module for recovering shitty realtek controller when it gets stuck in bugged power state

2026-10-15 10:48:08 +02:00
parent 74d3d6870d
commit ac7a47c26a
3 changed files with 148 additions and 0 deletions
@@ -0,0 +1,89 @@
+# Called once after boot or resume. No network access is required.
+set -euo pipefail
+nic="/sys/bus/pci/devices/$NIC_PCI"
+bridge="/sys/bus/pci/devices/$NIC_BRIDGE"
+log() { echo "rtl8125-recovery: $*"; }
+
+# Let normal PCI and network resume finish before diagnosing anything.
+sleep 10
+if [[ ! -d "$bridge" ]]; then
+    log "Parent bridge $NIC_BRIDGE is absent; refusing recovery."
+    exit 1
+fi
+if [[ $(cat "$bridge/vendor") != 0x1022 || $(cat "$bridge/device") != 0x43f5 ]]; then
+    log "Parent bridge identity changed; refusing recovery."
+    exit 1
+fi
+
+if [[ -d "$nic" ]]; then
+    if [[ $(cat "$nic/vendor") != 0x10ec || $(cat "$nic/device") != 0x8125 ]]; then
+        log "Device at $NIC_PCI is not an RTL8125; refusing recovery."
+        exit 1
+    fi
+    if [[ $(dirname "$(readlink -f "$nic")") != "$(readlink -f "$bridge")" ]]; then
+        log "NIC is no longer directly below the configured bridge; refusing recovery."
+        exit 1
+    fi
+    power=$(cat "$nic/power_state")
+    command=$(timeout 5 setpci -s "$NIC_PCI" COMMAND 2>/dev/null || true)
+    # Both memory decoding and bus mastering should be enabled on this host.
+    # Disconnected cables do not disable these PCI command bits.
+    if [[ "$power" != D3cold && "$command" =~ ^[[:xdigit:]]{4}$ ]] &&
+        (( (16#$command & 6) == 6 )); then
+        log "PCI state healthy ($power, COMMAND=$command); no recovery needed."
+        exit 0
+    fi
+    # An administratively disabled NIC should not be automatically reset.
+    netdirs=("$nic"/net/*)
+    if [[ -e "${netdirs[0]}/flags" ]]; then
+        flags=$(cat "${netdirs[0]}/flags")
+        if (( (flags & 1) == 0 )); then
+            log "Interface is administratively down; leaving it alone."
+            exit 0
+        fi
+    fi
+    log "Failed PCI state: power=$power COMMAND=$command."
+else
+    log "NIC missing; attempting one rescan of its known bridge."
+fi
+
+# Retain evidence from the latest failure without accumulating unlimited logs.
+evidence="${STATE_DIRECTORY:-/var/lib/rtl8125-recovery}/last-failure.log"
+mkdir -p "$(dirname "$evidence")"
+{
+    date -Is
+    timeout 5 lspci -s "$NIC_PCI" -vv || true
+    timeout 5 lspci -s "$NIC_BRIDGE" -vv || true
+    journalctl -b -k --since '10 minutes ago' -n 250 --no-pager || true
+} > "$evidence" 2>&1
+
+# This is the remove/rescan recovery that worked on this machine. Never reset
+# or remove the parent bridge, which could affect other hardware.
+if [[ -d "$nic" ]]; then
+    echo 1 > "$nic/remove"
+fi
+echo 1 > "$bridge/rescan"
+
+for ((attempt=0; attempt<20; attempt++)); do
+    netdirs=("$nic"/net/*)
+    if [[ -e "${netdirs[0]}/ifindex" ]]; then
+        iface=$(basename "${netdirs[0]}")
+        power=$(cat "$nic/power_state")
+        command=$(timeout 5 setpci -s "$NIC_PCI" COMMAND 2>/dev/null || true)
+        if [[ "$power" == D0 && "$command" =~ ^[[:xdigit:]]{4}$ ]] &&
+            (( (16#$command & 6) == 6 )); then
+            # NetworkManager also reapplies the saved profile. Explicitly
+            # restore WoL here so recovery does not depend on its timing.
+            ethtool -s "$iface" wol g
+            if ! ethtool "$iface" | grep -Eq '^[[:space:]]*Wake-on: g$'; then
+                log "PCI recovered, but magic-packet WoL could not be confirmed."
+                exit 1
+            fi
+            log "Recovered $iface in D0; magic-packet WoL enabled. Evidence: $evidence"
+            exit 0
+        fi
+    fi
+    sleep 1
+done
+log "Recovery did not restore operational PCI state; no further attempt this invocation."
+exit 1