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
  36 #include <asm/msr.h>
  37 #include <asm/io.h>
  38 #include <asm/delay.h>
  39
  40 #ifdef CONFIG_X86_POWERNOW_K8_ACPI
  41 #include <linux/acpi.h>
  42 #include <acpi/processor.h>
  43 #endif
  44
  45 #define PFX "powernow-k8: "
  46 #define BFX PFX "BIOS error: "
  47 #define VERSION "version 1.50.4"
  48 #include "powernow-k8.h"
  49
  50 /* serialize freq changes  */
  51 static DECLARE_MUTEX(fidvid_sem);
  52
  53 static struct powernow_k8_data *powernow_data[NR_CPUS];
  54
  55 #ifndef CONFIG_SMP
  56 static cpumask_t cpu_core_map[1];
  57 #endif
  58
  59 /* Return a frequency in MHz, given an input fid */
  60 static u32 find_freq_from_fid(u32 fid)
  61 {
  62         return 800 + (fid * 100);
  63 }
  64
  65 /* Return a frequency in KHz, given an input fid */
  66 static u32 find_khz_freq_from_fid(u32 fid)
  67 {
  68         return 1000 * find_freq_from_fid(fid);
  69 }
  70
  71 /* Return a voltage in miliVolts, given an input vid */
  72 static u32 find_millivolts_from_vid(struct powernow_k8_data *data, u32 vid)
  73 {
  74         return 1550-vid*25;
  75 }
  76
  77 /* Return the vco fid for an input fid
  78  *
  79  * Each "low" fid has corresponding "high" fid, and you can get to "low" fids
  80  * only from corresponding high fids. This returns "high" fid corresponding to
  81  * "low" one.
  82  */
  83 static u32 convert_fid_to_vco_fid(u32 fid)
  84 {
  85         if (fid < HI_FID_TABLE_BOTTOM) {
  86                 return 8 + (2 * fid);
  87         } else {
  88                 return fid;
  89         }
  90 }
  91
  92 /*
  93  * Return 1 if the pending bit is set. Unless we just instructed the processor
  94  * to transition to a new state, seeing this bit set is really bad news.
  95  */
  96 static int pending_bit_stuck(void)
  97 {
  98         u32 lo, hi;
  99
 100         rdmsr(MSR_FIDVID_STATUS, lo, hi);
 101         return lo & MSR_S_LO_CHANGE_PENDING ? 1 : 0;
 102 }
 103
 104 /*
 105  * Update the global current fid / vid values from the status msr.
 106  * Returns 1 on error.
 107  */
 108 static int query_current_values_with_pending_wait(struct powernow_k8_data *data)
 109 {
 110         u32 lo, hi;
 111         u32 i = 0;
 112
 113         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, 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                 if (reqfid > data->currfid) {
 362                         if (data->currfid > LO_FID_TABLE_TOP) {
 363                                 if (write_new_fid(data, data->currfid + 2)) {
 364                                         return 1;
 365                                 }
 366                         } else {
 367                                 if (write_new_fid
 368                                     (data, 2 + convert_fid_to_vco_fid(data->currfid))) {
 369                                         return 1;
 370                                 }
 371                         }
 372                 } else {
 373                         if (write_new_fid(data, data->currfid - 2))
 374                                 return 1;
 375                 }
 376
 377                 vcocurrfid = convert_fid_to_vco_fid(data->currfid);
 378                 vcofiddiff = vcocurrfid > vcoreqfid ? vcocurrfid - vcoreqfid
 379                     : vcoreqfid - vcocurrfid;
 380         }
 381
 382         if (write_new_fid(data, reqfid))
 383                 return 1;
 384
 385         if (query_current_values_with_pending_wait(data))
 386                 return 1;
 387
 388         if (data->currfid != reqfid) {
 389                 printk(KERN_ERR PFX
 390                         "ph2: mismatch, failed fid transition, curr 0x%x, req 0x%x\n",
 391                         data->currfid, reqfid);
 392                 return 1;
 393         }
 394
 395         if (savevid != data->currvid) {
 396                 printk(KERN_ERR PFX "ph2: vid changed, save 0x%x, curr 0x%x\n",
 397                         savevid, data->currvid);
 398                 return 1;
 399         }
 400
 401         dprintk("ph2 complete, currfid 0x%x, currvid 0x%x\n",
 402                 data->currfid, data->currvid);
 403
 404         return 0;
 405 }
 406
 407 /* Phase 3 - core voltage transition flow ... jump to the final vid. */
 408 static int core_voltage_post_transition(struct powernow_k8_data *data, u32 reqvid)
 409 {
 410         u32 savefid = data->currfid;
 411         u32 savereqvid = reqvid;
 412
 413         dprintk("ph3 (cpu%d): starting, currfid 0x%x, currvid 0x%x\n",
 414                 smp_processor_id(),
 415                 data->currfid, data->currvid);
 416
 417         if (reqvid != data->currvid) {
 418                 if (write_new_vid(data, reqvid))
 419                         return 1;
 420
 421                 if (savefid != data->currfid) {
 422                         printk(KERN_ERR PFX
 423                                "ph3: bad fid change, save 0x%x, curr 0x%x\n",
 424                                savefid, data->currfid);
 425                         return 1;
 426                 }
 427
 428                 if (data->currvid != reqvid) {
 429                         printk(KERN_ERR PFX
 430                                "ph3: failed vid transition\n, req 0x%x, curr 0x%x",
 431                                reqvid, data->currvid);
 432                         return 1;
 433                 }
 434         }
 435
 436         if (query_current_values_with_pending_wait(data))
 437                 return 1;
 438
 439         if (savereqvid != data->currvid) {
 440                 dprintk("ph3 failed, currvid 0x%x\n", data->currvid);
 441                 return 1;
 442         }
 443
 444         if (savefid != data->currfid) {
 445                 dprintk("ph3 failed, currfid changed 0x%x\n",
 446                         data->currfid);
 447                 return 1;
 448         }
 449
 450         dprintk("ph3 complete, currfid 0x%x, currvid 0x%x\n",
 451                 data->currfid, data->currvid);
 452
 453         return 0;
 454 }
 455
 456 static int check_supported_cpu(unsigned int cpu)
 457 {
 458         cpumask_t oldmask = CPU_MASK_ALL;
 459         u32 eax, ebx, ecx, edx;
 460         unsigned int rc = 0;
 461
 462         oldmask = current->cpus_allowed;
 463         set_cpus_allowed(current, cpumask_of_cpu(cpu));
 464         schedule();
 465
 466         if (smp_processor_id() != cpu) {
 467                 printk(KERN_ERR "limiting to cpu %u failed\n", cpu);
 468                 goto out;
 469         }
 470
 471         if (current_cpu_data.x86_vendor != X86_VENDOR_AMD)
 472                 goto out;
 473
 474         eax = cpuid_eax(CPUID_PROCESSOR_SIGNATURE);
 475         if (((eax & CPUID_USE_XFAM_XMOD) != CPUID_USE_XFAM_XMOD) ||
 476             ((eax & CPUID_XFAM) != CPUID_XFAM_K8) ||
 477             ((eax & CPUID_XMOD) > CPUID_XMOD_REV_F)) {
 478                 printk(KERN_INFO PFX "Processor cpuid %x not supported\n", eax);
 479                 goto out;
 480         }
 481
 482         eax = cpuid_eax(CPUID_GET_MAX_CAPABILITIES);
 483         if (eax < CPUID_FREQ_VOLT_CAPABILITIES) {
 484                 printk(KERN_INFO PFX
 485                        "No frequency change capabilities detected\n");
 486                 goto out;
 487         }
 488
 489         cpuid(CPUID_FREQ_VOLT_CAPABILITIES, &eax, &ebx, &ecx, &edx);
 490         if ((edx & P_STATE_TRANSITION_CAPABLE) != P_STATE_TRANSITION_CAPABLE) {
 491                 printk(KERN_INFO PFX "Power state transitions not supported\n");
 492                 goto out;
 493         }
 494
 495         rc = 1;
 496
 497 out:
 498         set_cpus_allowed(current, oldmask);
 499         schedule();
 500         return rc;
 501
 502 }
 503
 504 static int check_pst_table(struct powernow_k8_data *data, struct pst_s *pst, u8 maxvid)
 505 {
 506         unsigned int j;
 507         u8 lastfid = 0xff;
 508
 509         for (j = 0; j < data->numps; j++) {
 510                 if (pst[j].vid > LEAST_VID) {
 511                         printk(KERN_ERR PFX "vid %d invalid : 0x%x\n", j, pst[j].vid);
 512                         return -EINVAL;
 513                 }
 514                 if (pst[j].vid < data->rvo) {   /* vid + rvo >= 0 */
 515                         printk(KERN_ERR BFX "0 vid exceeded with pstate %d\n", j);
 516                         return -ENODEV;
 517                 }
 518                 if (pst[j].vid < maxvid + data->rvo) {  /* vid + rvo >= maxvid */
 519                         printk(KERN_ERR BFX "maxvid exceeded with pstate %d\n", j);
 520                         return -ENODEV;
 521                 }
 522                 if ((pst[j].fid > MAX_FID)
 523                     || (pst[j].fid & 1)
 524                     || (j && (pst[j].fid < HI_FID_TABLE_BOTTOM))) {
 525                         /* Only first fid is allowed to be in "low" range */
 526                         printk(KERN_ERR PFX "two low fids - %d : 0x%x\n", j, pst[j].fid);
 527                         return -EINVAL;
 528                 }
 529                 if (pst[j].fid < lastfid)
 530                         lastfid = pst[j].fid;
 531         }
 532         if (lastfid & 1) {
 533                 printk(KERN_ERR PFX "lastfid invalid\n");
 534                 return -EINVAL;
 535         }
 536         if (lastfid > LO_FID_TABLE_TOP)
 537                 printk(KERN_INFO PFX  "first fid not from lo freq table\n");
 538
 539         return 0;
 540 }
 541
 542 static void print_basics(struct powernow_k8_data *data)
 543 {
 544         int j;
 545         for (j = 0; j < data->numps; j++) {
 546                 if (data->powernow_table[j].frequency != CPUFREQ_ENTRY_INVALID)
 547                         printk(KERN_INFO PFX "   %d : fid 0x%x (%d MHz), vid 0x%x (%d mV)\n", j,
 548                                 data->powernow_table[j].index & 0xff,
 549                                 data->powernow_table[j].frequency/1000,
 550                                 data->powernow_table[j].index >> 8,
 551                                 find_millivolts_from_vid(data, data->powernow_table[j].index >> 8));
 552         }
 553         if (data->batps)
 554                 printk(KERN_INFO PFX "Only %d pstates on battery\n", data->batps);
 555 }
 556
 557 static int fill_powernow_table(struct powernow_k8_data *data, struct pst_s *pst, u8 maxvid)
 558 {
 559         struct cpufreq_frequency_table *powernow_table;
 560         unsigned int j;
 561
 562         if (data->batps) {    /* use ACPI support to get full speed on mains power */
 563                 printk(KERN_WARNING PFX "Only %d pstates usable (use ACPI driver for full range\n", data->batps);
 564                 data->numps = data->batps;
 565         }
 566
 567         for ( j=1; j<data->numps; j++ ) {
 568                 if (pst[j-1].fid >= pst[j].fid) {
 569                         printk(KERN_ERR PFX "PST out of sequence\n");
 570                         return -EINVAL;
 571                 }
 572         }
 573
 574         if (data->numps < 2) {
 575                 printk(KERN_ERR PFX "no p states to transition\n");
 576                 return -ENODEV;
 577         }
 578
 579         if (check_pst_table(data, pst, maxvid))
 580                 return -EINVAL;
 581
 582         powernow_table = kmalloc((sizeof(struct cpufreq_frequency_table)
 583                 * (data->numps + 1)), GFP_KERNEL);
 584         if (!powernow_table) {
 585                 printk(KERN_ERR PFX "powernow_table memory alloc failure\n");
 586                 return -ENOMEM;
 587         }
 588
 589         for (j = 0; j < data->numps; j++) {
 590                 powernow_table[j].index = pst[j].fid; /* lower 8 bits */
 591                 powernow_table[j].index |= (pst[j].vid << 8); /* upper 8 bits */
 592                 powernow_table[j].frequency = find_khz_freq_from_fid(pst[j].fid);
 593         }
 594         powernow_table[data->numps].frequency = CPUFREQ_TABLE_END;
 595         powernow_table[data->numps].index = 0;
 596
 597         if (query_current_values_with_pending_wait(data)) {
 598                 kfree(powernow_table);
 599                 return -EIO;
 600         }
 601
 602         dprintk("cfid 0x%x, cvid 0x%x\n", data->currfid, data->currvid);
 603         data->powernow_table = powernow_table;
 604         print_basics(data);
 605
 606         for (j = 0; j < data->numps; j++)
 607                 if ((pst[j].fid==data->currfid) && (pst[j].vid==data->currvid))
 608                         return 0;
 609
 610         dprintk("currfid/vid do not match PST, ignoring\n");
 611         return 0;
 612 }
 613
 614 /* Find and validate the PSB/PST table in BIOS. */
 615 static int find_psb_table(struct powernow_k8_data *data)
 616 {
 617         struct psb_s *psb;
 618         unsigned int i;
 619         u32 mvs;
 620         u8 maxvid;
 621         u32 cpst = 0;
 622         u32 thiscpuid;
 623
 624         for (i = 0xc0000; i < 0xffff0; i += 0x10) {
 625                 /* Scan BIOS looking for the signature. */
 626                 /* It can not be at ffff0 - it is too big. */
 627
 628                 psb = phys_to_virt(i);
 629                 if (memcmp(psb, PSB_ID_STRING, PSB_ID_STRING_LEN) != 0)
 630                         continue;
 631
 632                 dprintk("found PSB header at 0x%p\n", psb);
 633
 634                 dprintk("table vers: 0x%x\n", psb->tableversion);
 635                 if (psb->tableversion != PSB_VERSION_1_4) {
 636                         printk(KERN_INFO BFX "PSB table is not v1.4\n");
 637                         return -ENODEV;
 638                 }
 639
 640                 dprintk("flags: 0x%x\n", psb->flags1);
 641                 if (psb->flags1) {
 642                         printk(KERN_ERR BFX "unknown flags\n");
 643                         return -ENODEV;
 644                 }
 645
 646                 data->vstable = psb->vstable;
 647                 dprintk("voltage stabilization time: %d(*20us)\n", data->vstable);
 648
 649                 dprintk("flags2: 0x%x\n", psb->flags2);
 650                 data->rvo = psb->flags2 & 3;
 651                 data->irt = ((psb->flags2) >> 2) & 3;
 652                 mvs = ((psb->flags2) >> 4) & 3;
 653                 data->vidmvs = 1 << mvs;
 654                 data->batps = ((psb->flags2) >> 6) & 3;
 655
 656                 dprintk("ramp voltage offset: %d\n", data->rvo);
 657                 dprintk("isochronous relief time: %d\n", data->irt);
 658                 dprintk("maximum voltage step: %d - 0x%x\n", mvs, data->vidmvs);
 659
 660                 dprintk("numpst: 0x%x\n", psb->num_tables);
 661                 cpst = psb->num_tables;
 662                 if ((psb->cpuid == 0x00000fc0) || (psb->cpuid == 0x00000fe0) ){
 663                         thiscpuid = cpuid_eax(CPUID_PROCESSOR_SIGNATURE);
 664                         if ((thiscpuid == 0x00000fc0) || (thiscpuid == 0x00000fe0) ) {
 665                                 cpst = 1;
 666                         }
 667                 }
 668                 if (cpst != 1) {
 669                         printk(KERN_ERR BFX "numpst must be 1\n");
 670                         return -ENODEV;
 671                 }
 672
 673                 data->plllock = psb->plllocktime;
 674                 dprintk("plllocktime: 0x%x (units 1us)\n", psb->plllocktime);
 675                 dprintk("maxfid: 0x%x\n", psb->maxfid);
 676                 dprintk("maxvid: 0x%x\n", psb->maxvid);
 677                 maxvid = psb->maxvid;
 678
 679                 data->numps = psb->numps;
 680                 dprintk("numpstates: 0x%x\n", data->numps);
 681                 return fill_powernow_table(data, (struct pst_s *)(psb+1), maxvid);
 682         }
 683         /*
 684          * If you see this message, complain to BIOS manufacturer. If
 685          * he tells you "we do not support Linux" or some similar
 686          * nonsense, remember that Windows 2000 uses the same legacy
 687          * mechanism that the old Linux PSB driver uses. Tell them it
 688          * is broken with Windows 2000.
 689          *
 690          * The reference to the AMD documentation is chapter 9 in the
 691          * BIOS and Kernel Developer's Guide, which is available on
 692          * www.amd.com
 693          */
 694         printk(KERN_INFO PFX "BIOS error - no PSB or ACPI _PSS objects\n");
 695         return -ENODEV;
 696 }
 697
 698 #ifdef CONFIG_X86_POWERNOW_K8_ACPI
 699 static void powernow_k8_acpi_pst_values(struct powernow_k8_data *data, unsigned int index)
 700 {
 701         if (!data->acpi_data.state_count)
 702                 return;
 703
 704         data->irt = (data->acpi_data.states[index].control >> IRT_SHIFT) & IRT_MASK;
 705         data->rvo = (data->acpi_data.states[index].control >> RVO_SHIFT) & RVO_MASK;
 706         data->exttype = (data->acpi_data.states[index].control >> EXT_TYPE_SHIFT) & EXT_TYPE_MASK;
 707         data->plllock = (data->acpi_data.states[index].control >> PLL_L_SHIFT) & PLL_L_MASK;
 708         data->vidmvs = 1 << ((data->acpi_data.states[index].control >> MVS_SHIFT) & MVS_MASK);
 709         data->vstable = (data->acpi_data.states[index].control >> VST_SHIFT) & VST_MASK;
 710 }
 711
 712 static int powernow_k8_cpu_init_acpi(struct powernow_k8_data *data)
 713 {
 714         int i;
 715         int cntlofreq = 0;
 716         struct cpufreq_frequency_table *powernow_table;
 717
 718         if (acpi_processor_register_performance(&data->acpi_data, data->cpu)) {
 719                 dprintk("register performance failed: bad ACPI data\n");
 720                 return -EIO;
 721         }
 722
 723         /* verify the data contained in the ACPI structures */
 724         if (data->acpi_data.state_count <= 1) {
 725                 dprintk("No ACPI P-States\n");
 726                 goto err_out;
 727         }
 728
 729         if ((data->acpi_data.control_register.space_id != ACPI_ADR_SPACE_FIXED_HARDWARE) ||
 730                 (data->acpi_data.status_register.space_id != ACPI_ADR_SPACE_FIXED_HARDWARE)) {
 731                 dprintk("Invalid control/status registers (%x - %x)\n",
 732                         data->acpi_data.control_register.space_id,
 733                         data->acpi_data.status_register.space_id);
 734                 goto err_out;
 735         }
 736
 737         /* fill in data->powernow_table */
 738         powernow_table = kmalloc((sizeof(struct cpufreq_frequency_table)
 739                 * (data->acpi_data.state_count + 1)), GFP_KERNEL);
 740         if (!powernow_table) {
 741                 dprintk("powernow_table memory alloc failure\n");
 742                 goto err_out;
 743         }
 744
 745         for (i = 0; i < data->acpi_data.state_count; i++) {
 746                 u32 fid;
 747                 u32 vid;
 748
 749                 if (data->exttype) {
 750                         fid = data->acpi_data.states[i].status & FID_MASK;
 751                         vid = (data->acpi_data.states[i].status >> VID_SHIFT) & VID_MASK;
 752                 } else {
 753                         fid = data->acpi_data.states[i].control & FID_MASK;
 754                         vid = (data->acpi_data.states[i].control >> VID_SHIFT) & VID_MASK;
 755                 }
 756
 757                 dprintk("   %d : fid 0x%x, vid 0x%x\n", i, fid, vid);
 758
 759                 powernow_table[i].index = fid; /* lower 8 bits */
 760                 powernow_table[i].index |= (vid << 8); /* upper 8 bits */
 761                 powernow_table[i].frequency = find_khz_freq_from_fid(fid);
 762
 763                 /* verify frequency is OK */
 764                 if ((powernow_table[i].frequency > (MAX_FREQ * 1000)) ||
 765                         (powernow_table[i].frequency < (MIN_FREQ * 1000))) {
 766                         dprintk("invalid freq %u kHz, ignoring\n", powernow_table[i].frequency);
 767                         powernow_table[i].frequency = CPUFREQ_ENTRY_INVALID;
 768                         continue;
 769                 }
 770
 771                 /* verify voltage is OK - BIOSs are using "off" to indicate invalid */
 772                 if (vid == VID_OFF) {
 773                         dprintk("invalid vid %u, ignoring\n", vid);
 774                         powernow_table[i].frequency = CPUFREQ_ENTRY_INVALID;
 775                         continue;
 776                 }
 777
 778                 /* verify only 1 entry from the lo frequency table */
 779                 if (fid < HI_FID_TABLE_BOTTOM) {
 780                         if (cntlofreq) {
 781                                 /* if both entries are the same, ignore this
 782                                  * one...
 783                                  */
 784                                 if ((powernow_table[i].frequency != powernow_table[cntlofreq].frequency) ||
 785                                     (powernow_table[i].index != powernow_table[cntlofreq].index)) {
 786                                         printk(KERN_ERR PFX "Too many lo freq table entries\n");
 787                                         goto err_out_mem;
 788                                 }
 789
 790                                 dprintk("double low frequency table entry, ignoring it.\n");
 791                                 powernow_table[i].frequency = CPUFREQ_ENTRY_INVALID;
 792                                 continue;
 793                         } else
 794                                 cntlofreq = i;
 795                 }
 796
 797                 if (powernow_table[i].frequency != (data->acpi_data.states[i].core_frequency * 1000)) {
 798                         printk(KERN_INFO PFX "invalid freq entries %u kHz vs. %u kHz\n",
 799                                 powernow_table[i].frequency,
 800                                 (unsigned int) (data->acpi_data.states[i].core_frequency * 1000));
 801                         powernow_table[i].frequency = CPUFREQ_ENTRY_INVALID;
 802                         continue;
 803                 }
 804         }
 805
 806         powernow_table[data->acpi_data.state_count].frequency = CPUFREQ_TABLE_END;
 807         powernow_table[data->acpi_data.state_count].index = 0;
 808         data->powernow_table = powernow_table;
 809
 810         /* fill in data */
 811         data->numps = data->acpi_data.state_count;
 812         print_basics(data);
 813         powernow_k8_acpi_pst_values(data, 0);
 814
 815         /* notify BIOS that we exist */
 816         acpi_processor_notify_smm(THIS_MODULE);
 817
 818         return 0;
 819
 820 err_out_mem:
 821         kfree(powernow_table);
 822
 823 err_out:
 824         acpi_processor_unregister_performance(&data->acpi_data, data->cpu);
 825
 826         /* data->acpi_data.state_count informs us at ->exit() whether ACPI was used */
 827         data->acpi_data.state_count = 0;
 828
 829         return -ENODEV;
 830 }
 831
 832 static void powernow_k8_cpu_exit_acpi(struct powernow_k8_data *data)
 833 {
 834         if (data->acpi_data.state_count)
 835                 acpi_processor_unregister_performance(&data->acpi_data, data->cpu);
 836 }
 837
 838 #else
 839 static int powernow_k8_cpu_init_acpi(struct powernow_k8_data *data) { return -ENODEV; }
 840 static void powernow_k8_cpu_exit_acpi(struct powernow_k8_data *data) { return; }
 841 static void powernow_k8_acpi_pst_values(struct powernow_k8_data *data, unsigned int index) { return; }
 842 #endif /* CONFIG_X86_POWERNOW_K8_ACPI */
 843
 844 /* Take a frequency, and issue the fid/vid transition command */
 845 static int transition_frequency(struct powernow_k8_data *data, unsigned int index)
 846 {
 847         u32 fid;
 848         u32 vid;
 849         int res, i;
 850         struct cpufreq_freqs freqs;
 851
 852         dprintk("cpu %d transition to index %u\n", smp_processor_id(), index);
 853
 854         /* fid are the lower 8 bits of the index we stored into
 855          * the cpufreq frequency table in find_psb_table, vid are
 856          * the upper 8 bits.
 857          */
 858
 859         fid = data->powernow_table[index].index & 0xFF;
 860         vid = (data->powernow_table[index].index & 0xFF00) >> 8;
 861
 862         dprintk("table matched fid 0x%x, giving vid 0x%x\n", fid, vid);
 863
 864         if (query_current_values_with_pending_wait(data))
 865                 return 1;
 866
 867         if ((data->currvid == vid) && (data->currfid == fid)) {
 868                 dprintk("target matches current values (fid 0x%x, vid 0x%x)\n",
 869                         fid, vid);
 870                 return 0;
 871         }
 872
 873         if ((fid < HI_FID_TABLE_BOTTOM) && (data->currfid < HI_FID_TABLE_BOTTOM)) {
 874                 printk(KERN_ERR PFX
 875                        "ignoring illegal change in lo freq table-%x to 0x%x\n",
 876                        data->currfid, fid);
 877                 return 1;
 878         }
 879
 880         dprintk("cpu %d, changing to fid 0x%x, vid 0x%x\n",
 881                 smp_processor_id(), fid, vid);
 882
 883         freqs.cpu = data->cpu;
 884         freqs.old = find_khz_freq_from_fid(data->currfid);
 885         freqs.new = find_khz_freq_from_fid(fid);
 886         for_each_cpu_mask(i, cpu_core_map[data->cpu]) {
 887                 freqs.cpu = i;
 888                 cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
 889         }
 890
 891         res = transition_fid_vid(data, fid, vid);
 892
 893         freqs.new = find_khz_freq_from_fid(data->currfid);
 894         for_each_cpu_mask(i, cpu_core_map[data->cpu]) {
 895                 freqs.cpu = i;
 896                 cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
 897         }
 898         return res;
 899 }
 900
 901 /* Driver entry point to switch to the target frequency */
 902 static int powernowk8_target(struct cpufreq_policy *pol, unsigned targfreq, unsigned relation)
 903 {
 904         cpumask_t oldmask = CPU_MASK_ALL;
 905         struct powernow_k8_data *data = powernow_data[pol->cpu];
 906         u32 checkfid = data->currfid;
 907         u32 checkvid = data->currvid;
 908         unsigned int newstate;
 909         int ret = -EIO;
 910         int i;
 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         schedule();
 916
 917         if (smp_processor_id() != pol->cpu) {
 918                 printk(KERN_ERR "limiting to cpu %u failed\n", pol->cpu);
 919                 goto err_out;
 920         }
 921
 922         if (pending_bit_stuck()) {
 923                 printk(KERN_ERR PFX "failing targ, change pending bit set\n");
 924                 goto err_out;
 925         }
 926
 927         dprintk("targ: cpu %d, %d kHz, min %d, max %d, relation %d\n",
 928                 pol->cpu, targfreq, pol->min, pol->max, relation);
 929
 930         if (query_current_values_with_pending_wait(data)) {
 931                 ret = -EIO;
 932                 goto err_out;
 933         }
 934
 935         dprintk("targ: curr fid 0x%x, vid 0x%x\n",
 936                 data->currfid, data->currvid);
 937
 938         if ((checkvid != data->currvid) || (checkfid != data->currfid)) {
 939                 printk(KERN_INFO PFX
 940                         "error - out of sync, fix 0x%x 0x%x, vid 0x%x 0x%x\n",
 941                         checkfid, data->currfid, checkvid, data->currvid);
 942         }
 943
 944         if (cpufreq_frequency_table_target(pol, data->powernow_table, targfreq, relation, &newstate))
 945                 goto err_out;
 946
 947         down(&fidvid_sem);
 948
 949         powernow_k8_acpi_pst_values(data, newstate);
 950
 951         if (transition_frequency(data, newstate)) {
 952                 printk(KERN_ERR PFX "transition frequency failed\n");
 953                 ret = 1;
 954                 up(&fidvid_sem);
 955                 goto err_out;
 956         }
 957
 958         /* Update all the fid/vids of our siblings */
 959         for_each_cpu_mask(i, cpu_core_map[pol->cpu]) {
 960                 powernow_data[i]->currvid = data->currvid;
 961                 powernow_data[i]->currfid = data->currfid;
 962         }
 963         up(&fidvid_sem);
 964
 965         pol->cur = find_khz_freq_from_fid(data->currfid);
 966         ret = 0;
 967
 968 err_out:
 969         set_cpus_allowed(current, oldmask);
 970         schedule();
 971
 972         return ret;
 973 }
 974
 975 /* Driver entry point to verify the policy and range of frequencies */
 976 static int powernowk8_verify(struct cpufreq_policy *pol)
 977 {
 978         struct powernow_k8_data *data = powernow_data[pol->cpu];
 979
 980         return cpufreq_frequency_table_verify(pol, data->powernow_table);
 981 }
 982
 983 /* per CPU init entry point to the driver */
 984 static int __init powernowk8_cpu_init(struct cpufreq_policy *pol)
 985 {
 986         struct powernow_k8_data *data;
 987         cpumask_t oldmask = CPU_MASK_ALL;
 988         int rc, i;
 989
 990         if (!check_supported_cpu(pol->cpu))
 991                 return -ENODEV;
 992
 993         data = kmalloc(sizeof(struct powernow_k8_data), GFP_KERNEL);
 994         if (!data) {
 995                 printk(KERN_ERR PFX "unable to alloc powernow_k8_data");
 996                 return -ENOMEM;
 997         }
 998         memset(data,0,sizeof(struct powernow_k8_data));
 999
1000         data->cpu = pol->cpu;
1001
1002         if (powernow_k8_cpu_init_acpi(data)) {
1003                 /*
1004                  * Use the PSB BIOS structure. This is only availabe on
1005                  * an UP version, and is deprecated by AMD.
1006                  */
1007
1008                 if ((num_online_cpus() != 1) || (num_possible_cpus() != 1)) {
1009                         printk(KERN_ERR PFX "MP systems not supported by PSB BIOS structure\n");
1010                         kfree(data);
1011                         return -ENODEV;
1012                 }
1013                 if (pol->cpu != 0) {
1014                         printk(KERN_ERR PFX "init not cpu 0\n");
1015                         kfree(data);
1016                         return -ENODEV;
1017                 }
1018                 rc = find_psb_table(data);
1019                 if (rc) {
1020                         kfree(data);
1021                         return -ENODEV;
1022                 }
1023         }
1024
1025         /* only run on specific CPU from here on */
1026         oldmask = current->cpus_allowed;
1027         set_cpus_allowed(current, cpumask_of_cpu(pol->cpu));
1028         schedule();
1029
1030         if (smp_processor_id() != pol->cpu) {
1031                 printk(KERN_ERR "limiting to cpu %u failed\n", pol->cpu);
1032                 goto err_out;
1033         }
1034
1035         if (pending_bit_stuck()) {
1036                 printk(KERN_ERR PFX "failing init, change pending bit set\n");
1037                 goto err_out;
1038         }
1039
1040         if (query_current_values_with_pending_wait(data))
1041                 goto err_out;
1042
1043         fidvid_msr_init();
1044
1045         /* run on any CPU again */
1046         set_cpus_allowed(current, oldmask);
1047         schedule();
1048
1049         pol->governor = CPUFREQ_DEFAULT_GOVERNOR;
1050         pol->cpus = cpu_core_map[pol->cpu];
1051
1052         /* Take a crude guess here.
1053          * That guess was in microseconds, so multiply with 1000 */
1054         pol->cpuinfo.transition_latency = (((data->rvo + 8) * data->vstable * VST_UNITS_20US)
1055             + (3 * (1 << data->irt) * 10)) * 1000;
1056
1057         pol->cur = find_khz_freq_from_fid(data->currfid);
1058         dprintk("policy current frequency %d kHz\n", pol->cur);
1059
1060         /* min/max the cpu is capable of */
1061         if (cpufreq_frequency_table_cpuinfo(pol, data->powernow_table)) {
1062                 printk(KERN_ERR PFX "invalid powernow_table\n");
1063                 powernow_k8_cpu_exit_acpi(data);
1064                 kfree(data->powernow_table);
1065                 kfree(data);
1066                 return -EINVAL;
1067         }
1068
1069         cpufreq_frequency_table_get_attr(data->powernow_table, pol->cpu);
1070
1071         printk("cpu_init done, current fid 0x%x, vid 0x%x\n",
1072                data->currfid, data->currvid);
1073
1074         for_each_cpu_mask(i, cpu_core_map[pol->cpu]) {
1075                 powernow_data[i] = data;
1076         }
1077
1078         return 0;
1079
1080 err_out:
1081         set_cpus_allowed(current, oldmask);
1082         schedule();
1083         powernow_k8_cpu_exit_acpi(data);
1084
1085         kfree(data);
1086         return -ENODEV;
1087 }
1088
1089 static int __devexit powernowk8_cpu_exit (struct cpufreq_policy *pol)
1090 {
1091         struct powernow_k8_data *data = powernow_data[pol->cpu];
1092
1093         if (!data)
1094                 return -EINVAL;
1095
1096         powernow_k8_cpu_exit_acpi(data);
1097
1098         cpufreq_frequency_table_put_attr(pol->cpu);
1099
1100         kfree(data->powernow_table);
1101         kfree(data);
1102
1103         return 0;
1104 }
1105
1106 static unsigned int powernowk8_get (unsigned int cpu)
1107 {
1108         struct powernow_k8_data *data = powernow_data[cpu];
1109         cpumask_t oldmask = current->cpus_allowed;
1110         unsigned int khz = 0;
1111
1112         set_cpus_allowed(current, cpumask_of_cpu(cpu));
1113         if (smp_processor_id() != cpu) {
1114                 printk(KERN_ERR PFX "limiting to CPU %d failed in powernowk8_get\n", cpu);
1115                 set_cpus_allowed(current, oldmask);
1116                 return 0;
1117         }
1118         preempt_disable();
1119
1120         if (query_current_values_with_pending_wait(data))
1121                 goto out;
1122
1123         khz = find_khz_freq_from_fid(data->currfid);
1124
1125  out:
1126         preempt_enable_no_resched();
1127         set_cpus_allowed(current, oldmask);
1128
1129         return khz;
1130 }
1131
1132 static struct freq_attr* powernow_k8_attr[] = {
1133         &cpufreq_freq_attr_scaling_available_freqs,
1134         NULL,
1135 };
1136
1137 static struct cpufreq_driver cpufreq_amd64_driver = {
1138         .verify = powernowk8_verify,
1139         .target = powernowk8_target,
1140         .init = powernowk8_cpu_init,
1141         .exit = __devexit_p(powernowk8_cpu_exit),
1142         .get = powernowk8_get,
1143         .name = "powernow-k8",
1144         .owner = THIS_MODULE,
1145         .attr = powernow_k8_attr,
1146 };
1147
1148 /* driver entry point for init */
1149 static int __init powernowk8_init(void)
1150 {
1151         unsigned int i, supported_cpus = 0;
1152
1153         for (i=0; i<NR_CPUS; i++) {
1154                 if (!cpu_online(i))
1155                         continue;
1156                 if (check_supported_cpu(i))
1157                         supported_cpus++;
1158         }
1159
1160         if (supported_cpus == num_online_cpus()) {
1161                 printk(KERN_INFO PFX "Found %d AMD Athlon 64 / Opteron processors (" VERSION ")\n",
1162                         supported_cpus);
1163                 return cpufreq_register_driver(&cpufreq_amd64_driver);
1164         }
1165
1166         return -ENODEV;
1167 }
1168
1169 /* driver entry point for term */
1170 static void __exit powernowk8_exit(void)
1171 {
1172         dprintk("exit\n");
1173
1174         cpufreq_unregister_driver(&cpufreq_amd64_driver);
1175 }
1176
1177 MODULE_AUTHOR("Paul Devriendt <paul.devriendt@amd.com> and Mark Langsdorf <mark.langsdorf@amd.com.");
1178 MODULE_DESCRIPTION("AMD Athlon 64 and Opteron processor frequency driver.");
1179 MODULE_LICENSE("GPL");
1180
1181 late_initcall(powernowk8_init);
1182 module_exit(powernowk8_exit);
1183