Kernel compromise via shared perf sample buffer

A preemption race in affected Linux kernel builds allows local users to crash or compromise the kernel through sleepable uprobe BPF programs. bpf_perf_event_output leaves preemption enabled when invoked from bpf_prog_run_array_sleepable, so preempted tasks can reuse the same perf_sample_data buffer and corrupt kernel execution. Exploitation requires local code execution and a system-wide uprobe or USDT observability program using the helper; consequences include kernel-memory disclosure, control-flow hijacking, and repeated panics.