git.oblomov.eu Git - linux-2.6/blob - arch/x86/kernel/cpu/cpufreq/powernow-k8.c

   1 /*
   2  *   (c) 2003-2006 Advanced Micro Devices, Inc.
   3  *  Your use of this code is subject to the terms and conditions of the
   4  *  GNU general public license version 2. See "COPYING" or
   5  *  http://www.gnu.org/licenses/gpl.html
   6  *
   7  *  Support : mark.langsdorf@amd.com
   8  *
   9  *  Based on the powernow-k7.c module written by Dave Jones.
  10  *  (C) 2003 Dave Jones <davej@codemonkey.org.uk> on behalf of SuSE Labs
  11  *  (C) 2004 Dominik Brodowski <linux@brodo.de>
  12  *  (C) 2004 Pavel Machek <pavel@suse.cz>
  13  *  Licensed under the terms of the GNU GPL License version 2.
  14  *  Based upon datasheets & sample CPUs kindly provided by AMD.
  15  *
  16  *  Valuable input gratefully received from Dave Jones, Pavel Machek,
  17  *  Dominik Brodowski, 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
  37 #include <asm/msr.h>
  38 #include <asm/io.h>
  39 #include <asm/delay.h>
  40
  41 #ifdef CONFIG_X86_POWERNOW_K8_ACPI
  42 #include <linux/acpi.h>
  43 #include <linux/mutex.h>
  44 #include <acpi/processor.h>
  45 #endif
  46
  47 #define PFX "powernow-k8: "
  48 #define BFX PFX "BIOS error: "
  49 #define VERSION "version 2.00.00"
  50 #include "powernow-k8.h"
  51
  52 /* serialize freq changes  */
  53 static DEFINE_MUTEX(fidvid_mutex);
  54
  55 static struct powernow_k8_data *powernow_data[NR_CPUS];
  56
  57 static int cpu_family = CPU_OPTERON;
  58
  59 #ifndef CONFIG_SMP
  60 static cpumask_t cpu_core_map[1];
  61 #endif
  62
  63 /* Return a frequency in MHz, given an input fid */
  64 static u32 find_freq_from_fid(u32 fid)
  65 {
  66         return 800 + (fid * 100);
  67 }
  68
  69
  70 /* Return a frequency in KHz, given an input fid */
  71 static u32 find_khz_freq_from_fid(u32 fid)
  72 {
  73         return 1000 * find_freq_from_fid(fid);
  74 }
  75
  76 /* Return a frequency in MHz, given an input fid and did */
  77 static u32 find_freq_from_fiddid(u32 fid, u32 did)
  78 {
  79         if (current_cpu_data.x86 == 0x10)
  80                 return 100 * (fid + 0x10) >> did;
  81         else
  82                 return 100 * (fid + 0x8) >> did;
  83 }
  84
  85 static u32 find_khz_freq_from_fiddid(u32 fid, u32 did)
  86 {
  87         return 1000 * find_freq_from_fiddid(fid, did);
  88 }
  89
  90 static u32 find_fid_from_pstate(u32 pstate)
  91 {
  92         u32 hi, lo;
  93         rdmsr(MSR_PSTATE_DEF_BASE + pstate, lo, hi);
  94         return lo & HW_PSTATE_FID_MASK;
  95 }
  96
  97 static u32 find_did_from_pstate(u32 pstate)
  98 {
  99         u32 hi, lo;
 100         rdmsr(MSR_PSTATE_DEF_BASE + pstate, lo, hi);
 101         return (lo & HW_PSTATE_DID_MASK) >> HW_PSTATE_DID_SHIFT;
 102 }
 103
 104 /* Return the vco fid for an input fid
 105  *
 106  * Each "low" fid has corresponding "high" fid, and you can get to "low" fids
 107  * only from corresponding high fids. This returns "high" fid corresponding to
 108  * "low" one.
 109  */
 110 static u32 convert_fid_to_vco_fid(u32 fid)
 111 {
 112         if (fid < HI_FID_TABLE_BOTTOM)
 113                 return 8 + (2 * fid);
 114         else
 115                 return fid;
 116 }
 117
 118 /*
 119  * Return 1 if the pending bit is set. Unless we just instructed the processor
 120  * to transition to a new state, seeing this bit set is really bad news.
 121  */
 122 static int pending_bit_stuck(void)
 123 {
 124         u32 lo, hi;
 125
 126         if (cpu_family == CPU_HW_PSTATE)
 127                 return 0;
 128
 129         rdmsr(MSR_FIDVID_STATUS, lo, hi);
 130         return lo & MSR_S_LO_CHANGE_PENDING ? 1 : 0;
 131 }
 132
 133 /*
 134  * Update the global current fid / vid values from the status msr.
 135  * Returns 1 on error.
 136  */
 137 static int query_current_values_with_pending_wait(struct powernow_k8_data *data)
 138 {
 139         u32 lo, hi;
 140         u32 i = 0;
 141
 142         if (cpu_family == CPU_HW_PSTATE) {
 143                 rdmsr(MSR_PSTATE_STATUS, lo, hi);
 144                 i = lo & HW_PSTATE_MASK;
 145                 rdmsr(MSR_PSTATE_DEF_BASE + i, lo, hi);
 146                 data->currfid = lo & HW_PSTATE_FID_MASK;
 147                 data->currdid = (lo & HW_PSTATE_DID_MASK) >> HW_PSTATE_DID_SHIFT;
 148                 return 0;
 149         }
 150         do {
 151                 if (i++ > 10000) {
 152                         dprintk("detected change pending stuck\n");
 153                         return 1;
 154                 }
 155                 rdmsr(MSR_FIDVID_STATUS, lo, hi);
 156         } while (lo & MSR_S_LO_CHANGE_PENDING);
 157
 158         data->currvid = hi & MSR_S_HI_CURRENT_VID;
 159         data->currfid = lo & MSR_S_LO_CURRENT_FID;
 160
 161         return 0;
 162 }
 163
 164 /* the isochronous relief time */
 165 static void count_off_irt(struct powernow_k8_data *data)
 166 {
 167         udelay((1 << data->irt) * 10);
 168         return;
 169 }
 170
 171 /* the voltage stabalization time */
 172 static void count_off_vst(struct powernow_k8_data *data)
 173 {
 174         udelay(data->vstable * VST_UNITS_20US);
 175         return;
 176 }
 177
 178 /* need to init the control msr to a safe value (for each cpu) */
 179 static void fidvid_msr_init(void)
 180 {
 181         u32 lo, hi;
 182         u8 fid, vid;
 183
 184         rdmsr(MSR_FIDVID_STATUS, lo, hi);
 185         vid = hi & MSR_S_HI_CURRENT_VID;
 186         fid = lo & MSR_S_LO_CURRENT_FID;
 187         lo = fid | (vid << MSR_C_LO_VID_SHIFT);
 188         hi = MSR_C_HI_STP_GNT_BENIGN;
 189         dprintk("cpu%d, init lo 0x%x, hi 0x%x\n", smp_processor_id(), lo, hi);
 190         wrmsr(MSR_FIDVID_CTL, lo, hi);
 191 }
 192
 193
 194 /* write the new fid value along with the other control fields to the msr */
 195 static int write_new_fid(struct powernow_k8_data *data, u32 fid)
 196 {
 197         u32 lo;
 198         u32 savevid = data->currvid;
 199         u32 i = 0;
 200
 201         if ((fid & INVALID_FID_MASK) || (data->currvid & INVALID_VID_MASK)) {
 202                 printk(KERN_ERR PFX "internal error - overflow on fid write\n");
 203                 return 1;
 204         }
 205
 206         lo = fid | (data->currvid << MSR_C_LO_VID_SHIFT) | MSR_C_LO_INIT_FID_VID;
 207
 208         dprintk("writing fid 0x%x, lo 0x%x, hi 0x%x\n",
 209                 fid, lo, data->plllock * PLL_LOCK_CONVERSION);
 210
 211         do {
 212                 wrmsr(MSR_FIDVID_CTL, lo, data->plllock * PLL_LOCK_CONVERSION);
 213                 if (i++ > 100) {
 214                         printk(KERN_ERR PFX "Hardware error - pending bit very stuck - no further pstate changes possible\n");
 215                         return 1;
 216                 }
 217         } while (query_current_values_with_pending_wait(data));
 218
 219         count_off_irt(data);
 220
 221         if (savevid != data->currvid) {
 222                 printk(KERN_ERR PFX "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 "fid trans failed, fid 0x%x, curr 0x%x\n", fid,
 229                         data->currfid);
 230                 return 1;
 231         }
 232
 233         return 0;
 234 }
 235
 236 /* Write a new vid to the hardware */
 237 static int write_new_vid(struct powernow_k8_data *data, u32 vid)
 238 {
 239         u32 lo;
 240         u32 savefid = data->currfid;
 241         int i = 0;
 242
 243         if ((data->currfid & INVALID_FID_MASK) || (vid & INVALID_VID_MASK)) {
 244                 printk(KERN_ERR PFX "internal error - overflow on vid write\n");
 245                 return 1;
 246         }
 247
 248         lo = data->currfid | (vid << MSR_C_LO_VID_SHIFT) | MSR_C_LO_INIT_FID_VID;
 249
 250         dprintk("writing vid 0x%x, lo 0x%x, hi 0x%x\n",
 251                 vid, lo, STOP_GRANT_5NS);
 252
 253         do {
 254                 wrmsr(MSR_FIDVID_CTL, lo, STOP_GRANT_5NS);
 255                 if (i++ > 100) {
 256                         printk(KERN_ERR PFX "internal error - pending bit very stuck - no further pstate changes possible\n");
 257                         return 1;
 258                 }
 259         } while (query_current_values_with_pending_wait(data));
 260
 261         if (savefid != data->currfid) {
 262                 printk(KERN_ERR PFX "fid changed on vid trans, old 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, curr 0x%x\n", vid,
 269                                 data->currvid);
 270                 return 1;
 271         }
 272
 273         return 0;
 274 }
 275
 276 /*
 277  * Reduce the vid by the max of step or reqvid.
 278  * Decreasing vid codes represent increasing voltages:
 279  * vid of 0 is 1.550V, vid of 0x1e is 0.800V, vid of VID_OFF is off.
 280  */
 281 static int decrease_vid_code_by_step(struct powernow_k8_data *data, u32 reqvid, u32 step)
 282 {
 283         if ((data->currvid - reqvid) > step)
 284                 reqvid = data->currvid - step;
 285
 286         if (write_new_vid(data, reqvid))
 287                 return 1;
 288
 289         count_off_vst(data);
 290
 291         return 0;
 292 }
 293
 294 /* Change hardware pstate by single MSR write */
 295 static int transition_pstate(struct powernow_k8_data *data, u32 pstate)
 296 {
 297         wrmsr(MSR_PSTATE_CTRL, pstate, 0);
 298         data->currfid = find_fid_from_pstate(pstate);
 299         return 0;
 300 }
 301
 302 /* Change Opteron/Athlon64 fid and vid, by the 3 phases. */
 303 static int transition_fid_vid(struct powernow_k8_data *data, u32 reqfid, u32 reqvid)
 304 {
 305         if (core_voltage_pre_transition(data, reqvid))
 306                 return 1;
 307
 308         if (core_frequency_transition(data, reqfid))
 309                 return 1;
 310
 311         if (core_voltage_post_transition(data, reqvid))
 312                 return 1;
 313
 314         if (query_current_values_with_pending_wait(data))
 315                 return 1;
 316
 317         if ((reqfid != data->currfid) || (reqvid != data->currvid)) {
 318                 printk(KERN_ERR PFX "failed (cpu%d): req 0x%x 0x%x, curr 0x%x 0x%x\n",
 319                                 smp_processor_id(),
 320                                 reqfid, reqvid, data->currfid, data->currvid);
 321                 return 1;
 322         }
 323
 324         dprintk("transitioned (cpu%d): new fid 0x%x, vid 0x%x\n",
 325                 smp_processor_id(), data->currfid, data->currvid);
 326
 327         return 0;
 328 }
 329
 330 /* Phase 1 - core voltage transition ... setup voltage */
 331 static int core_voltage_pre_transition(struct powernow_k8_data *data, u32 reqvid)
 332 {
 333         u32 rvosteps = data->rvo;
 334         u32 savefid = data->currfid;
 335         u32 maxvid, lo;
 336
 337         dprintk("ph1 (cpu%d): start, currfid 0x%x, currvid 0x%x, reqvid 0x%x, rvo 0x%x\n",
 338                 smp_processor_id(),
 339                 data->currfid, data->currvid, reqvid, data->rvo);
 340
 341         rdmsr(MSR_FIDVID_STATUS, lo, maxvid);
 342         maxvid = 0x1f & (maxvid >> 16);
 343         dprintk("ph1 maxvid=0x%x\n", maxvid);
 344         if (reqvid < maxvid) /* lower numbers are higher voltages */
 345                 reqvid = maxvid;
 346
 347         while (data->currvid > reqvid) {
 348                 dprintk("ph1: curr 0x%x, req vid 0x%x\n",
 349                         data->currvid, reqvid);
 350                 if (decrease_vid_code_by_step(data, reqvid, data->vidmvs))
 351                         return 1;
 352         }
 353
 354         while ((rvosteps > 0) && ((data->rvo + data->currvid) > reqvid)) {
 355                 if (data->currvid == maxvid) {
 356                         rvosteps = 0;
 357                 } else {
 358                         dprintk("ph1: changing vid for rvo, req 0x%x\n",
 359                                 data->currvid - 1);
 360                         if (decrease_vid_code_by_step(data, data->currvid - 1, 1))
 361                                 return 1;
 362                         rvosteps--;
 363                 }
 364         }
 365
 366         if (query_current_values_with_pending_wait(data))
 367                 return 1;
 368
 369         if (savefid != data->currfid) {
 370                 printk(KERN_ERR PFX "ph1 err, currfid changed 0x%x\n", data->currfid);
 371                 return 1;
 372         }
 373
 374         dprintk("ph1 complete, currfid 0x%x, currvid 0x%x\n",
 375                 data->currfid, data->currvid);
 376
 377         return 0;
 378 }
 379
 380 /* Phase 2 - core frequency transition */
 381 static int core_frequency_transition(struct powernow_k8_data *data, u32 reqfid)
 382 {
 383         u32 vcoreqfid, vcocurrfid, vcofiddiff, fid_interval, savevid = data->currvid;
 384
 385         if ((reqfid < HI_FID_TABLE_BOTTOM) && (data->currfid < HI_FID_TABLE_BOTTOM)) {
 386                 printk(KERN_ERR PFX "ph2: illegal lo-lo transition 0x%x 0x%x\n",
 387                         reqfid, data->currfid);
 388                 return 1;
 389         }
 390
 391         if (data->currfid == reqfid) {
 392                 printk(KERN_ERR PFX "ph2 null fid transition 0x%x\n", data->currfid);
 393                 return 0;
 394         }
 395
 396         dprintk("ph2 (cpu%d): starting, currfid 0x%x, currvid 0x%x, reqfid 0x%x\n",
 397                 smp_processor_id(),
 398                 data->currfid, data->currvid, reqfid);
 399
 400         vcoreqfid = convert_fid_to_vco_fid(reqfid);
 401         vcocurrfid = convert_fid_to_vco_fid(data->currfid);
 402         vcofiddiff = vcocurrfid > vcoreqfid ? vcocurrfid - vcoreqfid
 403             : vcoreqfid - vcocurrfid;
 404
 405         while (vcofiddiff > 2) {
 406                 (data->currfid & 1) ? (fid_interval = 1) : (fid_interval = 2);
 407
 408                 if (reqfid > data->currfid) {
 409                         if (data->currfid > LO_FID_TABLE_TOP) {
 410                                 if (write_new_fid(data, data->currfid + fid_interval)) {
 411                                         return 1;
 412                                 }
 413                         } else {
 414                                 if (write_new_fid
 415                                     (data, 2 + convert_fid_to_vco_fid(data->currfid))) {
 416                                         return 1;
 417                                 }
 418                         }
 419                 } else {
 420                         if (write_new_fid(data, data->currfid - fid_interval))
 421                                 return 1;
 422                 }
 423
 424                 vcocurrfid = convert_fid_to_vco_fid(data->currfid);
 425                 vcofiddiff = vcocurrfid > vcoreqfid ? vcocurrfid - vcoreqfid
 426                     : vcoreqfid - vcocurrfid;
 427         }
 428
 429         if (write_new_fid(data, reqfid))
 430                 return 1;
 431
 432         if (query_current_values_with_pending_wait(data))
 433                 return 1;
 434
 435         if (data->currfid != reqfid) {
 436                 printk(KERN_ERR PFX
 437                         "ph2: mismatch, failed fid transition, curr 0x%x, req 0x%x\n",
 438                         data->currfid, reqfid);
 439                 return 1;
 440         }
 441
 442         if (savevid != data->currvid) {
 443                 printk(KERN_ERR PFX "ph2: vid changed, save 0x%x, curr 0x%x\n",
 444                         savevid, data->currvid);
 445                 return 1;
 446         }
 447
 448         dprintk("ph2 complete, currfid 0x%x, currvid 0x%x\n",
 449                 data->currfid, data->currvid);
 450
 451         return 0;
 452 }
 453
 454 /* Phase 3 - core voltage transition flow ... jump to the final vid. */
 455 static int core_voltage_post_transition(struct powernow_k8_data *data, u32 reqvid)
 456 {
 457         u32 savefid = data->currfid;
 458         u32 savereqvid = reqvid;
 459
 460         dprintk("ph3 (cpu%d): starting, currfid 0x%x, currvid 0x%x\n",
 461                 smp_processor_id(),
 462                 data->currfid, data->currvid);
 463
 464         if (reqvid != data->currvid) {
 465                 if (write_new_vid(data, reqvid))
 466                         return 1;
 467
 468                 if (savefid != data->currfid) {
 469                         printk(KERN_ERR PFX
 470                                "ph3: bad fid change, save 0x%x, curr 0x%x\n",
 471                                savefid, data->currfid);
 472                         return 1;
 473                 }
 474
 475                 if (data->currvid != reqvid) {
 476                         printk(KERN_ERR PFX
 477                                "ph3: failed vid transition\n, req 0x%x, curr 0x%x",
 478                                reqvid, data->currvid);
 479                         return 1;
 480                 }
 481         }
 482
 483         if (query_current_values_with_pending_wait(data))
 484                 return 1;
 485
 486         if (savereqvid != data->currvid) {
 487                 dprintk("ph3 failed, currvid 0x%x\n", data->currvid);
 488                 return 1;
 489         }
 490
 491         if (savefid != data->currfid) {
 492                 dprintk("ph3 failed, currfid changed 0x%x\n",
 493                         data->currfid);
 494                 return 1;
 495         }
 496
 497         dprintk("ph3 complete, currfid 0x%x, currvid 0x%x\n",
 498                 data->currfid, data->currvid);
 499
 500         return 0;
 501 }
 502
 503 static int check_supported_cpu(unsigned int cpu)
 504 {
 505         cpumask_t oldmask = CPU_MASK_ALL;
 506         u32 eax, ebx, ecx, edx;
 507         unsigned int rc = 0;
 508
 509         oldmask = current->cpus_allowed;
 510         set_cpus_allowed(current, cpumask_of_cpu(cpu));
 511
 512         if (smp_processor_id() != cpu) {
 513                 printk(KERN_ERR PFX "limiting to cpu %u failed\n", cpu);
 514                 goto out;
 515         }
 516
 517         if (current_cpu_data.x86_vendor != X86_VENDOR_AMD)
 518                 goto out;
 519
 520         eax = cpuid_eax(CPUID_PROCESSOR_SIGNATURE);
 521         if (((eax & CPUID_XFAM) != CPUID_XFAM_K8) &&
 522             ((eax & CPUID_XFAM) < CPUID_XFAM_10H))
 523                 goto out;
 524
 525         if ((eax & CPUID_XFAM) == CPUID_XFAM_K8) {
 526                 if (((eax & CPUID_USE_XFAM_XMOD) != CPUID_USE_XFAM_XMOD) ||
 527                     ((eax & CPUID_XMOD) > CPUID_XMOD_REV_MASK)) {
 528                         printk(KERN_INFO PFX "Processor cpuid %x not supported\n", eax);
 529                         goto out;
 530                 }
 531
 532                 eax = cpuid_eax(CPUID_GET_MAX_CAPABILITIES);
 533                 if (eax < CPUID_FREQ_VOLT_CAPABILITIES) {
 534                         printk(KERN_INFO PFX
 535                                "No frequency change capabilities detected\n");
 536                         goto out;
 537                 }
 538
 539                 cpuid(CPUID_FREQ_VOLT_CAPABILITIES, &eax, &ebx, &ecx, &edx);
 540                 if ((edx & P_STATE_TRANSITION_CAPABLE) != P_STATE_TRANSITION_CAPABLE) {
 541                         printk(KERN_INFO PFX "Power state transitions not supported\n");
 542                         goto out;
 543                 }
 544         } else { /* must be a HW Pstate capable processor */
 545                 cpuid(CPUID_FREQ_VOLT_CAPABILITIES, &eax, &ebx, &ecx, &edx);
 546                 if ((edx & USE_HW_PSTATE) == USE_HW_PSTATE)
 547                         cpu_family = CPU_HW_PSTATE;
 548                 else
 549                         goto out;
 550         }
 551
 552         rc = 1;
 553
 554 out:
 555         set_cpus_allowed(current, oldmask);
 556         return rc;
 557 }
 558
 559 static int check_pst_table(struct powernow_k8_data *data, struct pst_s *pst, u8 maxvid)
 560 {
 561         unsigned int j;
 562         u8 lastfid = 0xff;
 563
 564         for (j = 0; j < data->numps; j++) {
 565                 if (pst[j].vid > LEAST_VID) {
 566                         printk(KERN_ERR PFX "vid %d invalid : 0x%x\n", j, pst[j].vid);
 567                         return -EINVAL;
 568                 }
 569                 if (pst[j].vid < data->rvo) {   /* vid + rvo >= 0 */
 570                         printk(KERN_ERR BFX "0 vid exceeded with pstate %d\n", j);
 571                         return -ENODEV;
 572                 }
 573                 if (pst[j].vid < maxvid + data->rvo) {  /* vid + rvo >= maxvid */
 574                         printk(KERN_ERR BFX "maxvid exceeded with pstate %d\n", j);
 575                         return -ENODEV;
 576                 }
 577                 if (pst[j].fid > MAX_FID) {
 578                         printk(KERN_ERR BFX "maxfid exceeded with pstate %d\n", j);
 579                         return -ENODEV;
 580                 }
 581                 if (j && (pst[j].fid < HI_FID_TABLE_BOTTOM)) {
 582                         /* Only first fid is allowed to be in "low" range */
 583                         printk(KERN_ERR BFX "two low fids - %d : 0x%x\n", j, pst[j].fid);
 584                         return -EINVAL;
 585                 }
 586                 if (pst[j].fid < lastfid)
 587                         lastfid = pst[j].fid;
 588         }
 589         if (lastfid & 1) {
 590                 printk(KERN_ERR BFX "lastfid invalid\n");
 591                 return -EINVAL;
 592         }
 593         if (lastfid > LO_FID_TABLE_TOP)
 594                 printk(KERN_INFO BFX  "first fid not from lo freq table\n");
 595
 596         return 0;
 597 }
 598
 599 static void print_basics(struct powernow_k8_data *data)
 600 {
 601         int j;
 602         for (j = 0; j < data->numps; j++) {
 603                 if (data->powernow_table[j].frequency != CPUFREQ_ENTRY_INVALID) {
 604                         if (cpu_family == CPU_HW_PSTATE) {
 605                                 printk(KERN_INFO PFX "   %d : fid 0x%x did 0x%x (%d MHz)\n",
 606                                         j,
 607                                         (data->powernow_table[j].index & 0xff00) >> 8,
 608                                         (data->powernow_table[j].index & 0xff0000) >> 16,
 609                                         data->powernow_table[j].frequency/1000);
 610                         } else {
 611                                 printk(KERN_INFO PFX "   %d : fid 0x%x (%d MHz), vid 0x%x\n",
 612                                         j,
 613                                         data->powernow_table[j].index & 0xff,
 614                                         data->powernow_table[j].frequency/1000,
 615                                         data->powernow_table[j].index >> 8);
 616                         }
 617                 }
 618         }
 619         if (data->batps)
 620                 printk(KERN_INFO PFX "Only %d pstates on battery\n", data->batps);
 621 }
 622
 623 static int fill_powernow_table(struct powernow_k8_data *data, struct pst_s *pst, u8 maxvid)
 624 {
 625         struct cpufreq_frequency_table *powernow_table;
 626         unsigned int j;
 627
 628         if (data->batps) {    /* use ACPI support to get full speed on mains power */
 629                 printk(KERN_WARNING PFX "Only %d pstates usable (use ACPI driver for full range\n", data->batps);
 630                 data->numps = data->batps;
 631         }
 632
 633         for ( j=1; j<data->numps; j++ ) {
 634                 if (pst[j-1].fid >= pst[j].fid) {
 635                         printk(KERN_ERR PFX "PST out of sequence\n");
 636                         return -EINVAL;
 637                 }
 638         }
 639
 640         if (data->numps < 2) {
 641                 printk(KERN_ERR PFX "no p states to transition\n");
 642                 return -ENODEV;
 643         }
 644
 645         if (check_pst_table(data, pst, maxvid))
 646                 return -EINVAL;
 647
 648         powernow_table = kmalloc((sizeof(struct cpufreq_frequency_table)
 649                 * (data->numps + 1)), GFP_KERNEL);
 650         if (!powernow_table) {
 651                 printk(KERN_ERR PFX "powernow_table memory alloc failure\n");
 652                 return -ENOMEM;
 653         }
 654
 655         for (j = 0; j < data->numps; j++) {
 656                 powernow_table[j].index = pst[j].fid; /* lower 8 bits */
 657                 powernow_table[j].index |= (pst[j].vid << 8); /* upper 8 bits */
 658                 powernow_table[j].frequency = find_khz_freq_from_fid(pst[j].fid);
 659         }
 660         powernow_table[data->numps].frequency = CPUFREQ_TABLE_END;
 661         powernow_table[data->numps].index = 0;
 662
 663         if (query_current_values_with_pending_wait(data)) {
 664                 kfree(powernow_table);
 665                 return -EIO;
 666         }
 667
 668         dprintk("cfid 0x%x, cvid 0x%x\n", data->currfid, data->currvid);
 669         data->powernow_table = powernow_table;
 670         if (first_cpu(cpu_core_map[data->cpu]) == data->cpu)
 671                 print_basics(data);
 672
 673         for (j = 0; j < data->numps; j++)
 674                 if ((pst[j].fid==data->currfid) && (pst[j].vid==data->currvid))
 675                         return 0;
 676
 677         dprintk("currfid/vid do not match PST, ignoring\n");
 678         return 0;
 679 }
 680
 681 /* Find and validate the PSB/PST table in BIOS. */
 682 static int find_psb_table(struct powernow_k8_data *data)
 683 {
 684         struct psb_s *psb;
 685         unsigned int i;
 686         u32 mvs;
 687         u8 maxvid;
 688         u32 cpst = 0;
 689         u32 thiscpuid;
 690
 691         for (i = 0xc0000; i < 0xffff0; i += 0x10) {
 692                 /* Scan BIOS looking for the signature. */
 693                 /* It can not be at ffff0 - it is too big. */
 694
 695                 psb = phys_to_virt(i);
 696                 if (memcmp(psb, PSB_ID_STRING, PSB_ID_STRING_LEN) != 0)
 697                         continue;
 698
 699                 dprintk("found PSB header at 0x%p\n", psb);
 700
 701                 dprintk("table vers: 0x%x\n", psb->tableversion);
 702                 if (psb->tableversion != PSB_VERSION_1_4) {
 703                         printk(KERN_ERR BFX "PSB table is not v1.4\n");
 704                         return -ENODEV;
 705                 }
 706
 707                 dprintk("flags: 0x%x\n", psb->flags1);
 708                 if (psb->flags1) {
 709                         printk(KERN_ERR BFX "unknown flags\n");
 710                         return -ENODEV;
 711                 }
 712
 713                 data->vstable = psb->vstable;
 714                 dprintk("voltage stabilization time: %d(*20us)\n", data->vstable);
 715
 716                 dprintk("flags2: 0x%x\n", psb->flags2);
 717                 data->rvo = psb->flags2 & 3;
 718                 data->irt = ((psb->flags2) >> 2) & 3;
 719                 mvs = ((psb->flags2) >> 4) & 3;
 720                 data->vidmvs = 1 << mvs;
 721                 data->batps = ((psb->flags2) >> 6) & 3;
 722
 723                 dprintk("ramp voltage offset: %d\n", data->rvo);
 724                 dprintk("isochronous relief time: %d\n", data->irt);
 725                 dprintk("maximum voltage step: %d - 0x%x\n", mvs, data->vidmvs);
 726
 727                 dprintk("numpst: 0x%x\n", psb->num_tables);
 728                 cpst = psb->num_tables;
 729                 if ((psb->cpuid == 0x00000fc0) || (psb->cpuid == 0x00000fe0) ){
 730                         thiscpuid = cpuid_eax(CPUID_PROCESSOR_SIGNATURE);
 731                         if ((thiscpuid == 0x00000fc0) || (thiscpuid == 0x00000fe0) ) {
 732                                 cpst = 1;
 733                         }
 734                 }
 735                 if (cpst != 1) {
 736                         printk(KERN_ERR BFX "numpst must be 1\n");
 737                         return -ENODEV;
 738                 }
 739
 740                 data->plllock = psb->plllocktime;
 741                 dprintk("plllocktime: 0x%x (units 1us)\n", psb->plllocktime);
 742                 dprintk("maxfid: 0x%x\n", psb->maxfid);
 743                 dprintk("maxvid: 0x%x\n", psb->maxvid);
 744                 maxvid = psb->maxvid;
 745
 746                 data->numps = psb->numps;
 747                 dprintk("numpstates: 0x%x\n", data->numps);
 748                 return fill_powernow_table(data, (struct pst_s *)(psb+1), maxvid);
 749         }
 750         /*
 751          * If you see this message, complain to BIOS manufacturer. If
 752          * he tells you "we do not support Linux" or some similar
 753          * nonsense, remember that Windows 2000 uses the same legacy
 754          * mechanism that the old Linux PSB driver uses. Tell them it
 755          * is broken with Windows 2000.
 756          *
 757          * The reference to the AMD documentation is chapter 9 in the
 758          * BIOS and Kernel Developer's Guide, which is available on
 759          * www.amd.com
 760          */
 761         printk(KERN_ERR PFX "BIOS error - no PSB or ACPI _PSS objects\n");
 762         return -ENODEV;
 763 }
 764
 765 #ifdef CONFIG_X86_POWERNOW_K8_ACPI
 766 static void powernow_k8_acpi_pst_values(struct powernow_k8_data *data, unsigned int index)
 767 {
 768         if (!data->acpi_data.state_count || (cpu_family == CPU_HW_PSTATE))
 769                 return;
 770
 771         data->irt = (data->acpi_data.states[index].control >> IRT_SHIFT) & IRT_MASK;
 772         data->rvo = (data->acpi_data.states[index].control >> RVO_SHIFT) & RVO_MASK;
 773         data->exttype = (data->acpi_data.states[index].control >> EXT_TYPE_SHIFT) & EXT_TYPE_MASK;
 774         data->plllock = (data->acpi_data.states[index].control >> PLL_L_SHIFT) & PLL_L_MASK;
 775         data->vidmvs = 1 << ((data->acpi_data.states[index].control >> MVS_SHIFT) & MVS_MASK);
 776         data->vstable = (data->acpi_data.states[index].control >> VST_SHIFT) & VST_MASK;
 777 }
 778
 779 static int powernow_k8_cpu_init_acpi(struct powernow_k8_data *data)
 780 {
 781         struct cpufreq_frequency_table *powernow_table;
 782         int ret_val;
 783
 784         if (acpi_processor_register_performance(&data->acpi_data, data->cpu)) {
 785                 dprintk("register performance failed: bad ACPI data\n");
 786                 return -EIO;
 787         }
 788
 789         /* verify the data contained in the ACPI structures */
 790         if (data->acpi_data.state_count <= 1) {
 791                 dprintk("No ACPI P-States\n");
 792                 goto err_out;
 793         }
 794
 795         if ((data->acpi_data.control_register.space_id != ACPI_ADR_SPACE_FIXED_HARDWARE) ||
 796                 (data->acpi_data.status_register.space_id != ACPI_ADR_SPACE_FIXED_HARDWARE)) {
 797                 dprintk("Invalid control/status registers (%x - %x)\n",
 798                         data->acpi_data.control_register.space_id,
 799                         data->acpi_data.status_register.space_id);
 800                 goto err_out;
 801         }
 802
 803         /* fill in data->powernow_table */
 804         powernow_table = kmalloc((sizeof(struct cpufreq_frequency_table)
 805                 * (data->acpi_data.state_count + 1)), GFP_KERNEL);
 806         if (!powernow_table) {
 807                 dprintk("powernow_table memory alloc failure\n");
 808                 goto err_out;
 809         }
 810
 811         if (cpu_family == CPU_HW_PSTATE)
 812                 ret_val = fill_powernow_table_pstate(data, powernow_table);
 813         else
 814                 ret_val = fill_powernow_table_fidvid(data, powernow_table);
 815         if (ret_val)
 816                 goto err_out_mem;
 817
 818         powernow_table[data->acpi_data.state_count].frequency = CPUFREQ_TABLE_END;
 819         powernow_table[data->acpi_data.state_count].index = 0;
 820         data->powernow_table = powernow_table;
 821
 822         /* fill in data */
 823         data->numps = data->acpi_data.state_count;
 824         if (first_cpu(cpu_core_map[data->cpu]) == data->cpu)
 825                 print_basics(data);
 826         powernow_k8_acpi_pst_values(data, 0);
 827
 828         /* notify BIOS that we exist */
 829         acpi_processor_notify_smm(THIS_MODULE);
 830
 831         return 0;
 832
 833 err_out_mem:
 834         kfree(powernow_table);
 835
 836 err_out:
 837         acpi_processor_unregister_performance(&data->acpi_data, data->cpu);
 838
 839         /* data->acpi_data.state_count informs us at ->exit() whether ACPI was used */
 840         data->acpi_data.state_count = 0;
 841
 842         return -ENODEV;
 843 }
 844
 845 static int fill_powernow_table_pstate(struct powernow_k8_data *data, struct cpufreq_frequency_table *powernow_table)
 846 {
 847         int i;
 848
 849         for (i = 0; i < data->acpi_data.state_count; i++) {
 850                 u32 index;
 851                 u32 hi = 0, lo = 0;
 852                 u32 fid;
 853                 u32 did;
 854
 855                 index = data->acpi_data.states[i].control & HW_PSTATE_MASK;
 856                 if (index > MAX_HW_PSTATE) {
 857                         printk(KERN_ERR PFX "invalid pstate %d - bad value %d.\n", i, index);
 858                         printk(KERN_ERR PFX "Please report to BIOS manufacturer\n");
 859                 }
 860                 rdmsr(MSR_PSTATE_DEF_BASE + index, lo, hi);
 861                 if (!(hi & HW_PSTATE_VALID_MASK)) {
 862                         dprintk("invalid pstate %d, ignoring\n", index);
 863                         powernow_table[i].frequency = CPUFREQ_ENTRY_INVALID;
 864                         continue;
 865                 }
 866
 867                 fid = lo & HW_PSTATE_FID_MASK;
 868                 did = (lo & HW_PSTATE_DID_MASK) >> HW_PSTATE_DID_SHIFT;
 869
 870                 dprintk("   %d : fid 0x%x, did 0x%x\n", index, fid, did);
 871
 872                 powernow_table[i].index = index | (fid << HW_FID_INDEX_SHIFT) | (did << HW_DID_INDEX_SHIFT);
 873
 874                 powernow_table[i].frequency = find_khz_freq_from_fiddid(fid, did);
 875
 876                 if (powernow_table[i].frequency != (data->acpi_data.states[i].core_frequency * 1000)) {
 877                         printk(KERN_INFO PFX "invalid freq entries %u kHz vs. %u kHz\n",
 878                                 powernow_table[i].frequency,
 879                                 (unsigned int) (data->acpi_data.states[i].core_frequency * 1000));
 880                         powernow_table[i].frequency = CPUFREQ_ENTRY_INVALID;
 881                         continue;
 882                 }
 883         }
 884         return 0;
 885 }
 886
 887 static int fill_powernow_table_fidvid(struct powernow_k8_data *data, struct cpufreq_frequency_table *powernow_table)
 888 {
 889         int i;
 890         int cntlofreq = 0;
 891         for (i = 0; i < data->acpi_data.state_count; i++) {
 892                 u32 fid;
 893                 u32 vid;
 894
 895                 if (data->exttype) {
 896                         fid = data->acpi_data.states[i].status & EXT_FID_MASK;
 897                         vid = (data->acpi_data.states[i].status >> VID_SHIFT) & EXT_VID_MASK;
 898                 } else {
 899                         fid = data->acpi_data.states[i].control & FID_MASK;
 900                         vid = (data->acpi_data.states[i].control >> VID_SHIFT) & VID_MASK;
 901                 }
 902
 903                 dprintk("   %d : fid 0x%x, vid 0x%x\n", i, fid, vid);
 904
 905                 powernow_table[i].index = fid; /* lower 8 bits */
 906                 powernow_table[i].index |= (vid << 8); /* upper 8 bits */
 907                 powernow_table[i].frequency = find_khz_freq_from_fid(fid);
 908
 909                 /* verify frequency is OK */
 910                 if ((powernow_table[i].frequency > (MAX_FREQ * 1000)) ||
 911                         (powernow_table[i].frequency < (MIN_FREQ * 1000))) {
 912                         dprintk("invalid freq %u kHz, ignoring\n", powernow_table[i].frequency);
 913                         powernow_table[i].frequency = CPUFREQ_ENTRY_INVALID;
 914                         continue;
 915                 }
 916
 917                 /* verify voltage is OK - BIOSs are using "off" to indicate invalid */
 918                 if (vid == VID_OFF) {
 919                         dprintk("invalid vid %u, ignoring\n", vid);
 920                         powernow_table[i].frequency = CPUFREQ_ENTRY_INVALID;
 921                         continue;
 922                 }
 923
 924                 /* verify only 1 entry from the lo frequency table */
 925                 if (fid < HI_FID_TABLE_BOTTOM) {
 926                         if (cntlofreq) {
 927                                 /* if both entries are the same, ignore this one ... */
 928                                 if ((powernow_table[i].frequency != powernow_table[cntlofreq].frequency) ||
 929                                     (powernow_table[i].index != powernow_table[cntlofreq].index)) {
 930                                         printk(KERN_ERR PFX "Too many lo freq table entries\n");
 931                                         return 1;
 932                                 }
 933
 934                                 dprintk("double low frequency table entry, ignoring it.\n");
 935                                 powernow_table[i].frequency = CPUFREQ_ENTRY_INVALID;
 936                                 continue;
 937                         } else
 938                                 cntlofreq = i;
 939                 }
 940
 941                 if (powernow_table[i].frequency != (data->acpi_data.states[i].core_frequency * 1000)) {
 942                         printk(KERN_INFO PFX "invalid freq entries %u kHz vs. %u kHz\n",
 943                                 powernow_table[i].frequency,
 944                                 (unsigned int) (data->acpi_data.states[i].core_frequency * 1000));
 945                         powernow_table[i].frequency = CPUFREQ_ENTRY_INVALID;
 946                         continue;
 947                 }
 948         }
 949         return 0;
 950 }
 951
 952 static void powernow_k8_cpu_exit_acpi(struct powernow_k8_data *data)
 953 {
 954         if (data->acpi_data.state_count)
 955                 acpi_processor_unregister_performance(&data->acpi_data, data->cpu);
 956 }
 957
 958 #else
 959 static int powernow_k8_cpu_init_acpi(struct powernow_k8_data *data) { return -ENODEV; }
 960 static void powernow_k8_cpu_exit_acpi(struct powernow_k8_data *data) { return; }
 961 static void powernow_k8_acpi_pst_values(struct powernow_k8_data *data, unsigned int index) { return; }
 962 #endif /* CONFIG_X86_POWERNOW_K8_ACPI */
 963
 964 /* Take a frequency, and issue the fid/vid transition command */
 965 static int transition_frequency_fidvid(struct powernow_k8_data *data, unsigned int index)
 966 {
 967         u32 fid = 0;
 968         u32 vid = 0;
 969         int res, i;
 970         struct cpufreq_freqs freqs;
 971
 972         dprintk("cpu %d transition to index %u\n", smp_processor_id(), index);
 973
 974         /* fid/vid correctness check for k8 */
 975         /* fid are the lower 8 bits of the index we stored into
 976          * the cpufreq frequency table in find_psb_table, vid
 977          * are the upper 8 bits.
 978          */
 979         fid = data->powernow_table[index].index & 0xFF;
 980         vid = (data->powernow_table[index].index & 0xFF00) >> 8;
 981
 982         dprintk("table matched fid 0x%x, giving vid 0x%x\n", fid, vid);
 983
 984         if (query_current_values_with_pending_wait(data))
 985                 return 1;
 986
 987         if ((data->currvid == vid) && (data->currfid == fid)) {
 988                 dprintk("target matches current values (fid 0x%x, vid 0x%x)\n",
 989                         fid, vid);
 990                 return 0;
 991         }
 992
 993         if ((fid < HI_FID_TABLE_BOTTOM) && (data->currfid < HI_FID_TABLE_BOTTOM)) {
 994                 printk(KERN_ERR PFX
 995                        "ignoring illegal change in lo freq table-%x to 0x%x\n",
 996                        data->currfid, fid);
 997                 return 1;
 998         }
 999
1000         dprintk("cpu %d, changing to fid 0x%x, vid 0x%x\n",
1001                 smp_processor_id(), fid, vid);
1002         freqs.old = find_khz_freq_from_fid(data->currfid);
1003         freqs.new = find_khz_freq_from_fid(fid);
1004
1005         for_each_cpu_mask(i, *(data->available_cores)) {
1006                 freqs.cpu = i;
1007                 cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
1008         }
1009
1010         res = transition_fid_vid(data, fid, vid);
1011         freqs.new = find_khz_freq_from_fid(data->currfid);
1012
1013         for_each_cpu_mask(i, *(data->available_cores)) {
1014                 freqs.cpu = i;
1015                 cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
1016         }
1017         return res;
1018 }
1019
1020 /* Take a frequency, and issue the hardware pstate transition command */
1021 static int transition_frequency_pstate(struct powernow_k8_data *data, unsigned int index)
1022 {
1023         u32 fid = 0;
1024         u32 did = 0;
1025         u32 pstate = 0;
1026         int res, i;
1027         struct cpufreq_freqs freqs;
1028
1029         dprintk("cpu %d transition to index %u\n", smp_processor_id(), index);
1030
1031         /* get fid did for hardware pstate transition */
1032         pstate = index & HW_PSTATE_MASK;
1033         if (pstate > MAX_HW_PSTATE)
1034                 return 0;
1035         fid = (index & HW_FID_INDEX_MASK) >> HW_FID_INDEX_SHIFT;
1036         did = (index & HW_DID_INDEX_MASK) >> HW_DID_INDEX_SHIFT;
1037         freqs.old = find_khz_freq_from_fiddid(data->currfid, data->currdid);
1038         freqs.new = find_khz_freq_from_fiddid(fid, did);
1039
1040         for_each_cpu_mask(i, *(data->available_cores)) {
1041                 freqs.cpu = i;
1042                 cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
1043         }
1044
1045         res = transition_pstate(data, pstate);
1046         data->currfid = find_fid_from_pstate(pstate);
1047         data->currdid = find_did_from_pstate(pstate);
1048         freqs.new = find_khz_freq_from_fiddid(data->currfid, data->currdid);
1049
1050         for_each_cpu_mask(i, *(data->available_cores)) {
1051                 freqs.cpu = i;
1052                 cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
1053         }
1054         return res;
1055 }
1056
1057 /* Driver entry point to switch to the target frequency */
1058 static int powernowk8_target(struct cpufreq_policy *pol, unsigned targfreq, unsigned relation)
1059 {
1060         cpumask_t oldmask = CPU_MASK_ALL;
1061         struct powernow_k8_data *data = powernow_data[pol->cpu];
1062         u32 checkfid;
1063         u32 checkvid;
1064         unsigned int newstate;
1065         int ret = -EIO;
1066
1067         if (!data)
1068                 return -EINVAL;
1069
1070         checkfid = data->currfid;
1071         checkvid = data->currvid;
1072
1073         /* only run on specific CPU from here on */
1074         oldmask = current->cpus_allowed;
1075         set_cpus_allowed(current, cpumask_of_cpu(pol->cpu));
1076
1077         if (smp_processor_id() != pol->cpu) {
1078                 printk(KERN_ERR PFX "limiting to cpu %u failed\n", pol->cpu);
1079                 goto err_out;
1080         }
1081
1082         if (pending_bit_stuck()) {
1083                 printk(KERN_ERR PFX "failing targ, change pending bit set\n");
1084                 goto err_out;
1085         }
1086
1087         dprintk("targ: cpu %d, %d kHz, min %d, max %d, relation %d\n",
1088                 pol->cpu, targfreq, pol->min, pol->max, relation);
1089
1090         if (query_current_values_with_pending_wait(data))
1091                 goto err_out;
1092
1093         if (cpu_family == CPU_HW_PSTATE)
1094                 dprintk("targ: curr fid 0x%x, did 0x%x\n",
1095                         data->currfid, data->currdid);
1096         else {
1097                 dprintk("targ: curr fid 0x%x, vid 0x%x\n",
1098                 data->currfid, data->currvid);
1099
1100                 if ((checkvid != data->currvid) || (checkfid != data->currfid)) {
1101                         printk(KERN_INFO PFX
1102                                 "error - out of sync, fix 0x%x 0x%x, vid 0x%x 0x%x\n",
1103                                 checkfid, data->currfid, checkvid, data->currvid);
1104                 }
1105         }
1106
1107         if (cpufreq_frequency_table_target(pol, data->powernow_table, targfreq, relation, &newstate))
1108                 goto err_out;
1109
1110         mutex_lock(&fidvid_mutex);
1111
1112         powernow_k8_acpi_pst_values(data, newstate);
1113
1114         if (cpu_family == CPU_HW_PSTATE)
1115                 ret = transition_frequency_pstate(data, newstate);
1116         else
1117                 ret = transition_frequency_fidvid(data, newstate);
1118         if (ret) {
1119                 printk(KERN_ERR PFX "transition frequency failed\n");
1120                 ret = 1;
1121                 mutex_unlock(&fidvid_mutex);
1122                 goto err_out;
1123         }
1124         mutex_unlock(&fidvid_mutex);
1125
1126         if (cpu_family == CPU_HW_PSTATE)
1127                 pol->cur = find_khz_freq_from_fiddid(data->currfid, data->currdid);
1128         else
1129                 pol->cur = find_khz_freq_from_fid(data->currfid);
1130         ret = 0;
1131
1132 err_out:
1133         set_cpus_allowed(current, oldmask);
1134         return ret;
1135 }
1136
1137 /* Driver entry point to verify the policy and range of frequencies */
1138 static int powernowk8_verify(struct cpufreq_policy *pol)
1139 {
1140         struct powernow_k8_data *data = powernow_data[pol->cpu];
1141
1142         if (!data)
1143                 return -EINVAL;
1144
1145         return cpufreq_frequency_table_verify(pol, data->powernow_table);
1146 }
1147
1148 /* per CPU init entry point to the driver */
1149 static int __cpuinit powernowk8_cpu_init(struct cpufreq_policy *pol)
1150 {
1151         struct powernow_k8_data *data;
1152         cpumask_t oldmask = CPU_MASK_ALL;
1153         int rc;
1154
1155         if (!cpu_online(pol->cpu))
1156                 return -ENODEV;
1157
1158         if (!check_supported_cpu(pol->cpu))
1159                 return -ENODEV;
1160
1161         data = kzalloc(sizeof(struct powernow_k8_data), GFP_KERNEL);
1162         if (!data) {
1163                 printk(KERN_ERR PFX "unable to alloc powernow_k8_data");
1164                 return -ENOMEM;
1165         }
1166
1167         data->cpu = pol->cpu;
1168
1169         if (powernow_k8_cpu_init_acpi(data)) {
1170                 /*
1171                  * Use the PSB BIOS structure. This is only availabe on
1172                  * an UP version, and is deprecated by AMD.
1173                  */
1174                 if (num_online_cpus() != 1) {
1175                         printk(KERN_ERR PFX "MP systems not supported by PSB BIOS structure\n");
1176                         kfree(data);
1177                         return -ENODEV;
1178                 }
1179                 if (pol->cpu != 0) {
1180                         printk(KERN_ERR PFX "No _PSS objects for CPU other than CPU0\n");
1181                         kfree(data);
1182                         return -ENODEV;
1183                 }
1184                 rc = find_psb_table(data);
1185                 if (rc) {
1186                         kfree(data);
1187                         return -ENODEV;
1188                 }
1189         }
1190
1191         /* only run on specific CPU from here on */
1192         oldmask = current->cpus_allowed;
1193         set_cpus_allowed(current, cpumask_of_cpu(pol->cpu));
1194
1195         if (smp_processor_id() != pol->cpu) {
1196                 printk(KERN_ERR PFX "limiting to cpu %u failed\n", pol->cpu);
1197                 goto err_out;
1198         }
1199
1200         if (pending_bit_stuck()) {
1201                 printk(KERN_ERR PFX "failing init, change pending bit set\n");
1202                 goto err_out;
1203         }
1204
1205         if (query_current_values_with_pending_wait(data))
1206                 goto err_out;
1207
1208         if (cpu_family == CPU_OPTERON)
1209                 fidvid_msr_init();
1210
1211         /* run on any CPU again */
1212         set_cpus_allowed(current, oldmask);
1213
1214         if (cpu_family == CPU_HW_PSTATE)
1215                 pol->cpus = cpumask_of_cpu(pol->cpu);
1216         else
1217                 pol->cpus = cpu_core_map[pol->cpu];
1218         data->available_cores = &(pol->cpus);
1219
1220         /* Take a crude guess here.
1221          * That guess was in microseconds, so multiply with 1000 */
1222         pol->cpuinfo.transition_latency = (((data->rvo + 8) * data->vstable * VST_UNITS_20US)
1223             + (3 * (1 << data->irt) * 10)) * 1000;
1224
1225         if (cpu_family == CPU_HW_PSTATE)
1226                 pol->cur = find_khz_freq_from_fiddid(data->currfid, data->currdid);
1227         else
1228                 pol->cur = find_khz_freq_from_fid(data->currfid);
1229         dprintk("policy current frequency %d kHz\n", pol->cur);
1230
1231         /* min/max the cpu is capable of */
1232         if (cpufreq_frequency_table_cpuinfo(pol, data->powernow_table)) {
1233                 printk(KERN_ERR PFX "invalid powernow_table\n");
1234                 powernow_k8_cpu_exit_acpi(data);
1235                 kfree(data->powernow_table);
1236                 kfree(data);
1237                 return -EINVAL;
1238         }
1239
1240         cpufreq_frequency_table_get_attr(data->powernow_table, pol->cpu);
1241
1242         if (cpu_family == CPU_HW_PSTATE)
1243                 dprintk("cpu_init done, current fid 0x%x, did 0x%x\n",
1244                         data->currfid, data->currdid);
1245         else
1246                 dprintk("cpu_init done, current fid 0x%x, vid 0x%x\n",
1247                         data->currfid, data->currvid);
1248
1249         powernow_data[pol->cpu] = data;
1250
1251         return 0;
1252
1253 err_out:
1254         set_cpus_allowed(current, oldmask);
1255         powernow_k8_cpu_exit_acpi(data);
1256
1257         kfree(data);
1258         return -ENODEV;
1259 }
1260
1261 static int __devexit powernowk8_cpu_exit (struct cpufreq_policy *pol)
1262 {
1263         struct powernow_k8_data *data = powernow_data[pol->cpu];
1264
1265         if (!data)
1266                 return -EINVAL;
1267
1268         powernow_k8_cpu_exit_acpi(data);
1269
1270         cpufreq_frequency_table_put_attr(pol->cpu);
1271
1272         kfree(data->powernow_table);
1273         kfree(data);
1274
1275         return 0;
1276 }
1277
1278 static unsigned int powernowk8_get (unsigned int cpu)
1279 {
1280         struct powernow_k8_data *data;
1281         cpumask_t oldmask = current->cpus_allowed;
1282         unsigned int khz = 0;
1283
1284         data = powernow_data[first_cpu(cpu_core_map[cpu])];
1285
1286         if (!data)
1287                 return -EINVAL;
1288
1289         set_cpus_allowed(current, cpumask_of_cpu(cpu));
1290         if (smp_processor_id() != cpu) {
1291                 printk(KERN_ERR PFX "limiting to CPU %d failed in powernowk8_get\n", cpu);
1292                 set_cpus_allowed(current, oldmask);
1293                 return 0;
1294         }
1295
1296         if (query_current_values_with_pending_wait(data))
1297                 goto out;
1298
1299         if (cpu_family == CPU_HW_PSTATE)
1300                 khz = find_khz_freq_from_fiddid(data->currfid, data->currdid);
1301         else
1302                 khz = find_khz_freq_from_fid(data->currfid);
1303
1304
1305 out:
1306         set_cpus_allowed(current, oldmask);
1307         return khz;
1308 }
1309
1310 static struct freq_attr* powernow_k8_attr[] = {
1311         &cpufreq_freq_attr_scaling_available_freqs,
1312         NULL,
1313 };
1314
1315 static struct cpufreq_driver cpufreq_amd64_driver = {
1316         .verify = powernowk8_verify,
1317         .target = powernowk8_target,
1318         .init = powernowk8_cpu_init,
1319         .exit = __devexit_p(powernowk8_cpu_exit),
1320         .get = powernowk8_get,
1321         .name = "powernow-k8",
1322         .owner = THIS_MODULE,
1323         .attr = powernow_k8_attr,
1324 };
1325
1326 /* driver entry point for init */
1327 static int __cpuinit powernowk8_init(void)
1328 {
1329         unsigned int i, supported_cpus = 0;
1330
1331         for_each_online_cpu(i) {
1332                 if (check_supported_cpu(i))
1333                         supported_cpus++;
1334         }
1335
1336         if (supported_cpus == num_online_cpus()) {
1337                 printk(KERN_INFO PFX "Found %d %s "
1338                         "processors (%d cpu cores) (" VERSION ")\n",
1339                         num_online_nodes(),
1340                         boot_cpu_data.x86_model_id, supported_cpus);
1341                 return cpufreq_register_driver(&cpufreq_amd64_driver);
1342         }
1343
1344         return -ENODEV;
1345 }
1346
1347 /* driver entry point for term */
1348 static void __exit powernowk8_exit(void)
1349 {
1350         dprintk("exit\n");
1351
1352         cpufreq_unregister_driver(&cpufreq_amd64_driver);
1353 }
1354
1355 MODULE_AUTHOR("Paul Devriendt <paul.devriendt@amd.com> and Mark Langsdorf <mark.langsdorf@amd.com>");
1356 MODULE_DESCRIPTION("AMD Athlon 64 and Opteron processor frequency driver.");
1357 MODULE_LICENSE("GPL");
1358
1359 late_initcall(powernowk8_init);
1360 module_exit(powernowk8_exit);