1 diff --git a/libexec/rc/rc.conf b/libexec/rc/rc.conf
   2 index 824751078833..54bd19a7d42a 100644
   3 --- a/libexec/rc/rc.conf
   4 +++ b/libexec/rc/rc.conf
   5 @@ -718,8 +718,12 @@ devfs_set_rulesets=""	# A list of /mount/dev=ruleset_name settings to
   6  devfs_load_rulesets="YES"	# Enable to always load the default rulesets
   7  performance_cx_lowest="NONE"	# Online CPU idle state
   8  performance_cpu_freq="NONE"	# Online CPU frequency
   9 +performance_tz0_active="NONE"	# Online forced active cooling zone
  10 +performance_tz0_PSV="NONE"	# Online passive cooling threshold
  11  economy_cx_lowest="Cmax"	# Offline CPU idle state
  12  economy_cpu_freq="NONE"		# Offline CPU frequency
  13 +economy_tz0_active="NONE"	# Offline forced active cooling zone
  14 +economy_tz0_PSV="NONE"		# Offline passive cooling threshold
  15  virecover_enable="YES"	# Perform housekeeping for the vi(1) editor
  16  ugidfw_enable="NO"	# Load mac_bsdextended(4) rules on boot
  17  bsdextended_script="/etc/rc.bsdextended"	# Default mac_bsdextended(4)
  18 diff --git a/libexec/rc/rc.d/power_profile b/libexec/rc/rc.d/power_profile
  19 index 71f3f3ad8792..ec07eb920c93 100755
  20 --- a/libexec/rc/rc.d/power_profile
  21 +++ b/libexec/rc/rc.d/power_profile
  22 @@ -79,6 +79,14 @@ case ${state} in
  23  esac
  24  
  25  # Set the various sysctls based on the profile's values.
  26 +node="hw.acpi.thermal.tz0._PSV"
  27 +eval value=\$${profile}_tz0_PSV
  28 +sysctl_set
  29 +
  30 +node="hw.acpi.thermal.tz0.active"
  31 +eval value=\$${profile}_tz0_active
  32 +sysctl_set
  33 +
  34  node="hw.acpi.cpu.cx_lowest"
  35  highest_value="C1"
  36  lowest_value="Cmax"
  37 diff --git a/share/man/man4/acpi_thermal.4 b/share/man/man4/acpi_thermal.4
  38 index 3961ed36276c..256aa978b486 100644
  39 --- a/share/man/man4/acpi_thermal.4
  40 +++ b/share/man/man4/acpi_thermal.4
  41 @@ -22,7 +22,7 @@
  42  .\" OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  43  .\" SUCH DAMAGE.
  44  .\"
  45 -.Dd March 17, 2007
  46 +.Dd July 31, 2022
  47  .Dt ACPI_THERMAL 4
  48  .Os
  49  .Sh NAME
  50 @@ -82,6 +82,9 @@ Current temperature for this zone.
  51  .It Va hw.acpi.thermal.tz%d._PSV
  52  Temperature to start passive cooling by throttling down CPU, etc.
  53  This value can be overridden by the user.
  54 +.It Va hw.acpi.thermal.tz%d._CR3
  55 +Temperature to start critical suspend to RAM (S3).
  56 +This value can be overridden by the user.
  57  .It Va hw.acpi.thermal.tz%d._HOT
  58  Temperature to start critical suspend to disk (S4).
  59  This value can be overridden by the user.
  60 diff --git a/sys/dev/acpica/acpi_thermal.c b/sys/dev/acpica/acpi_thermal.c
  61 index c436390d84b4..fb7a478fd479 100644
  62 --- a/sys/dev/acpica/acpi_thermal.c
  63 +++ b/sys/dev/acpica/acpi_thermal.c
  64 @@ -58,7 +58,7 @@ ACPI_MODULE_NAME("THERMAL")
  65  #define TZ_NOTIFY_TEMPERATURE	0x80 /* Temperature changed. */
  66  #define TZ_NOTIFY_LEVELS	0x81 /* Cooling levels changed. */
  67  #define TZ_NOTIFY_DEVICES	0x82 /* Device lists changed. */
  68 -#define TZ_NOTIFY_CRITICAL	0xcc /* Fake notify that _CRT/_HOT reached. */
  69 +#define TZ_NOTIFY_CRITICAL	0xcc /* Fake notify that _CRT/_HOT/_CR3 reached. */
  70  
  71  /* Check for temperature changes every 10 seconds by default */
  72  #define TZ_POLLRATE	10
  73 @@ -76,6 +76,7 @@ struct acpi_tz_zone {
  74      ACPI_BUFFER	al[TZ_NUMLEVELS];
  75      int		crt;
  76      int		hot;
  77 +    int		cr3;
  78      ACPI_BUFFER	psl;
  79      int		psv;
  80      int		tc1;
  81 @@ -95,8 +96,9 @@ struct acpi_tz_softc {
  82      int				tz_thflags;	/*Current temp-related flags*/
  83  #define TZ_THFLAG_NONE		0
  84  #define TZ_THFLAG_PSV		(1<<0)
  85 -#define TZ_THFLAG_HOT		(1<<2)
  86 -#define TZ_THFLAG_CRT		(1<<3)
  87 +#define TZ_THFLAG_CR3		(1<<2)
  88 +#define TZ_THFLAG_HOT		(1<<3)
  89 +#define TZ_THFLAG_CRT		(1<<4)
  90      int				tz_flags;
  91  #define TZ_FLAG_NO_SCP		(1<<0)		/*No _SCP method*/
  92  #define TZ_FLAG_GETPROFILE	(1<<1)		/*Get power_profile in timeout*/
  93 @@ -280,6 +282,10 @@ acpi_tz_attach(device_t dev)
  94          OID_AUTO, "_PSV", CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, sc,
  95  	offsetof(struct acpi_tz_softc, tz_zone.psv), acpi_tz_temp_sysctl, "IK",
  96  	"passive cooling temp setpoint");
  97 +    SYSCTL_ADD_PROC(&sc->tz_sysctl_ctx, SYSCTL_CHILDREN(sc->tz_sysctl_tree),
  98 +        OID_AUTO, "_CR3", CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, sc,
  99 +	offsetof(struct acpi_tz_softc, tz_zone.cr3), acpi_tz_temp_sysctl, "IK",
 100 +	"too warm temp setpoint (standby now)");
 101      SYSCTL_ADD_PROC(&sc->tz_sysctl_ctx, SYSCTL_CHILDREN(sc->tz_sysctl_tree),
 102          OID_AUTO, "_HOT", CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, sc,
 103  	offsetof(struct acpi_tz_softc, tz_zone.hot), acpi_tz_temp_sysctl, "IK",
 104 @@ -419,6 +425,7 @@ acpi_tz_establish(struct acpi_tz_softc *sc)
 105      }
 106      acpi_tz_getparam(sc, "_CRT", &sc->tz_zone.crt);
 107      acpi_tz_getparam(sc, "_HOT", &sc->tz_zone.hot);
 108 +    acpi_tz_getparam(sc, "_CR3", &sc->tz_zone.cr3);
 109      sc->tz_zone.psl.Length = ACPI_ALLOCATE_BUFFER;
 110      sc->tz_zone.psl.Pointer = NULL;
 111      AcpiEvaluateObject(sc->tz_handle, "_PSL", NULL, &sc->tz_zone.psl);
 112 @@ -436,6 +443,7 @@ acpi_tz_establish(struct acpi_tz_softc *sc)
 113       */
 114      acpi_tz_sanity(sc, &sc->tz_zone.crt, "_CRT");
 115      acpi_tz_sanity(sc, &sc->tz_zone.hot, "_HOT");
 116 +    acpi_tz_sanity(sc, &sc->tz_zone.cr3, "_CR3");
 117      acpi_tz_sanity(sc, &sc->tz_zone.psv, "_PSV");
 118      for (i = 0; i < TZ_NUMLEVELS; i++)
 119  	acpi_tz_sanity(sc, &sc->tz_zone.ac[i], "_ACx");
 120 @@ -493,6 +501,7 @@ acpi_tz_get_temperature(struct acpi_tz_softc *sc)
 121  static void
 122  acpi_tz_monitor(void *Context)
 123  {
 124 +    struct acpi_softc	 *acpi_sc;
 125      struct acpi_tz_softc *sc;
 126      struct	timespec curtime;
 127      int		temp;
 128 @@ -543,6 +552,8 @@ acpi_tz_monitor(void *Context)
 129      newflags = TZ_THFLAG_NONE;
 130      if (sc->tz_zone.psv != -1 && temp >= sc->tz_zone.psv)
 131  	newflags |= TZ_THFLAG_PSV;
 132 +    if (sc->tz_zone.cr3 != -1 && temp >= sc->tz_zone.cr3)
 133 +	newflags |= TZ_THFLAG_CR3;
 134      if (sc->tz_zone.hot != -1 && temp >= sc->tz_zone.hot)
 135  	newflags |= TZ_THFLAG_HOT;
 136      if (sc->tz_zone.crt != -1 && temp >= sc->tz_zone.crt)
 137 @@ -593,7 +604,7 @@ acpi_tz_monitor(void *Context)
 138      /* XXX (de)activate any passive cooling that may be required. */
 139  
 140      /*
 141 -     * If the temperature is at _HOT or _CRT, increment our event count.
 142 +     * If the temperature is at _CR3, _HOT or _CRT, increment our event count.
 143       * If it has occurred enough times, shutdown the system.  This is
 144       * needed because some systems will report an invalid high temperature
 145       * for one poll cycle.  It is suspected this is due to the embedded
 146 @@ -602,13 +613,18 @@ acpi_tz_monitor(void *Context)
 147       *
 148       * If we're almost at that threshold, notify the user through devd(8).
 149       */
 150 -    if ((newflags & (TZ_THFLAG_HOT | TZ_THFLAG_CRT)) != 0) {
 151 +    if ((newflags & (TZ_THFLAG_CR3 | TZ_THFLAG_HOT | TZ_THFLAG_CRT)) != 0) {
 152  	sc->tz_validchecks++;
 153  	if (sc->tz_validchecks == TZ_VALIDCHECKS) {
 154  	    device_printf(sc->tz_dev,
 155  		"WARNING - current temperature (%d.%dC) exceeds safe limits\n",
 156  		TZ_KELVTOC(sc->tz_temperature));
 157 -	    shutdown_nice(RB_POWEROFF);
 158 +	    if ((newflags & (TZ_THFLAG_HOT | TZ_THFLAG_CRT)) != 0)
 159 +		shutdown_nice(RB_POWEROFF);
 160 +	    else {
 161 +		acpi_sc = acpi_device_get_parent_softc(sc->tz_dev);
 162 +		acpi_ReqSleepState(acpi_sc, ACPI_STATE_S3);
 163 +	    }
 164  	} else if (sc->tz_validchecks == TZ_NOTIFYCOUNT)
 165  	    acpi_UserNotify("Thermal", sc->tz_handle, TZ_NOTIFY_CRITICAL);
 166      } else {
 167 @@ -1004,51 +1020,68 @@ acpi_tz_thread(void *arg)
 168  static int
 169  acpi_tz_cpufreq_restore(struct acpi_tz_softc *sc)
 170  {
 171 -    device_t dev;
 172 -    int error;
 173 +    device_t *devs;
 174 +    int devcount, error, n;
 175  
 176      if (!sc->tz_cooling_updated)
 177  	return (0);
 178 -    if ((dev = devclass_get_device(devclass_find("cpufreq"), 0)) == NULL)
 179 +
 180 +    if ((error = devclass_get_devices(devclass_find("cpufreq"), &devs, &devcount))) {
 181 +	free(devs, M_TEMP);
 182  	return (ENXIO);
 183 +    }
 184 +
 185      ACPI_VPRINT(sc->tz_dev, acpi_device_get_parent_softc(sc->tz_dev),
 186  	"temperature %d.%dC: resuming previous clock speed (%d MHz)\n",
 187  	TZ_KELVTOC(sc->tz_temperature), sc->tz_cooling_saved_freq);
 188 -    error = CPUFREQ_SET(dev, NULL, CPUFREQ_PRIO_KERN);
 189 +    for (n = 0; n < devcount; n++) {
 190 +	if ((error = CPUFREQ_SET(devs[n], NULL, CPUFREQ_PRIO_KERN))) {
 191 +	    /*
 192 +	     * If you find yourself here:  cf_set_method() in kern_cpu.c
 193 +	     * requires CPUFREQ_GET() be run before CPUFREQ_SET(), otherwise on
 194 +	     * a restore sc->curr_level.total_set.freq = CPUFREQ_VAL_UNKNOWN,
 195 +	     * and then if it hasn't been updated the next save/restore will fail.
 196 +	     */
 197 +	    device_printf(sc->tz_dev, "cpu%d: failed to restore cpufreq priority\n", n);
 198 +	    break;
 199 +	}
 200 +    }
 201      if (error == 0)
 202  	sc->tz_cooling_updated = FALSE;
 203 +
 204 +    free(devs, M_TEMP);
 205      return (error);
 206  }
 207  
 208  static int
 209  acpi_tz_cpufreq_update(struct acpi_tz_softc *sc, int req)
 210  {
 211 -    device_t dev;
 212 +    device_t *devs;
 213      struct cf_level *levels;
 214 -    int num_levels, error, freq, desired_freq, perf, i;
 215 +    int num_levels, error, freq, desired_freq, perf, devcount, i, n;
 216  
 217      levels = malloc(CPUFREQ_MAX_LEVELS * sizeof(*levels), M_TEMP, M_NOWAIT);
 218      if (levels == NULL)
 219  	return (ENOMEM);
 220  
 221      /*
 222 -     * Find the main device, cpufreq0.  We don't yet support independent
 223 -     * CPU frequency control on SMP.
 224 +     * Find the main device, cpufreq0.  We make plans based upon cpu0, and
 225 +     * change all the cpu frequencies to the same level.
 226       */
 227 -    if ((dev = devclass_get_device(devclass_find("cpufreq"), 0)) == NULL) {
 228 -	error = ENXIO;
 229 +    if ((error = devclass_get_devices(devclass_find("cpufreq"), &devs, &devcount))) {
 230 +	free(devs, M_TEMP);
 231  	goto out;
 232      }
 233  
 234 -    /* Get the current frequency. */
 235 -    error = CPUFREQ_GET(dev, &levels[0]);
 236 +    /* Get the current frequency for cpu0. */
 237 +    error = CPUFREQ_GET(devs[0], &levels[0]);
 238      if (error)
 239  	goto out;
 240      freq = levels[0].total_set.freq;
 241  
 242 -    /* Get the current available frequency levels. */
 243 +    /* Get the current available frequency levels for cpu0. */
 244      num_levels = CPUFREQ_MAX_LEVELS;
 245 -    error = CPUFREQ_LEVELS(dev, levels, &num_levels);
 246 +    error = CPUFREQ_LEVELS(devs[0], levels, &num_levels);
 247      if (error) {
 248  	if (error == E2BIG)
 249  	    printf("cpufreq: need to increase CPUFREQ_MAX_LEVELS\n");
 250 @@ -1109,8 +1142,22 @@ acpi_tz_cpufreq_update(struct acpi_tz_softc *sc, int req)
 251  	    TZ_KELVTOC(sc->tz_temperature),
 252  	    (freq > levels[i].total_set.freq) ? "de" : "in",
 253  	    freq, levels[i].total_set.freq);
 254 -	error = CPUFREQ_SET(dev, &levels[i], CPUFREQ_PRIO_KERN);
 255 -	if (error == 0 && !sc->tz_cooling_updated) {
 256 +
 257 +	/* Change all the cpus to the chosen frequency. */
 258 +	for (n = 0; n < devcount; n++) {
 259 +	    /* CPUFREQ_SET() fails occasionally without this. */
 260 +	    if ((error = CPUFREQ_GET(devs[n], &levels[0])))
 261 +		goto out;
 262 +
 263 +	    if ((error = CPUFREQ_LEVELS(devs[n], levels, &num_levels))) {
 264 +		if (error == E2BIG)
 265 +		    printf("cpufreq: need to increase CPUFREQ_MAX_LEVELS\n");
 266 +		goto out;
 267 +	    }
 268 +	    if ((error = CPUFREQ_SET(devs[n], &levels[i], CPUFREQ_PRIO_KERN)))
 269 +		goto out;
 270 +	}
 271 +	if (!sc->tz_cooling_updated) {
 272  	    sc->tz_cooling_saved_freq = freq;
 273  	    sc->tz_cooling_updated = TRUE;
 274  	}
 275 @@ -1119,6 +1166,7 @@ acpi_tz_cpufreq_update(struct acpi_tz_softc *sc, int req)
 276  out:
 277      if (levels)
 278  	free(levels, M_TEMP);
 279 +    free(devs, M_TEMP);
 280      return (error);
 281  }
 282