uvm: fix NULL pointer dereference in uvm_hmm_unregister_gpu

[JCorners68]

Summary

Fixes #1082 — kernel NULL pointer dereference in uvm_hmm_unregister_gpu() during CUDA process exit.

gpu->parent->devmem is dereferenced unconditionally, but it is only initialized when uvm_pmm_devmem_init() succeeds during init_parent_gpu(). When it is NULL, the dereference crashes the kernel:

BUG: kernel NULL pointer dereference, address: 00000000000000a8
RIP: uvm_hmm_unregister_gpu+0x51/0x3a0 [nvidia_uvm]

The existing uvm_hmm_is_enabled(va_space) guard only checks the VA space — it does not guarantee the specific GPU has device-private memory initialized.

Fix

• Guard the devmem dereference and page-eviction loop with if (gpu->parent->devmem).
• When devmem is NULL there are no device-private pages to evict, so skip straight to HMM block unregistration.
• Move lock assertions above the devmem access so they fire before the NULL dereference (secondary bug: in the original code, the assertions were placed after the crash site and could never trigger).

Reproduction

Observed repeatedly on two different systems:

• RTX 5080 + RTX PRO 6000 / i9-14900KF / kernel 6.17.0-20-generic / driver 595.58.03
• RTX 3080 Mobile / i7-12700H / kernel 6.19.9 / driver 595.58.03 (original NULL pointer dereference in uvm_hmm_unregister_gpu during CUDA UVM VA space teardown on process exit #1082 report)

Triggered by any CUDA process exit (training jobs, nvtop, tokio async runtime, fossilize shader compilation). The kernel oops corrupts driver state and cascades to Xid 79 ("GPU has fallen off the bus") within minutes to hours, requiring a full node reboot.

Testing

• Confirmed options nvidia_uvm uvm_disable_hmm=1 (which skips this code path entirely) eliminates the crash, validating that the fault is in this function.
• The NULL guard is the minimal targeted fix for the same code path.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.

---

jcorners68 seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}