host_vars/localhost.yml

@@ -2,7 +2,7 @@
 # the profile tries to find the card with displays attached to apply these settings.
 # configuration of many GPUs not yet supported, one is assumed
 tuned_amdgpu_clock_min: "500"
-tuned_amdgpu_clock_max: "2725"
+tuned_amdgpu_clock_max: "2715"
 tuned_amdgpu_memclock_static: "1075"
 tuned_amdgpu_power_multi_def: 0.869969040247678  # 281W - real default
 tuned_amdgpu_power_multi_oc: 1.0  # full board power capability
@@ -11,25 +11,16 @@ tuned_amdgpu_power_multi_oc: 1.0  # full board power capability
 #   310:  0.959752321981424
 # sample worksheet in 'power_max multi tab calculator.ods'
 
-tuned_amdgpu_mv_offset: "+70" # add 70mV / 0.07V
+tuned_amdgpu_mv_offset: "+45" # add 45mV / 0.045V
 # '-50' undervolts GPU core voltage 50mV / 0.05V; warning: here be dragons/instability
 
 # 'tuned' plugins - used to set the kernel cmdline via bootloader... and sysctl tunables
 tuned_amdgpu_plugins:  # ref: https://github.com/redhat-performance/tuned/tree/master/tuned/plugins
-  rt_gamescope:
-    type: script
-    script: '/etc/tuned/rt_gamescope.sh'  # ensures gamescope has proper capability when profile is loaded
-  cpu:
-    boost: 1
-    governor: performance
-    energy_perf_bias: performance
   bootloader:  #  'cmdline' allows entries w/ a suffix, names should be unique across *all* profiles. values accept +/- operators
-    cmdline_amdgpu_general: "delayacct nowatchdog amdgpu.ppfeaturemask=0xfff7ffff"
+    cmdline_amdgpu_general: "delayacct nowatchdog kvm.ignore_msrs=1 kvm_amd.npt=1 amdgpu.ppfeaturemask=0xfff7ffff"
-    # cmdline_kvm: 'kvm.ignore_msrs=1 kvm_amd.npt=1'
-    cmdline_amdgpu_kvm: 'kvm_amd.npt=1'
     cmdline_amdgpu_hugepages: "default_hugepagesz=1G hugepagesz=1G hugepages=16"
     cmdline_amdgpu_iommu: "amd_iommu=on iommu=pt"
-    cmdline_amdgpu_nvme: "nvme_core.default_ps_max_latency_us=0"
+    cmdline_amdgpu_devs: "nvme_core.default_ps_max_latency_us=0 pci=realloc=off"
   sysctl:  # quote ints/cast to string to avoid surprises
     kernel.dmesg_restrict: '0'  # allow regular users to see the kernel ring buffer
     # net.core.default_qdisc: fq  # congestion control

roles/tuned_amdgpu/files/rt_gamescope.sh

@@ -1,33 +0,0 @@
-#!/bin/bash
-#
-# script run by 'tuned' to ensure gamescope has the proper capability
-
-# Check for arguments
-if [ $# -eq 0 ]; then
-    echo "Usage: $0 {verify|start}"
-    exit 1
-fi
-
-function verify_cap() {
-        /usr/sbin/getcap "$(which gamescope)" |& grep 'sys_nice=eip'
-        return $?
-}
-
-function set_cap() {
-        /usr/sbin/setcap 'CAP_SYS_NICE=eip' "$(which gamescope)"
-}
-
-# Handle arguments
-case "$1" in
-    verify)
-        verify_cap
-        ;;
-    start)
-        set_cap
-        ;;
-    *)
-        echo "Invalid argument. Use 'verify' or 'start'."
-        exit 1
-        ;;
-esac
-

roles/tuned_amdgpu/tasks/main.yml

@@ -22,12 +22,6 @@
     state: present
   become: true
 
-- name: Ensure 'tuned-ppd' is absent
-  ansible.builtin.package:
-    name: tuned-ppd
-    state: absent
-  become: true
-
 - name: Find bundled 'tuned' profiles
   ansible.builtin.find:
     paths:
@@ -91,15 +85,6 @@
   notify: Restart tuned
   become: true
 
-- name: Copy gamescope RT capability script
-  ansible.builtin.copy:
-    src: rt_gamescope.sh
-    dest: /etc/tuned/rt_gamescope.sh
-    owner: root
-    group: root
-    mode: '0755'
-  become: true
-
 - name: Template tuned.conf for custom profiles
   ansible.builtin.template:
     src: templates/tuned.conf.j2

roles/tuned_amdgpu/templates/amdgpu-profile.sh.j2

@@ -9,8 +9,6 @@
 # Variables shown below named '$TUNED_...' are given values by '/etc/tuned/amdgpu-profile-vars.conf'
 #
 # determine the connected GPU using the DRM subsystem. FIXME: assumes one card, make configurable
-# TODO: break this out into a role var; assume renderD128
-#       allow/use many configs
 CARD=$(/usr/bin/grep -ls ^connected /sys/class/drm/*/status | /usr/bin/grep -o 'card[0-9]' | /usr/bin/sort | /usr/bin/uniq | /usr/bin/sort -h | /usr/bin/tail -1)
 
 function get_hwmon_dir() {
@@ -23,6 +21,7 @@ function get_hwmon_dir() {
     done
 }
 
+
 # determine the hwmon directory
 HWMON_DIR=$(get_hwmon_dir "${CARD}")
 
@@ -52,16 +51,11 @@ function amdgpu_profile_reset() {
         echo "${PROF_DEFAULT_NUM}" | tee /sys/class/drm/"${CARD}"/device/pp_power_profile_mode
 
         # delay before returning - have mercy, may be followed with other profile function calls
-        sleep 0.25
+        sleep 0.5
 
 }
 
 function amdgpu_profile_overclock() {
-        # set manual control mode
-        #   allows control via 'pp_dpm_mclk', 'pp_dpm_sclk', 'pp_dpm_pcie', 'pp_dpm_fclk', and 'pp_power_profile_mode' files
-        echo 'manual' | tee /sys/class/drm/"${CARD}"/device/power_dpm_force_performance_level
-
-        # force GPU core and memory into highest clocks (fix flickering and poor power management)
         echo "s 0 ${TUNED_tuned_amdgpu_clock_min}" | tee /sys/class/drm/"${CARD}"/device/pp_od_clk_voltage
         echo "s 1 ${TUNED_tuned_amdgpu_clock_max}" | tee /sys/class/drm/"${CARD}"/device/pp_od_clk_voltage
         echo "m 1 ${TUNED_tuned_amdgpu_memclock_static}" | tee /sys/class/drm/"${CARD}"/device/pp_od_clk_voltage
@@ -74,6 +68,11 @@ function amdgpu_profile_overclock() {
         # commit the changes
         echo 'c' | tee /sys/class/drm/"${CARD}"/device/pp_od_clk_voltage
 
+        # force GPU core and memory into highest clocks (fix flickering and poor power management)
+        # set manual control mode
+        #   allows control via 'pp_dpm_mclk', 'pp_dpm_sclk', 'pp_dpm_pcie', 'pp_dpm_fclk', and 'pp_power_profile_mode' files
+        echo 'manual' | tee /sys/class/drm/"${CARD}"/device/power_dpm_force_performance_level
+
         # adjust power limit using multiplier against board capability
         POWER_LIM_OC=$(/usr/bin/awk -v m="$POWER_CAP" -v n="${TUNED_tuned_amdgpu_power_multi_oc}" 'BEGIN {printf "%.0f", (m*n)}')
         echo "$POWER_LIM_OC" | tee "${HWMON_DIR}/power1_cap"
@@ -89,7 +88,7 @@ function amdgpu_profile_overclock() {
         echo "${PROF_VR_NUM}" | tee /sys/class/drm/"${CARD}"/device/pp_power_profile_mode
 
         # delay before returning - have mercy, may be followed with other profile function calls
-        sleep 0.25
+        sleep 0.5
 
 }