New Linux 7.4 Feature For Reducing AI Noise Already Reverted Due To A "Crazy Beast"

A new feature merged last week for Linux 7.4 to help reduce AI/LLM generated noise from invalid/irrelevant testing configurations is now reverted due to a "crazy beast" in the VFIO subsystem. Greg Kroah-Hartman back in September crafted the TAINT_FORCED_BIND option to taint the running kernel if Linux's bind/unbind sysfs interface is used. This interface allows arbitrarily binding/unbinding hardware to a given driver. In the era of AI/LLM, this interface was abused by Syzbot and friends for trying to bind hardware to drivers not relevant to one another. This resulted in noise from irrelevant bug reports from these tools/bots that don't match any legitimate kernel use-case. So with the TAINT_FORCED_BIND, it would be clear to kernel developers in bug reports that sysfs bind/unbind was used and so helping quickly deduce it may not be a legitimate bug report. The feature was merged last week for Linux 7.4 for what Greg KH characterized as a recent major fuzzing "attack" against the kernel. Only after TAINT_FORCED_BIND was it realized VFIO code is relying on the bind/unbind behavior in legitimate practice. In the revert patch, Greg described VFIO as a "crazy beast" in its usage: "Turns out that vfio is a crazy beast and people have been "relying" on the bind/unbind debugging feature as a way to handle passing through devices to guest instances. Because of this (https://xkcd.com/1172/), let's revert the feature for now until we can figure out a more [agreeable] way to handle this such that normal workflows do not taint the kernel. Odds are something like what Uwe proposes at: https://lore.kernel.org/lkml/[email protected]/ will be the best way forward and will be revisited after for after 7.4. The feature was originally intended to stem the flow of unnecessary fuzzing reports of "broken" drivers where random devices were being bound to random drivers and of course things were going wrong. The main syzbot tool is now not doing this anymore, and hopefully no one else starts it up again (but in this madness of LLMs these days, you can't be sure.)" VFIO relies on bind/unbind for detaching PCIe devices from a Linux host driver and attaching to vfio-pci for use within VMs. The revert was merged on Friday night. Uwe Kleine-König's proposal is for using TAINT_DRIVER_OVERRIDE just in the usage of driver_override and disabling driver_override by default while allowing Linux driver the option of allowing themselves into it. We'll see if that ends up being adopted for Linux v7.5 but is now too late for v7.4 with the merge window expected to close tomorrow.