drivers/thermal/intel/x86_pkg_temp_thermal.c

   1 /*
   2  * x86_pkg_temp_thermal driver
   3  * Copyright (c) 2013, Intel Corporation.
   4  *
   5  * This program is free software; you can redistribute it and/or modify it
   6  * under the terms and conditions of the GNU General Public License,
   7  * version 2, as published by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  12  * more details.
  13  *
  14  * You should have received a copy of the GNU General Public License along with
  15  * this program; if not, write to the Free Software Foundation, Inc.
  16  *
  17  */
  18 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  19
  20 #include <linux/module.h>
  21 #include <linux/init.h>
  22 #include <linux/err.h>
  23 #include <linux/param.h>
  24 #include <linux/device.h>
  25 #include <linux/platform_device.h>
  26 #include <linux/cpu.h>
  27 #include <linux/smp.h>
  28 #include <linux/slab.h>
  29 #include <linux/pm.h>
  30 #include <linux/thermal.h>
  31 #include <linux/debugfs.h>
  32 #include <asm/cpu_device_id.h>
  33 #include <asm/mce.h>
  34
  35 /*
  36 * Rate control delay: Idea is to introduce denounce effect
  37 * This should be long enough to avoid reduce events, when
  38 * threshold is set to a temperature, which is constantly
  39 * violated, but at the short enough to take any action.
  40 * The action can be remove threshold or change it to next
  41 * interesting setting. Based on experiments, in around
  42 * every 5 seconds under load will give us a significant
  43 * temperature change.
  44 */
  45 #define PKG_TEMP_THERMAL_NOTIFY_DELAY   5000
  46 static int notify_delay_ms = PKG_TEMP_THERMAL_NOTIFY_DELAY;
  47 module_param(notify_delay_ms, int, 0644);
  48 MODULE_PARM_DESC(notify_delay_ms,
  49         "User space notification delay in milli seconds.");
  50
  51 /* Number of trip points in thermal zone. Currently it can't
  52 * be more than 2. MSR can allow setting and getting notifications
  53 * for only 2 thresholds. This define enforces this, if there
  54 * is some wrong values returned by cpuid for number of thresholds.
  55 */
  56 #define MAX_NUMBER_OF_TRIPS     2
  57
  58 struct pkg_device {
  59         int                             cpu;
  60         bool                            work_scheduled;
  61         u32                             tj_max;
  62         u32                             msr_pkg_therm_low;
  63         u32                             msr_pkg_therm_high;
  64         struct delayed_work             work;
  65         struct thermal_zone_device      *tzone;
  66         struct cpumask                  cpumask;
  67 };
  68
  69 static struct thermal_zone_params pkg_temp_tz_params = {
  70         .no_hwmon       = true,
  71 };
  72
  73 /* Keep track of how many package pointers we allocated in init() */
  74 static int max_packages __read_mostly;
  75 /* Array of package pointers */
  76 static struct pkg_device **packages;
  77 /* Serializes interrupt notification, work and hotplug */
  78 static DEFINE_SPINLOCK(pkg_temp_lock);
  79 /* Protects zone operation in the work function against hotplug removal */
  80 static DEFINE_MUTEX(thermal_zone_mutex);
  81
  82 /* The dynamically assigned cpu hotplug state for module_exit() */
  83 static enum cpuhp_state pkg_thermal_hp_state __read_mostly;
  84
  85 /* Debug counters to show using debugfs */
  86 static struct dentry *debugfs;
  87 static unsigned int pkg_interrupt_cnt;
  88 static unsigned int pkg_work_cnt;
  89
  90 static void pkg_temp_debugfs_init(void)
  91 {
  92         debugfs = debugfs_create_dir("pkg_temp_thermal", NULL);
  93
  94         debugfs_create_u32("pkg_thres_interrupt", S_IRUGO, debugfs,
  95                            &pkg_interrupt_cnt);
  96         debugfs_create_u32("pkg_thres_work", S_IRUGO, debugfs,
  97                            &pkg_work_cnt);
  98 }
  99
 100 /*
 101  * Protection:
 102  *
 103  * - cpu hotplug: Read serialized by cpu hotplug lock
 104  *                Write must hold pkg_temp_lock
 105  *
 106  * - Other callsites: Must hold pkg_temp_lock
 107  */
 108 static struct pkg_device *pkg_temp_thermal_get_dev(unsigned int cpu)
 109 {
 110         int pkgid = topology_logical_package_id(cpu);
 111
 112         if (pkgid >= 0 && pkgid < max_packages)
 113                 return packages[pkgid];
 114         return NULL;
 115 }
 116
 117 /*
 118 * tj-max is is interesting because threshold is set relative to this
 119 * temperature.
 120 */
 121 static int get_tj_max(int cpu, u32 *tj_max)
 122 {
 123         u32 eax, edx, val;
 124         int err;
 125
 126         err = rdmsr_safe_on_cpu(cpu, MSR_IA32_TEMPERATURE_TARGET, &eax, &edx);
 127         if (err)
 128                 return err;
 129
 130         val = (eax >> 16) & 0xff;
 131         *tj_max = val * 1000;
 132
 133         return val ? 0 : -EINVAL;
 134 }
 135
 136 static int sys_get_curr_temp(struct thermal_zone_device *tzd, int *temp)
 137 {
 138         struct pkg_device *pkgdev = tzd->devdata;
 139         u32 eax, edx;
 140
 141         rdmsr_on_cpu(pkgdev->cpu, MSR_IA32_PACKAGE_THERM_STATUS, &eax, &edx);
 142         if (eax & 0x80000000) {
 143                 *temp = pkgdev->tj_max - ((eax >> 16) & 0x7f) * 1000;
 144                 pr_debug("sys_get_curr_temp %d\n", *temp);
 145                 return 0;
 146         }
 147         return -EINVAL;
 148 }
 149
 150 static int sys_get_trip_temp(struct thermal_zone_device *tzd,
 151                              int trip, int *temp)
 152 {
 153         struct pkg_device *pkgdev = tzd->devdata;
 154         unsigned long thres_reg_value;
 155         u32 mask, shift, eax, edx;
 156         int ret;
 157
 158         if (trip >= MAX_NUMBER_OF_TRIPS)
 159                 return -EINVAL;
 160
 161         if (trip) {
 162                 mask = THERM_MASK_THRESHOLD1;
 163                 shift = THERM_SHIFT_THRESHOLD1;
 164         } else {
 165                 mask = THERM_MASK_THRESHOLD0;
 166                 shift = THERM_SHIFT_THRESHOLD0;
 167         }
 168
 169         ret = rdmsr_on_cpu(pkgdev->cpu, MSR_IA32_PACKAGE_THERM_INTERRUPT,
 170                            &eax, &edx);
 171         if (ret < 0)
 172                 return ret;
 173
 174         thres_reg_value = (eax & mask) >> shift;
 175         if (thres_reg_value)
 176                 *temp = pkgdev->tj_max - thres_reg_value * 1000;
 177         else
 178                 *temp = 0;
 179         pr_debug("sys_get_trip_temp %d\n", *temp);
 180
 181         return 0;
 182 }
 183
 184 static int
 185 sys_set_trip_temp(struct thermal_zone_device *tzd, int trip, int temp)
 186 {
 187         struct pkg_device *pkgdev = tzd->devdata;
 188         u32 l, h, mask, shift, intr;
 189         int ret;
 190
 191         if (trip >= MAX_NUMBER_OF_TRIPS || temp >= pkgdev->tj_max)
 192                 return -EINVAL;
 193
 194         ret = rdmsr_on_cpu(pkgdev->cpu, MSR_IA32_PACKAGE_THERM_INTERRUPT,
 195                            &l, &h);
 196         if (ret < 0)
 197                 return ret;
 198
 199         if (trip) {
 200                 mask = THERM_MASK_THRESHOLD1;
 201                 shift = THERM_SHIFT_THRESHOLD1;
 202                 intr = THERM_INT_THRESHOLD1_ENABLE;
 203         } else {
 204                 mask = THERM_MASK_THRESHOLD0;
 205                 shift = THERM_SHIFT_THRESHOLD0;
 206                 intr = THERM_INT_THRESHOLD0_ENABLE;
 207         }
 208         l &= ~mask;
 209         /*
 210         * When users space sets a trip temperature == 0, which is indication
 211         * that, it is no longer interested in receiving notifications.
 212         */
 213         if (!temp) {
 214                 l &= ~intr;
 215         } else {
 216                 l |= (pkgdev->tj_max - temp)/1000 << shift;
 217                 l |= intr;
 218         }
 219
 220         return wrmsr_on_cpu(pkgdev->cpu, MSR_IA32_PACKAGE_THERM_INTERRUPT, l, h);
 221 }
 222
 223 static int sys_get_trip_type(struct thermal_zone_device *thermal, int trip,
 224                              enum thermal_trip_type *type)
 225 {
 226         *type = THERMAL_TRIP_PASSIVE;
 227         return 0;
 228 }
 229
 230 /* Thermal zone callback registry */
 231 static struct thermal_zone_device_ops tzone_ops = {
 232         .get_temp = sys_get_curr_temp,
 233         .get_trip_temp = sys_get_trip_temp,
 234         .get_trip_type = sys_get_trip_type,
 235         .set_trip_temp = sys_set_trip_temp,
 236 };
 237
 238 static bool pkg_thermal_rate_control(void)
 239 {
 240         return true;
 241 }
 242
 243 /* Enable threshold interrupt on local package/cpu */
 244 static inline void enable_pkg_thres_interrupt(void)
 245 {
 246         u8 thres_0, thres_1;
 247         u32 l, h;
 248
 249         rdmsr(MSR_IA32_PACKAGE_THERM_INTERRUPT, l, h);
 250         /* only enable/disable if it had valid threshold value */
 251         thres_0 = (l & THERM_MASK_THRESHOLD0) >> THERM_SHIFT_THRESHOLD0;
 252         thres_1 = (l & THERM_MASK_THRESHOLD1) >> THERM_SHIFT_THRESHOLD1;
 253         if (thres_0)
 254                 l |= THERM_INT_THRESHOLD0_ENABLE;
 255         if (thres_1)
 256                 l |= THERM_INT_THRESHOLD1_ENABLE;
 257         wrmsr(MSR_IA32_PACKAGE_THERM_INTERRUPT, l, h);
 258 }
 259
 260 /* Disable threshold interrupt on local package/cpu */
 261 static inline void disable_pkg_thres_interrupt(void)
 262 {
 263         u32 l, h;
 264
 265         rdmsr(MSR_IA32_PACKAGE_THERM_INTERRUPT, l, h);
 266
 267         l &= ~(THERM_INT_THRESHOLD0_ENABLE | THERM_INT_THRESHOLD1_ENABLE);
 268         wrmsr(MSR_IA32_PACKAGE_THERM_INTERRUPT, l, h);
 269 }
 270
 271 static void pkg_temp_thermal_threshold_work_fn(struct work_struct *work)
 272 {
 273         struct thermal_zone_device *tzone = NULL;
 274         int cpu = smp_processor_id();
 275         struct pkg_device *pkgdev;
 276         u64 msr_val, wr_val;
 277
 278         mutex_lock(&thermal_zone_mutex);
 279         spin_lock_irq(&pkg_temp_lock);
 280         ++pkg_work_cnt;
 281
 282         pkgdev = pkg_temp_thermal_get_dev(cpu);
 283         if (!pkgdev) {
 284                 spin_unlock_irq(&pkg_temp_lock);
 285                 mutex_unlock(&thermal_zone_mutex);
 286                 return;
 287         }
 288         pkgdev->work_scheduled = false;
 289
 290         rdmsrl(MSR_IA32_PACKAGE_THERM_STATUS, msr_val);
 291         wr_val = msr_val & ~(THERM_LOG_THRESHOLD0 | THERM_LOG_THRESHOLD1);
 292         if (wr_val != msr_val) {
 293                 wrmsrl(MSR_IA32_PACKAGE_THERM_STATUS, wr_val);
 294                 tzone = pkgdev->tzone;
 295         }
 296
 297         enable_pkg_thres_interrupt();
 298         spin_unlock_irq(&pkg_temp_lock);
 299
 300         /*
 301          * If tzone is not NULL, then thermal_zone_mutex will prevent the
 302          * concurrent removal in the cpu offline callback.
 303          */
 304         if (tzone)
 305                 thermal_zone_device_update(tzone, THERMAL_EVENT_UNSPECIFIED);
 306
 307         mutex_unlock(&thermal_zone_mutex);
 308 }
 309
 310 static void pkg_thermal_schedule_work(int cpu, struct delayed_work *work)
 311 {
 312         unsigned long ms = msecs_to_jiffies(notify_delay_ms);
 313
 314         schedule_delayed_work_on(cpu, work, ms);
 315 }
 316
 317 static int pkg_thermal_notify(u64 msr_val)
 318 {
 319         int cpu = smp_processor_id();
 320         struct pkg_device *pkgdev;
 321         unsigned long flags;
 322
 323         spin_lock_irqsave(&pkg_temp_lock, flags);
 324         ++pkg_interrupt_cnt;
 325
 326         disable_pkg_thres_interrupt();
 327
 328         /* Work is per package, so scheduling it once is enough. */
 329         pkgdev = pkg_temp_thermal_get_dev(cpu);
 330         if (pkgdev && !pkgdev->work_scheduled) {
 331                 pkgdev->work_scheduled = true;
 332                 pkg_thermal_schedule_work(pkgdev->cpu, &pkgdev->work);
 333         }
 334
 335         spin_unlock_irqrestore(&pkg_temp_lock, flags);
 336         return 0;
 337 }
 338
 339 static int pkg_temp_thermal_device_add(unsigned int cpu)
 340 {
 341         int pkgid = topology_logical_package_id(cpu);
 342         u32 tj_max, eax, ebx, ecx, edx;
 343         struct pkg_device *pkgdev;
 344         int thres_count, err;
 345
 346         if (pkgid >= max_packages)
 347                 return -ENOMEM;
 348
 349         cpuid(6, &eax, &ebx, &ecx, &edx);
 350         thres_count = ebx & 0x07;
 351         if (!thres_count)
 352                 return -ENODEV;
 353
 354         thres_count = clamp_val(thres_count, 0, MAX_NUMBER_OF_TRIPS);
 355
 356         err = get_tj_max(cpu, &tj_max);
 357         if (err)
 358                 return err;
 359
 360         pkgdev = kzalloc(sizeof(*pkgdev), GFP_KERNEL);
 361         if (!pkgdev)
 362                 return -ENOMEM;
 363
 364         INIT_DELAYED_WORK(&pkgdev->work, pkg_temp_thermal_threshold_work_fn);
 365         pkgdev->cpu = cpu;
 366         pkgdev->tj_max = tj_max;
 367         pkgdev->tzone = thermal_zone_device_register("x86_pkg_temp",
 368                         thres_count,
 369                         (thres_count == MAX_NUMBER_OF_TRIPS) ? 0x03 : 0x01,
 370                         pkgdev, &tzone_ops, &pkg_temp_tz_params, 0, 0);
 371         if (IS_ERR(pkgdev->tzone)) {
 372                 err = PTR_ERR(pkgdev->tzone);
 373                 kfree(pkgdev);
 374                 return err;
 375         }
 376         /* Store MSR value for package thermal interrupt, to restore at exit */
 377         rdmsr(MSR_IA32_PACKAGE_THERM_INTERRUPT, pkgdev->msr_pkg_therm_low,
 378               pkgdev->msr_pkg_therm_high);
 379
 380         cpumask_set_cpu(cpu, &pkgdev->cpumask);
 381         spin_lock_irq(&pkg_temp_lock);
 382         packages[pkgid] = pkgdev;
 383         spin_unlock_irq(&pkg_temp_lock);
 384         return 0;
 385 }
 386
 387 static int pkg_thermal_cpu_offline(unsigned int cpu)
 388 {
 389         struct pkg_device *pkgdev = pkg_temp_thermal_get_dev(cpu);
 390         bool lastcpu, was_target;
 391         int target;
 392
 393         if (!pkgdev)
 394                 return 0;
 395
 396         target = cpumask_any_but(&pkgdev->cpumask, cpu);
 397         cpumask_clear_cpu(cpu, &pkgdev->cpumask);
 398         lastcpu = target >= nr_cpu_ids;
 399         /*
 400          * Remove the sysfs files, if this is the last cpu in the package
 401          * before doing further cleanups.
 402          */
 403         if (lastcpu) {
 404                 struct thermal_zone_device *tzone = pkgdev->tzone;
 405
 406                 /*
 407                  * We must protect against a work function calling
 408                  * thermal_zone_update, after/while unregister. We null out
 409                  * the pointer under the zone mutex, so the worker function
 410                  * won't try to call.
 411                  */
 412                 mutex_lock(&thermal_zone_mutex);
 413                 pkgdev->tzone = NULL;
 414                 mutex_unlock(&thermal_zone_mutex);
 415
 416                 thermal_zone_device_unregister(tzone);
 417         }
 418
 419         /* Protect against work and interrupts */
 420         spin_lock_irq(&pkg_temp_lock);
 421
 422         /*
 423          * Check whether this cpu was the current target and store the new
 424          * one. When we drop the lock, then the interrupt notify function
 425          * will see the new target.
 426          */
 427         was_target = pkgdev->cpu == cpu;
 428         pkgdev->cpu = target;
 429
 430         /*
 431          * If this is the last CPU in the package remove the package
 432          * reference from the array and restore the interrupt MSR. When we
 433          * drop the lock neither the interrupt notify function nor the
 434          * worker will see the package anymore.
 435          */
 436         if (lastcpu) {
 437                 packages[topology_logical_package_id(cpu)] = NULL;
 438                 /* After this point nothing touches the MSR anymore. */
 439                 wrmsr(MSR_IA32_PACKAGE_THERM_INTERRUPT,
 440                       pkgdev->msr_pkg_therm_low, pkgdev->msr_pkg_therm_high);
 441         }
 442
 443         /*
 444          * Check whether there is work scheduled and whether the work is
 445          * targeted at the outgoing CPU.
 446          */
 447         if (pkgdev->work_scheduled && was_target) {
 448                 /*
 449                  * To cancel the work we need to drop the lock, otherwise
 450                  * we might deadlock if the work needs to be flushed.
 451                  */
 452                 spin_unlock_irq(&pkg_temp_lock);
 453                 cancel_delayed_work_sync(&pkgdev->work);
 454                 spin_lock_irq(&pkg_temp_lock);
 455                 /*
 456                  * If this is not the last cpu in the package and the work
 457                  * did not run after we dropped the lock above, then we
 458                  * need to reschedule the work, otherwise the interrupt
 459                  * stays disabled forever.
 460                  */
 461                 if (!lastcpu && pkgdev->work_scheduled)
 462                         pkg_thermal_schedule_work(target, &pkgdev->work);
 463         }
 464
 465         spin_unlock_irq(&pkg_temp_lock);
 466
 467         /* Final cleanup if this is the last cpu */
 468         if (lastcpu)
 469                 kfree(pkgdev);
 470         return 0;
 471 }
 472
 473 static int pkg_thermal_cpu_online(unsigned int cpu)
 474 {
 475         struct pkg_device *pkgdev = pkg_temp_thermal_get_dev(cpu);
 476         struct cpuinfo_x86 *c = &cpu_data(cpu);
 477
 478         /* Paranoia check */
 479         if (!cpu_has(c, X86_FEATURE_DTHERM) || !cpu_has(c, X86_FEATURE_PTS))
 480                 return -ENODEV;
 481
 482         /* If the package exists, nothing to do */
 483         if (pkgdev) {
 484                 cpumask_set_cpu(cpu, &pkgdev->cpumask);
 485                 return 0;
 486         }
 487         return pkg_temp_thermal_device_add(cpu);
 488 }
 489
 490 static const struct x86_cpu_id __initconst pkg_temp_thermal_ids[] = {
 491         { X86_VENDOR_INTEL, X86_FAMILY_ANY, X86_MODEL_ANY, X86_FEATURE_PTS },
 492         {}
 493 };
 494 MODULE_DEVICE_TABLE(x86cpu, pkg_temp_thermal_ids);
 495
 496 static int __init pkg_temp_thermal_init(void)
 497 {
 498         int ret;
 499
 500         if (!x86_match_cpu(pkg_temp_thermal_ids))
 501                 return -ENODEV;
 502
 503         max_packages = topology_max_packages();
 504         packages = kcalloc(max_packages, sizeof(struct pkg_device *),
 505                            GFP_KERNEL);
 506         if (!packages)
 507                 return -ENOMEM;
 508
 509         ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "thermal/x86_pkg:online",
 510                                 pkg_thermal_cpu_online, pkg_thermal_cpu_offline);
 511         if (ret < 0)
 512                 goto err;
 513
 514         /* Store the state for module exit */
 515         pkg_thermal_hp_state = ret;
 516
 517         platform_thermal_package_notify = pkg_thermal_notify;
 518         platform_thermal_package_rate_control = pkg_thermal_rate_control;
 519
 520          /* Don't care if it fails */
 521         pkg_temp_debugfs_init();
 522         return 0;
 523
 524 err:
 525         kfree(packages);
 526         return ret;
 527 }
 528 module_init(pkg_temp_thermal_init)
 529
 530 static void __exit pkg_temp_thermal_exit(void)
 531 {
 532         platform_thermal_package_notify = NULL;
 533         platform_thermal_package_rate_control = NULL;
 534
 535         cpuhp_remove_state(pkg_thermal_hp_state);
 536         debugfs_remove_recursive(debugfs);
 537         kfree(packages);
 538 }
 539 module_exit(pkg_temp_thermal_exit)
 540
 541 MODULE_DESCRIPTION("X86 PKG TEMP Thermal Driver");
 542 MODULE_AUTHOR("Srinivas Pandruvada <srinivas.pandruvada@linux.intel.com>");
 543 MODULE_LICENSE("GPL v2");