Detect and quarantine wedged vGPU VFs - #435

Open

yummybomb wants to merge 20 commits into

hypeship/vendor-vfio-vgpufrom

hypeship/vgpu-wedge-quarantine

Commits on Aug 24, 2026

Quarantine wedged vGPU VFs convicted by the guest sentinel
yummybomb
committed
9809454
Report GPU init failure from the guest agent and harden quarantine
yummybomb
committed
c81b8c5
Close sentinel replay, fail-open, and persist-retry holes
yummybomb
committed
87427ff
Match only kernel-facility kmsg records in the GPU init watch
yummybomb
committed
c60db3b
Bound vGPU sentinel scans to the line cap
yummybomb
committed
e45eec3
Signal an unavailable VF health store
yummybomb
committed
3b6ac58
Repeat the GPU init-failure marker so printk splits cannot lose a report
yummybomb
committed
e572d84
Remove the conviction brake
yummybomb
committed
aae0848
Read /dev/kmsg with a record-sized buffer
yummybomb
committed
4bf5190
Fsync VF health state writes
yummybomb
committed
d4c82ff
Persist only the matched sentinel marker
yummybomb
committed
b21dfd9
Trim unused VF health and sentinel surface
yummybomb
committed
a79ddb7
Re-probe framework discovery while the sentinel gate is unknown
yummybomb
committed
9ab3db3
Revalidate the assignment before conviction and list targets strictly
yummybomb
committed
e059162
Document draining the parent GPU before an SR-IOV cycle
yummybomb
committed
cca5ec4
Deduplicate the wedge remediation docs
yummybomb
committed
b97ebe2
Convict on a released assignment instead of skipping
yummybomb
committed
381d2b6
Harden vGPU quarantine accounting
yummybomb
committed
f01a81a
Validate persisted VF health state
yummybomb
committed
da8eed0
Require full vGPU sentinel log lines
yummybomb
committed
b2d725f

Navigation Menu

Uh oh!

Detect and quarantine wedged vGPU VFs - #435

Detect and quarantine wedged vGPU VFs#435
yummybomb wants to merge 20 commits into
hypeship/vendor-vfio-vgpufrom
hypeship/vgpu-wedge-quarantine

Commits on Aug 24, 2026

Quarantine wedged vGPU VFs convicted by the guest sentinel

Report GPU init failure from the guest agent and harden quarantine

Close sentinel replay, fail-open, and persist-retry holes

Match only kernel-facility kmsg records in the GPU init watch

Bound vGPU sentinel scans to the line cap

Signal an unavailable VF health store

Repeat the GPU init-failure marker so printk splits cannot lose a report

Remove the conviction brake

Read /dev/kmsg with a record-sized buffer

Fsync VF health state writes

Persist only the matched sentinel marker

Trim unused VF health and sentinel surface

Re-probe framework discovery while the sentinel gate is unknown

Revalidate the assignment before conviction and list targets strictly

Document draining the parent GPU before an SR-IOV cycle

Deduplicate the wedge remediation docs

Convict on a released assignment instead of skipping

Harden vGPU quarantine accounting

Validate persisted VF health state

Require full vGPU sentinel log lines