arch/x86/kernel/cpu/cpufreq/powernow-k8.c

   1
   2 /*
   3  *   (c) 2003-2006 Advanced Micro Devices, Inc.
   4  *  Your use of this code is subject to the terms and conditions of the
   5  *  GNU general public license version 2. See "COPYING" or
   6  *  http://www.gnu.org/licenses/gpl.html
   7  *
   8  *  Support : mark.langsdorf@amd.com
   9  *
  10  *  Based on the powernow-k7.c module written by Dave Jones.
  11  *  (C) 2003 Dave Jones on behalf of SuSE Labs
  12  *  (C) 2004 Dominik Brodowski <linux@brodo.de>
  13  *  (C) 2004 Pavel Machek <pavel@suse.cz>
  14  *  Licensed under the terms of the GNU GPL License version 2.
  15  *  Based upon datasheets & sample CPUs kindly provided by AMD.
  16  *
  17  *  Valuable input gratefully received from Dave Jones, Pavel Machek,
  18  *  Dominik Brodowski, Jacob Shin, and others.
  19  *  Originally developed by Paul Devriendt.
  20  *  Processor information obtained from Chapter 9 (Power and Thermal Management)
  21  *  of the "BIOS and Kernel Developer's Guide for the AMD Athlon 64 and AMD
  22  *  Opteron Processors" available for download from www.amd.com
  23  *
  24  *  Tables for specific CPUs can be inferred from
  25  *     http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/30430.pdf
  26  */
  27
  28 #include <linux/kernel.h>
  29 #include <linux/smp.h>
  30 #include <linux/module.h>
  31 #include <linux/init.h>
  32 #include <linux/cpufreq.h>
  33 #include <linux/slab.h>
  34 #include <linux/string.h>
  35 #include <linux/cpumask.h>
  36 #include <linux/sched.h>        /* for current / set_cpus_allowed() */
  37 #include <linux/io.h>
  38 #include <linux/delay.h>
  39
  40 #include <asm/msr.h>
  41
  42 #include <linux/acpi.h>
  43 #include <linux/mutex.h>
  44 #include <acpi/processor.h>
  45
  46 #define PFX "powernow-k8: "
  47 #define VERSION "version 2.20.00"
  48 #include "powernow-k8.h"
  49
  50 /* serialize freq changes  */
  51 static DEFINE_MUTEX(fidvid_mutex);
  52
  53 static DEFINE_PER_CPU(struct powernow_k8_data *, powernow_data);
  54
  55 static int cpu_family = CPU_OPTERON;
  56
  57 #ifndef CONFIG_SMP
  58 static inline const struct cpumask *cpu_core_mask(int cpu)
  59 {
  60         return cpumask_of(0);
  61 }
  62 #endif
  63
  64 /* Return a frequency in MHz, given an input fid */
  65 static u32 find_freq_from_fid(u32 fid)
  66 {
  67         return 800 + (fid * 100);
  68 }
  69
  70 /* Return a frequency in KHz, given an input fid */
  71 static u32 find_khz_freq_from_fid(u32 fid)
  72 {
  73         return 1000 * find_freq_from_fid(fid);
  74 }
  75
  76 static u32 find_khz_freq_from_pstate(struct cpufreq_frequency_table *data,
  77                 u32 pstate)
  78 {
  79         return data[pstate].frequency;
  80 }
  81
  82 /* Return the vco fid for an input fid
  83  *
  84  * Each "low" fid has corresponding "high" fid, and you can get to "low" fids
  85  * only from corresponding high fids. This returns "high" fid corresponding to
  86  * "low" one.
  87  */
  88 static u32 convert_fid_to_vco_fid(u32 fid)
  89 {
  90         if (fid < HI_FID_TABLE_BOTTOM)
  91                 return 8 + (2 * fid);
  92         else
  93                 return fid;
  94 }
  95
  96 /*
  97  * Return 1 if the pending bit is set. Unless we just instructed the processor
  98  * to transition to a new state, seeing this bit set is really bad news.
  99  */
 100 static int pending_bit_stuck(void)
 101 {
 102         u32 lo, hi;
 103
 104         if (cpu_family == CPU_HW_PSTATE)
 105                 return 0;
 106
 107         rdmsr(MSR_FIDVID_STATUS, lo, hi);
 108         return lo & MSR_S_LO_CHANGE_PENDING ? 1 : 0;
 109 }
 110
 111 /*
 112  * Update the global current fid / vid values from the status msr.
 113  * Returns 1 on error.
 114  */
 115 static int query_current_values_with_pending_wait(struct powernow_k8_data *data)
 116 {
 117         u32 lo, hi;
 118         u32 i = 0;
 119
 120         if (cpu_family == CPU_HW_PSTATE) {
 121                 rdmsr(MSR_PSTATE_STATUS, lo, hi);
 122                 i = lo & HW_PSTATE_MASK;
 123                 data->currpstate = i;
 124
 125                 /*
 126                  * a workaround for family 11h erratum 311 might cause
 127                  * an "out-of-range Pstate if the core is in Pstate-0
 128                  */
 129                 if ((boot_cpu_data.x86 == 0x11) && (i >= data->numps))
 130                         data->currpstate = HW_PSTATE_0;
 131
 132                 return 0;
 133         }
 134         do {
 135                 if (i++ > 10000) {
 136                         dprintk("detected change pending stuck\n");
 137                         return 1;
 138                 }
 139                 rdmsr(MSR_FIDVID_STATUS, lo, hi);
 140         } while (lo & MSR_S_LO_CHANGE_PENDING);
 141
 142         data->currvid = hi & MSR_S_HI_CURRENT_VID;
 143         data->currfid = lo & MSR_S_LO_CURRENT_FID;
 144
 145         return 0;
 146 }
 147
 148 /* the isochronous relief time */
 149 static void count_off_irt(struct powernow_k8_data *data)
 150 {
 151         udelay((1 << data->irt) * 10);
 152         return;
 153 }
 154
 155 /* the voltage stabilization time */
 156 static void count_off_vst(struct powernow_k8_data *data)
 157 {
 158         udelay(data->vstable * VST_UNITS_20US);
 159         return;
 160 }
 161
 162 /* need to init the control msr to a safe value (for each cpu) */
 163 static void fidvid_msr_init(void)
 164 {
 165         u32 lo, hi;
 166         u8 fid, vid;
 167
 168         rdmsr(MSR_FIDVID_STATUS, lo, hi);
 169         vid = hi & MSR_S_HI_CURRENT_VID;
 170         fid = lo & MSR_S_LO_CURRENT_FID;
 171         lo = fid | (vid << MSR_C_LO_VID_SHIFT);
 172         hi = MSR_C_HI_STP_GNT_BENIGN;
 173         dprintk("cpu%d, init lo 0x%x, hi 0x%x\n", smp_processor_id(), lo, hi);
 174         wrmsr(MSR_FIDVID_CTL, lo, hi);
 175 }
 176
 177 /* write the new fid value along with the other control fields to the msr */
 178 static int write_new_fid(struct powernow_k8_data *data, u32 fid)
 179 {
 180         u32 lo;
 181         u32 savevid = data->currvid;
 182         u32 i = 0;
 183
 184         if ((fid & INVALID_FID_MASK) || (data->currvid & INVALID_VID_MASK)) {
 185                 printk(KERN_ERR PFX "internal error - overflow on fid write\n");
 186                 return 1;
 187         }
 188
 189         lo = fid;
 190         lo |= (data->currvid << MSR_C_LO_VID_SHIFT);
 191         lo |= MSR_C_LO_INIT_FID_VID;
 192
 193         dprintk("writing fid 0x%x, lo 0x%x, hi 0x%x\n",
 194                 fid, lo, data->plllock * PLL_LOCK_CONVERSION);
 195
 196         do {
 197                 wrmsr(MSR_FIDVID_CTL, lo, data->plllock * PLL_LOCK_CONVERSION);
 198                 if (i++ > 100) {
 199                         printk(KERN_ERR PFX
 200                                 "Hardware error - pending bit very stuck - "
 201                                 "no further pstate changes possible\n");
 202                         return 1;
 203                 }
 204         } while (query_current_values_with_pending_wait(data));
 205
 206         count_off_irt(data);
 207
 208         if (savevid != data->currvid) {
 209                 printk(KERN_ERR PFX
 210                         "vid change on fid trans, old 0x%x, new 0x%x\n",
 211                         savevid, data->currvid);
 212                 return 1;
 213         }
 214
 215         if (fid != data->currfid) {
 216                 printk(KERN_ERR PFX
 217                         "fid trans failed, fid 0x%x, curr 0x%x\n", fid,
 218                         data->currfid);
 219                 return 1;
 220         }
 221
 222         return 0;
 223 }
 224
 225 /* Write a new vid to the hardware */
 226 static int write_new_vid(struct powernow_k8_data *data, u32 vid)
 227 {
 228         u32 lo;
 229         u32 savefid = data->currfid;
 230         int i = 0;
 231
 232         if ((data->currfid & INVALID_FID_MASK) || (vid & INVALID_VID_MASK)) {
 233                 printk(KERN_ERR PFX "internal error - overflow on vid write\n");
 234                 return 1;
 235         }
 236
 237         lo = data->currfid;
 238         lo |= (vid << MSR_C_LO_VID_SHIFT);
 239         lo |= MSR_C_LO_INIT_FID_VID;
 240
 241         dprintk("writing vid 0x%x, lo 0x%x, hi 0x%x\n",
 242                 vid, lo, STOP_GRANT_5NS);
 243
 244         do {
 245                 wrmsr(MSR_FIDVID_CTL, lo, STOP_GRANT_5NS);
 246                 if (i++ > 100) {
 247                         printk(KERN_ERR PFX "internal error - pending bit "
 248                                         "very stuck - no further pstate "
 249                                         "changes possible\n");
 250                         return 1;
 251                 }
 252         } while (query_current_values_with_pending_wait(data));
 253
 254         if (savefid != data->currfid) {
 255                 printk(KERN_ERR PFX "fid changed on vid trans, old "
 256                         "0x%x new 0x%x\n",
 257                        savefid, data->currfid);
 258                 return 1;
 259         }
 260
 261         if (vid != data->currvid) {
 262                 printk(KERN_ERR PFX "vid trans failed, vid 0x%x, "
 263                                 "curr 0x%x\n",
 264                                 vid, data->currvid);
 265                 return 1;
 266         }
 267
 268         return 0;
 269 }
 270
 271 /*
 272  * Reduce the vid by the max of step or reqvid.
 273  * Decreasing vid codes represent increasing voltages:
 274  * vid of 0 is 1.550V, vid of 0x1e is 0.800V, vid of VID_OFF is off.
 275  */
 276 static int decrease_vid_code_by_step(struct powernow_k8_data *data,
 277                 u32 reqvid, u32 step)
 278 {
 279         if ((data->currvid - reqvid) > step)
 280                 reqvid = data->currvid - step;
 281
 282         if (write_new_vid(data, reqvid))
 283                 return 1;
 284
 285         count_off_vst(data);
 286
 287         return 0;
 288 }
 289
 290 /* Change hardware pstate by single MSR write */
 291 static int transition_pstate(struct powernow_k8_data *data, u32 pstate)
 292 {
 293         wrmsr(MSR_PSTATE_CTRL, pstate, 0);
 294         data->currpstate = pstate;
 295         return 0;
 296 }
 297
 298 /* Change Opteron/Athlon64 fid and vid, by the 3 phases. */
 299 static int transition_fid_vid(struct powernow_k8_data *data,
 300                 u32 reqfid, u32 reqvid)
 301 {
 302         if (core_voltage_pre_transition(data, reqvid))
 303                 return 1;
 304
 305         if (core_frequency_transition(data, reqfid))
 306                 return 1;
 307
 308         if (core_voltage_post_transition(data, reqvid))
 309                 return 1;
 310
 311         if (query_current_values_with_pending_wait(data))
 312                 return 1;
 313
 314         if ((reqfid != data->currfid) || (reqvid != data->currvid)) {
 315                 printk(KERN_ERR PFX "failed (cpu%d): req 0x%x 0x%x, "
 316                                 "curr 0x%x 0x%x\n",
 317                                 smp_processor_id(),
 318                                 reqfid, reqvid, data->currfid, data->currvid);
 319                 return 1;
 320         }
 321
 322         dprintk("transitioned (cpu%d): new fid 0x%x, vid 0x%x\n",
 323                 smp_processor_id(), data->currfid, data->currvid);
 324
 325         return 0;
 326 }
 327
 328 /* Phase 1 - core voltage transition ... setup voltage */
 329 static int core_voltage_pre_transition(struct powernow_k8_data *data,
 330                 u32 reqvid)
 331 {
 332         u32 rvosteps = data->rvo;
 333         u32 savefid = data->currfid;
 334         u32 maxvid, lo;
 335
 336         dprintk("ph1 (cpu%d): start, currfid 0x%x, currvid 0x%x, "
 337                 "reqvid 0x%x, rvo 0x%x\n",
 338                 smp_processor_id(),
 339                 data->currfid, data->currvid, reqvid, data->rvo);
 340
 341         rdmsr(MSR_FIDVID_STATUS, lo, maxvid);
 342         maxvid = 0x1f & (maxvid >> 16);
 343         dprintk("ph1 maxvid=0x%x\n", maxvid);
 344         if (reqvid < maxvid) /* lower numbers are higher voltages */
 345                 reqvid = maxvid;
 346
 347         while (data->currvid > reqvid) {
 348                 dprintk("ph1: curr 0x%x, req vid 0x%x\n",
 349                         data->currvid, reqvid);
 350                 if (decrease_vid_code_by_step(data, reqvid, data->vidmvs))
 351                         return 1;
 352         }
 353
 354         while ((rvosteps > 0) && ((data->rvo + data->currvid) > reqvid)) {
 355                 if (data->currvid == maxvid) {
 356                         rvosteps = 0;
 357                 } else {
 358                         dprintk("ph1: changing vid for rvo, req 0x%x\n",
 359                                 data->currvid - 1);
 360                         if (decrease_vid_code_by_step(data, data->currvid-1, 1))
 361                                 return 1;
 362                         rvosteps--;
 363                 }
 364         }
 365
 366         if (query_current_values_with_pending_wait(data))
 367                 return 1;
 368
 369         if (savefid != data->currfid) {
 370                 printk(KERN_ERR PFX "ph1 err, currfid changed 0x%x\n",
 371                                 data->currfid);
 372                 return 1;
 373         }
 374
 375         dprintk("ph1 complete, currfid 0x%x, currvid 0x%x\n",
 376                 data->currfid, data->currvid);
 377
 378         return 0;
 379 }
 380
 381 /* Phase 2 - core frequency transition */
 382 static int core_frequency_transition(struct powernow_k8_data *data, u32 reqfid)
 383 {
 384         u32 vcoreqfid, vcocurrfid, vcofiddiff;
 385         u32 fid_interval, savevid = data->currvid;
 386
 387         if ((reqfid < HI_FID_TABLE_BOTTOM) &&
 388             (data->currfid < HI_FID_TABLE_BOTTOM)) {
 389                 printk(KERN_ERR PFX "ph2: illegal lo-lo transition "
 390                                 "0x%x 0x%x\n", reqfid, data->currfid);
 391                 return 1;
 392         }
 393
 394         if (data->currfid == reqfid) {
 395                 printk(KERN_ERR PFX "ph2 null fid transition 0x%x\n",
 396                                 data->currfid);
 397                 return 0;
 398         }
 399
 400         dprintk("ph2 (cpu%d): starting, currfid 0x%x, currvid 0x%x, "
 401                 "reqfid 0x%x\n",
 402                 smp_processor_id(),
 403                 data->currfid, data->currvid, reqfid);
 404
 405         vcoreqfid = convert_fid_to_vco_fid(reqfid);
 406         vcocurrfid = convert_fid_to_vco_fid(data->currfid);
 407         vcofiddiff = vcocurrfid > vcoreqfid ? vcocurrfid - vcoreqfid
 408             : vcoreqfid - vcocurrfid;
 409
 410         while (vcofiddiff > 2) {
 411                 (data->currfid & 1) ? (fid_interval = 1) : (fid_interval = 2);
 412
 413                 if (reqfid > data->currfid) {
 414                         if (data->currfid > LO_FID_TABLE_TOP) {
 415                                 if (write_new_fid(data,
 416                                                 data->currfid + fid_interval))
 417                                         return 1;
 418                         } else {
 419                                 if (write_new_fid
 420                                     (data,
 421                                      2 + convert_fid_to_vco_fid(data->currfid)))
 422                                         return 1;
 423                         }
 424                 } else {
 425                         if (write_new_fid(data, data->currfid - fid_interval))
 426                                 return 1;
 427                 }
 428
 429                 vcocurrfid = convert_fid_to_vco_fid(data->currfid);
 430                 vcofiddiff = vcocurrfid > vcoreqfid ? vcocurrfid - vcoreqfid
 431                     : vcoreqfid - vcocurrfid;
 432         }
 433
 434         if (write_new_fid(data, reqfid))
 435                 return 1;
 436
 437         if (query_current_values_with_pending_wait(data))
 438                 return 1;
 439
 440         if (data->currfid != reqfid) {
 441                 printk(KERN_ERR PFX
 442                         "ph2: mismatch, failed fid transition, "
 443                         "curr 0x%x, req 0x%x\n",
 444                         data->currfid, reqfid);
 445                 return 1;
 446         }
 447
 448         if (savevid != data->currvid) {
 449                 printk(KERN_ERR PFX "ph2: vid changed, save 0x%x, curr 0x%x\n",
 450                         savevid, data->currvid);
 451                 return 1;
 452         }
 453
 454         dprintk("ph2 complete, currfid 0x%x, currvid 0x%x\n",
 455                 data->currfid, data->currvid);
 456
 457         return 0;
 458 }
 459
 460 /* Phase 3 - core voltage transition flow ... jump to the final vid. */
 461 static int core_voltage_post_transition(struct powernow_k8_data *data,
 462                 u32 reqvid)
 463 {
 464         u32 savefid = data->currfid;
 465         u32 savereqvid = reqvid;
 466
 467         dprintk("ph3 (cpu%d): starting, currfid 0x%x, currvid 0x%x\n",
 468                 smp_processor_id(),
 469                 data->currfid, data->currvid);
 470
 471         if (reqvid != data->currvid) {
 472                 if (write_new_vid(data, reqvid))
 473                         return 1;
 474
 475                 if (savefid != data->currfid) {
 476                         printk(KERN_ERR PFX
 477                                "ph3: bad fid change, save 0x%x, curr 0x%x\n",
 478                                savefid, data->currfid);
 479                         return 1;
 480                 }
 481
 482                 if (data->currvid != reqvid) {
 483                         printk(KERN_ERR PFX
 484                                "ph3: failed vid transition\n, "
 485                                "req 0x%x, curr 0x%x",
 486                                reqvid, data->currvid);
 487                         return 1;
 488                 }
 489         }
 490
 491         if (query_current_values_with_pending_wait(data))
 492                 return 1;
 493
 494         if (savereqvid != data->currvid) {
 495                 dprintk("ph3 failed, currvid 0x%x\n", data->currvid);
 496                 return 1;
 497         }
 498
 499         if (savefid != data->currfid) {
 500                 dprintk("ph3 failed, currfid changed 0x%x\n",
 501                         data->currfid);
 502                 return 1;
 503         }
 504
 505         dprintk("ph3 complete, currfid 0x%x, currvid 0x%x\n",
 506                 data->currfid, data->currvid);
 507
 508         return 0;
 509 }
 510
 511 static int check_supported_cpu(unsigned int cpu)
 512 {
 513         cpumask_t oldmask;
 514         u32 eax, ebx, ecx, edx;
 515         unsigned int rc = 0;
 516
 517         oldmask = current->cpus_allowed;
 518         set_cpus_allowed_ptr(current, &cpumask_of_cpu(cpu));
 519
 520         if (smp_processor_id() != cpu) {
 521                 printk(KERN_ERR PFX "limiting to cpu %u failed\n", cpu);
 522                 goto out;
 523         }
 524
 525         if (current_cpu_data.x86_vendor != X86_VENDOR_AMD)
 526                 goto out;
 527
 528         eax = cpuid_eax(CPUID_PROCESSOR_SIGNATURE);
 529         if (((eax & CPUID_XFAM) != CPUID_XFAM_K8) &&
 530             ((eax & CPUID_XFAM) < CPUID_XFAM_10H))
 531                 goto out;
 532
 533         if ((eax & CPUID_XFAM) == CPUID_XFAM_K8) {
 534                 if (((eax & CPUID_USE_XFAM_XMOD) != CPUID_USE_XFAM_XMOD) ||
 535                     ((eax & CPUID_XMOD) > CPUID_XMOD_REV_MASK)) {
 536                         printk(KERN_INFO PFX
 537                                 "Processor cpuid %x not supported\n", eax);
 538                         goto out;
 539                 }
 540
 541                 eax = cpuid_eax(CPUID_GET_MAX_CAPABILITIES);
 542                 if (eax < CPUID_FREQ_VOLT_CAPABILITIES) {
 543                         printk(KERN_INFO PFX
 544                                "No frequency change capabilities detected\n");
 545                         goto out;
 546                 }
 547
 548                 cpuid(CPUID_FREQ_VOLT_CAPABILITIES, &eax, &ebx, &ecx, &edx);
 549                 if ((edx & P_STATE_TRANSITION_CAPABLE)
 550                         != P_STATE_TRANSITION_CAPABLE) {
 551                         printk(KERN_INFO PFX
 552                                 "Power state transitions not supported\n");
 553                         goto out;
 554                 }
 555         } else { /* must be a HW Pstate capable processor */
 556                 cpuid(CPUID_FREQ_VOLT_CAPABILITIES, &eax, &ebx, &ecx, &edx);
 557                 if ((edx & USE_HW_PSTATE) == USE_HW_PSTATE)
 558                         cpu_family = CPU_HW_PSTATE;
 559                 else
 560                         goto out;
 561         }
 562
 563         rc = 1;
 564
 565 out:
 566         set_cpus_allowed_ptr(current, &oldmask);
 567         return rc;
 568 }
 569
 570 static int check_pst_table(struct powernow_k8_data *data, struct pst_s *pst,
 571                 u8 maxvid)
 572 {
 573         unsigned int j;
 574         u8 lastfid = 0xff;
 575
 576         for (j = 0; j < data->numps; j++) {
 577                 if (pst[j].vid > LEAST_VID) {
 578                         printk(KERN_ERR FW_BUG PFX "vid %d invalid : 0x%x\n",
 579                                j, pst[j].vid);
 580                         return -EINVAL;
 581                 }
 582                 if (pst[j].vid < data->rvo) {
 583                         /* vid + rvo >= 0 */
 584                         printk(KERN_ERR FW_BUG PFX "0 vid exceeded with pstate"
 585                                " %d\n", j);
 586                         return -ENODEV;
 587                 }
 588                 if (pst[j].vid < maxvid + data->rvo) {
 589                         /* vid + rvo >= maxvid */
 590                         printk(KERN_ERR FW_BUG PFX "maxvid exceeded with pstate"
 591                                " %d\n", j);
 592                         return -ENODEV;
 593                 }
 594                 if (pst[j].fid > MAX_FID) {
 595                         printk(KERN_ERR FW_BUG PFX "maxfid exceeded with pstate"
 596                                " %d\n", j);
 597                         return -ENODEV;
 598                 }
 599                 if (j && (pst[j].fid < HI_FID_TABLE_BOTTOM)) {
 600                         /* Only first fid is allowed to be in "low" range */
 601                         printk(KERN_ERR FW_BUG PFX "two low fids - %d : "
 602                                "0x%x\n", j, pst[j].fid);
 603                         return -EINVAL;
 604                 }
 605                 if (pst[j].fid < lastfid)
 606                         lastfid = pst[j].fid;
 607         }
 608         if (lastfid & 1) {
 609                 printk(KERN_ERR FW_BUG PFX "lastfid invalid\n");
 610                 return -EINVAL;
 611         }
 612         if (lastfid > LO_FID_TABLE_TOP)
 613                 printk(KERN_INFO FW_BUG PFX
 614                         "first fid not from lo freq table\n");
 615
 616         return 0;
 617 }
 618
 619 static void invalidate_entry(struct powernow_k8_data *data, unsigned int entry)
 620 {
 621         data->powernow_table[entry].frequency = CPUFREQ_ENTRY_INVALID;
 622 }
 623
 624 static void print_basics(struct powernow_k8_data *data)
 625 {
 626         int j;
 627         for (j = 0; j < data->numps; j++) {
 628                 if (data->powernow_table[j].frequency !=
 629                                 CPUFREQ_ENTRY_INVALID) {
 630                         if (cpu_family == CPU_HW_PSTATE) {
 631                                 printk(KERN_INFO PFX
 632                                         "   %d : pstate %d (%d MHz)\n", j,
 633                                         data->powernow_table[j].index,
 634                                         data->powernow_table[j].frequency/1000);
 635                         } else {
 636                                 printk(KERN_INFO PFX
 637                                         "   %d : fid 0x%x (%d MHz), vid 0x%x\n",
 638                                         j,
 639                                         data->powernow_table[j].index & 0xff,
 640                                         data->powernow_table[j].frequency/1000,
 641                                         data->powernow_table[j].index >> 8);
 642                         }
 643                 }
 644         }
 645         if (data->batps)
 646                 printk(KERN_INFO PFX "Only %d pstates on battery\n",
 647                                 data->batps);
 648 }
 649
 650 static u32 freq_from_fid_did(u32 fid, u32 did)
 651 {
 652         u32 mhz = 0;
 653
 654         if (boot_cpu_data.x86 == 0x10)
 655                 mhz = (100 * (fid + 0x10)) >> did;
 656         else if (boot_cpu_data.x86 == 0x11)
 657                 mhz = (100 * (fid + 8)) >> did;
 658         else
 659                 BUG();
 660
 661         return mhz * 1000;
 662 }
 663
 664 static int fill_powernow_table(struct powernow_k8_data *data,
 665                 struct pst_s *pst, u8 maxvid)
 666 {
 667         struct cpufreq_frequency_table *powernow_table;
 668         unsigned int j;
 669
 670         if (data->batps) {
 671                 /* use ACPI support to get full speed on mains power */
 672                 printk(KERN_WARNING PFX
 673                         "Only %d pstates usable (use ACPI driver for full "
 674                         "range\n", data->batps);
 675                 data->numps = data->batps;
 676         }
 677
 678         for (j = 1; j < data->numps; j++) {
 679                 if (pst[j-1].fid >= pst[j].fid) {
 680                         printk(KERN_ERR PFX "PST out of sequence\n");
 681                         return -EINVAL;
 682                 }
 683         }
 684
 685         if (data->numps < 2) {
 686                 printk(KERN_ERR PFX "no p states to transition\n");
 687                 return -ENODEV;
 688         }
 689
 690         if (check_pst_table(data, pst, maxvid))
 691                 return -EINVAL;
 692
 693         powernow_table = kmalloc((sizeof(struct cpufreq_frequency_table)
 694                 * (data->numps + 1)), GFP_KERNEL);
 695         if (!powernow_table) {
 696                 printk(KERN_ERR PFX "powernow_table memory alloc failure\n");
 697                 return -ENOMEM;
 698         }
 699
 700         for (j = 0; j < data->numps; j++) {
 701                 int freq;
 702                 powernow_table[j].index = pst[j].fid; /* lower 8 bits */
 703                 powernow_table[j].index |= (pst[j].vid << 8); /* upper 8 bits */
 704                 freq = find_khz_freq_from_fid(pst[j].fid);
 705                 powernow_table[j].frequency = freq;
 706         }
 707         powernow_table[data->numps].frequency = CPUFREQ_TABLE_END;
 708         powernow_table[data->numps].index = 0;
 709
 710         if (query_current_values_with_pending_wait(data)) {
 711                 kfree(powernow_table);
 712                 return -EIO;
 713         }
 714
 715         dprintk("cfid 0x%x, cvid 0x%x\n", data->currfid, data->currvid);
 716         data->powernow_table = powernow_table;
 717         if (cpumask_first(cpu_core_mask(data->cpu)) == data->cpu)
 718                 print_basics(data);
 719
 720         for (j = 0; j < data->numps; j++)
 721                 if ((pst[j].fid == data->currfid) &&
 722                     (pst[j].vid == data->currvid))
 723                         return 0;
 724
 725         dprintk("currfid/vid do not match PST, ignoring\n");
 726         return 0;
 727 }
 728
 729 /* Find and validate the PSB/PST table in BIOS. */
 730 static int find_psb_table(struct powernow_k8_data *data)
 731 {
 732         struct psb_s *psb;
 733         unsigned int i;
 734         u32 mvs;
 735         u8 maxvid;
 736         u32 cpst = 0;
 737         u32 thiscpuid;
 738
 739         for (i = 0xc0000; i < 0xffff0; i += 0x10) {
 740                 /* Scan BIOS looking for the signature. */
 741                 /* It can not be at ffff0 - it is too big. */
 742
 743                 psb = phys_to_virt(i);
 744                 if (memcmp(psb, PSB_ID_STRING, PSB_ID_STRING_LEN) != 0)
 745                         continue;
 746
 747                 dprintk("found PSB header at 0x%p\n", psb);
 748
 749                 dprintk("table vers: 0x%x\n", psb->tableversion);
 750                 if (psb->tableversion != PSB_VERSION_1_4) {
 751                         printk(KERN_ERR FW_BUG PFX "PSB table is not v1.4\n");
 752                         return -ENODEV;
 753                 }
 754
 755                 dprintk("flags: 0x%x\n", psb->flags1);
 756                 if (psb->flags1) {
 757                         printk(KERN_ERR FW_BUG PFX "unknown flags\n");
 758                         return -ENODEV;
 759                 }
 760
 761                 data->vstable = psb->vstable;
 762                 dprintk("voltage stabilization time: %d(*20us)\n",
 763                                 data->vstable);
 764
 765                 dprintk("flags2: 0x%x\n", psb->flags2);
 766                 data->rvo = psb->flags2 & 3;
 767                 data->irt = ((psb->flags2) >> 2) & 3;
 768                 mvs = ((psb->flags2) >> 4) & 3;
 769                 data->vidmvs = 1 << mvs;
 770                 data->batps = ((psb->flags2) >> 6) & 3;
 771
 772                 dprintk("ramp voltage offset: %d\n", data->rvo);
 773                 dprintk("isochronous relief time: %d\n", data->irt);
 774                 dprintk("maximum voltage step: %d - 0x%x\n", mvs, data->vidmvs);
 775
 776                 dprintk("numpst: 0x%x\n", psb->num_tables);
 777                 cpst = psb->num_tables;
 778                 if ((psb->cpuid == 0x00000fc0) ||
 779                     (psb->cpuid == 0x00000fe0)) {
 780                         thiscpuid = cpuid_eax(CPUID_PROCESSOR_SIGNATURE);
 781                         if ((thiscpuid == 0x00000fc0) ||
 782                             (thiscpuid == 0x00000fe0))
 783                                 cpst = 1;
 784                 }
 785                 if (cpst != 1) {
 786                         printk(KERN_ERR FW_BUG PFX "numpst must be 1\n");
 787                         return -ENODEV;
 788                 }
 789
 790                 data->plllock = psb->plllocktime;
 791                 dprintk("plllocktime: 0x%x (units 1us)\n", psb->plllocktime);
 792                 dprintk("maxfid: 0x%x\n", psb->maxfid);
 793                 dprintk("maxvid: 0x%x\n", psb->maxvid);
 794                 maxvid = psb->maxvid;
 795
 796                 data->numps = psb->numps;
 797                 dprintk("numpstates: 0x%x\n", data->numps);
 798                 return fill_powernow_table(data,
 799                                 (struct pst_s *)(psb+1), maxvid);
 800         }
 801         /*
 802          * If you see this message, complain to BIOS manufacturer. If
 803          * he tells you "we do not support Linux" or some similar
 804          * nonsense, remember that Windows 2000 uses the same legacy
 805          * mechanism that the old Linux PSB driver uses. Tell them it
 806          * is broken with Windows 2000.
 807          *
 808          * The reference to the AMD documentation is chapter 9 in the
 809          * BIOS and Kernel Developer's Guide, which is available on
 810          * www.amd.com
 811          */
 812         printk(KERN_ERR FW_BUG PFX "No PSB or ACPI _PSS objects\n");
 813         return -ENODEV;
 814 }
 815
 816 static void powernow_k8_acpi_pst_values(struct powernow_k8_data *data,
 817                 unsigned int index)
 818 {
 819         acpi_integer control;
 820
 821         if (!data->acpi_data.state_count || (cpu_family == CPU_HW_PSTATE))
 822                 return;
 823
 824         control = data->acpi_data.states[index].control;
 825         data->irt = (control >> IRT_SHIFT) & IRT_MASK;
 826         data->rvo = (control >> RVO_SHIFT) & RVO_MASK;
 827         data->exttype = (control >> EXT_TYPE_SHIFT) & EXT_TYPE_MASK;
 828         data->plllock = (control >> PLL_L_SHIFT) & PLL_L_MASK;
 829         data->vidmvs = 1 << ((control >> MVS_SHIFT) & MVS_MASK);
 830         data->vstable = (control >> VST_SHIFT) & VST_MASK;
 831 }
 832
 833 static int powernow_k8_cpu_init_acpi(struct powernow_k8_data *data)
 834 {
 835         struct cpufreq_frequency_table *powernow_table;
 836         int ret_val = -ENODEV;
 837         acpi_integer control, status;
 838
 839         if (acpi_processor_register_performance(&data->acpi_data, data->cpu)) {
 840                 dprintk("register performance failed: bad ACPI data\n");
 841                 return -EIO;
 842         }
 843
 844         /* verify the data contained in the ACPI structures */
 845         if (data->acpi_data.state_count <= 1) {
 846                 dprintk("No ACPI P-States\n");
 847                 goto err_out;
 848         }
 849
 850         control = data->acpi_data.control_register.space_id;
 851         status = data->acpi_data.status_register.space_id;
 852
 853         if ((control != ACPI_ADR_SPACE_FIXED_HARDWARE) ||
 854             (status != ACPI_ADR_SPACE_FIXED_HARDWARE)) {
 855                 dprintk("Invalid control/status registers (%x - %x)\n",
 856                         control, status);
 857                 goto err_out;
 858         }
 859
 860         /* fill in data->powernow_table */
 861         powernow_table = kmalloc((sizeof(struct cpufreq_frequency_table)
 862                 * (data->acpi_data.state_count + 1)), GFP_KERNEL);
 863         if (!powernow_table) {
 864                 dprintk("powernow_table memory alloc failure\n");
 865                 goto err_out;
 866         }
 867
 868         if (cpu_family == CPU_HW_PSTATE)
 869                 ret_val = fill_powernow_table_pstate(data, powernow_table);
 870         else
 871                 ret_val = fill_powernow_table_fidvid(data, powernow_table);
 872         if (ret_val)
 873                 goto err_out_mem;
 874
 875         powernow_table[data->acpi_data.state_count].frequency =
 876                 CPUFREQ_TABLE_END;
 877         powernow_table[data->acpi_data.state_count].index = 0;
 878         data->powernow_table = powernow_table;
 879
 880         /* fill in data */
 881         data->numps = data->acpi_data.state_count;
 882         if (cpumask_first(cpu_core_mask(data->cpu)) == data->cpu)
 883                 print_basics(data);
 884         powernow_k8_acpi_pst_values(data, 0);
 885
 886         /* notify BIOS that we exist */
 887         acpi_processor_notify_smm(THIS_MODULE);
 888
 889         if (!zalloc_cpumask_var(&data->acpi_data.shared_cpu_map, GFP_KERNEL)) {
 890                 printk(KERN_ERR PFX
 891                                 "unable to alloc powernow_k8_data cpumask\n");
 892                 ret_val = -ENOMEM;
 893                 goto err_out_mem;
 894         }
 895
 896         return 0;
 897
 898 err_out_mem:
 899         kfree(powernow_table);
 900
 901 err_out:
 902         acpi_processor_unregister_performance(&data->acpi_data, data->cpu);
 903
 904         /* data->acpi_data.state_count informs us at ->exit()
 905          * whether ACPI was used */
 906         data->acpi_data.state_count = 0;
 907
 908         return ret_val;
 909 }
 910
 911 static int fill_powernow_table_pstate(struct powernow_k8_data *data,
 912                 struct cpufreq_frequency_table *powernow_table)
 913 {
 914         int i;
 915         u32 hi = 0, lo = 0;
 916         rdmsr(MSR_PSTATE_CUR_LIMIT, hi, lo);
 917         data->max_hw_pstate = (hi & HW_PSTATE_MAX_MASK) >> HW_PSTATE_MAX_SHIFT;
 918
 919         for (i = 0; i < data->acpi_data.state_count; i++) {
 920                 u32 index;
 921
 922                 index = data->acpi_data.states[i].control & HW_PSTATE_MASK;
 923                 if (index > data->max_hw_pstate) {
 924                         printk(KERN_ERR PFX "invalid pstate %d - "
 925                                         "bad value %d.\n", i, index);
 926                         printk(KERN_ERR PFX "Please report to BIOS "
 927                                         "manufacturer\n");
 928                         invalidate_entry(data, i);
 929                         continue;
 930                 }
 931                 rdmsr(MSR_PSTATE_DEF_BASE + index, lo, hi);
 932                 if (!(hi & HW_PSTATE_VALID_MASK)) {
 933                         dprintk("invalid pstate %d, ignoring\n", index);
 934                         invalidate_entry(data, i);
 935                         continue;
 936                 }
 937
 938                 powernow_table[i].index = index;
 939
 940                 /* Frequency may be rounded for these */
 941                 if (boot_cpu_data.x86 == 0x10 || boot_cpu_data.x86 == 0x11) {
 942                         powernow_table[i].frequency =
 943                                 freq_from_fid_did(lo & 0x3f, (lo >> 6) & 7);
 944                 } else
 945                         powernow_table[i].frequency =
 946                                 data->acpi_data.states[i].core_frequency * 1000;
 947         }
 948         return 0;
 949 }
 950
 951 static int fill_powernow_table_fidvid(struct powernow_k8_data *data,
 952                 struct cpufreq_frequency_table *powernow_table)
 953 {
 954         int i;
 955         int cntlofreq = 0;
 956
 957         for (i = 0; i < data->acpi_data.state_count; i++) {
 958                 u32 fid;
 959                 u32 vid;
 960                 u32 freq, index;
 961                 acpi_integer status, control;
 962
 963                 if (data->exttype) {
 964                         status =  data->acpi_data.states[i].status;
 965                         fid = status & EXT_FID_MASK;
 966                         vid = (status >> VID_SHIFT) & EXT_VID_MASK;
 967                 } else {
 968                         control =  data->acpi_data.states[i].control;
 969                         fid = control & FID_MASK;
 970                         vid = (control >> VID_SHIFT) & VID_MASK;
 971                 }
 972
 973                 dprintk("   %d : fid 0x%x, vid 0x%x\n", i, fid, vid);
 974
 975                 index = fid | (vid<<8);
 976                 powernow_table[i].index = index;
 977
 978                 freq = find_khz_freq_from_fid(fid);
 979                 powernow_table[i].frequency = freq;
 980
 981                 /* verify frequency is OK */
 982                 if ((freq > (MAX_FREQ * 1000)) || (freq < (MIN_FREQ * 1000))) {
 983                         dprintk("invalid freq %u kHz, ignoring\n", freq);
 984                         invalidate_entry(data, i);
 985                         continue;
 986                 }
 987
 988                 /* verify voltage is OK -
 989                  * BIOSs are using "off" to indicate invalid */
 990                 if (vid == VID_OFF) {
 991                         dprintk("invalid vid %u, ignoring\n", vid);
 992                         invalidate_entry(data, i);
 993                         continue;
 994                 }
 995
 996                 /* verify only 1 entry from the lo frequency table */
 997                 if (fid < HI_FID_TABLE_BOTTOM) {
 998                         if (cntlofreq) {
 999                                 /* if both entries are the same,
1000                                  * ignore this one ... */
1001                                 if ((freq != powernow_table[cntlofreq].frequency) ||
1002                                     (index != powernow_table[cntlofreq].index)) {
1003                                         printk(KERN_ERR PFX
1004                                                 "Too many lo freq table "
1005                                                 "entries\n");
1006                                         return 1;
1007                                 }
1008
1009                                 dprintk("double low frequency table entry, "
1010                                                 "ignoring it.\n");
1011                                 invalidate_entry(data, i);
1012                                 continue;
1013                         } else
1014                                 cntlofreq = i;
1015                 }
1016
1017                 if (freq != (data->acpi_data.states[i].core_frequency * 1000)) {
1018                         printk(KERN_INFO PFX "invalid freq entries "
1019                                 "%u kHz vs. %u kHz\n", freq,
1020                                 (unsigned int)
1021                                 (data->acpi_data.states[i].core_frequency
1022                                  * 1000));
1023                         invalidate_entry(data, i);
1024                         continue;
1025                 }
1026         }
1027         return 0;
1028 }
1029
1030 static void powernow_k8_cpu_exit_acpi(struct powernow_k8_data *data)
1031 {
1032         if (data->acpi_data.state_count)
1033                 acpi_processor_unregister_performance(&data->acpi_data,
1034                                 data->cpu);
1035         free_cpumask_var(data->acpi_data.shared_cpu_map);
1036 }
1037
1038 static int get_transition_latency(struct powernow_k8_data *data)
1039 {
1040         int max_latency = 0;
1041         int i;
1042         for (i = 0; i < data->acpi_data.state_count; i++) {
1043                 int cur_latency = data->acpi_data.states[i].transition_latency
1044                         + data->acpi_data.states[i].bus_master_latency;
1045                 if (cur_latency > max_latency)
1046                         max_latency = cur_latency;
1047         }
1048         if (max_latency == 0) {
1049                 /*
1050                  * Fam 11h always returns 0 as transition latency.
1051                  * This is intended and means "very fast". While cpufreq core
1052                  * and governors currently can handle that gracefully, better
1053                  * set it to 1 to avoid problems in the future.
1054                  * For all others it's a BIOS bug.
1055                  */
1056                 if (!boot_cpu_data.x86 == 0x11)
1057                         printk(KERN_ERR FW_WARN PFX "Invalid zero transition "
1058                                 "latency\n");
1059                 max_latency = 1;
1060         }
1061         /* value in usecs, needs to be in nanoseconds */
1062         return 1000 * max_latency;
1063 }
1064
1065 /* Take a frequency, and issue the fid/vid transition command */
1066 static int transition_frequency_fidvid(struct powernow_k8_data *data,
1067                 unsigned int index)
1068 {
1069         u32 fid = 0;
1070         u32 vid = 0;
1071         int res, i;
1072         struct cpufreq_freqs freqs;
1073
1074         dprintk("cpu %d transition to index %u\n", smp_processor_id(), index);
1075
1076         /* fid/vid correctness check for k8 */
1077         /* fid are the lower 8 bits of the index we stored into
1078          * the cpufreq frequency table in find_psb_table, vid
1079          * are the upper 8 bits.
1080          */
1081         fid = data->powernow_table[index].index & 0xFF;
1082         vid = (data->powernow_table[index].index & 0xFF00) >> 8;
1083
1084         dprintk("table matched fid 0x%x, giving vid 0x%x\n", fid, vid);
1085
1086         if (query_current_values_with_pending_wait(data))
1087                 return 1;
1088
1089         if ((data->currvid == vid) && (data->currfid == fid)) {
1090                 dprintk("target matches current values (fid 0x%x, vid 0x%x)\n",
1091                         fid, vid);
1092                 return 0;
1093         }
1094
1095         if ((fid < HI_FID_TABLE_BOTTOM) &&
1096             (data->currfid < HI_FID_TABLE_BOTTOM)) {
1097                 printk(KERN_ERR PFX
1098                        "ignoring illegal change in lo freq table-%x to 0x%x\n",
1099                        data->currfid, fid);
1100                 return 1;
1101         }
1102
1103         dprintk("cpu %d, changing to fid 0x%x, vid 0x%x\n",
1104                 smp_processor_id(), fid, vid);
1105         freqs.old = find_khz_freq_from_fid(data->currfid);
1106         freqs.new = find_khz_freq_from_fid(fid);
1107
1108         for_each_cpu_mask_nr(i, *(data->available_cores)) {
1109                 freqs.cpu = i;
1110                 cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
1111         }
1112
1113         res = transition_fid_vid(data, fid, vid);
1114         freqs.new = find_khz_freq_from_fid(data->currfid);
1115
1116         for_each_cpu_mask_nr(i, *(data->available_cores)) {
1117                 freqs.cpu = i;
1118                 cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
1119         }
1120         return res;
1121 }
1122
1123 /* Take a frequency, and issue the hardware pstate transition command */
1124 static int transition_frequency_pstate(struct powernow_k8_data *data,
1125                 unsigned int index)
1126 {
1127         u32 pstate = 0;
1128         int res, i;
1129         struct cpufreq_freqs freqs;
1130
1131         dprintk("cpu %d transition to index %u\n", smp_processor_id(), index);
1132
1133         /* get MSR index for hardware pstate transition */
1134         pstate = index & HW_PSTATE_MASK;
1135         if (pstate > data->max_hw_pstate)
1136                 return 0;
1137         freqs.old = find_khz_freq_from_pstate(data->powernow_table,
1138                         data->currpstate);
1139         freqs.new = find_khz_freq_from_pstate(data->powernow_table, pstate);
1140
1141         for_each_cpu_mask_nr(i, *(data->available_cores)) {
1142                 freqs.cpu = i;
1143                 cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
1144         }
1145
1146         res = transition_pstate(data, pstate);
1147         freqs.new = find_khz_freq_from_pstate(data->powernow_table, pstate);
1148
1149         for_each_cpu_mask_nr(i, *(data->available_cores)) {
1150                 freqs.cpu = i;
1151                 cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
1152         }
1153         return res;
1154 }
1155
1156 /* Driver entry point to switch to the target frequency */
1157 static int powernowk8_target(struct cpufreq_policy *pol,
1158                 unsigned targfreq, unsigned relation)
1159 {
1160         cpumask_t oldmask;
1161         struct powernow_k8_data *data = per_cpu(powernow_data, pol->cpu);
1162         u32 checkfid;
1163         u32 checkvid;
1164         unsigned int newstate;
1165         int ret = -EIO;
1166
1167         if (!data)
1168                 return -EINVAL;
1169
1170         checkfid = data->currfid;
1171         checkvid = data->currvid;
1172
1173         /* only run on specific CPU from here on */
1174         oldmask = current->cpus_allowed;
1175         set_cpus_allowed_ptr(current, &cpumask_of_cpu(pol->cpu));
1176
1177         if (smp_processor_id() != pol->cpu) {
1178                 printk(KERN_ERR PFX "limiting to cpu %u failed\n", pol->cpu);
1179                 goto err_out;
1180         }
1181
1182         if (pending_bit_stuck()) {
1183                 printk(KERN_ERR PFX "failing targ, change pending bit set\n");
1184                 goto err_out;
1185         }
1186
1187         dprintk("targ: cpu %d, %d kHz, min %d, max %d, relation %d\n",
1188                 pol->cpu, targfreq, pol->min, pol->max, relation);
1189
1190         if (query_current_values_with_pending_wait(data))
1191                 goto err_out;
1192
1193         if (cpu_family != CPU_HW_PSTATE) {
1194                 dprintk("targ: curr fid 0x%x, vid 0x%x\n",
1195                 data->currfid, data->currvid);
1196
1197                 if ((checkvid != data->currvid) ||
1198                     (checkfid != data->currfid)) {
1199                         printk(KERN_INFO PFX
1200                                 "error - out of sync, fix 0x%x 0x%x, "
1201                                 "vid 0x%x 0x%x\n",
1202                                 checkfid, data->currfid,
1203                                 checkvid, data->currvid);
1204                 }
1205         }
1206
1207         if (cpufreq_frequency_table_target(pol, data->powernow_table,
1208                                 targfreq, relation, &newstate))
1209                 goto err_out;
1210
1211         mutex_lock(&fidvid_mutex);
1212
1213         powernow_k8_acpi_pst_values(data, newstate);
1214
1215         if (cpu_family == CPU_HW_PSTATE)
1216                 ret = transition_frequency_pstate(data, newstate);
1217         else
1218                 ret = transition_frequency_fidvid(data, newstate);
1219         if (ret) {
1220                 printk(KERN_ERR PFX "transition frequency failed\n");
1221                 ret = 1;
1222                 mutex_unlock(&fidvid_mutex);
1223                 goto err_out;
1224         }
1225         mutex_unlock(&fidvid_mutex);
1226
1227         if (cpu_family == CPU_HW_PSTATE)
1228                 pol->cur = find_khz_freq_from_pstate(data->powernow_table,
1229                                 newstate);
1230         else
1231                 pol->cur = find_khz_freq_from_fid(data->currfid);
1232         ret = 0;
1233
1234 err_out:
1235         set_cpus_allowed_ptr(current, &oldmask);
1236         return ret;
1237 }
1238
1239 /* Driver entry point to verify the policy and range of frequencies */
1240 static int powernowk8_verify(struct cpufreq_policy *pol)
1241 {
1242         struct powernow_k8_data *data = per_cpu(powernow_data, pol->cpu);
1243
1244         if (!data)
1245                 return -EINVAL;
1246
1247         return cpufreq_frequency_table_verify(pol, data->powernow_table);
1248 }
1249
1250 /* per CPU init entry point to the driver */
1251 static int __cpuinit powernowk8_cpu_init(struct cpufreq_policy *pol)
1252 {
1253         static const char ACPI_PSS_BIOS_BUG_MSG[] =
1254                 KERN_ERR FW_BUG PFX "No compatible ACPI _PSS objects found.\n"
1255                 KERN_ERR FW_BUG PFX "Try again with latest BIOS.\n";
1256         struct powernow_k8_data *data;
1257         cpumask_t oldmask;
1258         int rc;
1259
1260         if (!cpu_online(pol->cpu))
1261                 return -ENODEV;
1262
1263         if (!check_supported_cpu(pol->cpu))
1264                 return -ENODEV;
1265
1266         data = kzalloc(sizeof(struct powernow_k8_data), GFP_KERNEL);
1267         if (!data) {
1268                 printk(KERN_ERR PFX "unable to alloc powernow_k8_data");
1269                 return -ENOMEM;
1270         }
1271
1272         data->cpu = pol->cpu;
1273         data->currpstate = HW_PSTATE_INVALID;
1274
1275         if (powernow_k8_cpu_init_acpi(data)) {
1276                 /*
1277                  * Use the PSB BIOS structure. This is only availabe on
1278                  * an UP version, and is deprecated by AMD.
1279                  */
1280                 if (num_online_cpus() != 1) {
1281                         printk_once(ACPI_PSS_BIOS_BUG_MSG);
1282                         goto err_out;
1283                 }
1284                 if (pol->cpu != 0) {
1285                         printk(KERN_ERR FW_BUG PFX "No ACPI _PSS objects for "
1286                                "CPU other than CPU0. Complain to your BIOS "
1287                                "vendor.\n");
1288                         goto err_out;
1289                 }
1290                 rc = find_psb_table(data);
1291                 if (rc)
1292                         goto err_out;
1293
1294                 /* Take a crude guess here.
1295                  * That guess was in microseconds, so multiply with 1000 */
1296                 pol->cpuinfo.transition_latency = (
1297                          ((data->rvo + 8) * data->vstable * VST_UNITS_20US) +
1298                          ((1 << data->irt) * 30)) * 1000;
1299         } else /* ACPI _PSS objects available */
1300                 pol->cpuinfo.transition_latency = get_transition_latency(data);
1301
1302         /* only run on specific CPU from here on */
1303         oldmask = current->cpus_allowed;
1304         set_cpus_allowed_ptr(current, &cpumask_of_cpu(pol->cpu));
1305
1306         if (smp_processor_id() != pol->cpu) {
1307                 printk(KERN_ERR PFX "limiting to cpu %u failed\n", pol->cpu);
1308                 goto err_out_unmask;
1309         }
1310
1311         if (pending_bit_stuck()) {
1312                 printk(KERN_ERR PFX "failing init, change pending bit set\n");
1313                 goto err_out_unmask;
1314         }
1315
1316         if (query_current_values_with_pending_wait(data))
1317                 goto err_out_unmask;
1318
1319         if (cpu_family == CPU_OPTERON)
1320                 fidvid_msr_init();
1321
1322         /* run on any CPU again */
1323         set_cpus_allowed_ptr(current, &oldmask);
1324
1325         if (cpu_family == CPU_HW_PSTATE)
1326                 cpumask_copy(pol->cpus, cpumask_of(pol->cpu));
1327         else
1328                 cpumask_copy(pol->cpus, cpu_core_mask(pol->cpu));
1329         data->available_cores = pol->cpus;
1330
1331         if (cpu_family == CPU_HW_PSTATE)
1332                 pol->cur = find_khz_freq_from_pstate(data->powernow_table,
1333                                 data->currpstate);
1334         else
1335                 pol->cur = find_khz_freq_from_fid(data->currfid);
1336         dprintk("policy current frequency %d kHz\n", pol->cur);
1337
1338         /* min/max the cpu is capable of */
1339         if (cpufreq_frequency_table_cpuinfo(pol, data->powernow_table)) {
1340                 printk(KERN_ERR FW_BUG PFX "invalid powernow_table\n");
1341                 powernow_k8_cpu_exit_acpi(data);
1342                 kfree(data->powernow_table);
1343                 kfree(data);
1344                 return -EINVAL;
1345         }
1346
1347         cpufreq_frequency_table_get_attr(data->powernow_table, pol->cpu);
1348
1349         if (cpu_family == CPU_HW_PSTATE)
1350                 dprintk("cpu_init done, current pstate 0x%x\n",
1351                                 data->currpstate);
1352         else
1353                 dprintk("cpu_init done, current fid 0x%x, vid 0x%x\n",
1354                         data->currfid, data->currvid);
1355
1356         per_cpu(powernow_data, pol->cpu) = data;
1357
1358         return 0;
1359
1360 err_out_unmask:
1361         set_cpus_allowed_ptr(current, &oldmask);
1362         powernow_k8_cpu_exit_acpi(data);
1363
1364 err_out:
1365         kfree(data);
1366         return -ENODEV;
1367 }
1368
1369 static int __devexit powernowk8_cpu_exit(struct cpufreq_policy *pol)
1370 {
1371         struct powernow_k8_data *data = per_cpu(powernow_data, pol->cpu);
1372
1373         if (!data)
1374                 return -EINVAL;
1375
1376         powernow_k8_cpu_exit_acpi(data);
1377
1378         cpufreq_frequency_table_put_attr(pol->cpu);
1379
1380         kfree(data->powernow_table);
1381         kfree(data);
1382
1383         return 0;
1384 }
1385
1386 static unsigned int powernowk8_get(unsigned int cpu)
1387 {
1388         struct powernow_k8_data *data = per_cpu(powernow_data, cpu);
1389         cpumask_t oldmask = current->cpus_allowed;
1390         unsigned int khz = 0;
1391
1392         if (!data)
1393                 return -EINVAL;
1394
1395         set_cpus_allowed_ptr(current, &cpumask_of_cpu(cpu));
1396         if (smp_processor_id() != cpu) {
1397                 printk(KERN_ERR PFX
1398                         "limiting to CPU %d failed in powernowk8_get\n", cpu);
1399                 set_cpus_allowed_ptr(current, &oldmask);
1400                 return 0;
1401         }
1402
1403         if (query_current_values_with_pending_wait(data))
1404                 goto out;
1405
1406         if (cpu_family == CPU_HW_PSTATE)
1407                 khz = find_khz_freq_from_pstate(data->powernow_table,
1408                                                 data->currpstate);
1409         else
1410                 khz = find_khz_freq_from_fid(data->currfid);
1411
1412
1413 out:
1414         set_cpus_allowed_ptr(current, &oldmask);
1415         return khz;
1416 }
1417
1418 static struct freq_attr *powernow_k8_attr[] = {
1419         &cpufreq_freq_attr_scaling_available_freqs,
1420         NULL,
1421 };
1422
1423 static struct cpufreq_driver cpufreq_amd64_driver = {
1424         .verify = powernowk8_verify,
1425         .target = powernowk8_target,
1426         .init = powernowk8_cpu_init,
1427         .exit = __devexit_p(powernowk8_cpu_exit),
1428         .get = powernowk8_get,
1429         .name = "powernow-k8",
1430         .owner = THIS_MODULE,
1431         .attr = powernow_k8_attr,
1432 };
1433
1434 /* driver entry point for init */
1435 static int __cpuinit powernowk8_init(void)
1436 {
1437         unsigned int i, supported_cpus = 0;
1438
1439         for_each_online_cpu(i) {
1440                 if (check_supported_cpu(i))
1441                         supported_cpus++;
1442         }
1443
1444         if (supported_cpus == num_online_cpus()) {
1445                 printk(KERN_INFO PFX "Found %d %s "
1446                         "processors (%d cpu cores) (" VERSION ")\n",
1447                         num_online_nodes(),
1448                         boot_cpu_data.x86_model_id, supported_cpus);
1449                 return cpufreq_register_driver(&cpufreq_amd64_driver);
1450         }
1451
1452         return -ENODEV;
1453 }
1454
1455 /* driver entry point for term */
1456 static void __exit powernowk8_exit(void)
1457 {
1458         dprintk("exit\n");
1459
1460         cpufreq_unregister_driver(&cpufreq_amd64_driver);
1461 }
1462
1463 MODULE_AUTHOR("Paul Devriendt <paul.devriendt@amd.com> and "
1464                 "Mark Langsdorf <mark.langsdorf@amd.com>");
1465 MODULE_DESCRIPTION("AMD Athlon 64 and Opteron processor frequency driver.");
1466 MODULE_LICENSE("GPL");
1467
1468 late_initcall(powernowk8_init);
1469 module_exit(powernowk8_exit);