Slashing Watts on My Proxmox Homeserver

# set hdd spindown time
hdparm -B 127 /dev/sda
hdparm -S 120 /dev/sda
hdparm -B 127 /dev/sdb
hdparm -S 120 /dev/sdb
hdparm -B 127 /dev/sdc
hdparm -S 120 /dev/sdc
# Enable EEE on my LAN
# careful with this you may lose connection to your server
ethtool --set-eee enp12s0 eee on
ethtool -C enp12s0 rx-usecs 125
# setting govenour
apt install linux-cpupower tuned
cpupower frequency-set -g powersave
# no periodic wakeup by watchdog
sysfs_set /proc/sys/kernel/nmi_watchdog 0
sysfs_set /proc/sys/vm/dirty_writeback_centisecs 1500

#!/bin/bash
# pve-powersave.sh
#
# Re-applies my idle power settings on the Proxmox box
# (Ryzen 7 5825U, PVE 9 / Debian 13). Everything here is a runtime
# sysfs/ethtool/hdparm setting, nothing survives a reboot, so run it
# at boot (systemd oneshot or cron @reboot) and re-run whenever you like.
#
#   pve-powersave.sh            apply everything
#   pve-powersave.sh --status   just show what is set right now
#

set -u
PATH="/usr/sbin:/usr/bin:/sbin:/bin"

# ---- knobs, change here -------------------------------------------------
EPP="balance_power"       # power | balance_power | balance_performance
ASPM="powersave"          # powersupersave broke a NIC for me once, stay here
SATA_LPM="med_power_with_dipm"
HDD_STANDBY=180           # hdparm -S: 120 = 10 min, 0 = never spin down
HDD_APM=127               # 1-127 allows spindown, 128+ does not
EEE=1                     # set to 0 if the NIC link ever flaps
RX_USECS=125              # NIC interrupt coalescing, fewer wakeups
USB_AUTOSUSPEND=0         # 1 only if no UPS / keyboard / KVM on USB
PCI_RUNTIME_PM=1          # NICs and disk controllers are always skipped
RFKILL_WIFI_BT=1          # server is on ethernet, wifi + bluetooth are dead weight

# ---- helpers -------------------------------------------------------------

# print + syslog, so cron runs end up in journalctl -t pve-powersave
log()  { echo "$*";               logger -t pve-powersave -- "$*"; }
warn() { echo "WARN: $*" >&2;     logger -t pve-powersave -- "WARN: $*"; }

# write a value into a sysfs/proc file, fails quietly if the file is
# missing or the kernel refuses the value (caller decides if that matters)
sysfs_set() {
    [[ -w $1 ]] || return 1
    echo "$2" > "$1" 2>/dev/null
}

# ---- cpu -----------------------------------------------------------------
# With amd-pstate-epp the governor must be "powersave", the real tuning
# is then done via EPP. Anything else (acpi-cpufreq, passive amd-pstate)
# gets schedutil, because "powersave" there would pin the min frequency.
set_cpu() {
    local drv gov cpu
    drv=$(cat /sys/devices/system/cpu/cpu0/cpufreq/scaling_driver 2>/dev/null)

    if [[ $drv == amd-pstate-epp ]]; then
        gov=powersave
    else
        gov=schedutil
        warn "cpu driver is '$drv', not amd-pstate-epp. Add amd_pstate=active to the kernel cmdline and check CPPC is on in the BIOS."
    fi

    for cpu in /sys/devices/system/cpu/cpu[0-9]*/cpufreq; do
        sysfs_set "$cpu/scaling_governor" "$gov" || true
        # EPP only exists (and is only writable) with amd-pstate-epp + powersave
        sysfs_set "$cpu/energy_performance_preference" "$EPP" || true
    done

    log "cpu: $drv, governor=$(cat /sys/devices/system/cpu/cpu0/cpufreq/scaling_governor), epp=$(cat /sys/devices/system/cpu/cpu0/cpufreq/energy_performance_preference 2>/dev/null || echo n/a)"
}

# ---- pcie / sata / misc kernel knobs --------------------------------------
set_platform() {
    # PCIe link power states. Fails if the BIOS or cmdline locked it.
    sysfs_set /sys/module/pcie_aspm/parameters/policy "$ASPM" \
        && log "aspm: $ASPM" \
        || warn "aspm: could not set policy (locked by firmware/cmdline?)"

    # SATA link power management, one entry per AHCI port
    local host
    for host in /sys/class/scsi_host/host*/link_power_management_policy; do
        sysfs_set "$host" "$SATA_LPM" || warn "sata lpm: failed on $host"
    done
    log "sata lpm: $SATA_LPM"

    # Fewer periodic wakeups. nmi watchdog is a debug feature, writeback
    # every 15s instead of 5s is fine for a homelab.
    sysfs_set /proc/sys/kernel/nmi_watchdog 0 || true
    sysfs_set /proc/sys/vm/dirty_writeback_centisecs 1500 || true

    # HDMI audio on the iGPU, power it down after 10s idle
    sysfs_set /sys/module/snd_hda_intel/parameters/power_save 10 || true
    sysfs_set /sys/module/snd_hda_intel/parameters/power_save_controller Y || true

    # Wifi/BT card in the mini pc is never used, radios off saves ~0.5W
    if (( RFKILL_WIFI_BT )) && command -v rfkill >/dev/null; then
        rfkill block wifi bluetooth && log "rfkill: wifi + bluetooth off"
    fi
}

# ---- disks ---------------------------------------------------------------
# Spindown for real spinning disks only. SSD/NVMe do their own thing.
# Careful: do not spin down disks that are in a busy ZFS pool.
set_disks() {
    command -v hdparm >/dev/null || { warn "hdparm missing, skipping disks"; return; }
    local name type rota
    while read -r name type rota; do
        [[ $type == disk && $rota == 1 ]] || continue
        case $name in zram*|zd*|nbd*|rbd*) continue ;; esac   # not real disks
        # two calls so an unsupported APM does not stop the standby timer
        hdparm -B "$HDD_APM" "/dev/$name" >/dev/null 2>&1 || warn "hdd $name: APM not supported"
        hdparm -S "$HDD_STANDBY" "/dev/$name" >/dev/null 2>&1 || warn "hdd $name: standby failed"
        log "hdd $name: apm=$HDD_APM standby=$HDD_STANDBY"
    done < <(lsblk -dn -o NAME,TYPE,ROTA)
}

# ---- nics ----------------------------------------------------------------
# Only real NICs have a "device" link in sysfs. vmbr/tap/veth/fwpr do not.
physical_nics() {
    local n
    for n in /sys/class/net/*; do
        [[ -e $n/device ]] && basename "$n"
    done
}

set_nics() {
    command -v ethtool >/dev/null || { warn "ethtool missing, skipping nics"; return; }
    local nic eee_now was_up
    for nic in $(physical_nics); do
        # "EEE status: enabled - active" / "disabled" / "not supported"
        eee_now=$(ethtool --show-eee "$nic" 2>/dev/null | grep -o 'EEE status: [a-z]*' | awk '{print $3}')
        was_up=$(cat "/sys/class/net/$nic/operstate")

        if [[ -z $eee_now || $eee_now == not ]]; then
            log "nic $nic: no eee"
        elif (( EEE )) && [[ $eee_now != enabled ]]; then
            ethtool --set-eee "$nic" eee on 2>/dev/null
            # link renegotiates, give it a moment and check it came back
            sleep 3
            if [[ $was_up == up && $(cat "/sys/class/net/$nic/operstate") != up ]]; then
                ethtool --set-eee "$nic" eee off 2>/dev/null
                warn "nic $nic: eee dropped the link, turned it off again. Set EEE=0."
            else
                log "nic $nic: eee on"
            fi
        elif (( ! EEE )) && [[ $eee_now == enabled ]]; then
            ethtool --set-eee "$nic" eee off 2>/dev/null
            log "nic $nic: eee off"
        else
            log "nic $nic: eee already $eee_now"
        fi

        ethtool -C "$nic" rx-usecs "$RX_USECS" >/dev/null 2>&1 \
            && log "nic $nic: rx-usecs=$RX_USECS" \
            || log "nic $nic: driver does not take rx-usecs, ignoring"
    done
}

# ---- pci / usb / igpu runtime pm ----------------------------------------
set_runtime_pm() {
    local dev
    if (( PCI_RUNTIME_PM )); then
        for dev in /sys/bus/pci/devices/*; do
            # class 0x01 = storage, 0x02 = network. Leave those alone,
            # runtime PM on them causes latency spikes or worse.
            case $(cat "$dev/class") in 0x01*|0x02*) continue ;; esac
            sysfs_set "$dev/power/control" auto || true
        done
        log "pci runtime pm: auto (storage + network skipped)"
    fi

    if (( USB_AUTOSUSPEND )); then
        for dev in /sys/bus/usb/devices/*; do
            sysfs_set "$dev/power/autosuspend_delay_ms" 2000 || true
            sysfs_set "$dev/power/control" auto || true
        done
        log "usb autosuspend: on"
    fi

    # iGPU: let amdgpu pick its own clocks
    for dev in /sys/class/drm/card*/device/power_dpm_force_performance_level; do
        sysfs_set "$dev" auto && log "amdgpu: dpm=auto"
    done
}

# ---- status --------------------------------------------------------------
status() {
    echo "== cpu"
    echo "driver:   $(cat /sys/devices/system/cpu/cpu0/cpufreq/scaling_driver)"
    echo "governor: $(cat /sys/devices/system/cpu/cpu0/cpufreq/scaling_governor)"
    echo "epp:      $(cat /sys/devices/system/cpu/cpu0/cpufreq/energy_performance_preference 2>/dev/null || echo n/a)"
    # if this only shows C1/C2, the BIOS is holding back deep C-states
    echo "c-states: $(cat /sys/devices/system/cpu/cpu0/cpuidle/state*/name | tr '\n' ' ')"
    echo "== aspm:  $(cat /sys/module/pcie_aspm/parameters/policy)"
    echo "== sata lpm"
    grep -H . /sys/class/scsi_host/host*/link_power_management_policy 2>/dev/null
    echo "== disks"
    lsblk -dn -o NAME,TYPE,ROTA,TRAN,SIZE,MODEL
    echo "== nics"
    local nic
    for nic in $(physical_nics); do
        echo "$nic ($(cat /sys/class/net/$nic/operstate)): $(ethtool --show-eee "$nic" 2>/dev/null | grep 'EEE status' || echo 'eee n/a')"
    done
    echo "== pci runtime pm: $(grep -lx auto /sys/bus/pci/devices/*/power/control | wc -l) auto, $(grep -lx on /sys/bus/pci/devices/*/power/control | wc -l) on"
    echo "== nmi_watchdog: $(cat /proc/sys/kernel/nmi_watchdog), dirty_writeback_centisecs: $(cat /proc/sys/vm/dirty_writeback_centisecs)"
    echo "== rfkill"
    rfkill list 2>/dev/null | grep -E '^[0-9]|Soft' | paste - - || echo "rfkill n/a"
}

# ---- main ----------------------------------------------------------------
case "${1:-}" in
    --status|-s) status; exit 0 ;;
    "") ;;
    *) echo "usage: $0 [--status]"; exit 1 ;;
esac

[[ $EUID -eq 0 ]] || { echo "run as root"; exit 1; }

log "start"
set_cpu
set_platform
set_disks
set_nics
set_runtime_pm
log "done"