drivers/macintosh/therm_pm72.c

   1 /*
   2  * Device driver for the thermostats & fan controller of  the
   3  * Apple G5 "PowerMac7,2" desktop machines.
   4  *
   5  * (c) Copyright IBM Corp. 2003-2004
   6  *
   7  * Maintained by: Benjamin Herrenschmidt
   8  *                <benh@kernel.crashing.org>
   9  *
  10  *
  11  * The algorithm used is the PID control algorithm, used the same
  12  * way the published Darwin code does, using the same values that
  13  * are present in the Darwin 7.0 snapshot property lists.
  14  *
  15  * As far as the CPUs control loops are concerned, I use the
  16  * calibration & PID constants provided by the EEPROM,
  17  * I do _not_ embed any value from the property lists, as the ones
  18  * provided by Darwin 7.0 seem to always have an older version that
  19  * what I've seen on the actual computers.
  20  * It would be interesting to verify that though. Darwin has a
  21  * version code of 1.0.0d11 for all control loops it seems, while
  22  * so far, the machines EEPROMs contain a dataset versioned 1.0.0f
  23  *
  24  * Darwin doesn't provide source to all parts, some missing
  25  * bits like the AppleFCU driver or the actual scale of some
  26  * of the values returned by sensors had to be "guessed" some
  27  * way... or based on what Open Firmware does.
  28  *
  29  * I didn't yet figure out how to get the slots power consumption
  30  * out of the FCU, so that part has not been implemented yet and
  31  * the slots fan is set to a fixed 50% PWM, hoping this value is
  32  * safe enough ...
  33  *
  34  * Note: I have observed strange oscillations of the CPU control
  35  * loop on a dual G5 here. When idle, the CPU exhaust fan tend to
  36  * oscillates slowly (over several minutes) between the minimum
  37  * of 300RPMs and approx. 1000 RPMs. I don't know what is causing
  38  * this, it could be some incorrect constant or an error in the
  39  * way I ported the algorithm, or it could be just normal. I
  40  * don't have full understanding on the way Apple tweaked the PID
  41  * algorithm for the CPU control, it is definitely not a standard
  42  * implementation...
  43  *
  44  * TODO:  - Check MPU structure version/signature
  45  *        - Add things like /sbin/overtemp for non-critical
  46  *          overtemp conditions so userland can take some policy
  47  *          decisions, like slewing down CPUs
  48  *        - Deal with fan and i2c failures in a better way
  49  *        - Maybe do a generic PID based on params used for
  50  *          U3 and Drives ? Definitely need to factor code a bit
  51  *          bettter... also make sensor detection more robust using
  52  *          the device-tree to probe for them
  53  *        - Figure out how to get the slots consumption and set the
  54  *          slots fan accordingly
  55  *
  56  * History:
  57  *
  58  *  Nov. 13, 2003 : 0.5
  59  *      - First release
  60  *
  61  *  Nov. 14, 2003 : 0.6
  62  *      - Read fan speed from FCU, low level fan routines now deal
  63  *        with errors & check fan status, though higher level don't
  64  *        do much.
  65  *      - Move a bunch of definitions to .h file
  66  *
  67  *  Nov. 18, 2003 : 0.7
  68  *      - Fix build on ppc64 kernel
  69  *      - Move back statics definitions to .c file
  70  *      - Avoid calling schedule_timeout with a negative number
  71  *
  72  *  Dec. 18, 2003 : 0.8
  73  *      - Fix typo when reading back fan speed on 2 CPU machines
  74  *
  75  *  Mar. 11, 2004 : 0.9
  76  *      - Rework code accessing the ADC chips, make it more robust and
  77  *        closer to the chip spec. Also make sure it is configured properly,
  78  *        I've seen yet unexplained cases where on startup, I would have stale
  79  *        values in the configuration register
  80  *      - Switch back to use of target fan speed for PID, thus lowering
  81  *        pressure on i2c
  82  *
  83  *  Oct. 20, 2004 : 1.1
  84  *      - Add device-tree lookup for fan IDs, should detect liquid cooling
  85  *        pumps when present
  86  *      - Enable driver for PowerMac7,3 machines
  87  *      - Split the U3/Backside cooling on U3 & U3H versions as Darwin does
  88  *      - Add new CPU cooling algorithm for machines with liquid cooling
  89  *      - Workaround for some PowerMac7,3 with empty "fan" node in the devtree
  90  *      - Fix a signed/unsigned compare issue in some PID loops
  91  *
  92  *  Mar. 10, 2005 : 1.2
  93  *      - Add basic support for Xserve G5
  94  *      - Retreive pumps min/max from EEPROM image in device-tree (broken)
  95  *      - Use min/max macros here or there
  96  *      - Latest darwin updated U3H min fan speed to 20% PWM
  97  *
  98  */
  99
 100 #include <linux/config.h>
 101 #include <linux/types.h>
 102 #include <linux/module.h>
 103 #include <linux/errno.h>
 104 #include <linux/kernel.h>
 105 #include <linux/delay.h>
 106 #include <linux/sched.h>
 107 #include <linux/i2c.h>
 108 #include <linux/slab.h>
 109 #include <linux/init.h>
 110 #include <linux/spinlock.h>
 111 #include <linux/smp_lock.h>
 112 #include <linux/wait.h>
 113 #include <linux/reboot.h>
 114 #include <linux/kmod.h>
 115 #include <linux/i2c.h>
 116 #include <linux/i2c-dev.h>
 117 #include <asm/prom.h>
 118 #include <asm/machdep.h>
 119 #include <asm/io.h>
 120 #include <asm/system.h>
 121 #include <asm/sections.h>
 122 #include <asm/of_device.h>
 123
 124 #include "therm_pm72.h"
 125
 126 #define VERSION "1.2b2"
 127
 128 #undef DEBUG
 129
 130 #ifdef DEBUG
 131 #define DBG(args...)    printk(args)
 132 #else
 133 #define DBG(args...)    do { } while(0)
 134 #endif
 135
 136
 137 /*
 138  * Driver statics
 139  */
 140
 141 static struct of_device *               of_dev;
 142 static struct i2c_adapter *             u3_0;
 143 static struct i2c_adapter *             u3_1;
 144 static struct i2c_adapter *             k2;
 145 static struct i2c_client *              fcu;
 146 static struct cpu_pid_state             cpu_state[2];
 147 static struct basckside_pid_params      backside_params;
 148 static struct backside_pid_state        backside_state;
 149 static struct drives_pid_state          drives_state;
 150 static struct dimm_pid_state            dimms_state;
 151 static int                              state;
 152 static int                              cpu_count;
 153 static int                              cpu_pid_type;
 154 static pid_t                            ctrl_task;
 155 static struct completion                ctrl_complete;
 156 static int                              critical_state;
 157 static int                              rackmac;
 158 static s32                              dimm_output_clamp;
 159
 160 static DECLARE_MUTEX(driver_lock);
 161
 162 /*
 163  * We have 3 types of CPU PID control. One is "split" old style control
 164  * for intake & exhaust fans, the other is "combined" control for both
 165  * CPUs that also deals with the pumps when present. To be "compatible"
 166  * with OS X at this point, we only use "COMBINED" on the machines that
 167  * are identified as having the pumps (though that identification is at
 168  * least dodgy). Ultimately, we could probably switch completely to this
 169  * algorithm provided we hack it to deal with the UP case
 170  */
 171 #define CPU_PID_TYPE_SPLIT      0
 172 #define CPU_PID_TYPE_COMBINED   1
 173 #define CPU_PID_TYPE_RACKMAC    2
 174
 175 /*
 176  * This table describes all fans in the FCU. The "id" and "type" values
 177  * are defaults valid for all earlier machines. Newer machines will
 178  * eventually override the table content based on the device-tree
 179  */
 180 struct fcu_fan_table
 181 {
 182         char*   loc;    /* location code */
 183         int     type;   /* 0 = rpm, 1 = pwm, 2 = pump */
 184         int     id;     /* id or -1 */
 185 };
 186
 187 #define FCU_FAN_RPM             0
 188 #define FCU_FAN_PWM             1
 189
 190 #define FCU_FAN_ABSENT_ID       -1
 191
 192 #define FCU_FAN_COUNT           ARRAY_SIZE(fcu_fans)
 193
 194 struct fcu_fan_table    fcu_fans[] = {
 195         [BACKSIDE_FAN_PWM_INDEX] = {
 196                 .loc    = "BACKSIDE,SYS CTRLR FAN",
 197                 .type   = FCU_FAN_PWM,
 198                 .id     = BACKSIDE_FAN_PWM_DEFAULT_ID,
 199         },
 200         [DRIVES_FAN_RPM_INDEX] = {
 201                 .loc    = "DRIVE BAY",
 202                 .type   = FCU_FAN_RPM,
 203                 .id     = DRIVES_FAN_RPM_DEFAULT_ID,
 204         },
 205         [SLOTS_FAN_PWM_INDEX] = {
 206                 .loc    = "SLOT,PCI FAN",
 207                 .type   = FCU_FAN_PWM,
 208                 .id     = SLOTS_FAN_PWM_DEFAULT_ID,
 209         },
 210         [CPUA_INTAKE_FAN_RPM_INDEX] = {
 211                 .loc    = "CPU A INTAKE",
 212                 .type   = FCU_FAN_RPM,
 213                 .id     = CPUA_INTAKE_FAN_RPM_DEFAULT_ID,
 214         },
 215         [CPUA_EXHAUST_FAN_RPM_INDEX] = {
 216                 .loc    = "CPU A EXHAUST",
 217                 .type   = FCU_FAN_RPM,
 218                 .id     = CPUA_EXHAUST_FAN_RPM_DEFAULT_ID,
 219         },
 220         [CPUB_INTAKE_FAN_RPM_INDEX] = {
 221                 .loc    = "CPU B INTAKE",
 222                 .type   = FCU_FAN_RPM,
 223                 .id     = CPUB_INTAKE_FAN_RPM_DEFAULT_ID,
 224         },
 225         [CPUB_EXHAUST_FAN_RPM_INDEX] = {
 226                 .loc    = "CPU B EXHAUST",
 227                 .type   = FCU_FAN_RPM,
 228                 .id     = CPUB_EXHAUST_FAN_RPM_DEFAULT_ID,
 229         },
 230         /* pumps aren't present by default, have to be looked up in the
 231          * device-tree
 232          */
 233         [CPUA_PUMP_RPM_INDEX] = {
 234                 .loc    = "CPU A PUMP",
 235                 .type   = FCU_FAN_RPM,
 236                 .id     = FCU_FAN_ABSENT_ID,
 237         },
 238         [CPUB_PUMP_RPM_INDEX] = {
 239                 .loc    = "CPU B PUMP",
 240                 .type   = FCU_FAN_RPM,
 241                 .id     = FCU_FAN_ABSENT_ID,
 242         },
 243         /* Xserve fans */
 244         [CPU_A1_FAN_RPM_INDEX] = {
 245                 .loc    = "CPU A 1",
 246                 .type   = FCU_FAN_RPM,
 247                 .id     = FCU_FAN_ABSENT_ID,
 248         },
 249         [CPU_A2_FAN_RPM_INDEX] = {
 250                 .loc    = "CPU A 2",
 251                 .type   = FCU_FAN_RPM,
 252                 .id     = FCU_FAN_ABSENT_ID,
 253         },
 254         [CPU_A3_FAN_RPM_INDEX] = {
 255                 .loc    = "CPU A 3",
 256                 .type   = FCU_FAN_RPM,
 257                 .id     = FCU_FAN_ABSENT_ID,
 258         },
 259         [CPU_B1_FAN_RPM_INDEX] = {
 260                 .loc    = "CPU B 1",
 261                 .type   = FCU_FAN_RPM,
 262                 .id     = FCU_FAN_ABSENT_ID,
 263         },
 264         [CPU_B2_FAN_RPM_INDEX] = {
 265                 .loc    = "CPU B 2",
 266                 .type   = FCU_FAN_RPM,
 267                 .id     = FCU_FAN_ABSENT_ID,
 268         },
 269         [CPU_B3_FAN_RPM_INDEX] = {
 270                 .loc    = "CPU B 3",
 271                 .type   = FCU_FAN_RPM,
 272                 .id     = FCU_FAN_ABSENT_ID,
 273         },
 274 };
 275
 276 /*
 277  * i2c_driver structure to attach to the host i2c controller
 278  */
 279
 280 static int therm_pm72_attach(struct i2c_adapter *adapter);
 281 static int therm_pm72_detach(struct i2c_adapter *adapter);
 282
 283 static struct i2c_driver therm_pm72_driver =
 284 {
 285         .owner          = THIS_MODULE,
 286         .name           = "therm_pm72",
 287         .flags          = I2C_DF_NOTIFY,
 288         .attach_adapter = therm_pm72_attach,
 289         .detach_adapter = therm_pm72_detach,
 290 };
 291
 292 /*
 293  * Utility function to create an i2c_client structure and
 294  * attach it to one of u3 adapters
 295  */
 296 static struct i2c_client *attach_i2c_chip(int id, const char *name)
 297 {
 298         struct i2c_client *clt;
 299         struct i2c_adapter *adap;
 300
 301         if (id & 0x200)
 302                 adap = k2;
 303         else if (id & 0x100)
 304                 adap = u3_1;
 305         else
 306                 adap = u3_0;
 307         if (adap == NULL)
 308                 return NULL;
 309
 310         clt = kmalloc(sizeof(struct i2c_client), GFP_KERNEL);
 311         if (clt == NULL)
 312                 return NULL;
 313         memset(clt, 0, sizeof(struct i2c_client));
 314
 315         clt->addr = (id >> 1) & 0x7f;
 316         clt->adapter = adap;
 317         clt->driver = &therm_pm72_driver;
 318         strncpy(clt->name, name, I2C_NAME_SIZE-1);
 319
 320         if (i2c_attach_client(clt)) {
 321                 printk(KERN_ERR "therm_pm72: Failed to attach to i2c ID 0x%x\n", id);
 322                 kfree(clt);
 323                 return NULL;
 324         }
 325         return clt;
 326 }
 327
 328 /*
 329  * Utility function to get rid of the i2c_client structure
 330  * (will also detach from the adapter hopepfully)
 331  */
 332 static void detach_i2c_chip(struct i2c_client *clt)
 333 {
 334         i2c_detach_client(clt);
 335         kfree(clt);
 336 }
 337
 338 /*
 339  * Here are the i2c chip access wrappers
 340  */
 341
 342 static void initialize_adc(struct cpu_pid_state *state)
 343 {
 344         int rc;
 345         u8 buf[2];
 346
 347         /* Read ADC the configuration register and cache it. We
 348          * also make sure Config2 contains proper values, I've seen
 349          * cases where we got stale grabage in there, thus preventing
 350          * proper reading of conv. values
 351          */
 352
 353         /* Clear Config2 */
 354         buf[0] = 5;
 355         buf[1] = 0;
 356         i2c_master_send(state->monitor, buf, 2);
 357
 358         /* Read & cache Config1 */
 359         buf[0] = 1;
 360         rc = i2c_master_send(state->monitor, buf, 1);
 361         if (rc > 0) {
 362                 rc = i2c_master_recv(state->monitor, buf, 1);
 363                 if (rc > 0) {
 364                         state->adc_config = buf[0];
 365                         DBG("ADC config reg: %02x\n", state->adc_config);
 366                         /* Disable shutdown mode */
 367                         state->adc_config &= 0xfe;
 368                         buf[0] = 1;
 369                         buf[1] = state->adc_config;
 370                         rc = i2c_master_send(state->monitor, buf, 2);
 371                 }
 372         }
 373         if (rc <= 0)
 374                 printk(KERN_ERR "therm_pm72: Error reading ADC config"
 375                        " register !\n");
 376 }
 377
 378 static int read_smon_adc(struct cpu_pid_state *state, int chan)
 379 {
 380         int rc, data, tries = 0;
 381         u8 buf[2];
 382
 383         for (;;) {
 384                 /* Set channel */
 385                 buf[0] = 1;
 386                 buf[1] = (state->adc_config & 0x1f) | (chan << 5);
 387                 rc = i2c_master_send(state->monitor, buf, 2);
 388                 if (rc <= 0)
 389                         goto error;
 390                 /* Wait for convertion */
 391                 msleep(1);
 392                 /* Switch to data register */
 393                 buf[0] = 4;
 394                 rc = i2c_master_send(state->monitor, buf, 1);
 395                 if (rc <= 0)
 396                         goto error;
 397                 /* Read result */
 398                 rc = i2c_master_recv(state->monitor, buf, 2);
 399                 if (rc < 0)
 400                         goto error;
 401                 data = ((u16)buf[0]) << 8 | (u16)buf[1];
 402                 return data >> 6;
 403         error:
 404                 DBG("Error reading ADC, retrying...\n");
 405                 if (++tries > 10) {
 406                         printk(KERN_ERR "therm_pm72: Error reading ADC !\n");
 407                         return -1;
 408                 }
 409                 msleep(10);
 410         }
 411 }
 412
 413 static int read_lm87_reg(struct i2c_client * chip, int reg)
 414 {
 415         int rc, tries = 0;
 416         u8 buf;
 417
 418         for (;;) {
 419                 /* Set address */
 420                 buf = (u8)reg;
 421                 rc = i2c_master_send(chip, &buf, 1);
 422                 if (rc <= 0)
 423                         goto error;
 424                 rc = i2c_master_recv(chip, &buf, 1);
 425                 if (rc <= 0)
 426                         goto error;
 427                 return (int)buf;
 428         error:
 429                 DBG("Error reading LM87, retrying...\n");
 430                 if (++tries > 10) {
 431                         printk(KERN_ERR "therm_pm72: Error reading LM87 !\n");
 432                         return -1;
 433                 }
 434                 msleep(10);
 435         }
 436 }
 437
 438 static int fan_read_reg(int reg, unsigned char *buf, int nb)
 439 {
 440         int tries, nr, nw;
 441
 442         buf[0] = reg;
 443         tries = 0;
 444         for (;;) {
 445                 nw = i2c_master_send(fcu, buf, 1);
 446                 if (nw > 0 || (nw < 0 && nw != -EIO) || tries >= 100)
 447                         break;
 448                 msleep(10);
 449                 ++tries;
 450         }
 451         if (nw <= 0) {
 452                 printk(KERN_ERR "Failure writing address to FCU: %d", nw);
 453                 return -EIO;
 454         }
 455         tries = 0;
 456         for (;;) {
 457                 nr = i2c_master_recv(fcu, buf, nb);
 458                 if (nr > 0 || (nr < 0 && nr != ENODEV) || tries >= 100)
 459                         break;
 460                 msleep(10);
 461                 ++tries;
 462         }
 463         if (nr <= 0)
 464                 printk(KERN_ERR "Failure reading data from FCU: %d", nw);
 465         return nr;
 466 }
 467
 468 static int fan_write_reg(int reg, const unsigned char *ptr, int nb)
 469 {
 470         int tries, nw;
 471         unsigned char buf[16];
 472
 473         buf[0] = reg;
 474         memcpy(buf+1, ptr, nb);
 475         ++nb;
 476         tries = 0;
 477         for (;;) {
 478                 nw = i2c_master_send(fcu, buf, nb);
 479                 if (nw > 0 || (nw < 0 && nw != EIO) || tries >= 100)
 480                         break;
 481                 msleep(10);
 482                 ++tries;
 483         }
 484         if (nw < 0)
 485                 printk(KERN_ERR "Failure writing to FCU: %d", nw);
 486         return nw;
 487 }
 488
 489 static int start_fcu(void)
 490 {
 491         unsigned char buf = 0xff;
 492         int rc;
 493
 494         rc = fan_write_reg(0xe, &buf, 1);
 495         if (rc < 0)
 496                 return -EIO;
 497         rc = fan_write_reg(0x2e, &buf, 1);
 498         if (rc < 0)
 499                 return -EIO;
 500         return 0;
 501 }
 502
 503 static int set_rpm_fan(int fan_index, int rpm)
 504 {
 505         unsigned char buf[2];
 506         int rc, id;
 507
 508         if (fcu_fans[fan_index].type != FCU_FAN_RPM)
 509                 return -EINVAL;
 510         id = fcu_fans[fan_index].id;
 511         if (id == FCU_FAN_ABSENT_ID)
 512                 return -EINVAL;
 513
 514         if (rpm < 300)
 515                 rpm = 300;
 516         else if (rpm > 8191)
 517                 rpm = 8191;
 518         buf[0] = rpm >> 5;
 519         buf[1] = rpm << 3;
 520         rc = fan_write_reg(0x10 + (id * 2), buf, 2);
 521         if (rc < 0)
 522                 return -EIO;
 523         return 0;
 524 }
 525
 526 static int get_rpm_fan(int fan_index, int programmed)
 527 {
 528         unsigned char failure;
 529         unsigned char active;
 530         unsigned char buf[2];
 531         int rc, id, reg_base;
 532
 533         if (fcu_fans[fan_index].type != FCU_FAN_RPM)
 534                 return -EINVAL;
 535         id = fcu_fans[fan_index].id;
 536         if (id == FCU_FAN_ABSENT_ID)
 537                 return -EINVAL;
 538
 539         rc = fan_read_reg(0xb, &failure, 1);
 540         if (rc != 1)
 541                 return -EIO;
 542         if ((failure & (1 << id)) != 0)
 543                 return -EFAULT;
 544         rc = fan_read_reg(0xd, &active, 1);
 545         if (rc != 1)
 546                 return -EIO;
 547         if ((active & (1 << id)) == 0)
 548                 return -ENXIO;
 549
 550         /* Programmed value or real current speed */
 551         reg_base = programmed ? 0x10 : 0x11;
 552         rc = fan_read_reg(reg_base + (id * 2), buf, 2);
 553         if (rc != 2)
 554                 return -EIO;
 555
 556         return (buf[0] << 5) | buf[1] >> 3;
 557 }
 558
 559 static int set_pwm_fan(int fan_index, int pwm)
 560 {
 561         unsigned char buf[2];
 562         int rc, id;
 563
 564         if (fcu_fans[fan_index].type != FCU_FAN_PWM)
 565                 return -EINVAL;
 566         id = fcu_fans[fan_index].id;
 567         if (id == FCU_FAN_ABSENT_ID)
 568                 return -EINVAL;
 569
 570         if (pwm < 10)
 571                 pwm = 10;
 572         else if (pwm > 100)
 573                 pwm = 100;
 574         pwm = (pwm * 2559) / 1000;
 575         buf[0] = pwm;
 576         rc = fan_write_reg(0x30 + (id * 2), buf, 1);
 577         if (rc < 0)
 578                 return rc;
 579         return 0;
 580 }
 581
 582 static int get_pwm_fan(int fan_index)
 583 {
 584         unsigned char failure;
 585         unsigned char active;
 586         unsigned char buf[2];
 587         int rc, id;
 588
 589         if (fcu_fans[fan_index].type != FCU_FAN_PWM)
 590                 return -EINVAL;
 591         id = fcu_fans[fan_index].id;
 592         if (id == FCU_FAN_ABSENT_ID)
 593                 return -EINVAL;
 594
 595         rc = fan_read_reg(0x2b, &failure, 1);
 596         if (rc != 1)
 597                 return -EIO;
 598         if ((failure & (1 << id)) != 0)
 599                 return -EFAULT;
 600         rc = fan_read_reg(0x2d, &active, 1);
 601         if (rc != 1)
 602                 return -EIO;
 603         if ((active & (1 << id)) == 0)
 604                 return -ENXIO;
 605
 606         /* Programmed value or real current speed */
 607         rc = fan_read_reg(0x30 + (id * 2), buf, 1);
 608         if (rc != 1)
 609                 return -EIO;
 610
 611         return (buf[0] * 1000) / 2559;
 612 }
 613
 614 /*
 615  * Utility routine to read the CPU calibration EEPROM data
 616  * from the device-tree
 617  */
 618 static int read_eeprom(int cpu, struct mpu_data *out)
 619 {
 620         struct device_node *np;
 621         char nodename[64];
 622         u8 *data;
 623         int len;
 624
 625         /* prom.c routine for finding a node by path is a bit brain dead
 626          * and requires exact @xxx unit numbers. This is a bit ugly but
 627          * will work for these machines
 628          */
 629         sprintf(nodename, "/u3@0,f8000000/i2c@f8001000/cpuid@a%d", cpu ? 2 : 0);
 630         np = of_find_node_by_path(nodename);
 631         if (np == NULL) {
 632                 printk(KERN_ERR "therm_pm72: Failed to retreive cpuid node from device-tree\n");
 633                 return -ENODEV;
 634         }
 635         data = (u8 *)get_property(np, "cpuid", &len);
 636         if (data == NULL) {
 637                 printk(KERN_ERR "therm_pm72: Failed to retreive cpuid property from device-tree\n");
 638                 of_node_put(np);
 639                 return -ENODEV;
 640         }
 641         memcpy(out, data, sizeof(struct mpu_data));
 642         of_node_put(np);
 643
 644         return 0;
 645 }
 646
 647 static void fetch_cpu_pumps_minmax(void)
 648 {
 649         struct cpu_pid_state *state0 = &cpu_state[0];
 650         struct cpu_pid_state *state1 = &cpu_state[1];
 651         u16 pump_min = 0, pump_max = 0xffff;
 652         u16 tmp[4];
 653
 654         /* Try to fetch pumps min/max infos from eeprom */
 655
 656         memcpy(&tmp, &state0->mpu.processor_part_num, 8);
 657         if (tmp[0] != 0xffff && tmp[1] != 0xffff) {
 658                 pump_min = max(pump_min, tmp[0]);
 659                 pump_max = min(pump_max, tmp[1]);
 660         }
 661         if (tmp[2] != 0xffff && tmp[3] != 0xffff) {
 662                 pump_min = max(pump_min, tmp[2]);
 663                 pump_max = min(pump_max, tmp[3]);
 664         }
 665
 666         /* Double check the values, this _IS_ needed as the EEPROM on
 667          * some dual 2.5Ghz G5s seem, at least, to have both min & max
 668          * same to the same value ... (grrrr)
 669          */
 670         if (pump_min == pump_max || pump_min == 0 || pump_max == 0xffff) {
 671                 pump_min = CPU_PUMP_OUTPUT_MIN;
 672                 pump_max = CPU_PUMP_OUTPUT_MAX;
 673         }
 674
 675         state0->pump_min = state1->pump_min = pump_min;
 676         state0->pump_max = state1->pump_max = pump_max;
 677 }
 678
 679 /*
 680  * Now, unfortunately, sysfs doesn't give us a nice void * we could
 681  * pass around to the attribute functions, so we don't really have
 682  * choice but implement a bunch of them...
 683  *
 684  * That sucks a bit, we take the lock because FIX32TOPRINT evaluates
 685  * the input twice... I accept patches :)
 686  */
 687 #define BUILD_SHOW_FUNC_FIX(name, data)                         \
 688 static ssize_t show_##name(struct device *dev, char *buf)       \
 689 {                                                               \
 690         ssize_t r;                                              \
 691         down(&driver_lock);                                     \
 692         r = sprintf(buf, "%d.%03d", FIX32TOPRINT(data));        \
 693         up(&driver_lock);                                       \
 694         return r;                                               \
 695 }
 696 #define BUILD_SHOW_FUNC_INT(name, data)                         \
 697 static ssize_t show_##name(struct device *dev, char *buf)       \
 698 {                                                               \
 699         return sprintf(buf, "%d", data);                        \
 700 }
 701
 702 BUILD_SHOW_FUNC_FIX(cpu0_temperature, cpu_state[0].last_temp)
 703 BUILD_SHOW_FUNC_FIX(cpu0_voltage, cpu_state[0].voltage)
 704 BUILD_SHOW_FUNC_FIX(cpu0_current, cpu_state[0].current_a)
 705 BUILD_SHOW_FUNC_INT(cpu0_exhaust_fan_rpm, cpu_state[0].rpm)
 706 BUILD_SHOW_FUNC_INT(cpu0_intake_fan_rpm, cpu_state[0].intake_rpm)
 707
 708 BUILD_SHOW_FUNC_FIX(cpu1_temperature, cpu_state[1].last_temp)
 709 BUILD_SHOW_FUNC_FIX(cpu1_voltage, cpu_state[1].voltage)
 710 BUILD_SHOW_FUNC_FIX(cpu1_current, cpu_state[1].current_a)
 711 BUILD_SHOW_FUNC_INT(cpu1_exhaust_fan_rpm, cpu_state[1].rpm)
 712 BUILD_SHOW_FUNC_INT(cpu1_intake_fan_rpm, cpu_state[1].intake_rpm)
 713
 714 BUILD_SHOW_FUNC_FIX(backside_temperature, backside_state.last_temp)
 715 BUILD_SHOW_FUNC_INT(backside_fan_pwm, backside_state.pwm)
 716
 717 BUILD_SHOW_FUNC_FIX(drives_temperature, drives_state.last_temp)
 718 BUILD_SHOW_FUNC_INT(drives_fan_rpm, drives_state.rpm)
 719
 720 BUILD_SHOW_FUNC_FIX(dimms_temperature, dimms_state.last_temp)
 721
 722 static DEVICE_ATTR(cpu0_temperature,S_IRUGO,show_cpu0_temperature,NULL);
 723 static DEVICE_ATTR(cpu0_voltage,S_IRUGO,show_cpu0_voltage,NULL);
 724 static DEVICE_ATTR(cpu0_current,S_IRUGO,show_cpu0_current,NULL);
 725 static DEVICE_ATTR(cpu0_exhaust_fan_rpm,S_IRUGO,show_cpu0_exhaust_fan_rpm,NULL);
 726 static DEVICE_ATTR(cpu0_intake_fan_rpm,S_IRUGO,show_cpu0_intake_fan_rpm,NULL);
 727
 728 static DEVICE_ATTR(cpu1_temperature,S_IRUGO,show_cpu1_temperature,NULL);
 729 static DEVICE_ATTR(cpu1_voltage,S_IRUGO,show_cpu1_voltage,NULL);
 730 static DEVICE_ATTR(cpu1_current,S_IRUGO,show_cpu1_current,NULL);
 731 static DEVICE_ATTR(cpu1_exhaust_fan_rpm,S_IRUGO,show_cpu1_exhaust_fan_rpm,NULL);
 732 static DEVICE_ATTR(cpu1_intake_fan_rpm,S_IRUGO,show_cpu1_intake_fan_rpm,NULL);
 733
 734 static DEVICE_ATTR(backside_temperature,S_IRUGO,show_backside_temperature,NULL);
 735 static DEVICE_ATTR(backside_fan_pwm,S_IRUGO,show_backside_fan_pwm,NULL);
 736
 737 static DEVICE_ATTR(drives_temperature,S_IRUGO,show_drives_temperature,NULL);
 738 static DEVICE_ATTR(drives_fan_rpm,S_IRUGO,show_drives_fan_rpm,NULL);
 739
 740 static DEVICE_ATTR(dimms_temperature,S_IRUGO,show_dimms_temperature,NULL);
 741
 742 /*
 743  * CPUs fans control loop
 744  */
 745
 746 static int do_read_one_cpu_values(struct cpu_pid_state *state, s32 *temp, s32 *power)
 747 {
 748         s32 ltemp, volts, amps;
 749         int index, rc = 0;
 750
 751         /* Default (in case of error) */
 752         *temp = state->cur_temp;
 753         *power = state->cur_power;
 754
 755         if (cpu_pid_type == CPU_PID_TYPE_RACKMAC)
 756                 index = (state->index == 0) ?
 757                         CPU_A1_FAN_RPM_INDEX : CPU_B1_FAN_RPM_INDEX;
 758         else
 759                 index = (state->index == 0) ?
 760                         CPUA_EXHAUST_FAN_RPM_INDEX : CPUB_EXHAUST_FAN_RPM_INDEX;
 761
 762         /* Read current fan status */
 763         rc = get_rpm_fan(index, !RPM_PID_USE_ACTUAL_SPEED);
 764         if (rc < 0) {
 765                 /* XXX What do we do now ? Nothing for now, keep old value, but
 766                  * return error upstream
 767                  */
 768                 DBG("  cpu %d, fan reading error !\n", state->index);
 769         } else {
 770                 state->rpm = rc;
 771                 DBG("  cpu %d, exhaust RPM: %d\n", state->index, state->rpm);
 772         }
 773
 774         /* Get some sensor readings and scale it */
 775         ltemp = read_smon_adc(state, 1);
 776         if (ltemp == -1) {
 777                 /* XXX What do we do now ? */
 778                 state->overtemp++;
 779                 if (rc == 0)
 780                         rc = -EIO;
 781                 DBG("  cpu %d, temp reading error !\n", state->index);
 782         } else {
 783                 /* Fixup temperature according to diode calibration
 784                  */
 785                 DBG("  cpu %d, temp raw: %04x, m_diode: %04x, b_diode: %04x\n",
 786                     state->index,
 787                     ltemp, state->mpu.mdiode, state->mpu.bdiode);
 788                 *temp = ((s32)ltemp * (s32)state->mpu.mdiode + ((s32)state->mpu.bdiode << 12)) >> 2;
 789                 state->last_temp = *temp;
 790                 DBG("  temp: %d.%03d\n", FIX32TOPRINT((*temp)));
 791         }
 792
 793         /*
 794          * Read voltage & current and calculate power
 795          */
 796         volts = read_smon_adc(state, 3);
 797         amps = read_smon_adc(state, 4);
 798
 799         /* Scale voltage and current raw sensor values according to fixed scales
 800          * obtained in Darwin and calculate power from I and V
 801          */
 802         volts *= ADC_CPU_VOLTAGE_SCALE;
 803         amps *= ADC_CPU_CURRENT_SCALE;
 804         *power = (((u64)volts) * ((u64)amps)) >> 16;
 805         state->voltage = volts;
 806         state->current_a = amps;
 807         state->last_power = *power;
 808
 809         DBG("  cpu %d, current: %d.%03d, voltage: %d.%03d, power: %d.%03d W\n",
 810             state->index, FIX32TOPRINT(state->current_a),
 811             FIX32TOPRINT(state->voltage), FIX32TOPRINT(*power));
 812
 813         return 0;
 814 }
 815
 816 static void do_cpu_pid(struct cpu_pid_state *state, s32 temp, s32 power)
 817 {
 818         s32 power_target, integral, derivative, proportional, adj_in_target, sval;
 819         s64 integ_p, deriv_p, prop_p, sum;
 820         int i;
 821
 822         /* Calculate power target value (could be done once for all)
 823          * and convert to a 16.16 fp number
 824          */
 825         power_target = ((u32)(state->mpu.pmaxh - state->mpu.padjmax)) << 16;
 826         DBG("  power target: %d.%03d, error: %d.%03d\n",
 827             FIX32TOPRINT(power_target), FIX32TOPRINT(power_target - power));
 828
 829         /* Store temperature and power in history array */
 830         state->cur_temp = (state->cur_temp + 1) % CPU_TEMP_HISTORY_SIZE;
 831         state->temp_history[state->cur_temp] = temp;
 832         state->cur_power = (state->cur_power + 1) % state->count_power;
 833         state->power_history[state->cur_power] = power;
 834         state->error_history[state->cur_power] = power_target - power;
 835
 836         /* If first loop, fill the history table */
 837         if (state->first) {
 838                 for (i = 0; i < (state->count_power - 1); i++) {
 839                         state->cur_power = (state->cur_power + 1) % state->count_power;
 840                         state->power_history[state->cur_power] = power;
 841                         state->error_history[state->cur_power] = power_target - power;
 842                 }
 843                 for (i = 0; i < (CPU_TEMP_HISTORY_SIZE - 1); i++) {
 844                         state->cur_temp = (state->cur_temp + 1) % CPU_TEMP_HISTORY_SIZE;
 845                         state->temp_history[state->cur_temp] = temp;
 846                 }
 847                 state->first = 0;
 848         }
 849
 850         /* Calculate the integral term normally based on the "power" values */
 851         sum = 0;
 852         integral = 0;
 853         for (i = 0; i < state->count_power; i++)
 854                 integral += state->error_history[i];
 855         integral *= CPU_PID_INTERVAL;
 856         DBG("  integral: %08x\n", integral);
 857
 858         /* Calculate the adjusted input (sense value).
 859          *   G_r is 12.20
 860          *   integ is 16.16
 861          *   so the result is 28.36
 862          *
 863          * input target is mpu.ttarget, input max is mpu.tmax
 864          */
 865         integ_p = ((s64)state->mpu.pid_gr) * (s64)integral;
 866         DBG("   integ_p: %d\n", (int)(integ_p >> 36));
 867         sval = (state->mpu.tmax << 16) - ((integ_p >> 20) & 0xffffffff);
 868         adj_in_target = (state->mpu.ttarget << 16);
 869         if (adj_in_target > sval)
 870                 adj_in_target = sval;
 871         DBG("   adj_in_target: %d.%03d, ttarget: %d\n", FIX32TOPRINT(adj_in_target),
 872             state->mpu.ttarget);
 873
 874         /* Calculate the derivative term */
 875         derivative = state->temp_history[state->cur_temp] -
 876                 state->temp_history[(state->cur_temp + CPU_TEMP_HISTORY_SIZE - 1)
 877                                     % CPU_TEMP_HISTORY_SIZE];
 878         derivative /= CPU_PID_INTERVAL;
 879         deriv_p = ((s64)state->mpu.pid_gd) * (s64)derivative;
 880         DBG("   deriv_p: %d\n", (int)(deriv_p >> 36));
 881         sum += deriv_p;
 882
 883         /* Calculate the proportional term */
 884         proportional = temp - adj_in_target;
 885         prop_p = ((s64)state->mpu.pid_gp) * (s64)proportional;
 886         DBG("   prop_p: %d\n", (int)(prop_p >> 36));
 887         sum += prop_p;
 888
 889         /* Scale sum */
 890         sum >>= 36;
 891
 892         DBG("   sum: %d\n", (int)sum);
 893         state->rpm += (s32)sum;
 894 }
 895
 896 static void do_monitor_cpu_combined(void)
 897 {
 898         struct cpu_pid_state *state0 = &cpu_state[0];
 899         struct cpu_pid_state *state1 = &cpu_state[1];
 900         s32 temp0, power0, temp1, power1;
 901         s32 temp_combi, power_combi;
 902         int rc, intake, pump;
 903
 904         rc = do_read_one_cpu_values(state0, &temp0, &power0);
 905         if (rc < 0) {
 906                 /* XXX What do we do now ? */
 907         }
 908         state1->overtemp = 0;
 909         rc = do_read_one_cpu_values(state1, &temp1, &power1);
 910         if (rc < 0) {
 911                 /* XXX What do we do now ? */
 912         }
 913         if (state1->overtemp)
 914                 state0->overtemp++;
 915
 916         temp_combi = max(temp0, temp1);
 917         power_combi = max(power0, power1);
 918
 919         /* Check tmax, increment overtemp if we are there. At tmax+8, we go
 920          * full blown immediately and try to trigger a shutdown
 921          */
 922         if (temp_combi >= ((state0->mpu.tmax + 8) << 16)) {
 923                 printk(KERN_WARNING "Warning ! Temperature way above maximum (%d) !\n",
 924                        temp_combi >> 16);
 925                 state0->overtemp = CPU_MAX_OVERTEMP;
 926         } else if (temp_combi > (state0->mpu.tmax << 16))
 927                 state0->overtemp++;
 928         else
 929                 state0->overtemp = 0;
 930         if (state0->overtemp >= CPU_MAX_OVERTEMP)
 931                 critical_state = 1;
 932         if (state0->overtemp > 0) {
 933                 state0->rpm = state0->mpu.rmaxn_exhaust_fan;
 934                 state0->intake_rpm = intake = state0->mpu.rmaxn_intake_fan;
 935                 pump = state0->pump_min;
 936                 goto do_set_fans;
 937         }
 938
 939         /* Do the PID */
 940         do_cpu_pid(state0, temp_combi, power_combi);
 941
 942         /* Range check */
 943         state0->rpm = max(state0->rpm, (int)state0->mpu.rminn_exhaust_fan);
 944         state0->rpm = min(state0->rpm, (int)state0->mpu.rmaxn_exhaust_fan);
 945
 946         /* Calculate intake fan speed */
 947         intake = (state0->rpm * CPU_INTAKE_SCALE) >> 16;
 948         intake = max(intake, (int)state0->mpu.rminn_intake_fan);
 949         intake = min(intake, (int)state0->mpu.rmaxn_intake_fan);
 950         state0->intake_rpm = intake;
 951
 952         /* Calculate pump speed */
 953         pump = (state0->rpm * state0->pump_max) /
 954                 state0->mpu.rmaxn_exhaust_fan;
 955         pump = min(pump, state0->pump_max);
 956         pump = max(pump, state0->pump_min);
 957
 958  do_set_fans:
 959         /* We copy values from state 0 to state 1 for /sysfs */
 960         state1->rpm = state0->rpm;
 961         state1->intake_rpm = state0->intake_rpm;
 962
 963         DBG("** CPU %d RPM: %d Ex, %d, Pump: %d, In, overtemp: %d\n",
 964             state1->index, (int)state1->rpm, intake, pump, state1->overtemp);
 965
 966         /* We should check for errors, shouldn't we ? But then, what
 967          * do we do once the error occurs ? For FCU notified fan
 968          * failures (-EFAULT) we probably want to notify userland
 969          * some way...
 970          */
 971         set_rpm_fan(CPUA_INTAKE_FAN_RPM_INDEX, intake);
 972         set_rpm_fan(CPUA_EXHAUST_FAN_RPM_INDEX, state0->rpm);
 973         set_rpm_fan(CPUB_INTAKE_FAN_RPM_INDEX, intake);
 974         set_rpm_fan(CPUB_EXHAUST_FAN_RPM_INDEX, state0->rpm);
 975
 976         if (fcu_fans[CPUA_PUMP_RPM_INDEX].id != FCU_FAN_ABSENT_ID)
 977                 set_rpm_fan(CPUA_PUMP_RPM_INDEX, pump);
 978         if (fcu_fans[CPUB_PUMP_RPM_INDEX].id != FCU_FAN_ABSENT_ID)
 979                 set_rpm_fan(CPUB_PUMP_RPM_INDEX, pump);
 980 }
 981
 982 static void do_monitor_cpu_split(struct cpu_pid_state *state)
 983 {
 984         s32 temp, power;
 985         int rc, intake;
 986
 987         /* Read current fan status */
 988         rc = do_read_one_cpu_values(state, &temp, &power);
 989         if (rc < 0) {
 990                 /* XXX What do we do now ? */
 991         }
 992
 993         /* Check tmax, increment overtemp if we are there. At tmax+8, we go
 994          * full blown immediately and try to trigger a shutdown
 995          */
 996         if (temp >= ((state->mpu.tmax + 8) << 16)) {
 997                 printk(KERN_WARNING "Warning ! CPU %d temperature way above maximum"
 998                        " (%d) !\n",
 999                        state->index, temp >> 16);
1000                 state->overtemp = CPU_MAX_OVERTEMP;
1001         } else if (temp > (state->mpu.tmax << 16))
1002                 state->overtemp++;
1003         else
1004                 state->overtemp = 0;
1005         if (state->overtemp >= CPU_MAX_OVERTEMP)
1006                 critical_state = 1;
1007         if (state->overtemp > 0) {
1008                 state->rpm = state->mpu.rmaxn_exhaust_fan;
1009                 state->intake_rpm = intake = state->mpu.rmaxn_intake_fan;
1010                 goto do_set_fans;
1011         }
1012
1013         /* Do the PID */
1014         do_cpu_pid(state, temp, power);
1015
1016         /* Range check */
1017         state->rpm = max(state->rpm, (int)state->mpu.rminn_exhaust_fan);
1018         state->rpm = min(state->rpm, (int)state->mpu.rmaxn_exhaust_fan);
1019
1020         /* Calculate intake fan */
1021         intake = (state->rpm * CPU_INTAKE_SCALE) >> 16;
1022         intake = max(intake, (int)state->mpu.rminn_intake_fan);
1023         intake = min(intake, (int)state->mpu.rmaxn_intake_fan);
1024         state->intake_rpm = intake;
1025
1026  do_set_fans:
1027         DBG("** CPU %d RPM: %d Ex, %d In, overtemp: %d\n",
1028             state->index, (int)state->rpm, intake, state->overtemp);
1029
1030         /* We should check for errors, shouldn't we ? But then, what
1031          * do we do once the error occurs ? For FCU notified fan
1032          * failures (-EFAULT) we probably want to notify userland
1033          * some way...
1034          */
1035         if (state->index == 0) {
1036                 set_rpm_fan(CPUA_INTAKE_FAN_RPM_INDEX, intake);
1037                 set_rpm_fan(CPUA_EXHAUST_FAN_RPM_INDEX, state->rpm);
1038         } else {
1039                 set_rpm_fan(CPUB_INTAKE_FAN_RPM_INDEX, intake);
1040                 set_rpm_fan(CPUB_EXHAUST_FAN_RPM_INDEX, state->rpm);
1041         }
1042 }
1043
1044 static void do_monitor_cpu_rack(struct cpu_pid_state *state)
1045 {
1046         s32 temp, power, fan_min;
1047         int rc;
1048
1049         /* Read current fan status */
1050         rc = do_read_one_cpu_values(state, &temp, &power);
1051         if (rc < 0) {
1052                 /* XXX What do we do now ? */
1053         }
1054
1055         /* Check tmax, increment overtemp if we are there. At tmax+8, we go
1056          * full blown immediately and try to trigger a shutdown
1057          */
1058         if (temp >= ((state->mpu.tmax + 8) << 16)) {
1059                 printk(KERN_WARNING "Warning ! CPU %d temperature way above maximum"
1060                        " (%d) !\n",
1061                        state->index, temp >> 16);
1062                 state->overtemp = CPU_MAX_OVERTEMP;
1063         } else if (temp > (state->mpu.tmax << 16))
1064                 state->overtemp++;
1065         else
1066                 state->overtemp = 0;
1067         if (state->overtemp >= CPU_MAX_OVERTEMP)
1068                 critical_state = 1;
1069         if (state->overtemp > 0) {
1070                 state->rpm = state->intake_rpm = state->mpu.rmaxn_intake_fan;
1071                 goto do_set_fans;
1072         }
1073
1074         /* Do the PID */
1075         do_cpu_pid(state, temp, power);
1076
1077         /* Check clamp from dimms */
1078         fan_min = dimm_output_clamp;
1079         fan_min = max(fan_min, (int)state->mpu.rminn_intake_fan);
1080
1081         state->rpm = max(state->rpm, (int)fan_min);
1082         state->rpm = min(state->rpm, (int)state->mpu.rmaxn_intake_fan);
1083         state->intake_rpm = state->rpm;
1084
1085  do_set_fans:
1086         DBG("** CPU %d RPM: %d overtemp: %d\n",
1087             state->index, (int)state->rpm, state->overtemp);
1088
1089         /* We should check for errors, shouldn't we ? But then, what
1090          * do we do once the error occurs ? For FCU notified fan
1091          * failures (-EFAULT) we probably want to notify userland
1092          * some way...
1093          */
1094         if (state->index == 0) {
1095                 set_rpm_fan(CPU_A1_FAN_RPM_INDEX, state->rpm);
1096                 set_rpm_fan(CPU_A2_FAN_RPM_INDEX, state->rpm);
1097                 set_rpm_fan(CPU_A3_FAN_RPM_INDEX, state->rpm);
1098         } else {
1099                 set_rpm_fan(CPU_B1_FAN_RPM_INDEX, state->rpm);
1100                 set_rpm_fan(CPU_B2_FAN_RPM_INDEX, state->rpm);
1101                 set_rpm_fan(CPU_B3_FAN_RPM_INDEX, state->rpm);
1102         }
1103 }
1104
1105 /*
1106  * Initialize the state structure for one CPU control loop
1107  */
1108 static int init_cpu_state(struct cpu_pid_state *state, int index)
1109 {
1110         state->index = index;
1111         state->first = 1;
1112         state->rpm = (cpu_pid_type == CPU_PID_TYPE_RACKMAC) ? 4000 : 1000;
1113         state->overtemp = 0;
1114         state->adc_config = 0x00;
1115
1116
1117         if (index == 0)
1118                 state->monitor = attach_i2c_chip(SUPPLY_MONITOR_ID, "CPU0_monitor");
1119         else if (index == 1)
1120                 state->monitor = attach_i2c_chip(SUPPLY_MONITORB_ID, "CPU1_monitor");
1121         if (state->monitor == NULL)
1122                 goto fail;
1123
1124         if (read_eeprom(index, &state->mpu))
1125                 goto fail;
1126
1127         state->count_power = state->mpu.tguardband;
1128         if (state->count_power > CPU_POWER_HISTORY_SIZE) {
1129                 printk(KERN_WARNING "Warning ! too many power history slots\n");
1130                 state->count_power = CPU_POWER_HISTORY_SIZE;
1131         }
1132         DBG("CPU %d Using %d power history entries\n", index, state->count_power);
1133
1134         if (index == 0) {
1135                 device_create_file(&of_dev->dev, &dev_attr_cpu0_temperature);
1136                 device_create_file(&of_dev->dev, &dev_attr_cpu0_voltage);
1137                 device_create_file(&of_dev->dev, &dev_attr_cpu0_current);
1138                 device_create_file(&of_dev->dev, &dev_attr_cpu0_exhaust_fan_rpm);
1139                 device_create_file(&of_dev->dev, &dev_attr_cpu0_intake_fan_rpm);
1140         } else {
1141                 device_create_file(&of_dev->dev, &dev_attr_cpu1_temperature);
1142                 device_create_file(&of_dev->dev, &dev_attr_cpu1_voltage);
1143                 device_create_file(&of_dev->dev, &dev_attr_cpu1_current);
1144                 device_create_file(&of_dev->dev, &dev_attr_cpu1_exhaust_fan_rpm);
1145                 device_create_file(&of_dev->dev, &dev_attr_cpu1_intake_fan_rpm);
1146         }
1147
1148         return 0;
1149  fail:
1150         if (state->monitor)
1151                 detach_i2c_chip(state->monitor);
1152         state->monitor = NULL;
1153
1154         return -ENODEV;
1155 }
1156
1157 /*
1158  * Dispose of the state data for one CPU control loop
1159  */
1160 static void dispose_cpu_state(struct cpu_pid_state *state)
1161 {
1162         if (state->monitor == NULL)
1163                 return;
1164
1165         if (state->index == 0) {
1166                 device_remove_file(&of_dev->dev, &dev_attr_cpu0_temperature);
1167                 device_remove_file(&of_dev->dev, &dev_attr_cpu0_voltage);
1168                 device_remove_file(&of_dev->dev, &dev_attr_cpu0_current);
1169                 device_remove_file(&of_dev->dev, &dev_attr_cpu0_exhaust_fan_rpm);
1170                 device_remove_file(&of_dev->dev, &dev_attr_cpu0_intake_fan_rpm);
1171         } else {
1172                 device_remove_file(&of_dev->dev, &dev_attr_cpu1_temperature);
1173                 device_remove_file(&of_dev->dev, &dev_attr_cpu1_voltage);
1174                 device_remove_file(&of_dev->dev, &dev_attr_cpu1_current);
1175                 device_remove_file(&of_dev->dev, &dev_attr_cpu1_exhaust_fan_rpm);
1176                 device_remove_file(&of_dev->dev, &dev_attr_cpu1_intake_fan_rpm);
1177         }
1178
1179         detach_i2c_chip(state->monitor);
1180         state->monitor = NULL;
1181 }
1182
1183 /*
1184  * Motherboard backside & U3 heatsink fan control loop
1185  */
1186 static void do_monitor_backside(struct backside_pid_state *state)
1187 {
1188         s32 temp, integral, derivative, fan_min;
1189         s64 integ_p, deriv_p, prop_p, sum;
1190         int i, rc;
1191
1192         if (--state->ticks != 0)
1193                 return;
1194         state->ticks = backside_params.interval;
1195
1196         DBG("backside:\n");
1197
1198         /* Check fan status */
1199         rc = get_pwm_fan(BACKSIDE_FAN_PWM_INDEX);
1200         if (rc < 0) {
1201                 printk(KERN_WARNING "Error %d reading backside fan !\n", rc);
1202                 /* XXX What do we do now ? */
1203         } else
1204                 state->pwm = rc;
1205         DBG("  current pwm: %d\n", state->pwm);
1206
1207         /* Get some sensor readings */
1208         temp = i2c_smbus_read_byte_data(state->monitor, MAX6690_EXT_TEMP) << 16;
1209         state->last_temp = temp;
1210         DBG("  temp: %d.%03d, target: %d.%03d\n", FIX32TOPRINT(temp),
1211             FIX32TOPRINT(backside_params.input_target));
1212
1213         /* Store temperature and error in history array */
1214         state->cur_sample = (state->cur_sample + 1) % BACKSIDE_PID_HISTORY_SIZE;
1215         state->sample_history[state->cur_sample] = temp;
1216         state->error_history[state->cur_sample] = temp - backside_params.input_target;
1217
1218         /* If first loop, fill the history table */
1219         if (state->first) {
1220                 for (i = 0; i < (BACKSIDE_PID_HISTORY_SIZE - 1); i++) {
1221                         state->cur_sample = (state->cur_sample + 1) %
1222                                 BACKSIDE_PID_HISTORY_SIZE;
1223                         state->sample_history[state->cur_sample] = temp;
1224                         state->error_history[state->cur_sample] =
1225                                 temp - backside_params.input_target;
1226                 }
1227                 state->first = 0;
1228         }
1229
1230         /* Calculate the integral term */
1231         sum = 0;
1232         integral = 0;
1233         for (i = 0; i < BACKSIDE_PID_HISTORY_SIZE; i++)
1234                 integral += state->error_history[i];
1235         integral *= backside_params.interval;
1236         DBG("  integral: %08x\n", integral);
1237         integ_p = ((s64)backside_params.G_r) * (s64)integral;
1238         DBG("   integ_p: %d\n", (int)(integ_p >> 36));
1239         sum += integ_p;
1240
1241         /* Calculate the derivative term */
1242         derivative = state->error_history[state->cur_sample] -
1243                 state->error_history[(state->cur_sample + BACKSIDE_PID_HISTORY_SIZE - 1)
1244                                     % BACKSIDE_PID_HISTORY_SIZE];
1245         derivative /= backside_params.interval;
1246         deriv_p = ((s64)backside_params.G_d) * (s64)derivative;
1247         DBG("   deriv_p: %d\n", (int)(deriv_p >> 36));
1248         sum += deriv_p;
1249
1250         /* Calculate the proportional term */
1251         prop_p = ((s64)backside_params.G_p) * (s64)(state->error_history[state->cur_sample]);
1252         DBG("   prop_p: %d\n", (int)(prop_p >> 36));
1253         sum += prop_p;
1254
1255         /* Scale sum */
1256         sum >>= 36;
1257
1258         DBG("   sum: %d\n", (int)sum);
1259         if (backside_params.additive)
1260                 state->pwm += (s32)sum;
1261         else
1262                 state->pwm = sum;
1263
1264         /* Check for clamp */
1265         fan_min = (dimm_output_clamp * 100) / 14000;
1266         fan_min = max(fan_min, backside_params.output_min);
1267
1268         state->pwm = max(state->pwm, fan_min);
1269         state->pwm = min(state->pwm, backside_params.output_max);
1270
1271         DBG("** BACKSIDE PWM: %d\n", (int)state->pwm);
1272         set_pwm_fan(BACKSIDE_FAN_PWM_INDEX, state->pwm);
1273 }
1274
1275 /*
1276  * Initialize the state structure for the backside fan control loop
1277  */
1278 static int init_backside_state(struct backside_pid_state *state)
1279 {
1280         struct device_node *u3;
1281         int u3h = 1; /* conservative by default */
1282
1283         /*
1284          * There are different PID params for machines with U3 and machines
1285          * with U3H, pick the right ones now
1286          */
1287         u3 = of_find_node_by_path("/u3@0,f8000000");
1288         if (u3 != NULL) {
1289                 u32 *vers = (u32 *)get_property(u3, "device-rev", NULL);
1290                 if (vers)
1291                         if (((*vers) & 0x3f) < 0x34)
1292                                 u3h = 0;
1293                 of_node_put(u3);
1294         }
1295
1296         if (rackmac) {
1297                 backside_params.G_d = BACKSIDE_PID_RACK_G_d;
1298                 backside_params.input_target = BACKSIDE_PID_RACK_INPUT_TARGET;
1299                 backside_params.output_min = BACKSIDE_PID_U3H_OUTPUT_MIN;
1300                 backside_params.interval = BACKSIDE_PID_RACK_INTERVAL;
1301                 backside_params.G_p = BACKSIDE_PID_RACK_G_p;
1302                 backside_params.G_r = BACKSIDE_PID_G_r;
1303                 backside_params.output_max = BACKSIDE_PID_OUTPUT_MAX;
1304                 backside_params.additive = 0;
1305         } else if (u3h) {
1306                 backside_params.G_d = BACKSIDE_PID_U3H_G_d;
1307                 backside_params.input_target = BACKSIDE_PID_U3H_INPUT_TARGET;
1308                 backside_params.output_min = BACKSIDE_PID_U3H_OUTPUT_MIN;
1309                 backside_params.interval = BACKSIDE_PID_INTERVAL;
1310                 backside_params.G_p = BACKSIDE_PID_G_p;
1311                 backside_params.G_r = BACKSIDE_PID_G_r;
1312                 backside_params.output_max = BACKSIDE_PID_OUTPUT_MAX;
1313                 backside_params.additive = 1;
1314         } else {
1315                 backside_params.G_d = BACKSIDE_PID_U3_G_d;
1316                 backside_params.input_target = BACKSIDE_PID_U3_INPUT_TARGET;
1317                 backside_params.output_min = BACKSIDE_PID_U3_OUTPUT_MIN;
1318                 backside_params.interval = BACKSIDE_PID_INTERVAL;
1319                 backside_params.G_p = BACKSIDE_PID_G_p;
1320                 backside_params.G_r = BACKSIDE_PID_G_r;
1321                 backside_params.output_max = BACKSIDE_PID_OUTPUT_MAX;
1322                 backside_params.additive = 1;
1323         }
1324
1325         state->ticks = 1;
1326         state->first = 1;
1327         state->pwm = 50;
1328
1329         state->monitor = attach_i2c_chip(BACKSIDE_MAX_ID, "backside_temp");
1330         if (state->monitor == NULL)
1331                 return -ENODEV;
1332
1333         device_create_file(&of_dev->dev, &dev_attr_backside_temperature);
1334         device_create_file(&of_dev->dev, &dev_attr_backside_fan_pwm);
1335
1336         return 0;
1337 }
1338
1339 /*
1340  * Dispose of the state data for the backside control loop
1341  */
1342 static void dispose_backside_state(struct backside_pid_state *state)
1343 {
1344         if (state->monitor == NULL)
1345                 return;
1346
1347         device_remove_file(&of_dev->dev, &dev_attr_backside_temperature);
1348         device_remove_file(&of_dev->dev, &dev_attr_backside_fan_pwm);
1349
1350         detach_i2c_chip(state->monitor);
1351         state->monitor = NULL;
1352 }
1353
1354 /*
1355  * Drives bay fan control loop
1356  */
1357 static void do_monitor_drives(struct drives_pid_state *state)
1358 {
1359         s32 temp, integral, derivative;
1360         s64 integ_p, deriv_p, prop_p, sum;
1361         int i, rc;
1362
1363         if (--state->ticks != 0)
1364                 return;
1365         state->ticks = DRIVES_PID_INTERVAL;
1366
1367         DBG("drives:\n");
1368
1369         /* Check fan status */
1370         rc = get_rpm_fan(DRIVES_FAN_RPM_INDEX, !RPM_PID_USE_ACTUAL_SPEED);
1371         if (rc < 0) {
1372                 printk(KERN_WARNING "Error %d reading drives fan !\n", rc);
1373                 /* XXX What do we do now ? */
1374         } else
1375                 state->rpm = rc;
1376         DBG("  current rpm: %d\n", state->rpm);
1377
1378         /* Get some sensor readings */
1379         temp = le16_to_cpu(i2c_smbus_read_word_data(state->monitor, DS1775_TEMP)) << 8;
1380         state->last_temp = temp;
1381         DBG("  temp: %d.%03d, target: %d.%03d\n", FIX32TOPRINT(temp),
1382             FIX32TOPRINT(DRIVES_PID_INPUT_TARGET));
1383
1384         /* Store temperature and error in history array */
1385         state->cur_sample = (state->cur_sample + 1) % DRIVES_PID_HISTORY_SIZE;
1386         state->sample_history[state->cur_sample] = temp;
1387         state->error_history[state->cur_sample] = temp - DRIVES_PID_INPUT_TARGET;
1388
1389         /* If first loop, fill the history table */
1390         if (state->first) {
1391                 for (i = 0; i < (DRIVES_PID_HISTORY_SIZE - 1); i++) {
1392                         state->cur_sample = (state->cur_sample + 1) %
1393                                 DRIVES_PID_HISTORY_SIZE;
1394                         state->sample_history[state->cur_sample] = temp;
1395                         state->error_history[state->cur_sample] =
1396                                 temp - DRIVES_PID_INPUT_TARGET;
1397                 }
1398                 state->first = 0;
1399         }
1400
1401         /* Calculate the integral term */
1402         sum = 0;
1403         integral = 0;
1404         for (i = 0; i < DRIVES_PID_HISTORY_SIZE; i++)
1405                 integral += state->error_history[i];
1406         integral *= DRIVES_PID_INTERVAL;
1407         DBG("  integral: %08x\n", integral);
1408         integ_p = ((s64)DRIVES_PID_G_r) * (s64)integral;
1409         DBG("   integ_p: %d\n", (int)(integ_p >> 36));
1410         sum += integ_p;
1411
1412         /* Calculate the derivative term */
1413         derivative = state->error_history[state->cur_sample] -
1414                 state->error_history[(state->cur_sample + DRIVES_PID_HISTORY_SIZE - 1)
1415                                     % DRIVES_PID_HISTORY_SIZE];
1416         derivative /= DRIVES_PID_INTERVAL;
1417         deriv_p = ((s64)DRIVES_PID_G_d) * (s64)derivative;
1418         DBG("   deriv_p: %d\n", (int)(deriv_p >> 36));
1419         sum += deriv_p;
1420
1421         /* Calculate the proportional term */
1422         prop_p = ((s64)DRIVES_PID_G_p) * (s64)(state->error_history[state->cur_sample]);
1423         DBG("   prop_p: %d\n", (int)(prop_p >> 36));
1424         sum += prop_p;
1425
1426         /* Scale sum */
1427         sum >>= 36;
1428
1429         DBG("   sum: %d\n", (int)sum);
1430         state->rpm += (s32)sum;
1431
1432         state->rpm = max(state->rpm, DRIVES_PID_OUTPUT_MIN);
1433         state->rpm = min(state->rpm, DRIVES_PID_OUTPUT_MAX);
1434
1435         DBG("** DRIVES RPM: %d\n", (int)state->rpm);
1436         set_rpm_fan(DRIVES_FAN_RPM_INDEX, state->rpm);
1437 }
1438
1439 /*
1440  * Initialize the state structure for the drives bay fan control loop
1441  */
1442 static int init_drives_state(struct drives_pid_state *state)
1443 {
1444         state->ticks = 1;
1445         state->first = 1;
1446         state->rpm = 1000;
1447
1448         state->monitor = attach_i2c_chip(DRIVES_DALLAS_ID, "drives_temp");
1449         if (state->monitor == NULL)
1450                 return -ENODEV;
1451
1452         device_create_file(&of_dev->dev, &dev_attr_drives_temperature);
1453         device_create_file(&of_dev->dev, &dev_attr_drives_fan_rpm);
1454
1455         return 0;
1456 }
1457
1458 /*
1459  * Dispose of the state data for the drives control loop
1460  */
1461 static void dispose_drives_state(struct drives_pid_state *state)
1462 {
1463         if (state->monitor == NULL)
1464                 return;
1465
1466         device_remove_file(&of_dev->dev, &dev_attr_drives_temperature);
1467         device_remove_file(&of_dev->dev, &dev_attr_drives_fan_rpm);
1468
1469         detach_i2c_chip(state->monitor);
1470         state->monitor = NULL;
1471 }
1472
1473 /*
1474  * DIMMs temp control loop
1475  */
1476 static void do_monitor_dimms(struct dimm_pid_state *state)
1477 {
1478         s32 temp, integral, derivative, fan_min;
1479         s64 integ_p, deriv_p, prop_p, sum;
1480         int i;
1481
1482         if (--state->ticks != 0)
1483                 return;
1484         state->ticks = DIMM_PID_INTERVAL;
1485
1486         DBG("DIMM:\n");
1487
1488         DBG("  current value: %d\n", state->output);
1489
1490         temp = read_lm87_reg(state->monitor, LM87_INT_TEMP);
1491         if (temp < 0)
1492                 return;
1493         temp <<= 16;
1494         state->last_temp = temp;
1495         DBG("  temp: %d.%03d, target: %d.%03d\n", FIX32TOPRINT(temp),
1496             FIX32TOPRINT(DIMM_PID_INPUT_TARGET));
1497
1498         /* Store temperature and error in history array */
1499         state->cur_sample = (state->cur_sample + 1) % DIMM_PID_HISTORY_SIZE;
1500         state->sample_history[state->cur_sample] = temp;
1501         state->error_history[state->cur_sample] = temp - DIMM_PID_INPUT_TARGET;
1502
1503         /* If first loop, fill the history table */
1504         if (state->first) {
1505                 for (i = 0; i < (DIMM_PID_HISTORY_SIZE - 1); i++) {
1506                         state->cur_sample = (state->cur_sample + 1) %
1507                                 DIMM_PID_HISTORY_SIZE;
1508                         state->sample_history[state->cur_sample] = temp;
1509                         state->error_history[state->cur_sample] =
1510                                 temp - DIMM_PID_INPUT_TARGET;
1511                 }
1512                 state->first = 0;
1513         }
1514
1515         /* Calculate the integral term */
1516         sum = 0;
1517         integral = 0;
1518         for (i = 0; i < DIMM_PID_HISTORY_SIZE; i++)
1519                 integral += state->error_history[i];
1520         integral *= DIMM_PID_INTERVAL;
1521         DBG("  integral: %08x\n", integral);
1522         integ_p = ((s64)DIMM_PID_G_r) * (s64)integral;
1523         DBG("   integ_p: %d\n", (int)(integ_p >> 36));
1524         sum += integ_p;
1525
1526         /* Calculate the derivative term */
1527         derivative = state->error_history[state->cur_sample] -
1528                 state->error_history[(state->cur_sample + DIMM_PID_HISTORY_SIZE - 1)
1529                                     % DIMM_PID_HISTORY_SIZE];
1530         derivative /= DIMM_PID_INTERVAL;
1531         deriv_p = ((s64)DIMM_PID_G_d) * (s64)derivative;
1532         DBG("   deriv_p: %d\n", (int)(deriv_p >> 36));
1533         sum += deriv_p;
1534
1535         /* Calculate the proportional term */
1536         prop_p = ((s64)DIMM_PID_G_p) * (s64)(state->error_history[state->cur_sample]);
1537         DBG("   prop_p: %d\n", (int)(prop_p >> 36));
1538         sum += prop_p;
1539
1540         /* Scale sum */
1541         sum >>= 36;
1542
1543         DBG("   sum: %d\n", (int)sum);
1544         state->output = (s32)sum;
1545         state->output = max(state->output, DIMM_PID_OUTPUT_MIN);
1546         state->output = min(state->output, DIMM_PID_OUTPUT_MAX);
1547         dimm_output_clamp = state->output;
1548
1549         DBG("** DIMM clamp value: %d\n", (int)state->output);
1550
1551         /* Backside PID is only every 5 seconds, force backside fan clamping now */
1552         fan_min = (dimm_output_clamp * 100) / 14000;
1553         fan_min = max(fan_min, backside_params.output_min);
1554         if (backside_state.pwm < fan_min) {
1555                 backside_state.pwm = fan_min;
1556                 DBG(" -> applying clamp to backside fan now: %d  !\n", fan_min);
1557                 set_pwm_fan(BACKSIDE_FAN_PWM_INDEX, fan_min);
1558         }
1559 }
1560
1561 /*
1562  * Initialize the state structure for the DIMM temp control loop
1563  */
1564 static int init_dimms_state(struct dimm_pid_state *state)
1565 {
1566         state->ticks = 1;
1567         state->first = 1;
1568         state->output = 4000;
1569
1570         state->monitor = attach_i2c_chip(XSERVE_DIMMS_LM87, "dimms_temp");
1571         if (state->monitor == NULL)
1572                 return -ENODEV;
1573
1574         device_create_file(&of_dev->dev, &dev_attr_dimms_temperature);
1575
1576         return 0;
1577 }
1578
1579 /*
1580  * Dispose of the state data for the drives control loop
1581  */
1582 static void dispose_dimms_state(struct dimm_pid_state *state)
1583 {
1584         if (state->monitor == NULL)
1585                 return;
1586
1587         device_remove_file(&of_dev->dev, &dev_attr_dimms_temperature);
1588
1589         detach_i2c_chip(state->monitor);
1590         state->monitor = NULL;
1591 }
1592
1593 static int call_critical_overtemp(void)
1594 {
1595         char *argv[] = { critical_overtemp_path, NULL };
1596         static char *envp[] = { "HOME=/",
1597                                 "TERM=linux",
1598                                 "PATH=/sbin:/usr/sbin:/bin:/usr/bin",
1599                                 NULL };
1600
1601         return call_usermodehelper(critical_overtemp_path, argv, envp, 0);
1602 }
1603
1604
1605 /*
1606  * Here's the kernel thread that calls the various control loops
1607  */
1608 static int main_control_loop(void *x)
1609 {
1610         daemonize("kfand");
1611
1612         DBG("main_control_loop started\n");
1613
1614         down(&driver_lock);
1615
1616         if (start_fcu() < 0) {
1617                 printk(KERN_ERR "kfand: failed to start FCU\n");
1618                 up(&driver_lock);
1619                 goto out;
1620         }
1621
1622         /* Set the PCI fan once for now */
1623         set_pwm_fan(SLOTS_FAN_PWM_INDEX, SLOTS_FAN_DEFAULT_PWM);
1624
1625         /* Initialize ADCs */
1626         initialize_adc(&cpu_state[0]);
1627         if (cpu_state[1].monitor != NULL)
1628                 initialize_adc(&cpu_state[1]);
1629
1630         up(&driver_lock);
1631
1632         while (state == state_attached) {
1633                 unsigned long elapsed, start;
1634
1635                 start = jiffies;
1636
1637                 down(&driver_lock);
1638
1639                 /* First, we always calculate the new DIMMs state on an Xserve */
1640                 if (rackmac)
1641                         do_monitor_dimms(&dimms_state);
1642
1643                 /* Then, the CPUs */
1644                 if (cpu_pid_type == CPU_PID_TYPE_COMBINED)
1645                         do_monitor_cpu_combined();
1646                 else if (cpu_pid_type == CPU_PID_TYPE_RACKMAC) {
1647                         do_monitor_cpu_rack(&cpu_state[0]);
1648                         if (cpu_state[1].monitor != NULL)
1649                                 do_monitor_cpu_rack(&cpu_state[1]);
1650                         // better deal with UP
1651                 } else {
1652                         do_monitor_cpu_split(&cpu_state[0]);
1653                         if (cpu_state[1].monitor != NULL)
1654                                 do_monitor_cpu_split(&cpu_state[1]);
1655                         // better deal with UP
1656                 }
1657                 /* Then, the rest */
1658                 do_monitor_backside(&backside_state);
1659                 if (!rackmac)
1660                         do_monitor_drives(&drives_state);
1661                 up(&driver_lock);
1662
1663                 if (critical_state == 1) {
1664                         printk(KERN_WARNING "Temperature control detected a critical condition\n");
1665                         printk(KERN_WARNING "Attempting to shut down...\n");
1666                         if (call_critical_overtemp()) {
1667                                 printk(KERN_WARNING "Can't call %s, power off now!\n",
1668                                        critical_overtemp_path);
1669                                 machine_power_off();
1670                         }
1671                 }
1672                 if (critical_state > 0)
1673                         critical_state++;
1674                 if (critical_state > MAX_CRITICAL_STATE) {
1675                         printk(KERN_WARNING "Shutdown timed out, power off now !\n");
1676                         machine_power_off();
1677                 }
1678
1679                 // FIXME: Deal with signals
1680                 set_current_state(TASK_INTERRUPTIBLE);
1681                 elapsed = jiffies - start;
1682                 if (elapsed < HZ)
1683                         schedule_timeout(HZ - elapsed);
1684         }
1685
1686  out:
1687         DBG("main_control_loop ended\n");
1688
1689         ctrl_task = 0;
1690         complete_and_exit(&ctrl_complete, 0);
1691 }
1692
1693 /*
1694  * Dispose the control loops when tearing down
1695  */
1696 static void dispose_control_loops(void)
1697 {
1698         dispose_cpu_state(&cpu_state[0]);
1699         dispose_cpu_state(&cpu_state[1]);
1700         dispose_backside_state(&backside_state);
1701         dispose_drives_state(&drives_state);
1702         dispose_dimms_state(&dimms_state);
1703 }
1704
1705 /*
1706  * Create the control loops. U3-0 i2c bus is up, so we can now
1707  * get to the various sensors
1708  */
1709 static int create_control_loops(void)
1710 {
1711         struct device_node *np;
1712
1713         /* Count CPUs from the device-tree, we don't care how many are
1714          * actually used by Linux
1715          */
1716         cpu_count = 0;
1717         for (np = NULL; NULL != (np = of_find_node_by_type(np, "cpu"));)
1718                 cpu_count++;
1719
1720         DBG("counted %d CPUs in the device-tree\n", cpu_count);
1721
1722         /* Decide the type of PID algorithm to use based on the presence of
1723          * the pumps, though that may not be the best way, that is good enough
1724          * for now
1725          */
1726         if (rackmac)
1727                 cpu_pid_type = CPU_PID_TYPE_RACKMAC;
1728         else if (machine_is_compatible("PowerMac7,3")
1729             && (cpu_count > 1)
1730             && fcu_fans[CPUA_PUMP_RPM_INDEX].id != FCU_FAN_ABSENT_ID
1731             && fcu_fans[CPUB_PUMP_RPM_INDEX].id != FCU_FAN_ABSENT_ID) {
1732                 printk(KERN_INFO "Liquid cooling pumps detected, using new algorithm !\n");
1733                 cpu_pid_type = CPU_PID_TYPE_COMBINED;
1734         } else
1735                 cpu_pid_type = CPU_PID_TYPE_SPLIT;
1736
1737         /* Create control loops for everything. If any fail, everything
1738          * fails
1739          */
1740         if (init_cpu_state(&cpu_state[0], 0))
1741                 goto fail;
1742         if (cpu_pid_type == CPU_PID_TYPE_COMBINED)
1743                 fetch_cpu_pumps_minmax();
1744
1745         if (cpu_count > 1 && init_cpu_state(&cpu_state[1], 1))
1746                 goto fail;
1747         if (init_backside_state(&backside_state))
1748                 goto fail;
1749         if (rackmac && init_dimms_state(&dimms_state))
1750                 goto fail;
1751         if (!rackmac && init_drives_state(&drives_state))
1752                 goto fail;
1753
1754         DBG("all control loops up !\n");
1755
1756         return 0;
1757
1758  fail:
1759         DBG("failure creating control loops, disposing\n");
1760
1761         dispose_control_loops();
1762
1763         return -ENODEV;
1764 }
1765
1766 /*
1767  * Start the control loops after everything is up, that is create
1768  * the thread that will make them run
1769  */
1770 static void start_control_loops(void)
1771 {
1772         init_completion(&ctrl_complete);
1773
1774         ctrl_task = kernel_thread(main_control_loop, NULL, SIGCHLD | CLONE_KERNEL);
1775 }
1776
1777 /*
1778  * Stop the control loops when tearing down
1779  */
1780 static void stop_control_loops(void)
1781 {
1782         if (ctrl_task != 0)
1783                 wait_for_completion(&ctrl_complete);
1784 }
1785
1786 /*
1787  * Attach to the i2c FCU after detecting U3-1 bus
1788  */
1789 static int attach_fcu(void)
1790 {
1791         fcu = attach_i2c_chip(FAN_CTRLER_ID, "fcu");
1792         if (fcu == NULL)
1793                 return -ENODEV;
1794
1795         DBG("FCU attached\n");
1796
1797         return 0;
1798 }
1799
1800 /*
1801  * Detach from the i2c FCU when tearing down
1802  */
1803 static void detach_fcu(void)
1804 {
1805         if (fcu)
1806                 detach_i2c_chip(fcu);
1807         fcu = NULL;
1808 }
1809
1810 /*
1811  * Attach to the i2c controller. We probe the various chips based
1812  * on the device-tree nodes and build everything for the driver to
1813  * run, we then kick the driver monitoring thread
1814  */
1815 static int therm_pm72_attach(struct i2c_adapter *adapter)
1816 {
1817         down(&driver_lock);
1818
1819         /* Check state */
1820         if (state == state_detached)
1821                 state = state_attaching;
1822         if (state != state_attaching) {
1823                 up(&driver_lock);
1824                 return 0;
1825         }
1826
1827         /* Check if we are looking for one of these */
1828         if (u3_0 == NULL && !strcmp(adapter->name, "u3 0")) {
1829                 u3_0 = adapter;
1830                 DBG("found U3-0\n");
1831                 if (k2 || !rackmac)
1832                         if (create_control_loops())
1833                                 u3_0 = NULL;
1834         } else if (u3_1 == NULL && !strcmp(adapter->name, "u3 1")) {
1835                 u3_1 = adapter;
1836                 DBG("found U3-1, attaching FCU\n");
1837                 if (attach_fcu())
1838                         u3_1 = NULL;
1839         } else if (k2 == NULL && !strcmp(adapter->name, "mac-io 0")) {
1840                 k2 = adapter;
1841                 DBG("Found K2\n");
1842                 if (u3_0 && rackmac)
1843                         if (create_control_loops())
1844                                 k2 = NULL;
1845         }
1846         /* We got all we need, start control loops */
1847         if (u3_0 != NULL && u3_1 != NULL && (k2 || !rackmac)) {
1848                 DBG("everything up, starting control loops\n");
1849                 state = state_attached;
1850                 start_control_loops();
1851         }
1852         up(&driver_lock);
1853
1854         return 0;
1855 }
1856
1857 /*
1858  * Called on every adapter when the driver or the i2c controller
1859  * is going away.
1860  */
1861 static int therm_pm72_detach(struct i2c_adapter *adapter)
1862 {
1863         down(&driver_lock);
1864
1865         if (state != state_detached)
1866                 state = state_detaching;
1867
1868         /* Stop control loops if any */
1869         DBG("stopping control loops\n");
1870         up(&driver_lock);
1871         stop_control_loops();
1872         down(&driver_lock);
1873
1874         if (u3_0 != NULL && !strcmp(adapter->name, "u3 0")) {
1875                 DBG("lost U3-0, disposing control loops\n");
1876                 dispose_control_loops();
1877                 u3_0 = NULL;
1878         }
1879
1880         if (u3_1 != NULL && !strcmp(adapter->name, "u3 1")) {
1881                 DBG("lost U3-1, detaching FCU\n");
1882                 detach_fcu();
1883                 u3_1 = NULL;
1884         }
1885         if (u3_0 == NULL && u3_1 == NULL)
1886                 state = state_detached;
1887
1888         up(&driver_lock);
1889
1890         return 0;
1891 }
1892
1893 static int fan_check_loc_match(const char *loc, int fan)
1894 {
1895         char    tmp[64];
1896         char    *c, *e;
1897
1898         strlcpy(tmp, fcu_fans[fan].loc, 64);
1899
1900         c = tmp;
1901         for (;;) {
1902                 e = strchr(c, ',');
1903                 if (e)
1904                         *e = 0;
1905                 if (strcmp(loc, c) == 0)
1906                         return 1;
1907                 if (e == NULL)
1908                         break;
1909                 c = e + 1;
1910         }
1911         return 0;
1912 }
1913
1914 static void fcu_lookup_fans(struct device_node *fcu_node)
1915 {
1916         struct device_node *np = NULL;
1917         int i;
1918
1919         /* The table is filled by default with values that are suitable
1920          * for the old machines without device-tree informations. We scan
1921          * the device-tree and override those values with whatever is
1922          * there
1923          */
1924
1925         DBG("Looking up FCU controls in device-tree...\n");
1926
1927         while ((np = of_get_next_child(fcu_node, np)) != NULL) {
1928                 int type = -1;
1929                 char *loc;
1930                 u32 *reg;
1931
1932                 DBG(" control: %s, type: %s\n", np->name, np->type);
1933
1934                 /* Detect control type */
1935                 if (!strcmp(np->type, "fan-rpm-control") ||
1936                     !strcmp(np->type, "fan-rpm"))
1937                         type = FCU_FAN_RPM;
1938                 if (!strcmp(np->type, "fan-pwm-control") ||
1939                     !strcmp(np->type, "fan-pwm"))
1940                         type = FCU_FAN_PWM;
1941                 /* Only care about fans for now */
1942                 if (type == -1)
1943                         continue;
1944
1945                 /* Lookup for a matching location */
1946                 loc = (char *)get_property(np, "location", NULL);
1947                 reg = (u32 *)get_property(np, "reg", NULL);
1948                 if (loc == NULL || reg == NULL)
1949                         continue;
1950                 DBG(" matching location: %s, reg: 0x%08x\n", loc, *reg);
1951
1952                 for (i = 0; i < FCU_FAN_COUNT; i++) {
1953                         int fan_id;
1954
1955                         if (!fan_check_loc_match(loc, i))
1956                                 continue;
1957                         DBG(" location match, index: %d\n", i);
1958                         fcu_fans[i].id = FCU_FAN_ABSENT_ID;
1959                         if (type != fcu_fans[i].type) {
1960                                 printk(KERN_WARNING "therm_pm72: Fan type mismatch "
1961                                        "in device-tree for %s\n", np->full_name);
1962                                 break;
1963                         }
1964                         if (type == FCU_FAN_RPM)
1965                                 fan_id = ((*reg) - 0x10) / 2;
1966                         else
1967                                 fan_id = ((*reg) - 0x30) / 2;
1968                         if (fan_id > 7) {
1969                                 printk(KERN_WARNING "therm_pm72: Can't parse "
1970                                        "fan ID in device-tree for %s\n", np->full_name);
1971                                 break;
1972                         }
1973                         DBG(" fan id -> %d, type -> %d\n", fan_id, type);
1974                         fcu_fans[i].id = fan_id;
1975                 }
1976         }
1977
1978         /* Now dump the array */
1979         printk(KERN_INFO "Detected fan controls:\n");
1980         for (i = 0; i < FCU_FAN_COUNT; i++) {
1981                 if (fcu_fans[i].id == FCU_FAN_ABSENT_ID)
1982                         continue;
1983                 printk(KERN_INFO "  %d: %s fan, id %d, location: %s\n", i,
1984                        fcu_fans[i].type == FCU_FAN_RPM ? "RPM" : "PWM",
1985                        fcu_fans[i].id, fcu_fans[i].loc);
1986         }
1987 }
1988
1989 static int fcu_of_probe(struct of_device* dev, const struct of_match *match)
1990 {
1991         int rc;
1992
1993         state = state_detached;
1994
1995         /* Lookup the fans in the device tree */
1996         fcu_lookup_fans(dev->node);
1997
1998         /* Add the driver */
1999         rc = i2c_add_driver(&therm_pm72_driver);
2000         if (rc < 0)
2001                 return rc;
2002         return 0;
2003 }
2004
2005 static int fcu_of_remove(struct of_device* dev)
2006 {
2007         i2c_del_driver(&therm_pm72_driver);
2008
2009         return 0;
2010 }
2011
2012 static struct of_match fcu_of_match[] =
2013 {
2014         {
2015         .name           = OF_ANY_MATCH,
2016         .type           = "fcu",
2017         .compatible     = OF_ANY_MATCH
2018         },
2019         {},
2020 };
2021
2022 static struct of_platform_driver fcu_of_platform_driver =
2023 {
2024         .name           = "temperature",
2025         .match_table    = fcu_of_match,
2026         .probe          = fcu_of_probe,
2027         .remove         = fcu_of_remove
2028 };
2029
2030 /*
2031  * Check machine type, attach to i2c controller
2032  */
2033 static int __init therm_pm72_init(void)
2034 {
2035         struct device_node *np;
2036
2037         rackmac = machine_is_compatible("RackMac3,1");
2038
2039         if (!machine_is_compatible("PowerMac7,2") &&
2040             !machine_is_compatible("PowerMac7,3") &&
2041             !rackmac)
2042                 return -ENODEV;
2043
2044         printk(KERN_INFO "PowerMac G5 Thermal control driver %s\n", VERSION);
2045
2046         np = of_find_node_by_type(NULL, "fcu");
2047         if (np == NULL) {
2048                 /* Some machines have strangely broken device-tree */
2049                 np = of_find_node_by_path("/u3@0,f8000000/i2c@f8001000/fan@15e");
2050                 if (np == NULL) {
2051                             printk(KERN_ERR "Can't find FCU in device-tree !\n");
2052                             return -ENODEV;
2053                 }
2054         }
2055         of_dev = of_platform_device_create(np, "temperature");
2056         if (of_dev == NULL) {
2057                 printk(KERN_ERR "Can't register FCU platform device !\n");
2058                 return -ENODEV;
2059         }
2060
2061         of_register_driver(&fcu_of_platform_driver);
2062
2063         return 0;
2064 }
2065
2066 static void __exit therm_pm72_exit(void)
2067 {
2068         of_unregister_driver(&fcu_of_platform_driver);
2069
2070         if (of_dev)
2071                 of_device_unregister(of_dev);
2072 }
2073
2074 module_init(therm_pm72_init);
2075 module_exit(therm_pm72_exit);
2076
2077 MODULE_AUTHOR("Benjamin Herrenschmidt <benh@kernel.crashing.org>");
2078 MODULE_DESCRIPTION("Driver for Apple's PowerMac G5 thermal control");
2079 MODULE_LICENSE("GPL");
2080