi);
1519                 }
1520                 break;
1521
1522         case CPU_DOWN_PREPARE:
1523         case CPU_DOWN_PREPARE_FROZEN:
1524                 rdmsr_on_cpu(cpu, MSR_K7_HWCR, &lo, &hi);
1525                 lo &= ~BIT(25);
1526                 wrmsr_on_cpu(cpu, MSR_K7_HWCR, lo, hi);
1527                 break;
1528
1529         default:
1530                 break;
1531         }
1532
1533         return NOTIFY_OK;
1534 }
1535
1536 static struct notifier_block cpb_nb = {
1537         .notifier_call          = cpb_notify,
1538 };
1539
1540 /* driver entry point for init */
1541 static int __cpuinit powernowk8_init(void)
1542 {
1543         unsigned int i, supported_cpus = 0, cpu;
1544         int rv;
1545
1546         for_each_online_cpu(i) {
1547                 int rc;
1548                 smp_call_function_single(i, check_supported_cpu, &rc, 1);
1549                 if (rc == 0)
1550                         supported_cpus++;
1551         }
1552
1553         if (supported_cpus != num_online_cpus())
1554                 return -ENODEV;
1555
1556         printk(KERN_INFO PFX "Found %d %s (%d cpu cores) (" VERSION ")\n",
1557                 num_online_nodes(), boot_cpu_data.x86_model_id, supported_cpus);
1558
1559         if (boot_cpu_has(X86_FEATURE_CPB)) {
1560
1561                 cpb_capable = true;
1562
1563                 msrs = msrs_alloc();
1564                 if (!msrs) {
1565                         printk(KERN_ERR "%s: Error allocating msrs!\n", __func__);
1566                         return -ENOMEM;
1567                 }
1568
1569                 register_cpu_notifier(&cpb_nb);
1570
1571                 rdmsr_on_cpus(cpu_online_mask, MSR_K7_HWCR, msrs);
1572
1573                 for_each_cpu(cpu, cpu_online_mask) {
1574                         struct msr *reg = per_cpu_ptr(msrs, cpu);
1575                         cpb_enabled |= !(!!(reg->l & BIT(25)));
1576                 }
1577
1578                 printk(KERN_INFO PFX "Core Performance Boosting: %s.\n",
1579                         (cpb_enabled ? "on" : "off"));
1580         }
1581
1582         rv = cpufreq_register_driver(&cpufreq_amd64_driver);
1583         if (rv < 0 && boot_cpu_has(X86_FEATURE_CPB)) {
1584                 unregister_cpu_notifier(&cpb_nb);
1585                 msrs_free(msrs);
1586                 msrs = NULL;
1587         }
1588         return rv;
1589 }
1590
1591 /* driver entry point for term */
1592 static void __exit powernowk8_exit(void)
1593 {
1594         pr_debug("exit\n");
1595
1596         if (boot_cpu_has(X86_FEATURE_CPB)) {
1597                 msrs_free(msrs);
1598                 msrs = NULL;
1599
1600                 unregister_cpu_notifier(&cpb_nb);
1601         }
1602
1603         cpufreq_unregister_driver(&cpufreq_amd64_driver);
1604 }
1605
1606 MODULE_AUTHOR("Paul Devriendt <paul.devriendt@amd.com> and "
1607                 "Mark Langsdorf <mark.langsdorf@amd.com>");
1608 MODULE_DESCRIPTION("AMD Athlon 64 and Opteron processor frequency driver.");
1609 MODULE_LICENSE("GPL");
1610
1611 late_initcall(powernowk8_init);
1612 module_exit(powernowk8_exit);