Fix: Thunderbolt eGPU hot-unplug kernel support

[bdandy]

Thunderbolt eGPU Surprise Removal Support

Prevents kernel crashes when a Thunderbolt eGPU is unplugged unexpectedly.

Changes

Module	Files	Description
nvidia	nv-pci.c, nv.c, nv-acpi.c, nv-i2c.c, nv-rsync.c	Device removal tracking, nvInvalidateDeviceReferences()
nvidia-modeset	nvkms.c, nvkms-kapi.c, nvkms-evo.c, nvkms-dma.c	nvKmsIsDeviceValid() checks, safe event dispatch
nvidia-drm	nvidia-drm-drv.c, nvidia-drm-gem-*.c, nvidia-drm-fb.c	inSurpriseRemoval flag, skip nvKms calls on unplug
nvidia-uvm	uvm_gpu_isr.c, uvm_gpu.c, uvm_pmm_gpu.c, uvm_channel.c, uvm_gpu_semaphore.c	uvm_parent_gpu_is_accessible() guards in ISR, cleanup, and memory paths
RM core	nv_gpu_ops.c, rs_server.c, kernel_gsp.c, intr.c	gpuIsLost checks, graceful session teardown

Key Protections

• ISR handlers: Skip HAL calls when GPU not accessible
• Memory cleanup: Skip PMA/fault buffer operations for removed GPU
• Semaphore reads: Return cached values instead of reading GPU memory
• DRM objects: Skip nvKms free calls during surprise removal
• Session destroy: Warn instead of assert on orphaned devices

Testing

RTX 3060 + Thunderbolt 3: idle unplug, workload unplug, reconnect, module reload

fixes #842

[CLAassistant]

All committers have signed the CLA.

[roger-pmta]

Thanks! Fyi, this appears related to some of the crashes I was seeing on #979 and worked around in #984.

[bdandy]

I think that #984 is not related as it's about wrong detection of external gpu.

PR was tested on Thunderbolt 3 with 3060 GPU and everything working perfectly now (I was waiting for a fix more than few years).

As PR was initially created for 580.105.08 - merged it with master

Please review the changes and apply if possible!

Additionally created AUR package for those who need it right now https://aur.archlinux.org/packages/nvidia-open-egpu-dkms

[roger-pmta]

I think that #984 is not related as it's about wrong detection of external gpu.

Sorry, I should have been more clear: I'm also facing crashes on hot unplug and driver unload as well, mine just happened to be on a TB5 enclosure (which also happened to fail eGPU detection). Will try your patch alongside mine (on #984) when I get a moment and see if resolves crashes on TB5 as well 👍

[Hi-Angel]

I don't think this repository is maintained… I looked through the list of commits and PRs, and it turns out they never merge pull-requests, at least as far back as 2023 there wasn't a single one they merged.

[neon12345]

@bdandy Can you modify this patch to work without timeout on forced driver unbind?

[bdandy]

@neon12345 well the patch is not perfect, I guess there are some things added only by battle testing eGPU unplugging on my laptop many times (with games running etc) so it reflects issues found with it, however if you can explain the issue with timeout I will update the PR

[neon12345]

I switch between NVIDIA and vfio_pci for GPU passthrough on the host, and in my opinion, the driver should just unload as requested, no matter the usage count. I have now managed to get it to zero, but if you need to know what happened, you could in nv-pci.c:

1. Add a static variable to identify full driver unload (rmmod)
2. In the pci_remove function, check if the vendorId is 0xFFFFFFFF (hotplug)
3. If not, then the card is alive (unbind)