PCIe power management and error handling for a Thunderbolt eGPU: ASPM, D3cold, runtime PM, AER and DPC

PCIe power management and error handling for a Thunderbolt eGPU: ASPM, D3cold, runtime PM, AER and DPC

Which PCIe power-management and error-handling knobs actually matter for a Thunderbolt-tunnelled GPU on Linux — ASPM and pcie_aspm=off versus pcie_aspm.policy, port runtime PM and D3cold, NVIDIA’s runtime D3 parameters, USB4 CL states, reading AER and DPC evidence, and suspend/resume — separating the fix from the insurance from the cargo cult.

name: pcie-power-management-aer-dpc-egpu-linux title: PCIe Power Management, AER & DPC for Thunderbolt/USB4 eGPUs on Linux description: > Expert reference for PCIe power management and error handling of Thunderbolt/USB4-tunnelled devices (eGPUs) on Linux: ASPM (pcie_aspm=off vs pcie_aspm.policy), PCIe port runtime PM (pcie_port_pm, pci_bridge_d3_possible, D3cold, power/control), NVIDIA runtime D3 (NVreg_DynamicPowerManagement, RTD3), USB4 CL states (thunderbolt.clx), AER/DPC reading (native vs firmware-first, pci=noaer, UESta/CESta), pciehp+PM interactions, and eGPU suspend/resume. TRIGGER: eGPU falls off the bus, Xid 79, “D3cold to D0 device inaccessible”, which kernel params matter, AER/DPC log triage, s2idle hang with eGPU. SKIP: Thunderbolt authorization/boltctl basics, CUDA stack setup, Mac eGPU. verified-as-of: 2026-09-24

PCIe Power Management, AER & DPC for Thunderbolt/USB4 eGPUs on Linux

Verified-as-of 2026-09-24. Every claim is tagged [SOURCED url] (read from the cited page during research) or [INFERRED] (derived from sourced facts, code reading from memory, or the worked case; verify before relying on it). Kernel-parameter names, sysfs attribute names, and NVIDIA module-parameter names are quoted only where sourced.

Worked case (anchor)

Intel NUC 15 Pro (Arrow Lake-P, TB4) → Razer Core X V2 (Intel JHL9480 “Barlow Ridge” switch, PCI ID 8086:5786) → RTX 5080, Ubuntu 26.04.1, kernel 7.0.0-34, NVIDIA 610.57.04-open.

Working cmdline: thunderbolt.host_reset=0 pci=realloc=off pcie_ports=native pcie_port_pm=off pcie_aspm=off thunderbolt.clx=0 iommu=pt /etc/modprobe.d/nvidia-egpu-pm.conf: NVreg_DynamicPowerManagement=0x00 NVreg_PreserveVideoMemoryAllocations=0

Good-boot sysfs: GPU power/control=on, d3cold_allowed=1, runtime_status=active; switch bridges 02:00.0/03:00.0 control=auto; root port 00:07.0 LnkCtl: ASPM L1 Enabled while GPU and switch ports show ASPM Disabled. Failing boots: GPU audio function logged Unable to change power state from D3cold to D0, device inaccessible ~4.7 s into boot. pci=noaer was previously present (now removed).

Verdict (summary; details in each section):

Core Concepts

1. ASPM vs device power states. ASPM (L0s, L1, L1.1, L1.2) is link power management negotiated per link; D0–D3hot–D3cold are device states. D3cold means “Vcc removed” and “It is not possible to program a PCI device to go into D3cold” from config space; it needs a platform method (ACPI power resources, or the port upstream going to D3). [SOURCED https://www.kernel.org/doc/html/latest/power/pci.html]
2. PCIe port runtime PM. Since 4.7 the kernel may put PCIe ports in D3 when everything below is suspended (bridge_d3 flag). pci_bridge_d3_possible() gates this: false for conventional PCI bridges, false when pcie_port_pm=off, true when forced, false for native-hotplug ports on x86 (not validated until ~2018), and otherwise BIOS-year ≥2015 or a Thunderbolt-attached port. [SOURCED https://www.mail-archive.com/[email protected]/msg1614295.html; https://ratatoskr.run/lkml/2026/07/17337075/t; https://patchwork.kernel.org/project/linux-pci/patch/10206fe7f3967aa73ade5b24dc729de0e94c3b7f.1529173804.git.lukas@wunner.de/ (title only; page body was not readable during research)]
3. USB4 CL states. CL0s/CL1/CL2 are USB4 high-speed-lane low-power states enabled by the thunderbolt driver; clx module param (bool, default true) gates them. [SOURCED https://www.spinics.net/lists/linux-usb/msg219384.html; https://patchwork.kernel.org/project/linux-usb/patch/[email protected]/]
4. Host router reset at driver start. The host_reset module param in drivers/thunderbolt/nhi.c is static bool host_reset = true; module_param(host_reset, bool, 0444) with description “reset USB4 host router (default: true)”. The reset tears down tunnels created by boot firmware. [SOURCED https://github.com/torvalds/linux/blob/master/drivers/thunderbolt/nhi.c; https://ratatoskr.run/linux-usb/2026/08/17480231/t]
5. AER control is negotiated. “Linux does not handle AER events unless the firmware grants AER control to the OS via the ACPI _OSC method.” pcie_ports=native overrides that. DPC control is linked to AER control per PCIe r5.0 sec 6.2.10; pcie_ports=dpc-native decouples it. [SOURCED https://www.kernel.org/doc/html/latest/PCI/pcieaer-howto.html; https://www.spinics.net/lists/linux-pci/msg88329.html]
6. Error classes. Correctable errors “pose no impacts on the functionality of the interface” and are only logged/cleared; uncorrectable non-fatal → error_detected(dev, pci_channel_io_normal); fatal → error_detected(dev, pci_channel_io_frozen) and “performing a reset at upstream is necessary”. [SOURCED https://www.kernel.org/doc/html/latest/PCI/pcieaer-howto.html]
7. DPC = containment. “On platforms supporting Downstream Port Containment (PCIe r7.0 sec 6.2.11), the link to the sub-hierarchy with the faulting device is disabled”; devices are inaccessible until link reset/slot_reset. [SOURCED https://www.kernel.org/doc/html/latest/PCI/pci-error-recovery.html]
8. “D3cold → D0, device inaccessible” is a symptom, not a PM decision. pci_raw_set_power_state() uses the PM capability in config space; config reads that return ~0 were historically misread as “device in D3hot/D3cold”. The message means config space is unreachable — the device is gone (tunnel torn down, link down, port suspended). [SOURCED https://lkml.iu.edu/hypermail/linux/kernel/1908.2/07271.html]
9. Removable detection. dev_is_removable() depends on the ACPI ExternalFacingPort property on the upstream bridge; some firmware omits it on Thunderbolt root ports, so drivers miss eGPU-specific paths. A 2026 patch adds a pci_is_thunderbolt_attached() fallback. amdgpu now disables runtime PM when either is true (7.4). [SOURCED https://ratatoskr.run/lkml/2026/04/3535291/t; https://ratatoskr.run/amd-gfx/2026/08/17400523/t]
10. Why ASPM shows “Disabled” on the eGPU but “L1 Enabled” on the root port. For hotplugged Thunderbolt/USB4 devices “the BIOS may not have configured ASPM since the device wasn’t present at boot time”; Linux leaves them as found (Windows enables L1). A May-2026 patch to enable L0s/L1 for removable devices is still under discussion. [SOURCED https://ratatoskr.run/linux-pci/2026/05/3543749/t]

Knob Table

ASPM & CL States

• Read link state: lspci -vvs <BDF> | grep -E 'LnkCap|LnkCtl|LnkSta|L1SubCap|L1SubCtl'. LnkCtl: ASPM L1 Enabled = L1 negotiated; L1SubCtl1: ... L1_2 Enabled = L1.2 substate enabled; substates require CLKREQ# on both ends and both ends capable, else only legacy L1. [SOURCED https://www.synopsys.com/articles/reduce-power-consumption.html; https://semiengineering.com/using-pci-express-l1-sub-states-to-minimize-power-consumption-in-advanced-process-nodes/]
• L1.1 vs L1.2: L1.1 keeps common-mode voltage; L1.2 turns off all high-speed circuits — deeper, slower exit. [SOURCED same]
• Interpreting the worked case: root port L1 Enabled + GPU/switch Disabled is the normal Linux picture for a hotplugged Thunderbolt hierarchy (“Windows: LnkCtl: ASPM L1 Enabled” vs “Linux: LnkCtl: ASPM Disabled”). It is not caused by pcie_aspm=off, which by definition changes nothing. [SOURCED https://ratatoskr.run/linux-pci/2026/05/3543749/t; https://lkml.iu.edu/hypermail/linux/kernel/2404.3/06648.html]
• Interaction with hotplug/runtime PM: ThinkPad P16 (Meteor Lake) fails to power off with ASPM enabled — pciehp removal path triggers __pm_runtime_resume of the Thunderbolt domain during teardown (“interrupt for TX ring 0 is already enabled”); “strongly dependent on ASPM behavior”. Workaround was pcie_aspm=off. [SOURCED https://ratatoskr.run/linux-pci/2026/03/3483145/t]
• CL states: disable with thunderbolt.clx=0 when you see link-integrity precursors (below) or when the router is in the 2026 CLx quirk list. [SOURCED https://www.spinics.net/lists/linux-usb/msg219384.html; https://ratatoskr.run/linux-usb/2026/04/3533651]
• Coming default change: RFC v2 (2026-05-11, by the RFC author): BIOS ≥2025 → powersupersave if FADT allows; honoured exceptions: user policy or pcie_aspm param. If merged, hosts like this NUC (2025+ BIOS) will start enabling L1/L1.x on links Linux configures — consider pinning pcie_aspm.policy=performance or keeping pcie_aspm=off (which leaves the setting untouched) to prevent it. [SOURCED https://lkml.iu.edu/2605.1/06202.html]

Runtime PM & D3cold

• sysfs semantics: power/control = auto (“allow the device to be power managed at run time”) or on (“prevent the device from being power managed”); power/runtime_status ∈ suspended/suspending/resuming/active/error/unsupported; d3cold_allowed “If it is cleared, the device will never be put into D3Cold state”; power_state shows D0/D1/D2/D3hot/D3cold (read-only). [SOURCED https://www.kernel.org/doc/Documentation/ABI/testing/sysfs-devices-power; https://www.kernel.org/doc/Documentation/ABI/testing/sysfs-bus-pci]
• Port D3 gating (pci_bridge_d3_possible()): order of checks — pci_bridge_d3_disable → false; pci_bridge_d3_force → true; non-PCIe → false; not root/upstream/downstream → false; native hotplug port on x86 → false (validation only from ~2018); Thunderbolt-attached → allowed (Light Ridge 2010 already fine, “specifically including its hotplug ports”); else BIOS year ≥2015. [SOURCED https://www.mail-archive.com/[email protected]/msg1614295.html; https://ratatoskr.run/lkml/2026/07/17337075/t; https://patchwork.kernel.org/project/linux-pci/patch/10206fe7f3967aa73ade5b24dc729de0e94c3b7f.1529173804.git.lukas@wunner.de/]
• The CachyOS/Barlow Ridge loss (issue #1057): AMD Phoenix, USB4 PCIe tunnel port 0000:00:04.1 (1022:14ef) goes runtime_status=suspended with the TB5 dock attached; downstream JHL9480 (8086:5786) hierarchy is absent although the thunderbolt subsystem still sees the dock. Effective: udev power/control=on on that port, or pcie_port_pm=off; even lspci wakes it. Ineffective: xHCI control=on, embedded-controller (EC) reset. Reproduces on stock Arch kernels (upstream bug), open. [SOURCED https://github.com/CachyOS/linux-cachyos/issues/1057]
• The “becomes inaccessible later” variant (#1021): Strix Halo + JHL9480 + RTX 4090: bridge config space reads 0xff post-boot; snd_hda_intel ...: Unable to change power state from D0 to D0, device inaccessible; pcie_aspm=off needed to avoid crash on boot with eGPU attached. [SOURCED https://github.com/CachyOS/linux-cachyos/issues/1021]
• Regression on 6.19.10 (#794): Razer Core X + RX 6800 XT drops off with pciehp Link Down / Card not present. [SOURCED https://github.com/CachyOS/linux-cachyos/issues/794]
• Diagnosing D3cold vs “gone”: cat /sys/bus/pci/devices/<BDF>/power_state and runtime_status for GPU, its audio function, the switch ports, and the host root port. If the port is suspended and children are missing → port runtime PM (#1057 pattern; fix with the port rule). If the port is active but children read 0xff / “device inaccessible” → link/tunnel loss (host_reset, DPC, cable, PSU). [INFERRED from sourced cases]
• Why the worked-case audio message is not a PM bug: it appeared ~4.7 s into boot, right when the thunderbolt driver loads and (with host_reset=1) resets the host router; the audio function’s config space vanished mid-probe, which the PCI core reports as a failed D3cold→D0 transition. Removing host_reset removed the message. [INFERRED; mechanism SOURCED https://lkml.iu.edu/hypermail/linux/kernel/1908.2/07271.html + https://ratatoskr.run/linux-usb/2026/08/17480231/t]

NVIDIA Runtime D3

• Modes and defaults as in the knob table; feature “enabled by default on supported Ampere or newer notebook computers. In all other configurations, it is disabled by default”; needs kernel ≥4.18 (4.19–5.4 buggy, fixed 5.5). [SOURCED https://download.nvidia.com/XFree86/Linux-x86_64/580.65.06/README/dynamicpowermanagement.html]
• Driver-shipped udev rules (80-nvidia-pm.rules content): remove NVIDIA xHCI (class 0x0c0330), UCSI (0x0c8000) and Audio (0x040300) functions on add; on bind set power/control=auto for class 0x030000/0x030200; on unbind set on. [SOURCED same]
• eGPU implication: these rules are written for notebooks. On an eGPU: (a) removing the HDA function kills audio over the enclosure’s DP/HDMI; (b) control=auto invites the port-D3 path. With 0x00 the driver keeps the GPU active regardless — observed control=on, runtime_status=active on the worked case means the rules were not applied or the driver forbids runtime PM [INFERRED]. Do not install them for an eGPU unless you explicitly want RTD3.
• nvidia-persistenced with UVM persistence mode disables RTD3 outright — a good belt-and-braces for compute eGPUs. [SOURCED same]
• X and eGPUs: the X driver “does not configure X screens on external GPUs by default”; override with Option "AllowExternalGpus" (“system stability when an eGPU is unplugged while in use … is not guaranteed”). [SOURCED https://download.nvidia.com/XFree86/Linux-x86_64/515.65.01/README/egpu.html]

AER & DPC Reading Guide

Where to read

• lspci -vvvs <BDF>: DevSta: CorrErr+ NonFatalErr+ FatalErr+ UnsupReq+ (Device Status bits), UESta: (uncorrectable status: DLP, SDES = Surprise Down, TLP, FCP, CmpltTO, CmpltAbrt, UnxCmplt, RxOF, MalfTLP, ECRC, UnsupReq, ACSViol), CESta: (correctable status: RxErr, BadTLP, BadDLLP, Rollover, Timeout, AdvNonFatalErr, CorrIntErr, HeaderOF). Field names follow the kernel’s AER sysfs field list. [SOURCED https://www.kernel.org/doc/Documentation/ABI/testing/sysfs-bus-pci-devices-aer]
• sysfs counters: /sys/bus/pci/devices/<BDF>/aer_dev_correctable, aer_dev_nonfatal, aer_dev_fatal (per device, “seen and reported by this PCI device using ERR_COR/ERR_NONFATAL/ERR_FATAL”); root port: aer_rootport_total_err_cor|nonfatal|fatal. [SOURCED same]
• Kernel log: journalctl -k | grep -Ei 'aer|dpc|pciehp|thunderbolt|Xid|power state'. Native AER requires _OSC grant (ACPI: PCI: ... _OSC: OS now controls [... AER ... DPC]), or pcie_ports=native. [SOURCED https://www.kernel.org/doc/html/latest/PCI/pcieaer-howto.html]
• Injection for testing your alerting: aer-inject (https://github.com/intel/aer-inject.git). [SOURCED same]
• Stats daemon: rasdaemon (ras-mc-ctl --errors) persists AER events [INFERRED — standard tool, not verified this pass].

Three signatures

1. Link-integrity precursor. Growing CESta/aer_dev_correctable counts of RxErr, BadDLLP, BadTLP, Rollover, Timeout on the eGPU link (GPU function or the switch downstream port), often with LnkSta speed/width downtrain, minutes-to-hours before an Xid 79 or a fatal Surprise Down. Correctable errors are by definition recovered (“pose no impacts”), so the logger is your early warning. Actions: thunderbolt.clx=0, swap/shorten the cable, avoid L1 substates on that link (policy), reseat the enclosure card. [SOURCED AER classes: https://www.kernel.org/doc/html/latest/PCI/pcieaer-howto.html; field names: sysfs-bus-pci-devices-aer; forum practice of disabling ASPM for Xid 79: https://forums.developer.nvidia.com/t/xid-79-gpu-has-fallen-off-the-bus/359509] Precursor-to-Xid causality is [INFERRED].
2. Zero-precursor Xid 79. NVRM: Xid (PCI:...): 79, GPU has fallen off the bus with no AER history, typically with pciehp ...: Slot(N): Link Down / Card not present and follow-on Unable to change power state from D3cold to D0, device inaccessible. Means the device vanished electrically or logically: PSU/enclosure power, tunnel torn down, or the port runtime-suspended (see #1057). Happens under load (power) or at idle (PM). Actions: check power_state/runtime_status of the tunnel port, PSU, then host_reset/tunnel. [SOURCED Xid 79 = “GPU has fallen off the bus” and load/idle variants: https://forums.developer.nvidia.com/t/xid-79-gpu-has-fallen-off-the-bus/359509, https://github.com/NVIDIA/open-gpu-kernel-modules/issues/900; catalog index: https://docs.nvidia.com/deploy/xid-errors/index.html]
3. host_reset / tunnel rebuild at boot. Sequence: firmware-enumerated GPU visible at early boot → thunderbolt driver loads → USB4 host router reset tears down the firmware tunnel → pciehp Link Down on the TB downstream port → re-enumeration under a fresh bridge (and, with realloc, re-assigned windows) → driver probes race the rebuild → D3cold to D0, device inaccessible on a function, or “fallen off the bus” from the GPU driver. Fix: thunderbolt.host_reset=0 (+ pci=realloc=off when windows move). Exact log strings around the reset are [INFERRED]; the mechanism and the workaround are [SOURCED https://ratatoskr.run/linux-usb/2026/08/17480231/t].

DPC specifics

• DPC trigger disables the link below the port; pciehp deliberately ignores Link Down/Up caused by DPC (and by Secondary Bus Reset, D3cold suspend, FPGA/firmware update) when recovery succeeds, so the hotplugged device stays bound. A DPC that does not recover looks like surprise removal. [SOURCED https://www.kernel.org/doc/html/latest/PCI/pci-error-recovery.html; https://mailweb.openeuler.org/archives/list/[email protected]/message/4WFNKWQPUH7ZXD3QIGGT24YWPJX4F5WY/; https://ratatoskr.run/linux-pci/2026/09/17519182/t]
• Surprise Down on genuine hot removal is ignored by DPC (“PCI/DPC: Ignore Surprise Down error on hot removal”, v6 2023). [SOURCED https://lkml.iu.edu/2311.3/06842.html]
• pcie_ports=dpc-native only when the platform withholds AER but you still want containment; “May cause conflicts if firmware uses AER or DPC.” [SOURCED https://www.spinics.net/lists/linux-pci/msg88329.html]

Suspend/Resume

• Modes: s2idle is “pure software”; deep (S3) “everything … put into a low-power state, except for memory”. Check /sys/power/mem_sleep; override with mem_sleep_default=. [SOURCED https://www.kernel.org/doc/html/latest/admin-guide/pm/sleep-states.html]
• Known 7.x regressions with RTX 50 / eGPUs:
  • #1117: RTX 5070 Laptop, Ubuntu 26.04, kernel 7.0.0-14 hangs on s2idle resume; last line nvidia 0000:01:00.0: Enabling HDA controller; identical driver fine on 6.17; no fix listed. [SOURCED https://github.com/NVIDIA/open-gpu-kernel-modules/issues/1117]
  • #1376: RTX 5070 Ti eGPU (GB203) over TB5 enclosure, kernel 7.2.4, 615.71.09: pm_test=devices passes, pm_test=platform hangs; PreserveVideoMemoryAllocations=1 and procfs hooks did not help; open. [SOURCED https://github.com/NVIDIA/open-gpu-kernel-modules/issues/1376]
  • #1284: RTX 5090 desktop, 610.57.04 open: Xid 120 GSP page fault on suspend in both deep and s2idle. [SOURCED search result title https://github.com/NVIDIA/open-gpu-kernel-modules/issues/1284]
• Decision guide (eGPU attached):
  1. Does the box support deep? (cat /sys/power/mem_sleep). Arrow Lake NUCs are typically s2idle-only [INFERRED]. If deep exists and the eGPU hangs on s2idle, try mem_sleep_default=deep.
  2. If s2idle-only: keep NVreg_PreserveVideoMemoryAllocations=0, NVreg_DynamicPowerManagement=0x00; add a systemd-sleep pre hook that stops GPU users and unbinds nvidia from the eGPU (or de-authorises the Thunderbolt device, which “is the same as hot-remove”) and rebinds/re-authorises post-resume. [SOURCED de-auth semantics: https://www.kernel.org/doc/html/latest/admin-guide/thunderbolt.html] Hook itself is [INFERRED].
  3. If CUDA state must survive: PreserveVideoMemoryAllocations=1, enable nvidia-suspend/resume/hibernate services, size NVreg_TemporaryFilePath for full VRAM (16 GB for a 5080) on a fast filesystem; expect long suspend/resume through the tunnel. [SOURCED powermanagement.html]
  4. If it still hangs: bisect with pm_test (devices vs platform) as in #1376, and treat it as a driver-side regression to report; do not add PM knobs. [SOURCED #1376]
• Hibernate / S4: thunderbolt “Fix S4 resume incongruities” series (Jan 2026) — tunnels are rebuilt on resume; behaves like a re-plug. [SOURCED thread title https://ratatoskr.run/lkml/2026/01/3354066/t]

udev Templates

Find BDFs first: lspci -tv; the USB4 PCIe tunnel port is the host-side root port the enclosure hangs off (on the worked case: 00:07.0), the enclosure switch is 8086:5786.

A. Keep a tunnel/root port in D0 (CachyOS #1057 pattern). Needed when runtime_status=suspended on the port coincides with a missing downstream hierarchy; the surgical alternative to pcie_port_pm=off.

# /etc/udev/rules.d/90-egpu-port-pm.rules
# Host-side USB4 PCIe tunnel port (replace vendor/device with yours; AMD Phoenix example 1022:14ef)
ACTION=="add", SUBSYSTEM=="pci", ATTR{vendor}=="0x1022", ATTR{device}=="0x14ef", TEST=="power/control", ATTR{power/control}="on"
# Or pin by path (Intel NUC worked case root port 00:07.0)
ACTION=="add", SUBSYSTEM=="pci", KERNEL=="0000:00:07.0", TEST=="power/control", ATTR{power/control}="on"
# Barlow Ridge downstream/upstream ports inside the enclosure
ACTION=="add", SUBSYSTEM=="pci", ATTR{vendor}=="0x8086", ATTR{device}=="0x5786", TEST=="power/control", ATTR{power/control}="on"

[SOURCED effective workaround: https://github.com/CachyOS/linux-cachyos/issues/1057; attribute semantics: sysfs-devices-power]

B. Forbid D3cold for the eGPU and its functions. Needed when the GPU/audio function is being put in D3cold by the platform (ACPI power resources on the port) rather than by a suspended port — verify with power_state. Not needed when RTD3 is 0x00 and the port stays in D0 (worked case: d3cold_allowed=1 was fine).

# /etc/udev/rules.d/91-egpu-no-d3cold.rules
ACTION=="add", SUBSYSTEM=="pci", ATTR{vendor}=="0x10de", ATTR{class}=="0x030000", ATTR{d3cold_allowed}="0"
ACTION=="add", SUBSYSTEM=="pci", ATTR{vendor}=="0x10de", ATTR{class}=="0x040300", ATTR{d3cold_allowed}="0"

[SOURCED semantics: https://www.kernel.org/doc/Documentation/ABI/testing/sysfs-bus-pci; when-needed logic is INFERRED]

C. Neutralise the notebook-oriented NVIDIA PM rules on an eGPU box (if 80-nvidia-pm.rules is installed): copy it to /etc/udev/rules.d/80-nvidia-pm.rules with the audio-removal line deleted and the bind→auto lines changed to on, or mask it with an empty file of the same name. [SOURCED rule content: dynamicpowermanagement.html; the override technique is standard udev practice, INFERRED]

Apply: udevadm control --reload && udevadm trigger --subsystem-match=pci (or re-plug). Verify: grep . /sys/bus/pci/devices/<BDF>/power/{control,runtime_status} /sys/bus/pci/devices/<BDF>/{d3cold_allowed,power_state}.

Anti-patterns

1. pci=noaer as a “fix”. It removes the only precursor telemetry; the 2026 noaer_recovery/nodpc proposal exists precisely because people want logging without recovery, and even that got “we should fix them” pushback. [SOURCED https://ratatoskr.run/linux-doc/2026/07/17238362/t]
2. Believing pcie_aspm=off disables ASPM. It is documented as “leave any configuration done by firmware unchanged”. Want links in L0 only? pcie_aspm.policy=performance. [SOURCED https://lkml.iu.edu/hypermail/linux/kernel/2404.3/06648.html]
3. pcie_aspm=force on an eGPU chain. It enables ASPM on devices that do not advertise it, inviting instability. [SOURCED RHEL ASPM doc]
4. Global pcie_port_pm=off when one port rule suffices — CachyOS #1057 calls the global switch “undesirable”. [SOURCED https://github.com/CachyOS/linux-cachyos/issues/1057]
5. pcie_ports=native on a firmware-first platform without checking _OSC — documented “may cause conflicts”. [SOURCED kernel-parameters]
6. RTD3 (0x01/0x02) or power/control=auto on a tunnelled eGPU. The port-D3 path is the known loss mode; amdgpu now disables runtime PM for eGPUs outright. [SOURCED https://ratatoskr.run/amd-gfx/2026/08/17400523/t]
7. PreserveVideoMemoryAllocations=1 with a 16 GB eGPU and default /tmp. Saved VRAM must fit in NVreg_TemporaryFilePath, so size it for the full 16 GB first. [SOURCED powermanagement.html requirements]
8. Stacking every parameter from forum posts, then crediting the stack. In the worked case only two of seven cmdline knobs were causal; keep the rest as labelled insurance and re-test removal one at a time on a kernel upgrade.
9. Treating “D3cold to D0, device inaccessible” as a PM misconfiguration. It is the PCI core reporting unreachable config space. [SOURCED https://lkml.iu.edu/hypermail/linux/kernel/1908.2/07271.html]
10. Hot-unplugging an in-use eGPU expecting the driver to cope. NVIDIA: not guaranteed; the Thunderbolt driver maintainer: some GPU drivers “are not yet” prepared. [SOURCED egpu.html; host_reset thread]

Related references added later: egpu-idle-power-and-energy-accounting-linux.md (idle watts, persistence mode, power limits and energy cost).

Sources

• Kernel parameters (rendered): https://docs.kernel.org/admin-guide/kernel-parameters.html
• ASPM doc-fix patch (pcie_aspm=off semantics), PCI maintainer, Apr 2024: https://lkml.iu.edu/hypermail/linux/kernel/2404.3/06648.html
• RHEL ASPM guide (policies, force): https://docs.redhat.com/en/documentation/red_hat_enterprise_linux/7/html/power_management_guide/aspm
• RFC v2 “pcie/aspm: Enable all hardware power-saving states by default”, RFC author, 2026-05-11: https://lkml.iu.edu/2605.1/06202.html
• “PCI/ASPM: Enable L0s/L1 for removable devices” thread (Windows vs Linux eGPU lspci), patch author, May 2026: https://ratatoskr.run/linux-pci/2026/05/3543749/t
• PCI power management (power/pci.rst): https://www.kernel.org/doc/html/latest/power/pci.html
• sysfs-devices-power ABI: https://www.kernel.org/doc/Documentation/ABI/testing/sysfs-devices-power
• sysfs-bus-pci ABI (d3cold_allowed, power_state): https://www.kernel.org/doc/Documentation/ABI/testing/sysfs-bus-pci
• pci_bridge_d3_possible() context + pcie_port_pm patch: https://www.mail-archive.com/[email protected]/msg1614295.html
• “PCI: Allow D3 for native Hotplug” v2 (2026-07): https://ratatoskr.run/lkml/2026/07/17337075/t
• “PCI: Whitelist Thunderbolt ports for runtime D3” (runtime-PM patch author): https://patchwork.kernel.org/project/linux-pci/patch/10206fe7f3967aa73ade5b24dc729de0e94c3b7f.1529173804.git.lukas@wunner.de/
• CachyOS #1057 (Barlow Ridge tunnel runtime-suspend loss): https://github.com/CachyOS/linux-cachyos/issues/1057
• CachyOS #1021 (JHL9480 bridge inaccessible): https://github.com/CachyOS/linux-cachyos/issues/1021
• CachyOS #794 (Razer Core X regression 6.19.10): https://github.com/CachyOS/linux-cachyos/issues/794
• Thunderbolt host-reset regression thread (2026-08): https://ratatoskr.run/linux-usb/2026/08/17480231/t
• thunderbolt nhi.c (host_reset param): https://github.com/torvalds/linux/blob/master/drivers/thunderbolt/nhi.c
• “thunderbolt: Add kernel param for CLx disabling”: https://www.spinics.net/lists/linux-usb/msg219384.html
• “thunderbolt: Disable CL states on …” (2026-02): https://ratatoskr.run/linux-usb/2026/02/8892987/t
• Thunderbolt runtime-resume-during-hotplug bug (2026-03): https://ratatoskr.run/linux-pci/2026/03/3483145/t
• Thunderbolt admin guide: https://www.kernel.org/doc/html/latest/admin-guide/thunderbolt.html
• “PCI: mark Thunderbolt-attached devices removable” (2026-04): https://ratatoskr.run/lkml/2026/04/3535291/t
• “drm/amdgpu: Disable runtime PM for eGPU” (2026-08): https://ratatoskr.run/amd-gfx/2026/08/17400523/t
• AER HOWTO: https://www.kernel.org/doc/html/latest/PCI/pcieaer-howto.html
• PCI error recovery (DPC): https://www.kernel.org/doc/html/latest/PCI/pci-error-recovery.html
• AER sysfs ABI: https://www.kernel.org/doc/Documentation/ABI/testing/sysfs-bus-pci-devices-aer
• pcie_ports=dpc-native patch: https://www.spinics.net/lists/linux-pci/msg88329.html
• pci=noaer original patch: https://lkml.iu.edu/hypermail/linux/kernel/0710.0/1857.html
• pci=noaer_recovery / pci=nodpc proposal (2026-07): https://ratatoskr.run/linux-doc/2026/07/17238362/t
• pci=realloc patch (semantics): https://lkml.iu.edu/hypermail/linux/kernel/1202.2/03626.html
• “PCI / PM: Return error when changing power state from D3cold”: https://lkml.iu.edu/hypermail/linux/kernel/1908.2/07271.html
• pciehp ignore Link Down caused by DPC: https://mailweb.openeuler.org/archives/list/[email protected]/message/4WFNKWQPUH7ZXD3QIGGT24YWPJX4F5WY/
• PCI/DPC ignore Surprise Down on hot removal: https://lkml.iu.edu/2311.3/06842.html
• pciehp surprise-removal RFC (2026-09): https://ratatoskr.run/linux-pci/2026/09/17519182/t
• NVIDIA README ch. Runtime D3: https://download.nvidia.com/XFree86/Linux-x86_64/580.65.06/README/dynamicpowermanagement.html
• NVIDIA README ch. Power management (PreserveVideoMemoryAllocations, S0ix): https://download.nvidia.com/XFree86/Linux-x86_64/580.65.06/README/powermanagement.html
• NVIDIA README ch. External and Removable GPUs: https://download.nvidia.com/XFree86/Linux-x86_64/515.65.01/README/egpu.html
• NVIDIA Xid docs index: https://docs.nvidia.com/deploy/xid-errors/index.html
• NVIDIA open-gpu-kernel-modules #1117 (s2idle 7.0 regression): https://github.com/NVIDIA/open-gpu-kernel-modules/issues/1117
• #1376 (RTX 5070 Ti eGPU s2idle hang): https://github.com/NVIDIA/open-gpu-kernel-modules/issues/1376
• #900 (5090 Xid 79 over OCuLink): https://github.com/NVIDIA/open-gpu-kernel-modules/issues/900
• Xid 79 forum thread: https://forums.developer.nvidia.com/t/xid-79-gpu-has-fallen-off-the-bus/359509
• Sleep states doc: https://www.kernel.org/doc/html/latest/admin-guide/pm/sleep-states.html
• L1 substates explainers: https://www.synopsys.com/articles/reduce-power-consumption.html ; https://semiengineering.com/using-pci-express-l1-sub-states-to-minimize-power-consumption-in-advanced-process-nodes/