On Mon, Jul 2, 2018 at 12:00 PM, Mathieu Desnoyers <[email protected]> wrote: > ----- On Jul 2, 2018, at 1:11 PM, Andy Lutomirski [email protected] wrote: >> >> But I think that the limited solution of changing >> instruction_pointer_set() really is a sufficient >> architecture-dependent change to fully solve your problem. > > So let me recap with the changes I gather for 4.18 and 4.19: > > 4.18: > > * Change struct rseq_cs field types from LINUX_FIELD_u32_u64() to __u64 in > uapi/linux/rseq.h, > * Compare rseq->rseq_cs->abort_ip with TASK_SIZE before using it. Kill > offending > process if its value is over TASK_SIZE, > * Explicitly check that padding of rseq->rseq_cs is zero on 32-bit kernels > (#ifndef __LP64__). > > 4.19: > > * Introduce instruction_pointer_set() with input validation, use it when > setting > IP to abort_ip in rseq. This replaces the comparison of abort_ip with > TASK_SIZE. > > Is that consistent with what you have in mind ? >
Works for me. Linus, any objection?
