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

   1 /*
   2  *   (c) 2003, 2004, 2005 Advanced Micro Devices, Inc.
   3  *  Your use of this code is subject to the terms and conditions of the
   4  *  GNU general public license version 2. See "COPYING" or
   5  *  http://www.gnu.org/licenses/gpl.html
   6  *
   7  *  Support : mark.langsdorf@amd.com
   8  *
   9  *  Based on the powernow-k7.c module written by Dave Jones.
  10  *  (C) 2003 Dave Jones <davej@codemonkey.org.uk> on behalf of SuSE Labs
  11  *  (C) 2004 Dominik Brodowski <linux@brodo.de>
  12  *  (C) 2004 Pavel Machek <pavel@suse.cz>
  13  *  Licensed under the terms of the GNU GPL License version 2.
  14  *  Based upon datasheets & sample CPUs kindly provided by AMD.
  15  *
  16  *  Valuable input gratefully received from Dave Jones, Pavel Machek,
  17  *  Dominik Brodowski, and others.
  18  *  Originally developed by Paul Devriendt.
  19  *  Processor information obtained from Chapter 9 (Power and Thermal Management)
  20  *  of the "BIOS and Kernel Developer's Guide for the AMD Athlon 64 and AMD
  21  *  Opteron Processors" available for download from www.amd.com
  22  *
  23  *  Tables for specific CPUs can be infrerred from
  24  *     http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/30430.pdf
  25  */
  26
  27 #include <linux/kernel.h>
  28 #include <linux/smp.h>
  29 #include <linux/module.h>
  30 #include <linux/init.h>
  31 #include <linux/cpufreq.h>
  32 #include <linux/slab.h>
  33 #include <linux/string.h>
  34 #include <linux/cpumask.h>
  35 #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 <acpi/processor.h>
  44 #endif
  45
  46 #define PFX "powernow-k8: "
  47 #define BFX PFX "BIOS error: "
  48 #define VERSION "version 1.60.1"
  49 #include "powernow-k8.h"
  50
  51 /* serialize freq changes  */
  52 static DECLARE_MUTEX(fidvid_sem);
  53
  54 static struct powernow_k8_data *powernow_data[NR_CPUS];
  55
  56 #ifndef CONFIG_SMP
  57 static cpumask_t cpu_core_map[1];
  58 #endif
  59
  60 /* Return a frequency in MHz, given an input fid */
  61 static u32 find_freq_from_fid(u32 fid)
  62 {
  63         return 800 + (fid * 100);
  64 }
  65
  66 /* Return a frequency in KHz, given an input fid */
  67 static u32 find_khz_freq_from_fid(u32 fid)
  68 {
  69         return 1000 * find_freq_from_fid(fid);
  70 }
  71
  72 /* Return a voltage in miliVolts, given an input vid */
  73 static u32 find_millivolts_from_vid(struct powernow_k8_data *data, u32 vid)
  74 {
  75         return 1550-vid*25;
  76 }
  77
  78 /* Return the vco fid for an input fid
  79  *
  80  * Each "low" fid has corresponding "high" fid, and you can get to "low" fids
  81  * only from corresponding high fids. This returns "high" fid corresponding to
  82  * "low" one.
  83  */
  84 static u32 convert_fid_to_vco_fid(u32 fid)
  85 {
  86         if (fid < HI_FID_TABLE_BOTTOM)
  87                 return 8 + (2 * fid);
  88         else
  89                 return fid;
  90 }
  91
  92 /*
  93  * Return 1 if the pending bit is set. Unless we just instructed the processor
  94  * to transition to a new state, seeing this bit set is really bad news.
  95  */
  96 static int pending_bit_stuck(void)
  97 {
  98         u32 lo, hi;
  99
 100         rdmsr(MSR_FIDVID_STATUS, lo, hi);
 101         return lo & MSR_S_LO_CHANGE_PENDING ? 1 : 0;
 102 }
 103
 104 /*
 105  * Update the global current fid / vid values from the status msr.
 106  * Returns 1 on error.
 107  */
 108 static int query_current_values_with_pending_wait(struct powernow_k8_data *data)
 109 {
 110         u32 lo, hi;
 111         u32 i = 0;
 112
 113         do {
 114                 if (i++ > 10000) {
 115                         dprintk("detected change pending stuck\n");
 116                         return 1;
 117                 }
 118                 rdmsr(MSR_FIDVID_STATUS, lo, hi);
 119         } while (lo & MSR_S_LO_CHANGE_PENDING);
 120
 121         data->currvid = hi & MSR_S_HI_CURRENT_VID;
 122         data->currfid = lo & MSR_S_LO_CURRENT_FID;
 123
 124         return 0;
 125 }
 126
 127 /* the isochronous relief time */
 128 static void count_off_irt(struct powernow_k8_data *data)
 129 {
 130         udelay((1 << data->irt) * 10);
 131         return;
 132 }
 133
 134 /* the voltage stabalization time */
 135 static void count_off_vst(struct powernow_k8_data *data)
 136 {
 137         udelay(data->vstable * VST_UNITS_20US);
 138         return;
 139 }
 140
 141 /* need to init the control msr to a safe value (for each cpu) */
 142 static void fidvid_msr_init(void)
 143 {
 144         u32 lo, hi;
 145         u8 fid, vid;
 146
 147         rdmsr(MSR_FIDVID_STATUS, lo, hi);
 148         vid = hi & MSR_S_HI_CURRENT_VID;
 149         fid = lo & MSR_S_LO_CURRENT_FID;
 150         lo = fid | (vid << MSR_C_LO_VID_SHIFT);
 151         hi = MSR_C_HI_STP_GNT_BENIGN;
 152         dprintk("cpu%d, init lo 0x%x, hi 0x%x\n", smp_processor_id(), lo, hi);
 153         wrmsr(MSR_FIDVID_CTL, lo, hi);
 154 }
 155
 156
 157 /* write the new fid value along with the other control fields to the msr */
 158 static int write_new_fid(struct powernow_k8_data *data, u32 fid)
 159 {
 160         u32 lo;
 161         u32 savevid = data->currvid;
 162         u32 i = 0;
 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         do {
 175                 wrmsr(MSR_FIDVID_CTL, lo, data->plllock * PLL_LOCK_CONVERSION);
 176                 if (i++ > 100) {
 177                         printk(KERN_ERR PFX "internal error - pending bit very stuck - no further pstate changes possible\n");
 178                         return 1;
 179                 }
 180         } while (query_current_values_with_pending_wait(data));
 181
 182         count_off_irt(data);
 183
 184         if (savevid != data->currvid) {
 185                 printk(KERN_ERR PFX "vid change on fid trans, old 0x%x, new 0x%x\n",
 186                        savevid, data->currvid);
 187                 return 1;
 188         }
 189
 190         if (fid != data->currfid) {
 191                 printk(KERN_ERR PFX "fid trans failed, fid 0x%x, curr 0x%x\n", fid,
 192                         data->currfid);
 193                 return 1;
 194         }
 195
 196         return 0;
 197 }
 198
 199 /* Write a new vid to the hardware */
 200 static int write_new_vid(struct powernow_k8_data *data, u32 vid)
 201 {
 202         u32 lo;
 203         u32 savefid = data->currfid;
 204         int i = 0;
 205
 206         if ((data->currfid & INVALID_FID_MASK) || (vid & INVALID_VID_MASK)) {
 207                 printk(KERN_ERR PFX "internal error - overflow on vid write\n");
 208                 return 1;
 209         }
 210
 211         lo = data->currfid | (vid << MSR_C_LO_VID_SHIFT) | MSR_C_LO_INIT_FID_VID;
 212
 213         dprintk("writing vid 0x%x, lo 0x%x, hi 0x%x\n",
 214                 vid, lo, STOP_GRANT_5NS);
 215
 216         do {
 217                 wrmsr(MSR_FIDVID_CTL, lo, STOP_GRANT_5NS);
 218                 if (i++ > 100) {
 219                         printk(KERN_ERR PFX "internal error - pending bit very stuck - no further pstate changes possible\n");
 220                         return 1;
 221                 }
 222         } while (query_current_values_with_pending_wait(data));
 223
 224         if (savefid != data->currfid) {
 225                 printk(KERN_ERR PFX "fid changed on vid trans, old 0x%x new 0x%x\n",
 226                        savefid, data->currfid);
 227                 return 1;
 228         }
 229
 230         if (vid != data->currvid) {
 231                 printk(KERN_ERR PFX "vid trans failed, vid 0x%x, curr 0x%x\n", vid,
 232                                 data->currvid);
 233                 return 1;
 234         }
 235
 236         return 0;
 237 }
 238
 239 /*
 240  * Reduce the vid by the max of step or reqvid.
 241  * Decreasing vid codes represent increasing voltages:
 242  * vid of 0 is 1.550V, vid of 0x1e is 0.800V, vid of VID_OFF is off.
 243  */
 244 static int decrease_vid_code_by_step(struct powernow_k8_data *data, u32 reqvid, u32 step)
 245 {
 246         if ((data->currvid - reqvid) > step)
 247                 reqvid = data->currvid - step;
 248
 249         if (write_new_vid(data, reqvid))
 250                 return 1;
 251
 252         count_off_vst(data);
 253
 254         return 0;
 255 }
 256
 257 /* Change the fid and vid, by the 3 phases. */
 258 static int transition_fid_vid(struct powernow_k8_data *data, u32 reqfid, u32 reqvid)
 259 {
 260         if (core_voltage_pre_transition(data, reqvid))
 261                 return 1;
 262
 263         if (core_frequency_transition(data, reqfid))
 264                 return 1;
 265
 266         if (core_voltage_post_transition(data, reqvid))
 267                 return 1;
 268
 269         if (query_current_values_with_pending_wait(data))
 270                 return 1;
 271
 272         if ((reqfid != data->currfid) || (reqvid != data->currvid)) {
 273                 printk(KERN_ERR PFX "failed (cpu%d): req 0x%x 0x%x, curr 0x%x 0x%x\n",
 274                                 smp_processor_id(),
 275                                 reqfid, reqvid, data->currfid, data->currvid);
 276                 return 1;
 277         }
 278
 279         dprintk("transitioned (cpu%d): new fid 0x%x, vid 0x%x\n",
 280                 smp_processor_id(), data->currfid, data->currvid);
 281
 282         return 0;
 283 }
 284
 285 /* Phase 1 - core voltage transition ... setup voltage */
 286 static int core_voltage_pre_transition(struct powernow_k8_data *data, u32 reqvid)
 287 {
 288         u32 rvosteps = data->rvo;
 289         u32 savefid = data->currfid;
 290         u32 maxvid, lo;
 291
 292         dprintk("ph1 (cpu%d): start, currfid 0x%x, currvid 0x%x, reqvid 0x%x, rvo 0x%x\n",
 293                 smp_processor_id(),
 294                 data->currfid, data->currvid, reqvid, data->rvo);
 295
 296         rdmsr(MSR_FIDVID_STATUS, lo, maxvid);
 297         maxvid = 0x1f & (maxvid >> 16);
 298         dprintk("ph1 maxvid=0x%x\n", maxvid);
 299         if (reqvid < maxvid) /* lower numbers are higher voltages */
 300                 reqvid = maxvid;
 301
 302         while (data->currvid > reqvid) {
 303                 dprintk("ph1: curr 0x%x, req vid 0x%x\n",
 304                         data->currvid, reqvid);
 305                 if (decrease_vid_code_by_step(data, reqvid, data->vidmvs))
 306                         return 1;
 307         }
 308
 309         while ((rvosteps > 0) && ((data->rvo + data->currvid) > reqvid)) {
 310                 if (data->currvid == maxvid) {
 311                         rvosteps = 0;
 312                 } else {
 313                         dprintk("ph1: changing vid for rvo, req 0x%x\n",
 314                                 data->currvid - 1);
 315                         if (decrease_vid_code_by_step(data, data->currvid - 1, 1))
 316                                 return 1;
 317                         rvosteps--;
 318                 }
 319         }
 320
 321         if (query_current_values_with_pending_wait(data))
 322                 return 1;
 323
 324         if (savefid != data->currfid) {
 325                 printk(KERN_ERR PFX "ph1 err, currfid changed 0x%x\n", data->currfid);
 326                 return 1;
 327         }
 328
 329         dprintk("ph1 complete, currfid 0x%x, currvid 0x%x\n",
 330                 data->currfid, data->currvid);
 331
 332         return 0;
 333 }
 334
 335 /* Phase 2 - core frequency transition */
 336 static int core_frequency_transition(struct powernow_k8_data *data, u32 reqfid)
 337 {
 338         u32 vcoreqfid, vcocurrfid, vcofiddiff, fid_interval, savevid = data->currvid;
 339
 340         if ((reqfid < HI_FID_TABLE_BOTTOM) && (data->currfid < HI_FID_TABLE_BOTTOM)) {
 341                 printk(KERN_ERR PFX "ph2: illegal lo-lo transition 0x%x 0x%x\n",
 342                         reqfid, data->currfid);
 343                 return 1;
 344         }
 345
 346         if (data->currfid == reqfid) {
 347                 printk(KERN_ERR PFX "ph2 null fid transition 0x%x\n", data->currfid);
 348                 return 0;
 349         }
 350
 351         dprintk("ph2 (cpu%d): starting, currfid 0x%x, currvid 0x%x, reqfid 0x%x\n",
 352                 smp_processor_id(),
 353                 data->currfid, data->currvid, reqfid);
 354
 355         vcoreqfid = convert_fid_to_vco_fid(reqfid);
 356         vcocurrfid = convert_fid_to_vco_fid(data->currfid);
 357         vcofiddiff = vcocurrfid > vcoreqfid ? vcocurrfid - vcoreqfid
 358             : vcoreqfid - vcocurrfid;
 359
 360         while (vcofiddiff > 2) {
 361                 (data->currfid & 1) ? (fid_interval = 1) : (fid_interval = 2);
 362
 363                 if (reqfid > data->currfid) {
 364                         if (data->currfid > LO_FID_TABLE_TOP) {
 365                                 if (write_new_fid(data, data->currfid + fid_interval)) {
 366                                         return 1;
 367                                 }
 368                         } else {
 369                                 if (write_new_fid
 370                                     (data, 2 + convert_fid_to_vco_fid(data->currfid))) {
 371                                         return 1;
 372                                 }
 373                         }
 374                 } else {
 375                         if (write_new_fid(data, data->currfid - fid_interval))
 376                                 return 1;
 377                 }
 378
 379                 vcocurrfid = convert_fid_to_vco_fid(data->currfid);
 380                 vcofiddiff = vcocurrfid > vcoreqfid ? vcocurrfid - vcoreqfid
 381                     : vcoreqfid - vcocurrfid;
 382         }
 383
 384         if (write_new_fid(data, reqfid))
 385                 return 1;
 386
 387         if (query_current_values_with_pending_wait(data))
 388                 return 1;
 389
 390         if (data->currfid != reqfid) {
 391                 printk(KERN_ERR PFX
 392                         "ph2: mismatch, failed fid transition, curr 0x%x, req 0x%x\n",
 393                         data->currfid, reqfid);
 394                 return 1;
 395         }
 396
 397         if (savevid != data->currvid) {
 398                 printk(KERN_ERR PFX "ph2: vid changed, save 0x%x, curr 0x%x\n",
 399                         savevid, data->currvid);
 400                 return 1;
 401         }
 402
 403         dprintk("ph2 complete, currfid 0x%x, currvid 0x%x\n",
 404                 data->currfid, data->currvid);
 405
 406         return 0;
 407 }
 408
 409 /* Phase 3 - core voltage transition flow ... jump to the final vid. */
 410 static int core_voltage_post_transition(struct powernow_k8_data *data, u32 reqvid)
 411 {
 412         u32 savefid = data->currfid;
 413         u32 savereqvid = reqvid;
 414
 415         dprintk("ph3 (cpu%d): starting, currfid 0x%x, currvid 0x%x\n",
 416                 smp_processor_id(),
 417                 data->currfid, data->currvid);
 418
 419         if (reqvid != data->currvid) {
 420                 if (write_new_vid(data, reqvid))
 421                         return 1;
 422
 423                 if (savefid != data->currfid) {
 424                         printk(KERN_ERR PFX
 425                                "ph3: bad fid change, save 0x%x, curr 0x%x\n",
 426                                savefid, data->currfid);
 427                         return 1;
 428                 }
 429
 430                 if (data->currvid != reqvid) {
 431                         printk(KERN_ERR PFX
 432                                "ph3: failed vid transition\n, req 0x%x, curr 0x%x",
 433                                reqvid, data->currvid);
 434                         return 1;
 435                 }
 436         }
 437
 438         if (query_current_values_with_pending_wait(data))
 439                 return 1;
 440
 441         if (savereqvid != data->currvid) {
 442                 dprintk("ph3 failed, currvid 0x%x\n", data->currvid);
 443                 return 1;
 444         }
 445
 446         if (savefid != data->currfid) {
 447                 dprintk("ph3 failed, currfid changed 0x%x\n",
 448                         data->currfid);
 449                 return 1;
 450         }
 451
 452         dprintk("ph3 complete, currfid 0x%x, currvid 0x%x\n",
 453                 data->currfid, data->currvid);
 454
 455         return 0;
 456 }
 457
 458 static int check_supported_cpu(unsigned int cpu)
 459 {
 460         cpumask_t oldmask = CPU_MASK_ALL;
 461         u32 eax, ebx, ecx, edx;
 462         unsigned int rc = 0;
 463
 464         oldmask = current->cpus_allowed;
 465         set_cpus_allowed(current, cpumask_of_cpu(cpu));
 466
 467         if (smp_processor_id() != cpu) {
 468                 printk(KERN_ERR PFX "limiting to cpu %u failed\n", cpu);
 469                 goto out;
 470         }
 471
 472         if (current_cpu_data.x86_vendor != X86_VENDOR_AMD)
 473                 goto out;
 474
 475         eax = cpuid_eax(CPUID_PROCESSOR_SIGNATURE);
 476         if ((eax & CPUID_XFAM) != CPUID_XFAM_K8)
 477                 goto out;
 478
 479         if (((eax & CPUID_USE_XFAM_XMOD) != CPUID_USE_XFAM_XMOD) ||
 480             ((eax & CPUID_XMOD) > CPUID_XMOD_REV_G)) {
 481                 printk(KERN_INFO PFX "Processor cpuid %x not supported\n", eax);
 482                 goto out;
 483         }
 484
 485         eax = cpuid_eax(CPUID_GET_MAX_CAPABILITIES);
 486         if (eax < CPUID_FREQ_VOLT_CAPABILITIES) {
 487                 printk(KERN_INFO PFX
 488                        "No frequency change capabilities detected\n");
 489                 goto out;
 490         }
 491
 492         cpuid(CPUID_FREQ_VOLT_CAPABILITIES, &eax, &ebx, &ecx, &edx);
 493         if ((edx & P_STATE_TRANSITION_CAPABLE) != P_STATE_TRANSITION_CAPABLE) {
 494                 printk(KERN_INFO PFX "Power state transitions not supported\n");
 495                 goto out;
 496         }
 497
 498         rc = 1;
 499
 500 out:
 501         set_cpus_allowed(current, oldmask);
 502         return rc;
 503 }
 504
 505 static int check_pst_table(struct powernow_k8_data *data, struct pst_s *pst, u8 maxvid)
 506 {
 507         unsigned int j;
 508         u8 lastfid = 0xff;
 509
 510         for (j = 0; j < data->numps; j++) {
 511                 if (pst[j].vid > LEAST_VID) {
 512                         printk(KERN_ERR PFX "vid %d invalid : 0x%x\n", j, pst[j].vid);
 513                         return -EINVAL;
 514                 }
 515                 if (pst[j].vid < data->rvo) {   /* vid + rvo >= 0 */
 516                         printk(KERN_ERR BFX "0 vid exceeded with pstate %d\n", j);
 517                         return -ENODEV;
 518                 }
 519                 if (pst[j].vid < maxvid + data->rvo) {  /* vid + rvo >= maxvid */
 520                         printk(KERN_ERR BFX "maxvid exceeded with pstate %d\n", j);
 521                         return -ENODEV;
 522                 }
 523                 if (pst[j].fid > MAX_FID) {
 524                         printk(KERN_ERR BFX "maxfid exceeded with pstate %d\n", j);
 525                         return -ENODEV;
 526                 }
 527                 if (j && (pst[j].fid < HI_FID_TABLE_BOTTOM)) {
 528                         /* Only first fid is allowed to be in "low" range */
 529                         printk(KERN_ERR BFX "two low fids - %d : 0x%x\n", j, pst[j].fid);
 530                         return -EINVAL;
 531                 }
 532                 if (pst[j].fid < lastfid)
 533                         lastfid = pst[j].fid;
 534         }
 535         if (lastfid & 1) {
 536                 printk(KERN_ERR BFX "lastfid invalid\n");
 537                 return -EINVAL;
 538         }
 539         if (lastfid > LO_FID_TABLE_TOP)
 540                 printk(KERN_INFO BFX  "first fid not from lo freq table\n");
 541
 542         return 0;
 543 }
 544
 545 static void print_basics(struct powernow_k8_data *data)
 546 {
 547         int j;
 548         for (j = 0; j < data->numps; j++) {
 549                 if (data->powernow_table[j].frequency != CPUFREQ_ENTRY_INVALID)
 550                         printk(KERN_INFO PFX "   %d : fid 0x%x (%d MHz), vid 0x%x (%d mV)\n", j,
 551                                 data->powernow_table[j].index & 0xff,
 552                                 data->powernow_table[j].frequency/1000,
 553                                 data->powernow_table[j].index >> 8,
 554                                 find_millivolts_from_vid(data, data->powernow_table[j].index >> 8));
 555         }
 556         if (data->batps)
 557                 printk(KERN_INFO PFX "Only %d pstates on battery\n", data->batps);
 558 }
 559
 560 static int fill_powernow_table(struct powernow_k8_data *data, struct pst_s *pst, u8 maxvid)
 561 {
 562         struct cpufreq_frequency_table *powernow_table;
 563         unsigned int j;
 564
 565         if (data->batps) {    /* use ACPI support to get full speed on mains power */
 566                 printk(KERN_WARNING PFX "Only %d pstates usable (use ACPI driver for full range\n", data->batps);
 567                 data->numps = data->batps;
 568         }
 569
 570         for ( j=1; j<data->numps; j++ ) {
 571                 if (pst[j-1].fid >= pst[j].fid) {
 572                         printk(KERN_ERR PFX "PST out of sequence\n");
 573                         return -EINVAL;
 574                 }
 575         }
 576
 577         if (data->numps < 2) {
 578                 printk(KERN_ERR PFX "no p states to transition\n");
 579                 return -ENODEV;
 580         }
 581
 582         if (check_pst_table(data, pst, maxvid))
 583                 return -EINVAL;
 584
 585         powernow_table = kmalloc((sizeof(struct cpufreq_frequency_table)
 586                 * (data->numps + 1)), GFP_KERNEL);
 587         if (!powernow_table) {
 588                 printk(KERN_ERR PFX "powernow_table memory alloc failure\n");
 589                 return -ENOMEM;
 590         }
 591
 592         for (j = 0; j < data->numps; j++) {
 593                 powernow_table[j].index = pst[j].fid; /* lower 8 bits */
 594                 powernow_table[j].index |= (pst[j].vid << 8); /* upper 8 bits */
 595                 powernow_table[j].frequency = find_khz_freq_from_fid(pst[j].fid);
 596         }
 597         powernow_table[data->numps].frequency = CPUFREQ_TABLE_END;
 598         powernow_table[data->numps].index = 0;
 599
 600         if (query_current_values_with_pending_wait(data)) {
 601                 kfree(powernow_table);
 602                 return -EIO;
 603         }
 604
 605         dprintk("cfid 0x%x, cvid 0x%x\n", data->currfid, data->currvid);
 606         data->powernow_table = powernow_table;
 607         print_basics(data);
 608
 609         for (j = 0; j < data->numps; j++)
 610                 if ((pst[j].fid==data->currfid) && (pst[j].vid==data->currvid))
 611                         return 0;
 612
 613         dprintk("currfid/vid do not match PST, ignoring\n");
 614         return 0;
 615 }
 616
 617 /* Find and validate the PSB/PST table in BIOS. */
 618 static int find_psb_table(struct powernow_k8_data *data)
 619 {
 620         struct psb_s *psb;
 621         unsigned int i;
 622         u32 mvs;
 623         u8 maxvid;
 624         u32 cpst = 0;
 625         u32 thiscpuid;
 626
 627         for (i = 0xc0000; i < 0xffff0; i += 0x10) {
 628                 /* Scan BIOS looking for the signature. */
 629                 /* It can not be at ffff0 - it is too big. */
 630
 631                 psb = phys_to_virt(i);
 632                 if (memcmp(psb, PSB_ID_STRING, PSB_ID_STRING_LEN) != 0)
 633                         continue;
 634
 635                 dprintk("found PSB header at 0x%p\n", psb);
 636
 637                 dprintk("table vers: 0x%x\n", psb->tableversion);
 638                 if (psb->tableversion != PSB_VERSION_1_4) {
 639                         printk(KERN_ERR BFX "PSB table is not v1.4\n");
 640                         return -ENODEV;
 641                 }
 642
 643                 dprintk("flags: 0x%x\n", psb->flags1);
 644                 if (psb->flags1) {
 645                         printk(KERN_ERR BFX "unknown flags\n");
 646                         return -ENODEV;
 647                 }
 648
 649                 data->vstable = psb->vstable;
 650                 dprintk("voltage stabilization time: %d(*20us)\n", data->vstable);
 651
 652                 dprintk("flags2: 0x%x\n", psb->flags2);
 653                 data->rvo = psb->flags2 & 3;
 654                 data->irt = ((psb->flags2) >> 2) & 3;
 655                 mvs = ((psb->flags2) >> 4) & 3;
 656                 data->vidmvs = 1 << mvs;
 657                 data->batps = ((psb->flags2) >> 6) & 3;
 658
 659                 dprintk("ramp voltage offset: %d\n", data->rvo);
 660                 dprintk("isochronous relief time: %d\n", data->irt);
 661                 dprintk("maximum voltage step: %d - 0x%x\n", mvs, data->vidmvs);
 662
 663                 dprintk("numpst: 0x%x\n", psb->num_tables);
 664                 cpst = psb->num_tables;
 665                 if ((psb->cpuid == 0x00000fc0) || (psb->cpuid == 0x00000fe0) ){
 666                         thiscpuid = cpuid_eax(CPUID_PROCESSOR_SIGNATURE);
 667                         if ((thiscpuid == 0x00000fc0) || (thiscpuid == 0x00000fe0) ) {
 668                                 cpst = 1;
 669                         }
 670                 }
 671                 if (cpst != 1) {
 672                         printk(KERN_ERR BFX "numpst must be 1\n");
 673                         return -ENODEV;
 674                 }
 675
 676                 data->plllock = psb->plllocktime;
 677                 dprintk("plllocktime: 0x%x (units 1us)\n", psb->plllocktime);
 678                 dprintk("maxfid: 0x%x\n", psb->maxfid);
 679                 dprintk("maxvid: 0x%x\n", psb->maxvid);
 680                 maxvid = psb->maxvid;
 681
 682                 data->numps = psb->numps;
 683                 dprintk("numpstates: 0x%x\n", data->numps);
 684                 return fill_powernow_table(data, (struct pst_s *)(psb+1), maxvid);
 685         }
 686         /*
 687          * If you see this message, complain to BIOS manufacturer. If
 688          * he tells you "we do not support Linux" or some similar
 689          * nonsense, remember that Windows 2000 uses the same legacy
 690          * mechanism that the old Linux PSB driver uses. Tell them it
 691          * is broken with Windows 2000.
 692          *
 693          * The reference to the AMD documentation is chapter 9 in the
 694          * BIOS and Kernel Developer's Guide, which is available on
 695          * www.amd.com
 696          */
 697         printk(KERN_ERR PFX "BIOS error - no PSB or ACPI _PSS objects\n");
 698         return -ENODEV;
 699 }
 700
 701 #ifdef CONFIG_X86_POWERNOW_K8_ACPI
 702 static void powernow_k8_acpi_pst_values(struct powernow_k8_data *data, unsigned int index)
 703 {
 704         if (!data->acpi_data.state_count)
 705                 return;
 706
 707         data->irt = (data->acpi_data.states[index].control >> IRT_SHIFT) & IRT_MASK;
 708         data->rvo = (data->acpi_data.states[index].control >> RVO_SHIFT) & RVO_MASK;
 709         data->exttype = (data->acpi_data.states[index].control >> EXT_TYPE_SHIFT) & EXT_TYPE_MASK;
 710         data->plllock = (data->acpi_data.states[index].control >> PLL_L_SHIFT) & PLL_L_MASK;
 711         data->vidmvs = 1 << ((data->acpi_data.states[index].control >> MVS_SHIFT) & MVS_MASK);
 712         data->vstable = (data->acpi_data.states[index].control >> VST_SHIFT) & VST_MASK;
 713 }
 714
 715 static int powernow_k8_cpu_init_acpi(struct powernow_k8_data *data)
 716 {
 717         int i;
 718         int cntlofreq = 0;
 719         struct cpufreq_frequency_table *powernow_table;
 720
 721         if (acpi_processor_register_performance(&data->acpi_data, data->cpu)) {
 722                 dprintk("register performance failed: bad ACPI data\n");
 723                 return -EIO;
 724         }
 725
 726         /* verify the data contained in the ACPI structures */
 727         if (data->acpi_data.state_count <= 1) {
 728                 dprintk("No ACPI P-States\n");
 729                 goto err_out;
 730         }
 731
 732         if ((data->acpi_data.control_register.space_id != ACPI_ADR_SPACE_FIXED_HARDWARE) ||
 733                 (data->acpi_data.status_register.space_id != ACPI_ADR_SPACE_FIXED_HARDWARE)) {
 734                 dprintk("Invalid control/status registers (%x - %x)\n",
 735                         data->acpi_data.control_register.space_id,
 736                         data->acpi_data.status_register.space_id);
 737                 goto err_out;
 738         }
 739
 740         /* fill in data->powernow_table */
 741         powernow_table = kmalloc((sizeof(struct cpufreq_frequency_table)
 742                 * (data->acpi_data.state_count + 1)), GFP_KERNEL);
 743         if (!powernow_table) {
 744                 dprintk("powernow_table memory alloc failure\n");
 745                 goto err_out;
 746         }
 747
 748         for (i = 0; i < data->acpi_data.state_count; i++) {
 749                 u32 fid;
 750                 u32 vid;
 751
 752                 if (data->exttype) {
 753                         fid = data->acpi_data.states[i].status & FID_MASK;
 754                         vid = (data->acpi_data.states[i].status >> VID_SHIFT) & VID_MASK;
 755                 } else {
 756                         fid = data->acpi_data.states[i].control & FID_MASK;
 757                         vid = (data->acpi_data.states[i].control >> VID_SHIFT) & VID_MASK;
 758                 }
 759
 760                 dprintk("   %d : fid 0x%x, vid 0x%x\n", i, fid, vid);
 761
 762                 powernow_table[i].index = fid; /* lower 8 bits */
 763                 powernow_table[i].index |= (vid << 8); /* upper 8 bits */
 764                 powernow_table[i].frequency = find_khz_freq_from_fid(fid);
 765
 766                 /* verify frequency is OK */
 767                 if ((powernow_table[i].frequency > (MAX_FREQ * 1000)) ||
 768                         (powernow_table[i].frequency < (MIN_FREQ * 1000))) {
 769                         dprintk("invalid freq %u kHz, ignoring\n", powernow_table[i].frequency);
 770                         powernow_table[i].frequency = CPUFREQ_ENTRY_INVALID;
 771                         continue;
 772                 }
 773
 774                 /* verify voltage is OK - BIOSs are using "off" to indicate invalid */
 775                 if (vid == VID_OFF) {
 776                         dprintk("invalid vid %u, ignoring\n", vid);
 777                         powernow_table[i].frequency = CPUFREQ_ENTRY_INVALID;
 778                         continue;
 779                 }
 780
 781                 /* verify only 1 entry from the lo frequency table */
 782                 if (fid < HI_FID_TABLE_BOTTOM) {
 783                         if (cntlofreq) {
 784                                 /* if both entries are the same, ignore this one ... */
 785                                 if ((powernow_table[i].frequency != powernow_table[cntlofreq].frequency) ||
 786                                     (powernow_table[i].index != powernow_table[cntlofreq].index)) {
 787                                         printk(KERN_ERR PFX "Too many lo freq table entries\n");
 788                                         goto err_out_mem;
 789                                 }
 790
 791                                 dprintk("double low frequency table entry, ignoring it.\n");
 792                                 powernow_table[i].frequency = CPUFREQ_ENTRY_INVALID;
 793                                 continue;
 794                         } else
 795                                 cntlofreq = i;
 796                 }
 797
 798                 if (powernow_table[i].frequency != (data->acpi_data.states[i].core_frequency * 1000)) {
 799                         printk(KERN_INFO PFX "invalid freq entries %u kHz vs. %u kHz\n",
 800                                 powernow_table[i].frequency,
 801                                 (unsigned int) (data->acpi_data.states[i].core_frequency * 1000));
 802                         powernow_table[i].frequency = CPUFREQ_ENTRY_INVALID;
 803                         continue;
 804                 }
 805         }
 806
 807         powernow_table[data->acpi_data.state_count].frequency = CPUFREQ_TABLE_END;
 808         powernow_table[data->acpi_data.state_count].index = 0;
 809         data->powernow_table = powernow_table;
 810
 811         /* fill in data */
 812         data->numps = data->acpi_data.state_count;
 813         print_basics(data);
 814         powernow_k8_acpi_pst_values(data, 0);
 815
 816         /* notify BIOS that we exist */
 817         acpi_processor_notify_smm(THIS_MODULE);
 818
 819         return 0;
 820
 821 err_out_mem:
 822         kfree(powernow_table);
 823
 824 err_out:
 825         acpi_processor_unregister_performance(&data->acpi_data, data->cpu);
 826
 827         /* data->acpi_data.state_count informs us at ->exit() whether ACPI was used */
 828         data->acpi_data.state_count = 0;
 829
 830         return -ENODEV;
 831 }
 832
 833 static void powernow_k8_cpu_exit_acpi(struct powernow_k8_data *data)
 834 {
 835         if (data->acpi_data.state_count)
 836                 acpi_processor_unregister_performance(&data->acpi_data, data->cpu);
 837 }
 838
 839 #else
 840 static int powernow_k8_cpu_init_acpi(struct powernow_k8_data *data) { return -ENODEV; }
 841 static void powernow_k8_cpu_exit_acpi(struct powernow_k8_data *data) { return; }
 842 static void powernow_k8_acpi_pst_values(struct powernow_k8_data *data, unsigned int index) { return; }
 843 #endif /* CONFIG_X86_POWERNOW_K8_ACPI */
 844
 845 /* Take a frequency, and issue the fid/vid transition command */
 846 static int transition_frequency(struct powernow_k8_data *data, unsigned int index)
 847 {
 848         u32 fid;
 849         u32 vid;
 850         int res, i;
 851         struct cpufreq_freqs freqs;
 852
 853         dprintk("cpu %d transition to index %u\n", smp_processor_id(), index);
 854
 855         /* fid are the lower 8 bits of the index we stored into
 856          * the cpufreq frequency table in find_psb_table, vid are
 857          * the upper 8 bits.
 858          */
 859
 860         fid = data->powernow_table[index].index & 0xFF;
 861         vid = (data->powernow_table[index].index & 0xFF00) >> 8;
 862
 863         dprintk("table matched fid 0x%x, giving vid 0x%x\n", fid, vid);
 864
 865         if (query_current_values_with_pending_wait(data))
 866                 return 1;
 867
 868         if ((data->currvid == vid) && (data->currfid == fid)) {
 869                 dprintk("target matches current values (fid 0x%x, vid 0x%x)\n",
 870                         fid, vid);
 871                 return 0;
 872         }
 873
 874         if ((fid < HI_FID_TABLE_BOTTOM) && (data->currfid < HI_FID_TABLE_BOTTOM)) {
 875                 printk(KERN_ERR PFX
 876                        "ignoring illegal change in lo freq table-%x to 0x%x\n",
 877                        data->currfid, fid);
 878                 return 1;
 879         }
 880
 881         dprintk("cpu %d, changing to fid 0x%x, vid 0x%x\n",
 882                 smp_processor_id(), fid, vid);
 883
 884         freqs.cpu = data->cpu;
 885         freqs.old = find_khz_freq_from_fid(data->currfid);
 886         freqs.new = find_khz_freq_from_fid(fid);
 887         for_each_cpu_mask(i, cpu_core_map[data->cpu]) {
 888                 freqs.cpu = i;
 889                 cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
 890         }
 891
 892         res = transition_fid_vid(data, fid, vid);
 893
 894         freqs.new = find_khz_freq_from_fid(data->currfid);
 895         for_each_cpu_mask(i, cpu_core_map[data->cpu]) {
 896                 freqs.cpu = i;
 897                 cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
 898         }
 899         return res;
 900 }
 901
 902 /* Driver entry point to switch to the target frequency */
 903 static int powernowk8_target(struct cpufreq_policy *pol, unsigned targfreq, unsigned relation)
 904 {
 905         cpumask_t oldmask = CPU_MASK_ALL;
 906         struct powernow_k8_data *data = powernow_data[pol->cpu];
 907         u32 checkfid = data->currfid;
 908         u32 checkvid = data->currvid;
 909         unsigned int newstate;
 910         int ret = -EIO;
 911
 912         /* only run on specific CPU from here on */
 913         oldmask = current->cpus_allowed;
 914         set_cpus_allowed(current, cpumask_of_cpu(pol->cpu));
 915
 916         if (smp_processor_id() != pol->cpu) {
 917                 printk(KERN_ERR PFX "limiting to cpu %u failed\n", pol->cpu);
 918                 goto err_out;
 919         }
 920
 921         if (pending_bit_stuck()) {
 922                 printk(KERN_ERR PFX "failing targ, change pending bit set\n");
 923                 goto err_out;
 924         }
 925
 926         dprintk("targ: cpu %d, %d kHz, min %d, max %d, relation %d\n",
 927                 pol->cpu, targfreq, pol->min, pol->max, relation);
 928
 929         if (query_current_values_with_pending_wait(data)) {
 930                 ret = -EIO;
 931                 goto err_out;
 932         }
 933
 934         dprintk("targ: curr fid 0x%x, vid 0x%x\n",
 935                 data->currfid, data->currvid);
 936
 937         if ((checkvid != data->currvid) || (checkfid != data->currfid)) {
 938                 printk(KERN_INFO PFX
 939                         "error - out of sync, fix 0x%x 0x%x, vid 0x%x 0x%x\n",
 940                         checkfid, data->currfid, checkvid, data->currvid);
 941         }
 942
 943         if (cpufreq_frequency_table_target(pol, data->powernow_table, targfreq, relation, &newstate))
 944                 goto err_out;
 945
 946         down(&fidvid_sem);
 947
 948         powernow_k8_acpi_pst_values(data, newstate);
 949
 950         if (transition_frequency(data, newstate)) {
 951                 printk(KERN_ERR PFX "transition frequency failed\n");
 952                 ret = 1;
 953                 up(&fidvid_sem);
 954                 goto err_out;
 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         return ret;
 964 }
 965
 966 /* Driver entry point to verify the policy and range of frequencies */
 967 static int powernowk8_verify(struct cpufreq_policy *pol)
 968 {
 969         struct powernow_k8_data *data = powernow_data[pol->cpu];
 970
 971         return cpufreq_frequency_table_verify(pol, data->powernow_table);
 972 }
 973
 974 /* per CPU init entry point to the driver */
 975 static int __cpuinit powernowk8_cpu_init(struct cpufreq_policy *pol)
 976 {
 977         struct powernow_k8_data *data;
 978         cpumask_t oldmask = CPU_MASK_ALL;
 979         int rc;
 980
 981         if (!cpu_online(pol->cpu))
 982                 return -ENODEV;
 983
 984         if (!check_supported_cpu(pol->cpu))
 985                 return -ENODEV;
 986
 987         data = kzalloc(sizeof(struct powernow_k8_data), GFP_KERNEL);
 988         if (!data) {
 989                 printk(KERN_ERR PFX "unable to alloc powernow_k8_data");
 990                 return -ENOMEM;
 991         }
 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
1022         if (smp_processor_id() != pol->cpu) {
1023                 printk(KERN_ERR PFX "limiting to cpu %u failed\n", pol->cpu);
1024                 goto err_out;
1025         }
1026
1027         if (pending_bit_stuck()) {
1028                 printk(KERN_ERR PFX "failing init, change pending bit set\n");
1029                 goto err_out;
1030         }
1031
1032         if (query_current_values_with_pending_wait(data))
1033                 goto err_out;
1034
1035         fidvid_msr_init();
1036
1037         /* run on any CPU again */
1038         set_cpus_allowed(current, oldmask);
1039
1040         pol->governor = CPUFREQ_DEFAULT_GOVERNOR;
1041         pol->cpus = cpu_core_map[pol->cpu];
1042
1043         /* Take a crude guess here.
1044          * That guess was in microseconds, so multiply with 1000 */
1045         pol->cpuinfo.transition_latency = (((data->rvo + 8) * data->vstable * VST_UNITS_20US)
1046             + (3 * (1 << data->irt) * 10)) * 1000;
1047
1048         pol->cur = find_khz_freq_from_fid(data->currfid);
1049         dprintk("policy current frequency %d kHz\n", pol->cur);
1050
1051         /* min/max the cpu is capable of */
1052         if (cpufreq_frequency_table_cpuinfo(pol, data->powernow_table)) {
1053                 printk(KERN_ERR PFX "invalid powernow_table\n");
1054                 powernow_k8_cpu_exit_acpi(data);
1055                 kfree(data->powernow_table);
1056                 kfree(data);
1057                 return -EINVAL;
1058         }
1059
1060         cpufreq_frequency_table_get_attr(data->powernow_table, pol->cpu);
1061
1062         printk("cpu_init done, current fid 0x%x, vid 0x%x\n",
1063                data->currfid, data->currvid);
1064
1065         powernow_data[pol->cpu] = data;
1066
1067         return 0;
1068
1069 err_out:
1070         set_cpus_allowed(current, oldmask);
1071         powernow_k8_cpu_exit_acpi(data);
1072
1073         kfree(data);
1074         return -ENODEV;
1075 }
1076
1077 static int __devexit powernowk8_cpu_exit (struct cpufreq_policy *pol)
1078 {
1079         struct powernow_k8_data *data = powernow_data[pol->cpu];
1080
1081         if (!data)
1082                 return -EINVAL;
1083
1084         powernow_k8_cpu_exit_acpi(data);
1085
1086         cpufreq_frequency_table_put_attr(pol->cpu);
1087
1088         kfree(data->powernow_table);
1089         kfree(data);
1090
1091         return 0;
1092 }
1093
1094 static unsigned int powernowk8_get (unsigned int cpu)
1095 {
1096         struct powernow_k8_data *data = powernow_data[cpu];
1097         cpumask_t oldmask = current->cpus_allowed;
1098         unsigned int khz = 0;
1099
1100         set_cpus_allowed(current, cpumask_of_cpu(cpu));
1101         if (smp_processor_id() != cpu) {
1102                 printk(KERN_ERR PFX "limiting to CPU %d failed in powernowk8_get\n", cpu);
1103                 set_cpus_allowed(current, oldmask);
1104                 return 0;
1105         }
1106
1107         if (query_current_values_with_pending_wait(data))
1108                 goto out;
1109
1110         khz = find_khz_freq_from_fid(data->currfid);
1111
1112 out:
1113         set_cpus_allowed(current, oldmask);
1114         return khz;
1115 }
1116
1117 static struct freq_attr* powernow_k8_attr[] = {
1118         &cpufreq_freq_attr_scaling_available_freqs,
1119         NULL,
1120 };
1121
1122 static struct cpufreq_driver cpufreq_amd64_driver = {
1123         .verify = powernowk8_verify,
1124         .target = powernowk8_target,
1125         .init = powernowk8_cpu_init,
1126         .exit = __devexit_p(powernowk8_cpu_exit),
1127         .get = powernowk8_get,
1128         .name = "powernow-k8",
1129         .owner = THIS_MODULE,
1130         .attr = powernow_k8_attr,
1131 };
1132
1133 /* driver entry point for init */
1134 static int __cpuinit powernowk8_init(void)
1135 {
1136         unsigned int i, supported_cpus = 0;
1137
1138         for (i=0; i<NR_CPUS; i++) {
1139                 if (!cpu_online(i))
1140                         continue;
1141                 if (check_supported_cpu(i))
1142                         supported_cpus++;
1143         }
1144
1145         if (supported_cpus == num_online_cpus()) {
1146                 printk(KERN_INFO PFX "Found %d AMD Athlon 64 / Opteron processors (" VERSION ")\n",
1147                         supported_cpus);
1148                 return cpufreq_register_driver(&cpufreq_amd64_driver);
1149         }
1150
1151         return -ENODEV;
1152 }
1153
1154 /* driver entry point for term */
1155 static void __exit powernowk8_exit(void)
1156 {
1157         dprintk("exit\n");
1158
1159         cpufreq_unregister_driver(&cpufreq_amd64_driver);
1160 }
1161
1162 MODULE_AUTHOR("Paul Devriendt <paul.devriendt@amd.com> and Mark Langsdorf <mark.langsdorf@amd.com>");
1163 MODULE_DESCRIPTION("AMD Athlon 64 and Opteron processor frequency driver.");
1164 MODULE_LICENSE("GPL");
1165
1166 late_initcall(powernowk8_init);
1167 module_exit(powernowk8_exit);