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

   1 /*
   2  *   (c) 2003, 2004, 2005 Advanced Micro Devices, Inc.
   3  *  Your use of this code is subject to the terms and conditions of the
   4  *  GNU general public license version 2. See "COPYING" or
   5  *  http://www.gnu.org/licenses/gpl.html
   6  *
   7  *  Support : mark.langsdorf@amd.com
   8  *
   9  *  Based on the powernow-k7.c module written by Dave Jones.
  10  *  (C) 2003 Dave Jones <davej@codemonkey.org.uk> on behalf of SuSE Labs
  11  *  (C) 2004 Dominik Brodowski <linux@brodo.de>
  12  *  (C) 2004 Pavel Machek <pavel@suse.cz>
  13  *  Licensed under the terms of the GNU GPL License version 2.
  14  *  Based upon datasheets & sample CPUs kindly provided by AMD.
  15  *
  16  *  Valuable input gratefully received from Dave Jones, Pavel Machek,
  17  *  Dominik Brodowski, and others.
  18  *  Originally developed by Paul Devriendt.
  19  *  Processor information obtained from Chapter 9 (Power and Thermal Management)
  20  *  of the "BIOS and Kernel Developer's Guide for the AMD Athlon 64 and AMD
  21  *  Opteron Processors" available for download from www.amd.com
  22  *
  23  *  Tables for specific CPUs can be infrerred from
  24  *     http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/30430.pdf
  25  */
  26
  27 #include <linux/kernel.h>
  28 #include <linux/smp.h>
  29 #include <linux/module.h>
  30 #include <linux/init.h>
  31 #include <linux/cpufreq.h>
  32 #include <linux/slab.h>
  33 #include <linux/string.h>
  34 #include <linux/cpumask.h>
  35 #include <linux/sched.h>        /* for current / set_cpus_allowed() */
  36
  37 #include <asm/msr.h>
  38 #include <asm/io.h>
  39 #include <asm/delay.h>
  40
  41 #ifdef CONFIG_X86_POWERNOW_K8_ACPI
  42 #include <linux/acpi.h>
  43 #include <acpi/processor.h>
  44 #endif
  45
  46 #define PFX "powernow-k8: "
  47 #define BFX PFX "BIOS error: "
  48 #define VERSION "version 1.50.4"
  49 #include "powernow-k8.h"
  50
  51 /* serialize freq changes  */
  52 static DECLARE_MUTEX(fidvid_sem);
  53
  54 static struct powernow_k8_data *powernow_data[NR_CPUS];
  55
  56 #ifndef CONFIG_SMP
  57 static cpumask_t cpu_core_map[1];
  58 #endif
  59
  60 /* Return a frequency in MHz, given an input fid */
  61 static u32 find_freq_from_fid(u32 fid)
  62 {
  63         return 800 + (fid * 100);
  64 }
  65
  66 /* Return a frequency in KHz, given an input fid */
  67 static u32 find_khz_freq_from_fid(u32 fid)
  68 {
  69         return 1000 * find_freq_from_fid(fid);
  70 }
  71
  72 /* Return a voltage in miliVolts, given an input vid */
  73 static u32 find_millivolts_from_vid(struct powernow_k8_data *data, u32 vid)
  74 {
  75         return 1550-vid*25;
  76 }
  77
  78 /* Return the vco fid for an input fid
  79  *
  80  * Each "low" fid has corresponding "high" fid, and you can get to "low" fids
  81  * only from corresponding high fids. This returns "high" fid corresponding to
  82  * "low" one.
  83  */
  84 static u32 convert_fid_to_vco_fid(u32 fid)
  85 {
  86         if (fid < HI_FID_TABLE_BOTTOM) {
  87                 return 8 + (2 * fid);
  88         } else {
  89                 return fid;
  90         }
  91 }
  92
  93 /*
  94  * Return 1 if the pending bit is set. Unless we just instructed the processor
  95  * to transition to a new state, seeing this bit set is really bad news.
  96  */
  97 static int pending_bit_stuck(void)
  98 {
  99         u32 lo, hi;
 100
 101         rdmsr(MSR_FIDVID_STATUS, lo, hi);
 102         return lo & MSR_S_LO_CHANGE_PENDING ? 1 : 0;
 103 }
 104
 105 /*
 106  * Update the global current fid / vid values from the status msr.
 107  * Returns 1 on error.
 108  */
 109 static int query_current_values_with_pending_wait(struct powernow_k8_data *data)
 110 {
 111         u32 lo, hi;
 112         u32 i = 0;
 113
 114         do {
 115                 if (i++ > 10000) {
 116                         dprintk("detected change pending stuck\n");
 117                         return 1;
 118                 }
 119                 rdmsr(MSR_FIDVID_STATUS, lo, hi);
 120         } while (lo & MSR_S_LO_CHANGE_PENDING);
 121
 122         data->currvid = hi & MSR_S_HI_CURRENT_VID;
 123         data->currfid = lo & MSR_S_LO_CURRENT_FID;
 124
 125         return 0;
 126 }
 127
 128 /* the isochronous relief time */
 129 static void count_off_irt(struct powernow_k8_data *data)
 130 {
 131         udelay((1 << data->irt) * 10);
 132         return;
 133 }
 134
 135 /* the voltage stabalization time */
 136 static void count_off_vst(struct powernow_k8_data *data)
 137 {
 138         udelay(data->vstable * VST_UNITS_20US);
 139         return;
 140 }
 141
 142 /* need to init the control msr to a safe value (for each cpu) */
 143 static void fidvid_msr_init(void)
 144 {
 145         u32 lo, hi;
 146         u8 fid, vid;
 147
 148         rdmsr(MSR_FIDVID_STATUS, lo, hi);
 149         vid = hi & MSR_S_HI_CURRENT_VID;
 150         fid = lo & MSR_S_LO_CURRENT_FID;
 151         lo = fid | (vid << MSR_C_LO_VID_SHIFT);
 152         hi = MSR_C_HI_STP_GNT_BENIGN;
 153         dprintk("cpu%d, init lo 0x%x, hi 0x%x\n", smp_processor_id(), lo, hi);
 154         wrmsr(MSR_FIDVID_CTL, lo, hi);
 155 }
 156
 157
 158 /* write the new fid value along with the other control fields to the msr */
 159 static int write_new_fid(struct powernow_k8_data *data, u32 fid)
 160 {
 161         u32 lo;
 162         u32 savevid = data->currvid;
 163         u32 i = 0;
 164
 165         if ((fid & INVALID_FID_MASK) || (data->currvid & INVALID_VID_MASK)) {
 166                 printk(KERN_ERR PFX "internal error - overflow on fid write\n");
 167                 return 1;
 168         }
 169
 170         lo = fid | (data->currvid << MSR_C_LO_VID_SHIFT) | MSR_C_LO_INIT_FID_VID;
 171
 172         dprintk("writing fid 0x%x, lo 0x%x, hi 0x%x\n",
 173                 fid, lo, data->plllock * PLL_LOCK_CONVERSION);
 174
 175         do {
 176                 wrmsr(MSR_FIDVID_CTL, lo, data->plllock * PLL_LOCK_CONVERSION);
 177                 if (i++ > 100) {
 178                         printk(KERN_ERR PFX "internal error - pending bit very stuck - no further pstate changes possible\n");
 179                         return 1;
 180                 }
 181         } while (query_current_values_with_pending_wait(data));
 182
 183         count_off_irt(data);
 184
 185         if (savevid != data->currvid) {
 186                 printk(KERN_ERR PFX "vid change on fid trans, old 0x%x, new 0x%x\n",
 187                        savevid, data->currvid);
 188                 return 1;
 189         }
 190
 191         if (fid != data->currfid) {
 192                 printk(KERN_ERR PFX "fid trans failed, fid 0x%x, curr 0x%x\n", fid,
 193                         data->currfid);
 194                 return 1;
 195         }
 196
 197         return 0;
 198 }
 199
 200 /* Write a new vid to the hardware */
 201 static int write_new_vid(struct powernow_k8_data *data, u32 vid)
 202 {
 203         u32 lo;
 204         u32 savefid = data->currfid;
 205         int i = 0;
 206
 207         if ((data->currfid & INVALID_FID_MASK) || (vid & INVALID_VID_MASK)) {
 208                 printk(KERN_ERR PFX "internal error - overflow on vid write\n");
 209                 return 1;
 210         }
 211
 212         lo = data->currfid | (vid << MSR_C_LO_VID_SHIFT) | MSR_C_LO_INIT_FID_VID;
 213
 214         dprintk("writing vid 0x%x, lo 0x%x, hi 0x%x\n",
 215                 vid, lo, STOP_GRANT_5NS);
 216
 217         do {
 218                 wrmsr(MSR_FIDVID_CTL, lo, STOP_GRANT_5NS);
 219                 if (i++ > 100) {
 220                         printk(KERN_ERR PFX "internal error - pending bit very stuck - no further pstate changes possible\n");
 221                         return 1;
 222                 }
 223         } while (query_current_values_with_pending_wait(data));
 224
 225         if (savefid != data->currfid) {
 226                 printk(KERN_ERR PFX "fid changed on vid trans, old 0x%x new 0x%x\n",
 227                        savefid, data->currfid);
 228                 return 1;
 229         }
 230
 231         if (vid != data->currvid) {
 232                 printk(KERN_ERR PFX "vid trans failed, vid 0x%x, curr 0x%x\n", vid,
 233                                 data->currvid);
 234                 return 1;
 235         }
 236
 237         return 0;
 238 }
 239
 240 /*
 241  * Reduce the vid by the max of step or reqvid.
 242  * Decreasing vid codes represent increasing voltages:
 243  * vid of 0 is 1.550V, vid of 0x1e is 0.800V, vid of VID_OFF is off.
 244  */
 245 static int decrease_vid_code_by_step(struct powernow_k8_data *data, u32 reqvid, u32 step)
 246 {
 247         if ((data->currvid - reqvid) > step)
 248                 reqvid = data->currvid - step;
 249
 250         if (write_new_vid(data, reqvid))
 251                 return 1;
 252
 253         count_off_vst(data);
 254
 255         return 0;
 256 }
 257
 258 /* Change the fid and vid, by the 3 phases. */
 259 static int transition_fid_vid(struct powernow_k8_data *data, u32 reqfid, u32 reqvid)
 260 {
 261         if (core_voltage_pre_transition(data, reqvid))
 262                 return 1;
 263
 264         if (core_frequency_transition(data, reqfid))
 265                 return 1;
 266
 267         if (core_voltage_post_transition(data, reqvid))
 268                 return 1;
 269
 270         if (query_current_values_with_pending_wait(data))
 271                 return 1;
 272
 273         if ((reqfid != data->currfid) || (reqvid != data->currvid)) {
 274                 printk(KERN_ERR PFX "failed (cpu%d): req 0x%x 0x%x, curr 0x%x 0x%x\n",
 275                                 smp_processor_id(),
 276                                 reqfid, reqvid, data->currfid, data->currvid);
 277                 return 1;
 278         }
 279
 280         dprintk("transitioned (cpu%d): new fid 0x%x, vid 0x%x\n",
 281                 smp_processor_id(), data->currfid, data->currvid);
 282
 283         return 0;
 284 }
 285
 286 /* Phase 1 - core voltage transition ... setup voltage */
 287 static int core_voltage_pre_transition(struct powernow_k8_data *data, u32 reqvid)
 288 {
 289         u32 rvosteps = data->rvo;
 290         u32 savefid = data->currfid;
 291         u32 maxvid, lo;
 292
 293         dprintk("ph1 (cpu%d): start, currfid 0x%x, currvid 0x%x, reqvid 0x%x, rvo 0x%x\n",
 294                 smp_processor_id(),
 295                 data->currfid, data->currvid, reqvid, data->rvo);
 296
 297         rdmsr(MSR_FIDVID_STATUS, lo, maxvid);
 298         maxvid = 0x1f & (maxvid >> 16);
 299         dprintk("ph1 maxvid=0x%x\n", maxvid);
 300         if (reqvid < maxvid) /* lower numbers are higher voltages */
 301                 reqvid = maxvid;
 302
 303         while (data->currvid > reqvid) {
 304                 dprintk("ph1: curr 0x%x, req vid 0x%x\n",
 305                         data->currvid, reqvid);
 306                 if (decrease_vid_code_by_step(data, reqvid, data->vidmvs))
 307                         return 1;
 308         }
 309
 310         while ((rvosteps > 0) && ((data->rvo + data->currvid) > reqvid)) {
 311                 if (data->currvid == maxvid) {
 312                         rvosteps = 0;
 313                 } else {
 314                         dprintk("ph1: changing vid for rvo, req 0x%x\n",
 315                                 data->currvid - 1);
 316                         if (decrease_vid_code_by_step(data, data->currvid - 1, 1))
 317                                 return 1;
 318                         rvosteps--;
 319                 }
 320         }
 321
 322         if (query_current_values_with_pending_wait(data))
 323                 return 1;
 324
 325         if (savefid != data->currfid) {
 326                 printk(KERN_ERR PFX "ph1 err, currfid changed 0x%x\n", data->currfid);
 327                 return 1;
 328         }
 329
 330         dprintk("ph1 complete, currfid 0x%x, currvid 0x%x\n",
 331                 data->currfid, data->currvid);
 332
 333         return 0;
 334 }
 335
 336 /* Phase 2 - core frequency transition */
 337 static int core_frequency_transition(struct powernow_k8_data *data, u32 reqfid)
 338 {
 339         u32 vcoreqfid, vcocurrfid, vcofiddiff, savevid = data->currvid;
 340
 341         if ((reqfid < HI_FID_TABLE_BOTTOM) && (data->currfid < HI_FID_TABLE_BOTTOM)) {
 342                 printk(KERN_ERR PFX "ph2: illegal lo-lo transition 0x%x 0x%x\n",
 343                         reqfid, data->currfid);
 344                 return 1;
 345         }
 346
 347         if (data->currfid == reqfid) {
 348                 printk(KERN_ERR PFX "ph2 null fid transition 0x%x\n", data->currfid);
 349                 return 0;
 350         }
 351
 352         dprintk("ph2 (cpu%d): starting, currfid 0x%x, currvid 0x%x, reqfid 0x%x\n",
 353                 smp_processor_id(),
 354                 data->currfid, data->currvid, reqfid);
 355
 356         vcoreqfid = convert_fid_to_vco_fid(reqfid);
 357         vcocurrfid = convert_fid_to_vco_fid(data->currfid);
 358         vcofiddiff = vcocurrfid > vcoreqfid ? vcocurrfid - vcoreqfid
 359             : vcoreqfid - vcocurrfid;
 360
 361         while (vcofiddiff > 2) {
 362                 if (reqfid > data->currfid) {
 363                         if (data->currfid > LO_FID_TABLE_TOP) {
 364                                 if (write_new_fid(data, data->currfid + 2)) {
 365                                         return 1;
 366                                 }
 367                         } else {
 368                                 if (write_new_fid
 369                                     (data, 2 + convert_fid_to_vco_fid(data->currfid))) {
 370                                         return 1;
 371                                 }
 372                         }
 373                 } else {
 374                         if (write_new_fid(data, data->currfid - 2))
 375                                 return 1;
 376                 }
 377
 378                 vcocurrfid = convert_fid_to_vco_fid(data->currfid);
 379                 vcofiddiff = vcocurrfid > vcoreqfid ? vcocurrfid - vcoreqfid
 380                     : vcoreqfid - vcocurrfid;
 381         }
 382
 383         if (write_new_fid(data, reqfid))
 384                 return 1;
 385
 386         if (query_current_values_with_pending_wait(data))
 387                 return 1;
 388
 389         if (data->currfid != reqfid) {
 390                 printk(KERN_ERR PFX
 391                         "ph2: mismatch, failed fid transition, curr 0x%x, req 0x%x\n",
 392                         data->currfid, reqfid);
 393                 return 1;
 394         }
 395
 396         if (savevid != data->currvid) {
 397                 printk(KERN_ERR PFX "ph2: vid changed, save 0x%x, curr 0x%x\n",
 398                         savevid, data->currvid);
 399                 return 1;
 400         }
 401
 402         dprintk("ph2 complete, currfid 0x%x, currvid 0x%x\n",
 403                 data->currfid, data->currvid);
 404
 405         return 0;
 406 }
 407
 408 /* Phase 3 - core voltage transition flow ... jump to the final vid. */
 409 static int core_voltage_post_transition(struct powernow_k8_data *data, u32 reqvid)
 410 {
 411         u32 savefid = data->currfid;
 412         u32 savereqvid = reqvid;
 413
 414         dprintk("ph3 (cpu%d): starting, currfid 0x%x, currvid 0x%x\n",
 415                 smp_processor_id(),
 416                 data->currfid, data->currvid);
 417
 418         if (reqvid != data->currvid) {
 419                 if (write_new_vid(data, reqvid))
 420                         return 1;
 421
 422                 if (savefid != data->currfid) {
 423                         printk(KERN_ERR PFX
 424                                "ph3: bad fid change, save 0x%x, curr 0x%x\n",
 425                                savefid, data->currfid);
 426                         return 1;
 427                 }
 428
 429                 if (data->currvid != reqvid) {
 430                         printk(KERN_ERR PFX
 431                                "ph3: failed vid transition\n, req 0x%x, curr 0x%x",
 432                                reqvid, data->currvid);
 433                         return 1;
 434                 }
 435         }
 436
 437         if (query_current_values_with_pending_wait(data))
 438                 return 1;
 439
 440         if (savereqvid != data->currvid) {
 441                 dprintk("ph3 failed, currvid 0x%x\n", data->currvid);
 442                 return 1;
 443         }
 444
 445         if (savefid != data->currfid) {
 446                 dprintk("ph3 failed, currfid changed 0x%x\n",
 447                         data->currfid);
 448                 return 1;
 449         }
 450
 451         dprintk("ph3 complete, currfid 0x%x, currvid 0x%x\n",
 452                 data->currfid, data->currvid);
 453
 454         return 0;
 455 }
 456
 457 static int check_supported_cpu(unsigned int cpu)
 458 {
 459         cpumask_t oldmask = CPU_MASK_ALL;
 460         u32 eax, ebx, ecx, edx;
 461         unsigned int rc = 0;
 462
 463         oldmask = current->cpus_allowed;
 464         set_cpus_allowed(current, cpumask_of_cpu(cpu));
 465         schedule();
 466
 467         if (smp_processor_id() != cpu) {
 468                 printk(KERN_ERR "limiting to cpu %u failed\n", cpu);
 469                 goto out;
 470         }
 471
 472         if (current_cpu_data.x86_vendor != X86_VENDOR_AMD)
 473                 goto out;
 474
 475         eax = cpuid_eax(CPUID_PROCESSOR_SIGNATURE);
 476         if (((eax & CPUID_USE_XFAM_XMOD) != CPUID_USE_XFAM_XMOD) ||
 477             ((eax & CPUID_XFAM) != CPUID_XFAM_K8) ||
 478             ((eax & CPUID_XMOD) > CPUID_XMOD_REV_F)) {
 479                 printk(KERN_INFO PFX "Processor cpuid %x not supported\n", eax);
 480                 goto out;
 481         }
 482
 483         eax = cpuid_eax(CPUID_GET_MAX_CAPABILITIES);
 484         if (eax < CPUID_FREQ_VOLT_CAPABILITIES) {
 485                 printk(KERN_INFO PFX
 486                        "No frequency change capabilities detected\n");
 487                 goto out;
 488         }
 489
 490         cpuid(CPUID_FREQ_VOLT_CAPABILITIES, &eax, &ebx, &ecx, &edx);
 491         if ((edx & P_STATE_TRANSITION_CAPABLE) != P_STATE_TRANSITION_CAPABLE) {
 492                 printk(KERN_INFO PFX "Power state transitions not supported\n");
 493                 goto out;
 494         }
 495
 496         rc = 1;
 497
 498 out:
 499         set_cpus_allowed(current, oldmask);
 500         schedule();
 501         return rc;
 502
 503 }
 504
 505 static int check_pst_table(struct powernow_k8_data *data, struct pst_s *pst, u8 maxvid)
 506 {
 507         unsigned int j;
 508         u8 lastfid = 0xff;
 509
 510         for (j = 0; j < data->numps; j++) {
 511                 if (pst[j].vid > LEAST_VID) {
 512                         printk(KERN_ERR PFX "vid %d invalid : 0x%x\n", j, pst[j].vid);
 513                         return -EINVAL;
 514                 }
 515                 if (pst[j].vid < data->rvo) {   /* vid + rvo >= 0 */
 516                         printk(KERN_ERR BFX "0 vid exceeded with pstate %d\n", j);
 517                         return -ENODEV;
 518                 }
 519                 if (pst[j].vid < maxvid + data->rvo) {  /* vid + rvo >= maxvid */
 520                         printk(KERN_ERR BFX "maxvid exceeded with pstate %d\n", j);
 521                         return -ENODEV;
 522                 }
 523                 if ((pst[j].fid > MAX_FID)
 524                     || (pst[j].fid & 1)
 525                     || (j && (pst[j].fid < HI_FID_TABLE_BOTTOM))) {
 526                         /* Only first fid is allowed to be in "low" range */
 527                         printk(KERN_ERR PFX "two low fids - %d : 0x%x\n", j, pst[j].fid);
 528                         return -EINVAL;
 529                 }
 530                 if (pst[j].fid < lastfid)
 531                         lastfid = pst[j].fid;
 532         }
 533         if (lastfid & 1) {
 534                 printk(KERN_ERR PFX "lastfid invalid\n");
 535                 return -EINVAL;
 536         }
 537         if (lastfid > LO_FID_TABLE_TOP)
 538                 printk(KERN_INFO PFX  "first fid not from lo freq table\n");
 539
 540         return 0;
 541 }
 542
 543 static void print_basics(struct powernow_k8_data *data)
 544 {
 545         int j;
 546         for (j = 0; j < data->numps; j++) {
 547                 if (data->powernow_table[j].frequency != CPUFREQ_ENTRY_INVALID)
 548                         printk(KERN_INFO PFX "   %d : fid 0x%x (%d MHz), vid 0x%x (%d mV)\n", j,
 549                                 data->powernow_table[j].index & 0xff,
 550                                 data->powernow_table[j].frequency/1000,
 551                                 data->powernow_table[j].index >> 8,
 552                                 find_millivolts_from_vid(data, data->powernow_table[j].index >> 8));
 553         }
 554         if (data->batps)
 555                 printk(KERN_INFO PFX "Only %d pstates on battery\n", data->batps);
 556 }
 557
 558 static int fill_powernow_table(struct powernow_k8_data *data, struct pst_s *pst, u8 maxvid)
 559 {
 560         struct cpufreq_frequency_table *powernow_table;
 561         unsigned int j;
 562
 563         if (data->batps) {    /* use ACPI support to get full speed on mains power */
 564                 printk(KERN_WARNING PFX "Only %d pstates usable (use ACPI driver for full range\n", data->batps);
 565                 data->numps = data->batps;
 566         }
 567
 568         for ( j=1; j<data->numps; j++ ) {
 569                 if (pst[j-1].fid >= pst[j].fid) {
 570                         printk(KERN_ERR PFX "PST out of sequence\n");
 571                         return -EINVAL;
 572                 }
 573         }
 574
 575         if (data->numps < 2) {
 576                 printk(KERN_ERR PFX "no p states to transition\n");
 577                 return -ENODEV;
 578         }
 579
 580         if (check_pst_table(data, pst, maxvid))
 581                 return -EINVAL;
 582
 583         powernow_table = kmalloc((sizeof(struct cpufreq_frequency_table)
 584                 * (data->numps + 1)), GFP_KERNEL);
 585         if (!powernow_table) {
 586                 printk(KERN_ERR PFX "powernow_table memory alloc failure\n");
 587                 return -ENOMEM;
 588         }
 589
 590         for (j = 0; j < data->numps; j++) {
 591                 powernow_table[j].index = pst[j].fid; /* lower 8 bits */
 592                 powernow_table[j].index |= (pst[j].vid << 8); /* upper 8 bits */
 593                 powernow_table[j].frequency = find_khz_freq_from_fid(pst[j].fid);
 594         }
 595         powernow_table[data->numps].frequency = CPUFREQ_TABLE_END;
 596         powernow_table[data->numps].index = 0;
 597
 598         if (query_current_values_with_pending_wait(data)) {
 599                 kfree(powernow_table);
 600                 return -EIO;
 601         }
 602
 603         dprintk("cfid 0x%x, cvid 0x%x\n", data->currfid, data->currvid);
 604         data->powernow_table = powernow_table;
 605         print_basics(data);
 606
 607         for (j = 0; j < data->numps; j++)
 608                 if ((pst[j].fid==data->currfid) && (pst[j].vid==data->currvid))
 609                         return 0;
 610
 611         dprintk("currfid/vid do not match PST, ignoring\n");
 612         return 0;
 613 }
 614
 615 /* Find and validate the PSB/PST table in BIOS. */
 616 static int find_psb_table(struct powernow_k8_data *data)
 617 {
 618         struct psb_s *psb;
 619         unsigned int i;
 620         u32 mvs;
 621         u8 maxvid;
 622         u32 cpst = 0;
 623         u32 thiscpuid;
 624
 625         for (i = 0xc0000; i < 0xffff0; i += 0x10) {
 626                 /* Scan BIOS looking for the signature. */
 627                 /* It can not be at ffff0 - it is too big. */
 628
 629                 psb = phys_to_virt(i);
 630                 if (memcmp(psb, PSB_ID_STRING, PSB_ID_STRING_LEN) != 0)
 631                         continue;
 632
 633                 dprintk("found PSB header at 0x%p\n", psb);
 634
 635                 dprintk("table vers: 0x%x\n", psb->tableversion);
 636                 if (psb->tableversion != PSB_VERSION_1_4) {
 637                         printk(KERN_INFO BFX "PSB table is not v1.4\n");
 638                         return -ENODEV;
 639                 }
 640
 641                 dprintk("flags: 0x%x\n", psb->flags1);
 642                 if (psb->flags1) {
 643                         printk(KERN_ERR BFX "unknown flags\n");
 644                         return -ENODEV;
 645                 }
 646
 647                 data->vstable = psb->vstable;
 648                 dprintk("voltage stabilization time: %d(*20us)\n", data->vstable);
 649
 650                 dprintk("flags2: 0x%x\n", psb->flags2);
 651                 data->rvo = psb->flags2 & 3;
 652                 data->irt = ((psb->flags2) >> 2) & 3;
 653                 mvs = ((psb->flags2) >> 4) & 3;
 654                 data->vidmvs = 1 << mvs;
 655                 data->batps = ((psb->flags2) >> 6) & 3;
 656
 657                 dprintk("ramp voltage offset: %d\n", data->rvo);
 658                 dprintk("isochronous relief time: %d\n", data->irt);
 659                 dprintk("maximum voltage step: %d - 0x%x\n", mvs, data->vidmvs);
 660
 661                 dprintk("numpst: 0x%x\n", psb->num_tables);
 662                 cpst = psb->num_tables;
 663                 if ((psb->cpuid == 0x00000fc0) || (psb->cpuid == 0x00000fe0) ){
 664                         thiscpuid = cpuid_eax(CPUID_PROCESSOR_SIGNATURE);
 665                         if ((thiscpuid == 0x00000fc0) || (thiscpuid == 0x00000fe0) ) {
 666                                 cpst = 1;
 667                         }
 668                 }
 669                 if (cpst != 1) {
 670                         printk(KERN_ERR BFX "numpst must be 1\n");
 671                         return -ENODEV;
 672                 }
 673
 674                 data->plllock = psb->plllocktime;
 675                 dprintk("plllocktime: 0x%x (units 1us)\n", psb->plllocktime);
 676                 dprintk("maxfid: 0x%x\n", psb->maxfid);
 677                 dprintk("maxvid: 0x%x\n", psb->maxvid);
 678                 maxvid = psb->maxvid;
 679
 680                 data->numps = psb->numps;
 681                 dprintk("numpstates: 0x%x\n", data->numps);
 682                 return fill_powernow_table(data, (struct pst_s *)(psb+1), maxvid);
 683         }
 684         /*
 685          * If you see this message, complain to BIOS manufacturer. If
 686          * he tells you "we do not support Linux" or some similar
 687          * nonsense, remember that Windows 2000 uses the same legacy
 688          * mechanism that the old Linux PSB driver uses. Tell them it
 689          * is broken with Windows 2000.
 690          *
 691          * The reference to the AMD documentation is chapter 9 in the
 692          * BIOS and Kernel Developer's Guide, which is available on
 693          * www.amd.com
 694          */
 695         printk(KERN_INFO PFX "BIOS error - no PSB or ACPI _PSS objects\n");
 696         return -ENODEV;
 697 }
 698
 699 #ifdef CONFIG_X86_POWERNOW_K8_ACPI
 700 static void powernow_k8_acpi_pst_values(struct powernow_k8_data *data, unsigned int index)
 701 {
 702         if (!data->acpi_data.state_count)
 703                 return;
 704
 705         data->irt = (data->acpi_data.states[index].control >> IRT_SHIFT) & IRT_MASK;
 706         data->rvo = (data->acpi_data.states[index].control >> RVO_SHIFT) & RVO_MASK;
 707         data->exttype = (data->acpi_data.states[index].control >> EXT_TYPE_SHIFT) & EXT_TYPE_MASK;
 708         data->plllock = (data->acpi_data.states[index].control >> PLL_L_SHIFT) & PLL_L_MASK;
 709         data->vidmvs = 1 << ((data->acpi_data.states[index].control >> MVS_SHIFT) & MVS_MASK);
 710         data->vstable = (data->acpi_data.states[index].control >> VST_SHIFT) & VST_MASK;
 711 }
 712
 713 static int powernow_k8_cpu_init_acpi(struct powernow_k8_data *data)
 714 {
 715         int i;
 716         int cntlofreq = 0;
 717         struct cpufreq_frequency_table *powernow_table;
 718
 719         if (acpi_processor_register_performance(&data->acpi_data, data->cpu)) {
 720                 dprintk("register performance failed: bad ACPI data\n");
 721                 return -EIO;
 722         }
 723
 724         /* verify the data contained in the ACPI structures */
 725         if (data->acpi_data.state_count <= 1) {
 726                 dprintk("No ACPI P-States\n");
 727                 goto err_out;
 728         }
 729
 730         if ((data->acpi_data.control_register.space_id != ACPI_ADR_SPACE_FIXED_HARDWARE) ||
 731                 (data->acpi_data.status_register.space_id != ACPI_ADR_SPACE_FIXED_HARDWARE)) {
 732                 dprintk("Invalid control/status registers (%x - %x)\n",
 733                         data->acpi_data.control_register.space_id,
 734                         data->acpi_data.status_register.space_id);
 735                 goto err_out;
 736         }
 737
 738         /* fill in data->powernow_table */
 739         powernow_table = kmalloc((sizeof(struct cpufreq_frequency_table)
 740                 * (data->acpi_data.state_count + 1)), GFP_KERNEL);
 741         if (!powernow_table) {
 742                 dprintk("powernow_table memory alloc failure\n");
 743                 goto err_out;
 744         }
 745
 746         for (i = 0; i < data->acpi_data.state_count; i++) {
 747                 u32 fid;
 748                 u32 vid;
 749
 750                 if (data->exttype) {
 751                         fid = data->acpi_data.states[i].status & FID_MASK;
 752                         vid = (data->acpi_data.states[i].status >> VID_SHIFT) & VID_MASK;
 753                 } else {
 754                         fid = data->acpi_data.states[i].control & FID_MASK;
 755                         vid = (data->acpi_data.states[i].control >> VID_SHIFT) & VID_MASK;
 756                 }
 757
 758                 dprintk("   %d : fid 0x%x, vid 0x%x\n", i, fid, vid);
 759
 760                 powernow_table[i].index = fid; /* lower 8 bits */
 761                 powernow_table[i].index |= (vid << 8); /* upper 8 bits */
 762                 powernow_table[i].frequency = find_khz_freq_from_fid(fid);
 763
 764                 /* verify frequency is OK */
 765                 if ((powernow_table[i].frequency > (MAX_FREQ * 1000)) ||
 766                         (powernow_table[i].frequency < (MIN_FREQ * 1000))) {
 767                         dprintk("invalid freq %u kHz, ignoring\n", powernow_table[i].frequency);
 768                         powernow_table[i].frequency = CPUFREQ_ENTRY_INVALID;
 769                         continue;
 770                 }
 771
 772                 /* verify voltage is OK - BIOSs are using "off" to indicate invalid */
 773                 if (vid == VID_OFF) {
 774                         dprintk("invalid vid %u, ignoring\n", vid);
 775                         powernow_table[i].frequency = CPUFREQ_ENTRY_INVALID;
 776                         continue;
 777                 }
 778
 779                 /* verify only 1 entry from the lo frequency table */
 780                 if (fid < HI_FID_TABLE_BOTTOM) {
 781                         if (cntlofreq) {
 782                                 /* if both entries are the same, ignore this
 783                                  * one...
 784                                  */
 785                                 if ((powernow_table[i].frequency != powernow_table[cntlofreq].frequency) ||
 786                                     (powernow_table[i].index != powernow_table[cntlofreq].index)) {
 787                                         printk(KERN_ERR PFX "Too many lo freq table entries\n");
 788                                         goto err_out_mem;
 789                                 }
 790
 791                                 dprintk("double low frequency table entry, ignoring it.\n");
 792                                 powernow_table[i].frequency = CPUFREQ_ENTRY_INVALID;
 793                                 continue;
 794                         } else
 795                                 cntlofreq = i;
 796                 }
 797
 798                 if (powernow_table[i].frequency != (data->acpi_data.states[i].core_frequency * 1000)) {
 799                         printk(KERN_INFO PFX "invalid freq entries %u kHz vs. %u kHz\n",
 800                                 powernow_table[i].frequency,
 801                                 (unsigned int) (data->acpi_data.states[i].core_frequency * 1000));
 802                         powernow_table[i].frequency = CPUFREQ_ENTRY_INVALID;
 803                         continue;
 804                 }
 805         }
 806
 807         powernow_table[data->acpi_data.state_count].frequency = CPUFREQ_TABLE_END;
 808         powernow_table[data->acpi_data.state_count].index = 0;
 809         data->powernow_table = powernow_table;
 810
 811         /* fill in data */
 812         data->numps = data->acpi_data.state_count;
 813         print_basics(data);
 814         powernow_k8_acpi_pst_values(data, 0);
 815
 816         /* notify BIOS that we exist */
 817         acpi_processor_notify_smm(THIS_MODULE);
 818
 819         return 0;
 820
 821 err_out_mem:
 822         kfree(powernow_table);
 823
 824 err_out:
 825         acpi_processor_unregister_performance(&data->acpi_data, data->cpu);
 826
 827         /* data->acpi_data.state_count informs us at ->exit() whether ACPI was used */
 828         data->acpi_data.state_count = 0;
 829
 830         return -ENODEV;
 831 }
 832
 833 static void powernow_k8_cpu_exit_acpi(struct powernow_k8_data *data)
 834 {
 835         if (data->acpi_data.state_count)
 836                 acpi_processor_unregister_performance(&data->acpi_data, data->cpu);
 837 }
 838
 839 #else
 840 static int powernow_k8_cpu_init_acpi(struct powernow_k8_data *data) { return -ENODEV; }
 841 static void powernow_k8_cpu_exit_acpi(struct powernow_k8_data *data) { return; }
 842 static void powernow_k8_acpi_pst_values(struct powernow_k8_data *data, unsigned int index) { return; }
 843 #endif /* CONFIG_X86_POWERNOW_K8_ACPI */
 844
 845 /* Take a frequency, and issue the fid/vid transition command */
 846 static int transition_frequency(struct powernow_k8_data *data, unsigned int index)
 847 {
 848         u32 fid;
 849         u32 vid;
 850         int res, i;
 851         struct cpufreq_freqs freqs;
 852
 853         dprintk("cpu %d transition to index %u\n", smp_processor_id(), index);
 854
 855         /* fid are the lower 8 bits of the index we stored into
 856          * the cpufreq frequency table in find_psb_table, vid are
 857          * the upper 8 bits.
 858          */
 859
 860         fid = data->powernow_table[index].index & 0xFF;
 861         vid = (data->powernow_table[index].index & 0xFF00) >> 8;
 862
 863         dprintk("table matched fid 0x%x, giving vid 0x%x\n", fid, vid);
 864
 865         if (query_current_values_with_pending_wait(data))
 866                 return 1;
 867
 868         if ((data->currvid == vid) && (data->currfid == fid)) {
 869                 dprintk("target matches current values (fid 0x%x, vid 0x%x)\n",
 870                         fid, vid);
 871                 return 0;
 872         }
 873
 874         if ((fid < HI_FID_TABLE_BOTTOM) && (data->currfid < HI_FID_TABLE_BOTTOM)) {
 875                 printk(KERN_ERR PFX
 876                        "ignoring illegal change in lo freq table-%x to 0x%x\n",
 877                        data->currfid, fid);
 878                 return 1;
 879         }
 880
 881         dprintk("cpu %d, changing to fid 0x%x, vid 0x%x\n",
 882                 smp_processor_id(), fid, vid);
 883
 884         freqs.cpu = data->cpu;
 885         freqs.old = find_khz_freq_from_fid(data->currfid);
 886         freqs.new = find_khz_freq_from_fid(fid);
 887         for_each_cpu_mask(i, cpu_core_map[data->cpu]) {
 888                 freqs.cpu = i;
 889                 cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
 890         }
 891
 892         res = transition_fid_vid(data, fid, vid);
 893
 894         freqs.new = find_khz_freq_from_fid(data->currfid);
 895         for_each_cpu_mask(i, cpu_core_map[data->cpu]) {
 896                 freqs.cpu = i;
 897                 cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
 898         }
 899         return res;
 900 }
 901
 902 /* Driver entry point to switch to the target frequency */
 903 static int powernowk8_target(struct cpufreq_policy *pol, unsigned targfreq, unsigned relation)
 904 {
 905         cpumask_t oldmask = CPU_MASK_ALL;
 906         struct powernow_k8_data *data = powernow_data[pol->cpu];
 907         u32 checkfid = data->currfid;
 908         u32 checkvid = data->currvid;
 909         unsigned int newstate;
 910         int ret = -EIO;
 911         int i;
 912
 913         /* only run on specific CPU from here on */
 914         oldmask = current->cpus_allowed;
 915         set_cpus_allowed(current, cpumask_of_cpu(pol->cpu));
 916         schedule();
 917
 918         if (smp_processor_id() != pol->cpu) {
 919                 printk(KERN_ERR "limiting to cpu %u failed\n", pol->cpu);
 920                 goto err_out;
 921         }
 922
 923         if (pending_bit_stuck()) {
 924                 printk(KERN_ERR PFX "failing targ, change pending bit set\n");
 925                 goto err_out;
 926         }
 927
 928         dprintk("targ: cpu %d, %d kHz, min %d, max %d, relation %d\n",
 929                 pol->cpu, targfreq, pol->min, pol->max, relation);
 930
 931         if (query_current_values_with_pending_wait(data)) {
 932                 ret = -EIO;
 933                 goto err_out;
 934         }
 935
 936         dprintk("targ: curr fid 0x%x, vid 0x%x\n",
 937                 data->currfid, data->currvid);
 938
 939         if ((checkvid != data->currvid) || (checkfid != data->currfid)) {
 940                 printk(KERN_INFO PFX
 941                         "error - out of sync, fix 0x%x 0x%x, vid 0x%x 0x%x\n",
 942                         checkfid, data->currfid, checkvid, data->currvid);
 943         }
 944
 945         if (cpufreq_frequency_table_target(pol, data->powernow_table, targfreq, relation, &newstate))
 946                 goto err_out;
 947
 948         down(&fidvid_sem);
 949
 950         powernow_k8_acpi_pst_values(data, newstate);
 951
 952         if (transition_frequency(data, newstate)) {
 953                 printk(KERN_ERR PFX "transition frequency failed\n");
 954                 ret = 1;
 955                 up(&fidvid_sem);
 956                 goto err_out;
 957         }
 958
 959         /* Update all the fid/vids of our siblings */
 960         for_each_cpu_mask(i, cpu_core_map[pol->cpu]) {
 961                 powernow_data[i]->currvid = data->currvid;
 962                 powernow_data[i]->currfid = data->currfid;
 963         }
 964         up(&fidvid_sem);
 965
 966         pol->cur = find_khz_freq_from_fid(data->currfid);
 967         ret = 0;
 968
 969 err_out:
 970         set_cpus_allowed(current, oldmask);
 971         schedule();
 972
 973         return ret;
 974 }
 975
 976 /* Driver entry point to verify the policy and range of frequencies */
 977 static int powernowk8_verify(struct cpufreq_policy *pol)
 978 {
 979         struct powernow_k8_data *data = powernow_data[pol->cpu];
 980
 981         return cpufreq_frequency_table_verify(pol, data->powernow_table);
 982 }
 983
 984 /* per CPU init entry point to the driver */
 985 static int __init powernowk8_cpu_init(struct cpufreq_policy *pol)
 986 {
 987         struct powernow_k8_data *data;
 988         cpumask_t oldmask = CPU_MASK_ALL;
 989         int rc, i;
 990
 991         if (!check_supported_cpu(pol->cpu))
 992                 return -ENODEV;
 993
 994         data = kmalloc(sizeof(struct powernow_k8_data), GFP_KERNEL);
 995         if (!data) {
 996                 printk(KERN_ERR PFX "unable to alloc powernow_k8_data");
 997                 return -ENOMEM;
 998         }
 999         memset(data,0,sizeof(struct powernow_k8_data));
1000
1001         data->cpu = pol->cpu;
1002
1003         if (powernow_k8_cpu_init_acpi(data)) {
1004                 /*
1005                  * Use the PSB BIOS structure. This is only availabe on
1006                  * an UP version, and is deprecated by AMD.
1007                  */
1008
1009                 if ((num_online_cpus() != 1) || (num_possible_cpus() != 1)) {
1010                         printk(KERN_ERR PFX "MP systems not supported by PSB BIOS structure\n");
1011                         kfree(data);
1012                         return -ENODEV;
1013                 }
1014                 if (pol->cpu != 0) {
1015                         printk(KERN_ERR PFX "init not cpu 0\n");
1016                         kfree(data);
1017                         return -ENODEV;
1018                 }
1019                 rc = find_psb_table(data);
1020                 if (rc) {
1021                         kfree(data);
1022                         return -ENODEV;
1023                 }
1024         }
1025
1026         /* only run on specific CPU from here on */
1027         oldmask = current->cpus_allowed;
1028         set_cpus_allowed(current, cpumask_of_cpu(pol->cpu));
1029         schedule();
1030
1031         if (smp_processor_id() != pol->cpu) {
1032                 printk(KERN_ERR "limiting to cpu %u failed\n", pol->cpu);
1033                 goto err_out;
1034         }
1035
1036         if (pending_bit_stuck()) {
1037                 printk(KERN_ERR PFX "failing init, change pending bit set\n");
1038                 goto err_out;
1039         }
1040
1041         if (query_current_values_with_pending_wait(data))
1042                 goto err_out;
1043
1044         fidvid_msr_init();
1045
1046         /* run on any CPU again */
1047         set_cpus_allowed(current, oldmask);
1048         schedule();
1049
1050         pol->governor = CPUFREQ_DEFAULT_GOVERNOR;
1051         pol->cpus = cpu_core_map[pol->cpu];
1052
1053         /* Take a crude guess here.
1054          * That guess was in microseconds, so multiply with 1000 */
1055         pol->cpuinfo.transition_latency = (((data->rvo + 8) * data->vstable * VST_UNITS_20US)
1056             + (3 * (1 << data->irt) * 10)) * 1000;
1057
1058         pol->cur = find_khz_freq_from_fid(data->currfid);
1059         dprintk("policy current frequency %d kHz\n", pol->cur);
1060
1061         /* min/max the cpu is capable of */
1062         if (cpufreq_frequency_table_cpuinfo(pol, data->powernow_table)) {
1063                 printk(KERN_ERR PFX "invalid powernow_table\n");
1064                 powernow_k8_cpu_exit_acpi(data);
1065                 kfree(data->powernow_table);
1066                 kfree(data);
1067                 return -EINVAL;
1068         }
1069
1070         cpufreq_frequency_table_get_attr(data->powernow_table, pol->cpu);
1071
1072         printk("cpu_init done, current fid 0x%x, vid 0x%x\n",
1073                data->currfid, data->currvid);
1074
1075         for_each_cpu_mask(i, cpu_core_map[pol->cpu]) {
1076                 powernow_data[i] = data;
1077         }
1078
1079         return 0;
1080
1081 err_out:
1082         set_cpus_allowed(current, oldmask);
1083         schedule();
1084         powernow_k8_cpu_exit_acpi(data);
1085
1086         kfree(data);
1087         return -ENODEV;
1088 }
1089
1090 static int __devexit powernowk8_cpu_exit (struct cpufreq_policy *pol)
1091 {
1092         struct powernow_k8_data *data = powernow_data[pol->cpu];
1093
1094         if (!data)
1095                 return -EINVAL;
1096
1097         powernow_k8_cpu_exit_acpi(data);
1098
1099         cpufreq_frequency_table_put_attr(pol->cpu);
1100
1101         kfree(data->powernow_table);
1102         kfree(data);
1103
1104         return 0;
1105 }
1106
1107 static unsigned int powernowk8_get (unsigned int cpu)
1108 {
1109         struct powernow_k8_data *data = powernow_data[cpu];
1110         cpumask_t oldmask = current->cpus_allowed;
1111         unsigned int khz = 0;
1112
1113         set_cpus_allowed(current, cpumask_of_cpu(cpu));
1114         if (smp_processor_id() != cpu) {
1115                 printk(KERN_ERR PFX "limiting to CPU %d failed in powernowk8_get\n", cpu);
1116                 set_cpus_allowed(current, oldmask);
1117                 return 0;
1118         }
1119         preempt_disable();
1120
1121         if (query_current_values_with_pending_wait(data))
1122                 goto out;
1123
1124         khz = find_khz_freq_from_fid(data->currfid);
1125
1126  out:
1127         preempt_enable_no_resched();
1128         set_cpus_allowed(current, oldmask);
1129
1130         return khz;
1131 }
1132
1133 static struct freq_attr* powernow_k8_attr[] = {
1134         &cpufreq_freq_attr_scaling_available_freqs,
1135         NULL,
1136 };
1137
1138 static struct cpufreq_driver cpufreq_amd64_driver = {
1139         .verify = powernowk8_verify,
1140         .target = powernowk8_target,
1141         .init = powernowk8_cpu_init,
1142         .exit = __devexit_p(powernowk8_cpu_exit),
1143         .get = powernowk8_get,
1144         .name = "powernow-k8",
1145         .owner = THIS_MODULE,
1146         .attr = powernow_k8_attr,
1147 };
1148
1149 /* driver entry point for init */
1150 static int __init powernowk8_init(void)
1151 {
1152         unsigned int i, supported_cpus = 0;
1153
1154         for (i=0; i<NR_CPUS; i++) {
1155                 if (!cpu_online(i))
1156                         continue;
1157                 if (check_supported_cpu(i))
1158                         supported_cpus++;
1159         }
1160
1161         if (supported_cpus == num_online_cpus()) {
1162                 printk(KERN_INFO PFX "Found %d AMD Athlon 64 / Opteron processors (" VERSION ")\n",
1163                         supported_cpus);
1164                 return cpufreq_register_driver(&cpufreq_amd64_driver);
1165         }
1166
1167         return -ENODEV;
1168 }
1169
1170 /* driver entry point for term */
1171 static void __exit powernowk8_exit(void)
1172 {
1173         dprintk("exit\n");
1174
1175         cpufreq_unregister_driver(&cpufreq_amd64_driver);
1176 }
1177
1178 MODULE_AUTHOR("Paul Devriendt <paul.devriendt@amd.com> and Mark Langsdorf <mark.langsdorf@amd.com.");
1179 MODULE_DESCRIPTION("AMD Athlon 64 and Opteron processor frequency driver.");
1180 MODULE_LICENSE("GPL");
1181
1182 late_initcall(powernowk8_init);
1183 module_exit(powernowk8_exit);
1184