Revert "KVM: Prevent module exit until all VMs are freed"
Revert KVM's misguided attempt to "fix" a use-after-module-unload bug that was actually due to failure to flush a workqueue, not a lack of module refcounting. Pinning the KVM module until kvm_vm_destroy() doesn't prevent use-after-free due to the module being unloaded, as userspace can invoke delete_module() the instant the last reference to KVM is put, i.e. can cause all KVM code to be unmapped while KVM is actively executing said code. Generally speaking, the many instances of module_put(THIS_MODULE) notwithstanding, outside of a few special paths, a module can never safely put the last reference to itself without creating deadlock, i.e. something external to the module *must* put the last reference. In other words, having VMs grab a reference to the KVM module is futile, pointless, and as evidenced by the now-reverted commit 70375c2d ("Revert "KVM: set owner of cpu and vm file operations""), actively dangerous. This reverts commit 405294f2 and commit 5f6de5cb. Fixes: 405294f2 ("KVM: Unconditionally get a ref to /dev/kvm module when creating a VM") Fixes: 5f6de5cb ("KVM: Prevent module exit until all VMs are freed") Link: https://lore.kernel.org/r/20231018204624.1905300-4-seanjc@google.comSigned-off-by: Sean Christopherson <seanjc@google.com>
Showing
Please register or sign in to comment