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

   1 /*
   2  *   (c) 2003, 2004 Advanced Micro Devices, Inc.
   3  *  Your use of this code is subject to the terms and conditions of the
   4  *  GNU general public license version 2. See "COPYING" or
   5  *  http://www.gnu.org/licenses/gpl.html
   6  *
   7  *  Support : mark.langsdorf@amd.com
   8  *
   9  *  Based on the powernow-k7.c module written by Dave Jones.
  10  *  (C) 2003 Dave Jones <davej@codemonkey.org.uk> on behalf of SuSE Labs
  11  *  (C) 2004 Dominik Brodowski <linux@brodo.de>
  12  *  (C) 2004 Pavel Machek <pavel@suse.cz>
  13  *  Licensed under the terms of the GNU GPL License version 2.
  14  *  Based upon datasheets & sample CPUs kindly provided by AMD.
  15  *
  16  *  Valuable input gratefully received from Dave Jones, Pavel Machek,
  17  *  Dominik Brodowski, and others.
  18  *  Originally developed by Paul Devriendt.
  19  *  Processor information obtained from Chapter 9 (Power and Thermal Management)
  20  *  of the "BIOS and Kernel Developer's Guide for the AMD Athlon 64 and AMD
  21  *  Opteron Processors" available for download from www.amd.com
  22  *
  23  *  Tables for specific CPUs can be infrerred from
  24  *     http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/30430.pdf
  25  */
  26
  27 #include <linux/kernel.h>
  28 #include <linux/smp.h>
  29 #include <linux/module.h>
  30 #include <linux/init.h>
  31 #include <linux/cpufreq.h>
  32 #include <linux/slab.h>
  33 #include <linux/string.h>
  34 #include <linux/cpumask.h>
  35
  36 #include <asm/msr.h>
  37 #include <asm/io.h>
  38 #include <asm/delay.h>
  39
  40 #ifdef CONFIG_X86_POWERNOW_K8_ACPI
  41 #include <linux/acpi.h>
  42 #include <acpi/processor.h>
  43 #endif
  44
  45 #define PFX "powernow-k8: "
  46 #define BFX PFX "BIOS error: "
  47 #define VERSION "version 1.40.2"
  48 #include "powernow-k8.h"
  49
  50 /* serialize freq changes  */
  51 static DECLARE_MUTEX(fidvid_sem);
  52
  53 static struct powernow_k8_data *powernow_data[NR_CPUS];
  54
  55 #ifndef CONFIG_SMP
  56 static cpumask_t cpu_core_map[1];
  57 #endif
  58
  59 /* Return a frequency in MHz, given an input fid */
  60 static u32 find_freq_from_fid(u32 fid)
  61 {
  62         return 800 + (fid * 100);
  63 }
  64
  65 /* Return a frequency in KHz, given an input fid */
  66 static u32 find_khz_freq_from_fid(u32 fid)
  67 {
  68         return 1000 * find_freq_from_fid(fid);
  69 }
  70
  71 /* Return a voltage in miliVolts, given an input vid */
  72 static u32 find_millivolts_from_vid(struct powernow_k8_data *data, u32 vid)
  73 {
  74         return 1550-vid*25;
  75 }
  76
  77 /* Return the vco fid for an input fid
  78  *
  79  * Each "low" fid has corresponding "high" fid, and you can get to "low" fids
  80  * only from corresponding high fids. This returns "high" fid corresponding to
  81  * "low" one.
  82  */
  83 static u32 convert_fid_to_vco_fid(u32 fid)
  84 {
  85         if (fid < HI_FID_TABLE_BOTTOM) {
  86                 return 8 + (2 * fid);
  87         } else {
  88                 return fid;
  89         }
  90 }
  91
  92 /*
  93  * Return 1 if the pending bit is set. Unless we just instructed the processor
  94  * to transition to a new state, seeing this bit set is really bad news.
  95  */
  96 static int pending_bit_stuck(void)
  97 {
  98         u32 lo, hi;
  99
 100         rdmsr(MSR_FIDVID_STATUS, lo, hi);
 101         return lo & MSR_S_LO_CHANGE_PENDING ? 1 : 0;
 102 }
 103
 104 /*
 105  * Update the global current fid / vid values from the status msr.
 106  * Returns 1 on error.
 107  */
 108 static int query_current_values_with_pending_wait(struct powernow_k8_data *data)
 109 {
 110         u32 lo, hi;
 111         u32 i = 0;
 112
 113         lo = MSR_S_LO_CHANGE_PENDING;
 114         while (lo & MSR_S_LO_CHANGE_PENDING) {
 115                 if (i++ > 0x1000000) {
 116                         printk(KERN_ERR PFX "detected change pending stuck\n");
 117                         return 1;
 118                 }
 119                 rdmsr(MSR_FIDVID_STATUS, lo, hi);
 120         }
 121
 122         data->currvid = hi & MSR_S_HI_CURRENT_VID;
 123         data->currfid = lo & MSR_S_LO_CURRENT_FID;
 124
 125         return 0;
 126 }
 127
 128 /* the isochronous relief time */
 129 static void count_off_irt(struct powernow_k8_data *data)
 130 {
 131         udelay((1 << data->irt) * 10);
 132         return;
 133 }
 134
 135 /* the voltage stabalization time */
 136 static void count_off_vst(struct powernow_k8_data *data)
 137 {
 138         udelay(data->vstable * VST_UNITS_20US);
 139         return;
 140 }
 141
 142 /* need to init the control msr to a safe value (for each cpu) */
 143 static void fidvid_msr_init(void)
 144 {
 145         u32 lo, hi;
 146         u8 fid, vid;
 147
 148         rdmsr(MSR_FIDVID_STATUS, lo, hi);
 149         vid = hi & MSR_S_HI_CURRENT_VID;
 150         fid = lo & MSR_S_LO_CURRENT_FID;
 151         lo = fid | (vid << MSR_C_LO_VID_SHIFT);
 152         hi = MSR_C_HI_STP_GNT_BENIGN;
 153         dprintk("cpu%d, init lo 0x%x, hi 0x%x\n", smp_processor_id(), lo, hi);
 154         wrmsr(MSR_FIDVID_CTL, lo, hi);
 155 }
 156
 157
 158 /* write the new fid value along with the other control fields to the msr */
 159 static int write_new_fid(struct powernow_k8_data *data, u32 fid)
 160 {
 161         u32 lo;
 162         u32 savevid = data->currvid;
 163
 164         if ((fid & INVALID_FID_MASK) || (data->currvid & INVALID_VID_MASK)) {
 165                 printk(KERN_ERR PFX "internal error - overflow on fid write\n");
 166                 return 1;
 167         }
 168
 169         lo = fid | (data->currvid << MSR_C_LO_VID_SHIFT) | MSR_C_LO_INIT_FID_VID;
 170
 171         dprintk("writing fid 0x%x, lo 0x%x, hi 0x%x\n",
 172                 fid, lo, data->plllock * PLL_LOCK_CONVERSION);
 173
 174         wrmsr(MSR_FIDVID_CTL, lo, data->plllock * PLL_LOCK_CONVERSION);
 175
 176         if (query_current_values_with_pending_wait(data))
 177                 return 1;
 178
 179         count_off_irt(data);
 180
 181         if (savevid != data->currvid) {
 182                 printk(KERN_ERR PFX "vid change on fid trans, old 0x%x, new 0x%x\n",
 183                        savevid, data->currvid);
 184                 return 1;
 185         }
 186
 187         if (fid != data->currfid) {
 188                 printk(KERN_ERR PFX "fid trans failed, fid 0x%x, curr 0x%x\n", fid,
 189                         data->currfid);
 190                 return 1;
 191         }
 192
 193         return 0;
 194 }
 195
 196 /* Write a new vid to the hardware */
 197 static int write_new_vid(struct powernow_k8_data *data, u32 vid)
 198 {
 199         u32 lo;
 200         u32 savefid = data->currfid;
 201
 202         if ((data->currfid & INVALID_FID_MASK) || (vid & INVALID_VID_MASK)) {
 203                 printk(KERN_ERR PFX "internal error - overflow on vid write\n");
 204                 return 1;
 205         }
 206
 207         lo = data->currfid | (vid << MSR_C_LO_VID_SHIFT) | MSR_C_LO_INIT_FID_VID;
 208
 209         dprintk("writing vid 0x%x, lo 0x%x, hi 0x%x\n",
 210                 vid, lo, STOP_GRANT_5NS);
 211
 212         wrmsr(MSR_FIDVID_CTL, lo, STOP_GRANT_5NS);
 213
 214         if (query_current_values_with_pending_wait(data))
 215                 return 1;
 216
 217         if (savefid != data->currfid) {
 218                 printk(KERN_ERR PFX "fid changed on vid trans, old 0x%x new 0x%x\n",
 219                        savefid, data->currfid);
 220                 return 1;
 221         }
 222
 223         if (vid != data->currvid) {
 224                 printk(KERN_ERR PFX "vid trans failed, vid 0x%x, curr 0x%x\n", vid,
 225                                 data->currvid);
 226                 return 1;
 227         }
 228
 229         return 0;
 230 }
 231
 232 /*
 233  * Reduce the vid by the max of step or reqvid.
 234  * Decreasing vid codes represent increasing voltages:
 235  * vid of 0 is 1.550V, vid of 0x1e is 0.800V, vid of 0x1f is off.
 236  */
 237 static int decrease_vid_code_by_step(struct powernow_k8_data *data, u32 reqvid, u32 step)
 238 {
 239         if ((data->currvid - reqvid) > step)
 240                 reqvid = data->currvid - step;
 241
 242         if (write_new_vid(data, reqvid))
 243                 return 1;
 244
 245         count_off_vst(data);
 246
 247         return 0;
 248 }
 249
 250 /* Change the fid and vid, by the 3 phases. */
 251 static int transition_fid_vid(struct powernow_k8_data *data, u32 reqfid, u32 reqvid)
 252 {
 253         if (core_voltage_pre_transition(data, reqvid))
 254                 return 1;
 255
 256         if (core_frequency_transition(data, reqfid))
 257                 return 1;
 258
 259         if (core_voltage_post_transition(data, reqvid))
 260                 return 1;
 261
 262         if (query_current_values_with_pending_wait(data))
 263                 return 1;
 264
 265         if ((reqfid != data->currfid) || (reqvid != data->currvid)) {
 266                 printk(KERN_ERR PFX "failed (cpu%d): req 0x%x 0x%x, curr 0x%x 0x%x\n",
 267                                 smp_processor_id(),
 268                                 reqfid, reqvid, data->currfid, data->currvid);
 269                 return 1;
 270         }
 271
 272         dprintk("transitioned (cpu%d): new fid 0x%x, vid 0x%x\n",
 273                 smp_processor_id(), data->currfid, data->currvid);
 274
 275         return 0;
 276 }
 277
 278 /* Phase 1 - core voltage transition ... setup voltage */
 279 static int core_voltage_pre_transition(struct powernow_k8_data *data, u32 reqvid)
 280 {
 281         u32 rvosteps = data->rvo;
 282         u32 savefid = data->currfid;
 283         u32 maxvid, lo;
 284
 285         dprintk("ph1 (cpu%d): start, currfid 0x%x, currvid 0x%x, reqvid 0x%x, rvo 0x%x\n",
 286                 smp_processor_id(),
 287                 data->currfid, data->currvid, reqvid, data->rvo);
 288
 289         rdmsr(MSR_FIDVID_STATUS, lo, maxvid);
 290         maxvid = 0x1f & (maxvid >> 16);
 291         dprintk("ph1 maxvid=0x%x\n", maxvid);
 292         if (reqvid < maxvid) /* lower numbers are higher voltages */
 293                 reqvid = maxvid;
 294
 295         while (data->currvid > reqvid) {
 296                 dprintk("ph1: curr 0x%x, req vid 0x%x\n",
 297                         data->currvid, reqvid);
 298                 if (decrease_vid_code_by_step(data, reqvid, data->vidmvs))
 299                         return 1;
 300         }
 301
 302         while ((rvosteps > 0) && ((data->rvo + data->currvid) > reqvid)) {
 303                 if (data->currvid == maxvid) {
 304                         rvosteps = 0;
 305                 } else {
 306                         dprintk("ph1: changing vid for rvo, req 0x%x\n",
 307                                 data->currvid - 1);
 308                         if (decrease_vid_code_by_step(data, data->currvid - 1, 1))
 309                                 return 1;
 310                         rvosteps--;
 311                 }
 312         }
 313
 314         if (query_current_values_with_pending_wait(data))
 315                 return 1;
 316
 317         if (savefid != data->currfid) {
 318                 printk(KERN_ERR PFX "ph1 err, currfid changed 0x%x\n", data->currfid);
 319                 return 1;
 320         }
 321
 322         dprintk("ph1 complete, currfid 0x%x, currvid 0x%x\n",
 323                 data->currfid, data->currvid);
 324
 325         return 0;
 326 }
 327
 328 /* Phase 2 - core frequency transition */
 329 static int core_frequency_transition(struct powernow_k8_data *data, u32 reqfid)
 330 {
 331         u32 vcoreqfid, vcocurrfid, vcofiddiff, savevid = data->currvid;
 332
 333         if ((reqfid < HI_FID_TABLE_BOTTOM) && (data->currfid < HI_FID_TABLE_BOTTOM)) {
 334                 printk(KERN_ERR PFX "ph2: illegal lo-lo transition 0x%x 0x%x\n",
 335                         reqfid, data->currfid);
 336                 return 1;
 337         }
 338
 339         if (data->currfid == reqfid) {
 340                 printk(KERN_ERR PFX "ph2 null fid transition 0x%x\n", data->currfid);
 341                 return 0;
 342         }
 343
 344         dprintk("ph2 (cpu%d): starting, currfid 0x%x, currvid 0x%x, reqfid 0x%x\n",
 345                 smp_processor_id(),
 346                 data->currfid, data->currvid, reqfid);
 347
 348         vcoreqfid = convert_fid_to_vco_fid(reqfid);
 349         vcocurrfid = convert_fid_to_vco_fid(data->currfid);
 350         vcofiddiff = vcocurrfid > vcoreqfid ? vcocurrfid - vcoreqfid
 351             : vcoreqfid - vcocurrfid;
 352
 353         while (vcofiddiff > 2) {
 354                 if (reqfid > data->currfid) {
 355                         if (data->currfid > LO_FID_TABLE_TOP) {
 356                                 if (write_new_fid(data, data->currfid + 2)) {
 357                                         return 1;
 358                                 }
 359                         } else {
 360                                 if (write_new_fid
 361                                     (data, 2 + convert_fid_to_vco_fid(data->currfid))) {
 362                                         return 1;
 363                                 }
 364                         }
 365                 } else {
 366                         if (write_new_fid(data, data->currfid - 2))
 367                                 return 1;
 368                 }
 369
 370                 vcocurrfid = convert_fid_to_vco_fid(data->currfid);
 371                 vcofiddiff = vcocurrfid > vcoreqfid ? vcocurrfid - vcoreqfid
 372                     : vcoreqfid - vcocurrfid;
 373         }
 374
 375         if (write_new_fid(data, reqfid))
 376                 return 1;
 377
 378         if (query_current_values_with_pending_wait(data))
 379                 return 1;
 380
 381         if (data->currfid != reqfid) {
 382                 printk(KERN_ERR PFX
 383                         "ph2: mismatch, failed fid transition, curr 0x%x, req 0x%x\n",
 384                         data->currfid, reqfid);
 385                 return 1;
 386         }
 387
 388         if (savevid != data->currvid) {
 389                 printk(KERN_ERR PFX "ph2: vid changed, save 0x%x, curr 0x%x\n",
 390                         savevid, data->currvid);
 391                 return 1;
 392         }
 393
 394         dprintk("ph2 complete, currfid 0x%x, currvid 0x%x\n",
 395                 data->currfid, data->currvid);
 396
 397         return 0;
 398 }
 399
 400 /* Phase 3 - core voltage transition flow ... jump to the final vid. */
 401 static int core_voltage_post_transition(struct powernow_k8_data *data, u32 reqvid)
 402 {
 403         u32 savefid = data->currfid;
 404         u32 savereqvid = reqvid;
 405
 406         dprintk("ph3 (cpu%d): starting, currfid 0x%x, currvid 0x%x\n",
 407                 smp_processor_id(),
 408                 data->currfid, data->currvid);
 409
 410         if (reqvid != data->currvid) {
 411                 if (write_new_vid(data, reqvid))
 412                         return 1;
 413
 414                 if (savefid != data->currfid) {
 415                         printk(KERN_ERR PFX
 416                                "ph3: bad fid change, save 0x%x, curr 0x%x\n",
 417                                savefid, data->currfid);
 418                         return 1;
 419                 }
 420
 421                 if (data->currvid != reqvid) {
 422                         printk(KERN_ERR PFX
 423                                "ph3: failed vid transition\n, req 0x%x, curr 0x%x",
 424                                reqvid, data->currvid);
 425                         return 1;
 426                 }
 427         }
 428
 429         if (query_current_values_with_pending_wait(data))
 430                 return 1;
 431
 432         if (savereqvid != data->currvid) {
 433                 dprintk("ph3 failed, currvid 0x%x\n", data->currvid);
 434                 return 1;
 435         }
 436
 437         if (savefid != data->currfid) {
 438                 dprintk("ph3 failed, currfid changed 0x%x\n",
 439                         data->currfid);
 440                 return 1;
 441         }
 442
 443         dprintk("ph3 complete, currfid 0x%x, currvid 0x%x\n",
 444                 data->currfid, data->currvid);
 445
 446         return 0;
 447 }
 448
 449 static int check_supported_cpu(unsigned int cpu)
 450 {
 451         cpumask_t oldmask = CPU_MASK_ALL;
 452         u32 eax, ebx, ecx, edx;
 453         unsigned int rc = 0;
 454
 455         oldmask = current->cpus_allowed;
 456         set_cpus_allowed(current, cpumask_of_cpu(cpu));
 457         schedule();
 458
 459         if (smp_processor_id() != cpu) {
 460                 printk(KERN_ERR "limiting to cpu %u failed\n", cpu);
 461                 goto out;
 462         }
 463
 464         if (current_cpu_data.x86_vendor != X86_VENDOR_AMD)
 465                 goto out;
 466
 467         eax = cpuid_eax(CPUID_PROCESSOR_SIGNATURE);
 468         if (((eax & CPUID_USE_XFAM_XMOD) != CPUID_USE_XFAM_XMOD) ||
 469             ((eax & CPUID_XFAM) != CPUID_XFAM_K8) ||
 470             ((eax & CPUID_XMOD) > CPUID_XMOD_REV_E)) {
 471                 printk(KERN_INFO PFX "Processor cpuid %x not supported\n", eax);
 472                 goto out;
 473         }
 474
 475         eax = cpuid_eax(CPUID_GET_MAX_CAPABILITIES);
 476         if (eax < CPUID_FREQ_VOLT_CAPABILITIES) {
 477                 printk(KERN_INFO PFX
 478                        "No frequency change capabilities detected\n");
 479                 goto out;
 480         }
 481
 482         cpuid(CPUID_FREQ_VOLT_CAPABILITIES, &eax, &ebx, &ecx, &edx);
 483         if ((edx & P_STATE_TRANSITION_CAPABLE) != P_STATE_TRANSITION_CAPABLE) {
 484                 printk(KERN_INFO PFX "Power state transitions not supported\n");
 485                 goto out;
 486         }
 487
 488         rc = 1;
 489
 490 out:
 491         set_cpus_allowed(current, oldmask);
 492         schedule();
 493         return rc;
 494
 495 }
 496
 497 static int check_pst_table(struct powernow_k8_data *data, struct pst_s *pst, u8 maxvid)
 498 {
 499         unsigned int j;
 500         u8 lastfid = 0xff;
 501
 502         for (j = 0; j < data->numps; j++) {
 503                 if (pst[j].vid > LEAST_VID) {
 504                         printk(KERN_ERR PFX "vid %d invalid : 0x%x\n", j, pst[j].vid);
 505                         return -EINVAL;
 506                 }
 507                 if (pst[j].vid < data->rvo) {   /* vid + rvo >= 0 */
 508                         printk(KERN_ERR BFX "0 vid exceeded with pstate %d\n", j);
 509                         return -ENODEV;
 510                 }
 511                 if (pst[j].vid < maxvid + data->rvo) {  /* vid + rvo >= maxvid */
 512                         printk(KERN_ERR BFX "maxvid exceeded with pstate %d\n", j);
 513                         return -ENODEV;
 514                 }
 515                 if ((pst[j].fid > MAX_FID)
 516                     || (pst[j].fid & 1)
 517                     || (j && (pst[j].fid < HI_FID_TABLE_BOTTOM))) {
 518                         /* Only first fid is allowed to be in "low" range */
 519                         printk(KERN_ERR PFX "two low fids - %d : 0x%x\n", j, pst[j].fid);
 520                         return -EINVAL;
 521                 }
 522                 if (pst[j].fid < lastfid)
 523                         lastfid = pst[j].fid;
 524         }
 525         if (lastfid & 1) {
 526                 printk(KERN_ERR PFX "lastfid invalid\n");
 527                 return -EINVAL;
 528         }
 529         if (lastfid > LO_FID_TABLE_TOP)
 530                 printk(KERN_INFO PFX  "first fid not from lo freq table\n");
 531
 532         return 0;
 533 }
 534
 535 static void print_basics(struct powernow_k8_data *data)
 536 {
 537         int j;
 538         for (j = 0; j < data->numps; j++) {
 539                 if (data->powernow_table[j].frequency != CPUFREQ_ENTRY_INVALID)
 540                         printk(KERN_INFO PFX "   %d : fid 0x%x (%d MHz), vid 0x%x (%d mV)\n", j,
 541                                 data->powernow_table[j].index & 0xff,
 542                                 data->powernow_table[j].frequency/1000,
 543                                 data->powernow_table[j].index >> 8,
 544                                 find_millivolts_from_vid(data, data->powernow_table[j].index >> 8));
 545         }
 546         if (data->batps)
 547                 printk(KERN_INFO PFX "Only %d pstates on battery\n", data->batps);
 548 }
 549
 550 static int fill_powernow_table(struct powernow_k8_data *data, struct pst_s *pst, u8 maxvid)
 551 {
 552         struct cpufreq_frequency_table *powernow_table;
 553         unsigned int j;
 554
 555         if (data->batps) {    /* use ACPI support to get full speed on mains power */
 556                 printk(KERN_WARNING PFX "Only %d pstates usable (use ACPI driver for full range\n", data->batps);
 557                 data->numps = data->batps;
 558         }
 559
 560         for ( j=1; j<data->numps; j++ ) {
 561                 if (pst[j-1].fid >= pst[j].fid) {
 562                         printk(KERN_ERR PFX "PST out of sequence\n");
 563                         return -EINVAL;
 564                 }
 565         }
 566
 567         if (data->numps < 2) {
 568                 printk(KERN_ERR PFX "no p states to transition\n");
 569                 return -ENODEV;
 570         }
 571
 572         if (check_pst_table(data, pst, maxvid))
 573                 return -EINVAL;
 574
 575         powernow_table = kmalloc((sizeof(struct cpufreq_frequency_table)
 576                 * (data->numps + 1)), GFP_KERNEL);
 577         if (!powernow_table) {
 578                 printk(KERN_ERR PFX "powernow_table memory alloc failure\n");
 579                 return -ENOMEM;
 580         }
 581
 582         for (j = 0; j < data->numps; j++) {
 583                 powernow_table[j].index = pst[j].fid; /* lower 8 bits */
 584                 powernow_table[j].index |= (pst[j].vid << 8); /* upper 8 bits */
 585                 powernow_table[j].frequency = find_khz_freq_from_fid(pst[j].fid);
 586         }
 587         powernow_table[data->numps].frequency = CPUFREQ_TABLE_END;
 588         powernow_table[data->numps].index = 0;
 589
 590         if (query_current_values_with_pending_wait(data)) {
 591                 kfree(powernow_table);
 592                 return -EIO;
 593         }
 594
 595         dprintk("cfid 0x%x, cvid 0x%x\n", data->currfid, data->currvid);
 596         data->powernow_table = powernow_table;
 597         print_basics(data);
 598
 599         for (j = 0; j < data->numps; j++)
 600                 if ((pst[j].fid==data->currfid) && (pst[j].vid==data->currvid))
 601                         return 0;
 602
 603         dprintk("currfid/vid do not match PST, ignoring\n");
 604         return 0;
 605 }
 606
 607 /* Find and validate the PSB/PST table in BIOS. */
 608 static int find_psb_table(struct powernow_k8_data *data)
 609 {
 610         struct psb_s *psb;
 611         unsigned int i;
 612         u32 mvs;
 613         u8 maxvid;
 614         u32 cpst = 0;
 615         u32 thiscpuid;
 616
 617         for (i = 0xc0000; i < 0xffff0; i += 0x10) {
 618                 /* Scan BIOS looking for the signature. */
 619                 /* It can not be at ffff0 - it is too big. */
 620
 621                 psb = phys_to_virt(i);
 622                 if (memcmp(psb, PSB_ID_STRING, PSB_ID_STRING_LEN) != 0)
 623                         continue;
 624
 625                 dprintk("found PSB header at 0x%p\n", psb);
 626
 627                 dprintk("table vers: 0x%x\n", psb->tableversion);
 628                 if (psb->tableversion != PSB_VERSION_1_4) {
 629                         printk(KERN_INFO BFX "PSB table is not v1.4\n");
 630                         return -ENODEV;
 631                 }
 632
 633                 dprintk("flags: 0x%x\n", psb->flags1);
 634                 if (psb->flags1) {
 635                         printk(KERN_ERR BFX "unknown flags\n");
 636                         return -ENODEV;
 637                 }
 638
 639                 data->vstable = psb->vstable;
 640                 dprintk("voltage stabilization time: %d(*20us)\n", data->vstable);
 641
 642                 dprintk("flags2: 0x%x\n", psb->flags2);
 643                 data->rvo = psb->flags2 & 3;
 644                 data->irt = ((psb->flags2) >> 2) & 3;
 645                 mvs = ((psb->flags2) >> 4) & 3;
 646                 data->vidmvs = 1 << mvs;
 647                 data->batps = ((psb->flags2) >> 6) & 3;
 648
 649                 dprintk("ramp voltage offset: %d\n", data->rvo);
 650                 dprintk("isochronous relief time: %d\n", data->irt);
 651                 dprintk("maximum voltage step: %d - 0x%x\n", mvs, data->vidmvs);
 652
 653                 dprintk("numpst: 0x%x\n", psb->num_tables);
 654                 cpst = psb->num_tables;
 655                 if ((psb->cpuid == 0x00000fc0) || (psb->cpuid == 0x00000fe0) ){
 656                         thiscpuid = cpuid_eax(CPUID_PROCESSOR_SIGNATURE);
 657                         if ((thiscpuid == 0x00000fc0) || (thiscpuid == 0x00000fe0) ) {
 658                                 cpst = 1;
 659                         }
 660                 }
 661                 if (cpst != 1) {
 662                         printk(KERN_ERR BFX "numpst must be 1\n");
 663                         return -ENODEV;
 664                 }
 665
 666                 data->plllock = psb->plllocktime;
 667                 dprintk("plllocktime: 0x%x (units 1us)\n", psb->plllocktime);
 668                 dprintk("maxfid: 0x%x\n", psb->maxfid);
 669                 dprintk("maxvid: 0x%x\n", psb->maxvid);
 670                 maxvid = psb->maxvid;
 671
 672                 data->numps = psb->numps;
 673                 dprintk("numpstates: 0x%x\n", data->numps);
 674                 return fill_powernow_table(data, (struct pst_s *)(psb+1), maxvid);
 675         }
 676         /*
 677          * If you see this message, complain to BIOS manufacturer. If
 678          * he tells you "we do not support Linux" or some similar
 679          * nonsense, remember that Windows 2000 uses the same legacy
 680          * mechanism that the old Linux PSB driver uses. Tell them it
 681          * is broken with Windows 2000.
 682          *
 683          * The reference to the AMD documentation is chapter 9 in the
 684          * BIOS and Kernel Developer's Guide, which is available on
 685          * www.amd.com
 686          */
 687         printk(KERN_INFO PFX "BIOS error - no PSB or ACPI _PSS objects\n");
 688         return -ENODEV;
 689 }
 690
 691 #ifdef CONFIG_X86_POWERNOW_K8_ACPI
 692 static void powernow_k8_acpi_pst_values(struct powernow_k8_data *data, unsigned int index)
 693 {
 694         if (!data->acpi_data.state_count)
 695                 return;
 696
 697         data->irt = (data->acpi_data.states[index].control >> IRT_SHIFT) & IRT_MASK;
 698         data->rvo = (data->acpi_data.states[index].control >> RVO_SHIFT) & RVO_MASK;
 699         data->plllock = (data->acpi_data.states[index].control >> PLL_L_SHIFT) & PLL_L_MASK;
 700         data->vidmvs = 1 << ((data->acpi_data.states[index].control >> MVS_SHIFT) & MVS_MASK);
 701         data->vstable = (data->acpi_data.states[index].control >> VST_SHIFT) & VST_MASK;
 702 }
 703
 704 static int powernow_k8_cpu_init_acpi(struct powernow_k8_data *data)
 705 {
 706         int i;
 707         int cntlofreq = 0;
 708         struct cpufreq_frequency_table *powernow_table;
 709
 710         if (acpi_processor_register_performance(&data->acpi_data, data->cpu)) {
 711                 dprintk("register performance failed: bad ACPI data\n");
 712                 return -EIO;
 713         }
 714
 715         /* verify the data contained in the ACPI structures */
 716         if (data->acpi_data.state_count <= 1) {
 717                 dprintk("No ACPI P-States\n");
 718                 goto err_out;
 719         }
 720
 721         if ((data->acpi_data.control_register.space_id != ACPI_ADR_SPACE_FIXED_HARDWARE) ||
 722                 (data->acpi_data.status_register.space_id != ACPI_ADR_SPACE_FIXED_HARDWARE)) {
 723                 dprintk("Invalid control/status registers (%x - %x)\n",
 724                         data->acpi_data.control_register.space_id,
 725                         data->acpi_data.status_register.space_id);
 726                 goto err_out;
 727         }
 728
 729         /* fill in data->powernow_table */
 730         powernow_table = kmalloc((sizeof(struct cpufreq_frequency_table)
 731                 * (data->acpi_data.state_count + 1)), GFP_KERNEL);
 732         if (!powernow_table) {
 733                 dprintk("powernow_table memory alloc failure\n");
 734                 goto err_out;
 735         }
 736
 737         for (i = 0; i < data->acpi_data.state_count; i++) {
 738                 u32 fid = data->acpi_data.states[i].control & FID_MASK;
 739                 u32 vid = (data->acpi_data.states[i].control >> VID_SHIFT) & VID_MASK;
 740
 741                 dprintk("   %d : fid 0x%x, vid 0x%x\n", i, fid, vid);
 742
 743                 powernow_table[i].index = fid; /* lower 8 bits */
 744                 powernow_table[i].index |= (vid << 8); /* upper 8 bits */
 745                 powernow_table[i].frequency = find_khz_freq_from_fid(fid);
 746
 747                 /* verify frequency is OK */
 748                 if ((powernow_table[i].frequency > (MAX_FREQ * 1000)) ||
 749                         (powernow_table[i].frequency < (MIN_FREQ * 1000))) {
 750                         dprintk("invalid freq %u kHz, ignoring\n", powernow_table[i].frequency);
 751                         powernow_table[i].frequency = CPUFREQ_ENTRY_INVALID;
 752                         continue;
 753                 }
 754
 755                 /* verify voltage is OK - BIOSs are using "off" to indicate invalid */
 756                 if (vid == 0x1f) {
 757                         dprintk("invalid vid %u, ignoring\n", vid);
 758                         powernow_table[i].frequency = CPUFREQ_ENTRY_INVALID;
 759                         continue;
 760                 }
 761
 762                 /* verify only 1 entry from the lo frequency table */
 763                 if (fid < HI_FID_TABLE_BOTTOM) {
 764                         if (cntlofreq) {
 765                                 /* if both entries are the same, ignore this
 766                                  * one...
 767                                  */
 768                                 if ((powernow_table[i].frequency != powernow_table[cntlofreq].frequency) ||
 769                                     (powernow_table[i].index != powernow_table[cntlofreq].index)) {
 770                                         printk(KERN_ERR PFX "Too many lo freq table entries\n");
 771                                         goto err_out_mem;
 772                                 }
 773
 774                                 dprintk("double low frequency table entry, ignoring it.\n");
 775                                 powernow_table[i].frequency = CPUFREQ_ENTRY_INVALID;
 776                                 continue;
 777                         } else
 778                                 cntlofreq = i;
 779                 }
 780
 781                 if (powernow_table[i].frequency != (data->acpi_data.states[i].core_frequency * 1000)) {
 782                         printk(KERN_INFO PFX "invalid freq entries %u kHz vs. %u kHz\n",
 783                                 powernow_table[i].frequency,
 784                                 (unsigned int) (data->acpi_data.states[i].core_frequency * 1000));
 785                         powernow_table[i].frequency = CPUFREQ_ENTRY_INVALID;
 786                         continue;
 787                 }
 788         }
 789
 790         powernow_table[data->acpi_data.state_count].frequency = CPUFREQ_TABLE_END;
 791         powernow_table[data->acpi_data.state_count].index = 0;
 792         data->powernow_table = powernow_table;
 793
 794         /* fill in data */
 795         data->numps = data->acpi_data.state_count;
 796         print_basics(data);
 797         powernow_k8_acpi_pst_values(data, 0);
 798
 799         /* notify BIOS that we exist */
 800         acpi_processor_notify_smm(THIS_MODULE);
 801
 802         return 0;
 803
 804 err_out_mem:
 805         kfree(powernow_table);
 806
 807 err_out:
 808         acpi_processor_unregister_performance(&data->acpi_data, data->cpu);
 809
 810         /* data->acpi_data.state_count informs us at ->exit() whether ACPI was used */
 811         data->acpi_data.state_count = 0;
 812
 813         return -ENODEV;
 814 }
 815
 816 static void powernow_k8_cpu_exit_acpi(struct powernow_k8_data *data)
 817 {
 818         if (data->acpi_data.state_count)
 819                 acpi_processor_unregister_performance(&data->acpi_data, data->cpu);
 820 }
 821
 822 #else
 823 static int powernow_k8_cpu_init_acpi(struct powernow_k8_data *data) { return -ENODEV; }
 824 static void powernow_k8_cpu_exit_acpi(struct powernow_k8_data *data) { return; }
 825 static void powernow_k8_acpi_pst_values(struct powernow_k8_data *data, unsigned int index) { return; }
 826 #endif /* CONFIG_X86_POWERNOW_K8_ACPI */
 827
 828 /* Take a frequency, and issue the fid/vid transition command */
 829 static int transition_frequency(struct powernow_k8_data *data, unsigned int index)
 830 {
 831         u32 fid;
 832         u32 vid;
 833         int res, i;
 834         struct cpufreq_freqs freqs;
 835
 836         dprintk("cpu %d transition to index %u\n", smp_processor_id(), index);
 837
 838         /* fid are the lower 8 bits of the index we stored into
 839          * the cpufreq frequency table in find_psb_table, vid are
 840          * the upper 8 bits.
 841          */
 842
 843         fid = data->powernow_table[index].index & 0xFF;
 844         vid = (data->powernow_table[index].index & 0xFF00) >> 8;
 845
 846         dprintk("table matched fid 0x%x, giving vid 0x%x\n", fid, vid);
 847
 848         if (query_current_values_with_pending_wait(data))
 849                 return 1;
 850
 851         if ((data->currvid == vid) && (data->currfid == fid)) {
 852                 dprintk("target matches current values (fid 0x%x, vid 0x%x)\n",
 853                         fid, vid);
 854                 return 0;
 855         }
 856
 857         if ((fid < HI_FID_TABLE_BOTTOM) && (data->currfid < HI_FID_TABLE_BOTTOM)) {
 858                 printk(KERN_ERR PFX
 859                        "ignoring illegal change in lo freq table-%x to 0x%x\n",
 860                        data->currfid, fid);
 861                 return 1;
 862         }
 863
 864         dprintk("cpu %d, changing to fid 0x%x, vid 0x%x\n",
 865                 smp_processor_id(), fid, vid);
 866
 867         freqs.cpu = data->cpu;
 868         freqs.old = find_khz_freq_from_fid(data->currfid);
 869         freqs.new = find_khz_freq_from_fid(fid);
 870         for_each_cpu_mask(i, cpu_core_map[data->cpu]) {
 871                 freqs.cpu = i;
 872                 cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
 873         }
 874
 875         res = transition_fid_vid(data, fid, vid);
 876
 877         freqs.new = find_khz_freq_from_fid(data->currfid);
 878         for_each_cpu_mask(i, cpu_core_map[data->cpu]) {
 879                 freqs.cpu = i;
 880                 cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
 881         }
 882         return res;
 883 }
 884
 885 /* Driver entry point to switch to the target frequency */
 886 static int powernowk8_target(struct cpufreq_policy *pol, unsigned targfreq, unsigned relation)
 887 {
 888         cpumask_t oldmask = CPU_MASK_ALL;
 889         struct powernow_k8_data *data = powernow_data[pol->cpu];
 890         u32 checkfid = data->currfid;
 891         u32 checkvid = data->currvid;
 892         unsigned int newstate;
 893         int ret = -EIO;
 894         int i;
 895
 896         /* only run on specific CPU from here on */
 897         oldmask = current->cpus_allowed;
 898         set_cpus_allowed(current, cpumask_of_cpu(pol->cpu));
 899         schedule();
 900
 901         if (smp_processor_id() != pol->cpu) {
 902                 printk(KERN_ERR "limiting to cpu %u failed\n", pol->cpu);
 903                 goto err_out;
 904         }
 905
 906         if (pending_bit_stuck()) {
 907                 printk(KERN_ERR PFX "failing targ, change pending bit set\n");
 908                 goto err_out;
 909         }
 910
 911         dprintk("targ: cpu %d, %d kHz, min %d, max %d, relation %d\n",
 912                 pol->cpu, targfreq, pol->min, pol->max, relation);
 913
 914         if (query_current_values_with_pending_wait(data)) {
 915                 ret = -EIO;
 916                 goto err_out;
 917         }
 918
 919         dprintk("targ: curr fid 0x%x, vid 0x%x\n",
 920                 data->currfid, data->currvid);
 921
 922         if ((checkvid != data->currvid) || (checkfid != data->currfid)) {
 923                 printk(KERN_INFO PFX
 924                         "error - out of sync, fix 0x%x 0x%x, vid 0x%x 0x%x\n",
 925                         checkfid, data->currfid, checkvid, data->currvid);
 926         }
 927
 928         if (cpufreq_frequency_table_target(pol, data->powernow_table, targfreq, relation, &newstate))
 929                 goto err_out;
 930
 931         down(&fidvid_sem);
 932
 933         for_each_cpu_mask(i, cpu_core_map[pol->cpu]) {
 934                 /* make sure the sibling is initialized */
 935                 if (!powernow_data[i]) {
 936                         ret = 0;
 937                         up(&fidvid_sem);
 938                         goto err_out;
 939                 }
 940         }
 941
 942         powernow_k8_acpi_pst_values(data, newstate);
 943
 944         if (transition_frequency(data, newstate)) {
 945                 printk(KERN_ERR PFX "transition frequency failed\n");
 946                 ret = 1;
 947                 up(&fidvid_sem);
 948                 goto err_out;
 949         }
 950
 951         /* Update all the fid/vids of our siblings */
 952         for_each_cpu_mask(i, cpu_core_map[pol->cpu]) {
 953                 powernow_data[i]->currvid = data->currvid;
 954                 powernow_data[i]->currfid = data->currfid;
 955         }
 956         up(&fidvid_sem);
 957
 958         pol->cur = find_khz_freq_from_fid(data->currfid);
 959         ret = 0;
 960
 961 err_out:
 962         set_cpus_allowed(current, oldmask);
 963         schedule();
 964
 965         return ret;
 966 }
 967
 968 /* Driver entry point to verify the policy and range of frequencies */
 969 static int powernowk8_verify(struct cpufreq_policy *pol)
 970 {
 971         struct powernow_k8_data *data = powernow_data[pol->cpu];
 972
 973         return cpufreq_frequency_table_verify(pol, data->powernow_table);
 974 }
 975
 976 /* per CPU init entry point to the driver */
 977 static int __init powernowk8_cpu_init(struct cpufreq_policy *pol)
 978 {
 979         struct powernow_k8_data *data;
 980         cpumask_t oldmask = CPU_MASK_ALL;
 981         int rc;
 982
 983         if (!check_supported_cpu(pol->cpu))
 984                 return -ENODEV;
 985
 986         data = kmalloc(sizeof(struct powernow_k8_data), GFP_KERNEL);
 987         if (!data) {
 988                 printk(KERN_ERR PFX "unable to alloc powernow_k8_data");
 989                 return -ENOMEM;
 990         }
 991         memset(data,0,sizeof(struct powernow_k8_data));
 992
 993         data->cpu = pol->cpu;
 994
 995         if (powernow_k8_cpu_init_acpi(data)) {
 996                 /*
 997                  * Use the PSB BIOS structure. This is only availabe on
 998                  * an UP version, and is deprecated by AMD.
 999                  */
1000
1001                 if ((num_online_cpus() != 1) || (num_possible_cpus() != 1)) {
1002                         printk(KERN_ERR PFX "MP systems not supported by PSB BIOS structure\n");
1003                         kfree(data);
1004                         return -ENODEV;
1005                 }
1006                 if (pol->cpu != 0) {
1007                         printk(KERN_ERR PFX "init not cpu 0\n");
1008                         kfree(data);
1009                         return -ENODEV;
1010                 }
1011                 rc = find_psb_table(data);
1012                 if (rc) {
1013                         kfree(data);
1014                         return -ENODEV;
1015                 }
1016         }
1017
1018         /* only run on specific CPU from here on */
1019         oldmask = current->cpus_allowed;
1020         set_cpus_allowed(current, cpumask_of_cpu(pol->cpu));
1021         schedule();
1022
1023         if (smp_processor_id() != pol->cpu) {
1024                 printk(KERN_ERR "limiting to cpu %u failed\n", pol->cpu);
1025                 goto err_out;
1026         }
1027
1028         if (pending_bit_stuck()) {
1029                 printk(KERN_ERR PFX "failing init, change pending bit set\n");
1030                 goto err_out;
1031         }
1032
1033         if (query_current_values_with_pending_wait(data))
1034                 goto err_out;
1035
1036         fidvid_msr_init();
1037
1038         /* run on any CPU again */
1039         set_cpus_allowed(current, oldmask);
1040         schedule();
1041
1042         pol->governor = CPUFREQ_DEFAULT_GOVERNOR;
1043         pol->cpus = cpu_core_map[pol->cpu];
1044
1045         /* Take a crude guess here.
1046          * That guess was in microseconds, so multiply with 1000 */
1047         pol->cpuinfo.transition_latency = (((data->rvo + 8) * data->vstable * VST_UNITS_20US)
1048             + (3 * (1 << data->irt) * 10)) * 1000;
1049
1050         pol->cur = find_khz_freq_from_fid(data->currfid);
1051         dprintk("policy current frequency %d kHz\n", pol->cur);
1052
1053         /* min/max the cpu is capable of */
1054         if (cpufreq_frequency_table_cpuinfo(pol, data->powernow_table)) {
1055                 printk(KERN_ERR PFX "invalid powernow_table\n");
1056                 powernow_k8_cpu_exit_acpi(data);
1057                 kfree(data->powernow_table);
1058                 kfree(data);
1059                 return -EINVAL;
1060         }
1061
1062         cpufreq_frequency_table_get_attr(data->powernow_table, pol->cpu);
1063
1064         printk("cpu_init done, current fid 0x%x, vid 0x%x\n",
1065                data->currfid, data->currvid);
1066
1067         powernow_data[pol->cpu] = data;
1068
1069         return 0;
1070
1071 err_out:
1072         set_cpus_allowed(current, oldmask);
1073         schedule();
1074         powernow_k8_cpu_exit_acpi(data);
1075
1076         kfree(data);
1077         return -ENODEV;
1078 }
1079
1080 static int __devexit powernowk8_cpu_exit (struct cpufreq_policy *pol)
1081 {
1082         struct powernow_k8_data *data = powernow_data[pol->cpu];
1083
1084         if (!data)
1085                 return -EINVAL;
1086
1087         powernow_k8_cpu_exit_acpi(data);
1088
1089         cpufreq_frequency_table_put_attr(pol->cpu);
1090
1091         kfree(data->powernow_table);
1092         kfree(data);
1093
1094         return 0;
1095 }
1096
1097 static unsigned int powernowk8_get (unsigned int cpu)
1098 {
1099         struct powernow_k8_data *data = powernow_data[cpu];
1100         cpumask_t oldmask = current->cpus_allowed;
1101         unsigned int khz = 0;
1102
1103         set_cpus_allowed(current, cpumask_of_cpu(cpu));
1104         if (smp_processor_id() != cpu) {
1105                 printk(KERN_ERR PFX "limiting to CPU %d failed in powernowk8_get\n", cpu);
1106                 set_cpus_allowed(current, oldmask);
1107                 return 0;
1108         }
1109         preempt_disable();
1110
1111         if (query_current_values_with_pending_wait(data))
1112                 goto out;
1113
1114         khz = find_khz_freq_from_fid(data->currfid);
1115
1116  out:
1117         preempt_enable_no_resched();
1118         set_cpus_allowed(current, oldmask);
1119
1120         return khz;
1121 }
1122
1123 static struct freq_attr* powernow_k8_attr[] = {
1124         &cpufreq_freq_attr_scaling_available_freqs,
1125         NULL,
1126 };
1127
1128 static struct cpufreq_driver cpufreq_amd64_driver = {
1129         .verify = powernowk8_verify,
1130         .target = powernowk8_target,
1131         .init = powernowk8_cpu_init,
1132         .exit = __devexit_p(powernowk8_cpu_exit),
1133         .get = powernowk8_get,
1134         .name = "powernow-k8",
1135         .owner = THIS_MODULE,
1136         .attr = powernow_k8_attr,
1137 };
1138
1139 /* driver entry point for init */
1140 static int __init powernowk8_init(void)
1141 {
1142         unsigned int i, supported_cpus = 0;
1143
1144         for (i=0; i<NR_CPUS; i++) {
1145                 if (!cpu_online(i))
1146                         continue;
1147                 if (check_supported_cpu(i))
1148                         supported_cpus++;
1149         }
1150
1151         if (supported_cpus == num_online_cpus()) {
1152                 printk(KERN_INFO PFX "Found %d AMD Athlon 64 / Opteron processors (" VERSION ")\n",
1153                         supported_cpus);
1154                 return cpufreq_register_driver(&cpufreq_amd64_driver);
1155         }
1156
1157         return -ENODEV;
1158 }
1159
1160 /* driver entry point for term */
1161 static void __exit powernowk8_exit(void)
1162 {
1163         dprintk("exit\n");
1164
1165         cpufreq_unregister_driver(&cpufreq_amd64_driver);
1166 }
1167
1168 MODULE_AUTHOR("Paul Devriendt <paul.devriendt@amd.com> and Mark Langsdorf <mark.langsdorf@amd.com.");
1169 MODULE_DESCRIPTION("AMD Athlon 64 and Opteron processor frequency driver.");
1170 MODULE_LICENSE("GPL");
1171
1172 late_initcall(powernowk8_init);
1173 module_exit(powernowk8_exit);
1174