|
|
|
@@ -25,7 +25,14 @@ log_event() {
|
|
|
|
|
local msg="$1"
|
|
|
|
|
mkdir -p "$(dirname "${GPUDRIVER_LOG}")" 2>/dev/null
|
|
|
|
|
echo "$(date '+%Y-%m-%d %H:%M:%S') gpudriver: ${msg}" >>"${GPUDRIVER_LOG}" 2>/dev/null
|
|
|
|
|
logger -t gpudriver "${msg}" 2>/dev/null
|
|
|
|
|
# Intentionally do NOT call `logger` here. The parallel autostart Phase B
|
|
|
|
|
# bursts hundreds of log messages at once; the journald socket fills, and
|
|
|
|
|
# `logger` (a synchronous binary) blocks waiting for the daemon to drain.
|
|
|
|
|
# That deadlock pinned gpudriver between "calling load_driver panfrost"
|
|
|
|
|
# and the next log step on R36S 20260515 builds — the script never
|
|
|
|
|
# returned, the autostart wait never finished, and the watchdog reset
|
|
|
|
|
# the board. The on-disk echo above is enough for our post-mortem
|
|
|
|
|
# diagnostic needs.
|
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
# Verify the GPU node ended up bound to the requested kernel driver AND
|
|
|
|
@@ -102,7 +109,13 @@ load_driver() {
|
|
|
|
|
|
|
|
|
|
case ${DRIVER_TO_LOAD} in
|
|
|
|
|
"libmali")
|
|
|
|
|
modprobe -r @PAN@ 2>/dev/null
|
|
|
|
|
# Symmetric safety: only rmmod @PAN@ if loaded. In-tree panfrost
|
|
|
|
|
# is safer to unload than mali_kbase but skip anyway — a reboot
|
|
|
|
|
# gives a cleaner switch and matches the user's expectation that
|
|
|
|
|
# changing GPU driver requires a restart.
|
|
|
|
|
if lsmod 2>/dev/null | grep -q "^@PAN@ "; then
|
|
|
|
|
log_event "libmali: @PAN@ loaded — switch needs reboot to take effect"
|
|
|
|
|
fi
|
|
|
|
|
modprobe mali_kbase
|
|
|
|
|
|
|
|
|
|
# Bind-mount mali GLES libraries over mesa if not already done.
|
|
|
|
@@ -121,8 +134,75 @@ load_driver() {
|
|
|
|
|
done
|
|
|
|
|
;;
|
|
|
|
|
"panfrost")
|
|
|
|
|
modprobe -r mali_kbase 2>/dev/null
|
|
|
|
|
modprobe @PAN@
|
|
|
|
|
# Snapshot ANY kernel messages before we touch the driver. If the
|
|
|
|
|
# system is already in a degraded state by the time gpudriver
|
|
|
|
|
# runs (e.g. an earlier autostart script tickled something), we
|
|
|
|
|
# want to see it in the post-mortem rather than blaming the
|
|
|
|
|
# subsequent modprobe.
|
|
|
|
|
log_event "panfrost: step 0a - capturing pre-state dmesg"
|
|
|
|
|
dmesg -T 2>/dev/null | tail -100 >>"${GPUDRIVER_LOG}.dmesg-pre" 2>/dev/null
|
|
|
|
|
log_event "panfrost: step 0b - dmesg pre-state captured"
|
|
|
|
|
|
|
|
|
|
# mali_kbase is the ARM vendor out-of-tree blob. Its driver-exit
|
|
|
|
|
# path double-frees the GPU regulator and clock at unbind time
|
|
|
|
|
# (observed kernel Oops in __clk_put at POISON_LIST_END +
|
|
|
|
|
# _regulator_put WARN, cascading into further oopses until the
|
|
|
|
|
# board watchdog resets — confirmed on R36S with kernel 6.12.79).
|
|
|
|
|
# Removing it in-place is therefore unsafe: skip the rmmod
|
|
|
|
|
# entirely if the module is loaded and let the user reboot to
|
|
|
|
|
# land in a fresh kernel where the module never came up. If the
|
|
|
|
|
# module is NOT loaded we can proceed; this is the normal case
|
|
|
|
|
# on a fresh boot where gpu.driver=panfrost.
|
|
|
|
|
if lsmod 2>/dev/null | grep -q "^mali_kbase "; then
|
|
|
|
|
log_event "panfrost: step 1 - mali_kbase IS LOADED, skipping rmmod (would crash kernel) — reboot required for switch to take effect"
|
|
|
|
|
else
|
|
|
|
|
log_event "panfrost: step 1 - mali_kbase not loaded, no rmmod needed"
|
|
|
|
|
fi
|
|
|
|
|
log_event "panfrost: step 2 - rmmod check done"
|
|
|
|
|
|
|
|
|
|
# Snapshot the last dmesg line so we can quote only what the
|
|
|
|
|
# panfrost probe added.
|
|
|
|
|
PRE_DMESG_COUNT=$(dmesg 2>/dev/null | wc -l)
|
|
|
|
|
log_event "panfrost: step 3 - dmesg snapshot taken (count=${PRE_DMESG_COUNT})"
|
|
|
|
|
|
|
|
|
|
# Launch modprobe asynchronously. `timeout` would normally cap
|
|
|
|
|
# this, but the panfrost probe on certain RK3326 clones blocks
|
|
|
|
|
# the modprobe process inside an uninterruptible kernel syscall
|
|
|
|
|
# (D state) — neither SIGTERM from `timeout` nor SIGKILL from a
|
|
|
|
|
# subsequent kill can release it. The script must never block
|
|
|
|
|
# waiting for modprobe; instead we poll for the platform bind to
|
|
|
|
|
# appear, give up after 30s, and let gpu_is_bound conclude with
|
|
|
|
|
# a failure that triggers the libmali fallback.
|
|
|
|
|
modprobe @PAN@ 2>/dev/null &
|
|
|
|
|
MODPROBE_PID=$!
|
|
|
|
|
log_event "panfrost: step 4 - modprobe launched pid=${MODPROBE_PID}"
|
|
|
|
|
|
|
|
|
|
_iter=0
|
|
|
|
|
for _wait in $(seq 1 30); do
|
|
|
|
|
_iter=$_wait
|
|
|
|
|
for g in /sys/bus/platform/drivers/panfrost/[a-f0-9]*.gpu; do
|
|
|
|
|
[ -e "$g" ] && break 2
|
|
|
|
|
done
|
|
|
|
|
kill -0 ${MODPROBE_PID} 2>/dev/null || break
|
|
|
|
|
sleep 1
|
|
|
|
|
done
|
|
|
|
|
log_event "panfrost: step 5 - poll loop exited at iter=${_iter}"
|
|
|
|
|
|
|
|
|
|
if ls /sys/bus/platform/drivers/panfrost/[a-f0-9]*.gpu >/dev/null 2>&1; then
|
|
|
|
|
log_event "panfrost: step 6a - platform bind observed"
|
|
|
|
|
else
|
|
|
|
|
if kill -0 ${MODPROBE_PID} 2>/dev/null; then
|
|
|
|
|
log_event "panfrost: step 6b - modprobe still running, probe stuck"
|
|
|
|
|
else
|
|
|
|
|
log_event "panfrost: step 6c - modprobe exited without bind"
|
|
|
|
|
fi
|
|
|
|
|
fi
|
|
|
|
|
|
|
|
|
|
# Drop the kernel messages emitted during the probe to the log so
|
|
|
|
|
# the failure mode is recoverable without a serial console.
|
|
|
|
|
log_event "panfrost: step 7 - capturing dmesg delta..."
|
|
|
|
|
dmesg 2>/dev/null | tail -n +$((PRE_DMESG_COUNT + 1)) | head -40 >>"${GPUDRIVER_LOG}" 2>/dev/null
|
|
|
|
|
log_event "panfrost: step 8 - dmesg captured"
|
|
|
|
|
|
|
|
|
|
# Reverse every bind the libmali path did. Order matters: undo the
|
|
|
|
|
# libEGL/libGLESv2/etc. binds from /usr/lib/mali first, then the
|
|
|
|
@@ -167,7 +247,37 @@ case "$1" in
|
|
|
|
|
"--start")
|
|
|
|
|
get_current_driver
|
|
|
|
|
log_event "boot: selected driver=${CONFDRIVER}"
|
|
|
|
|
|
|
|
|
|
# Boot-loop protection. Some hardware variants kernel-panic on
|
|
|
|
|
# `modprobe panfrost` (observed on RK3326 clones), which the watchdog
|
|
|
|
|
# then resets — so the system never reaches gpu_is_bound to record
|
|
|
|
|
# the failure, and the next boot tries panfrost again, ad infinitum.
|
|
|
|
|
# We drop a sentinel on /storage before the modprobe and only clear
|
|
|
|
|
# it once we know the driver bound successfully. On the next boot
|
|
|
|
|
# we look for the sentinel: if present, the previous attempt with
|
|
|
|
|
# this driver crashed before recovery — silently force libmali and
|
|
|
|
|
# rewrite the setting so the user can pick again from a working UI.
|
|
|
|
|
SENTINEL="/storage/.config/gpudriver.attempting"
|
|
|
|
|
if [ -f "${SENTINEL}" ]; then
|
|
|
|
|
PREV="$(cat "${SENTINEL}" 2>/dev/null)"
|
|
|
|
|
log_event "previous boot crashed while loading ${PREV} — forcing libmali"
|
|
|
|
|
rm -f "${SENTINEL}"
|
|
|
|
|
if [ "${CONFDRIVER}" = "${PREV}" ] && [ "${PREV}" != "libmali" ]; then
|
|
|
|
|
CONFDRIVER="libmali"
|
|
|
|
|
set_setting ${GPU_DRIVER_SETTING_KEY} libmali
|
|
|
|
|
fi
|
|
|
|
|
fi
|
|
|
|
|
|
|
|
|
|
# Use dd with oflag=sync so the sentinel is on the SD (not in page
|
|
|
|
|
# cache) BEFORE we touch the GPU driver. A subsequent kernel hang
|
|
|
|
|
# or watchdog reset would otherwise drop the sentinel and the boot
|
|
|
|
|
# after that would try panfrost again, looping.
|
|
|
|
|
echo "${CONFDRIVER}" | dd of="${SENTINEL}" oflag=sync 2>/dev/null
|
|
|
|
|
log_event "sentinel persisted: ${CONFDRIVER}"
|
|
|
|
|
|
|
|
|
|
log_event "calling load_driver ${CONFDRIVER}"
|
|
|
|
|
load_driver ${CONFDRIVER}
|
|
|
|
|
log_event "load_driver returned"
|
|
|
|
|
# If the requested driver didn't actually bind, fall back to the
|
|
|
|
|
# other one so the user gets a graphical system either way. Mali-G31
|
|
|
|
|
# on RK3326 is normally driven by libmali; if mali_kbase decides the
|
|
|
|
@@ -182,17 +292,35 @@ case "$1" in
|
|
|
|
|
load_driver "${ALT}" 2>/dev/null
|
|
|
|
|
if gpu_is_bound "${ALT}"; then
|
|
|
|
|
log_event "fallback to ${ALT} succeeded"
|
|
|
|
|
# Persist the working driver as the new default so the user is
|
|
|
|
|
# not asked to confirm again on next reboot.
|
|
|
|
|
set_setting ${GPU_DRIVER_SETTING_KEY} "${ALT}"
|
|
|
|
|
else
|
|
|
|
|
log_event "fallback to ${ALT} ALSO FAILED — system will run software-rendered"
|
|
|
|
|
fi
|
|
|
|
|
fi
|
|
|
|
|
|
|
|
|
|
# Driver bound (or fallback ran). Clear the sentinel so the next boot
|
|
|
|
|
# does NOT think the current boot crashed.
|
|
|
|
|
rm -f "${SENTINEL}" 2>/dev/null
|
|
|
|
|
sync 2>/dev/null
|
|
|
|
|
;;
|
|
|
|
|
"libmali")
|
|
|
|
|
set_setting ${GPU_DRIVER_SETTING_KEY} $1
|
|
|
|
|
# Force the setting onto disk before we return control to the UI,
|
|
|
|
|
# which may immediately trigger a reboot.
|
|
|
|
|
sync
|
|
|
|
|
@DTB_OVERLAY_UNLOAD@
|
|
|
|
|
;;
|
|
|
|
|
"panfrost")
|
|
|
|
|
set_setting ${GPU_DRIVER_SETTING_KEY} $1
|
|
|
|
|
# Force the setting onto disk before we return control to the UI,
|
|
|
|
|
# which may immediately trigger a reboot ("Reboot now" in the popup
|
|
|
|
|
# is dispatched synchronously; ext4 might still have the new value
|
|
|
|
|
# in the page cache, not on the SD). Without this sync the next
|
|
|
|
|
# boot could still read gpu.driver=libmali and the user perceives
|
|
|
|
|
# the switch as silently broken.
|
|
|
|
|
sync
|
|
|
|
|
@DTB_OVERLAY_LOAD@
|
|
|
|
|
;;
|
|
|
|
|
"")
|
|
|
|
|
Douglas Teles