On 05/23/2016 02:01 PM, Vlastimil Babka wrote: >> if I read the report correctly it's not about bpf, but rather points to >> the issue inside percpu logic. >> First __alloc_percpu_gfp() is called, then the memory is freed with >> free_percpu() which triggers async pcpu_balance_work and then >> pcpu_extend_area_map is hitting use-after-free. >> I guess bpf percpu array map is stressing this logic the most. > > I've been staring at it for a while (not knowing the code at all) and > the first thing that struck me is that pcpu_extend_area_map() is done > outside of pcpu_lock. So what prevents the chunk from being freed during > the extend? Erm to be precise, pcpu_lock is unlocked just before calling pcpu_extend_area_map(), which relocks it after an allocation, and assumes the chunk still exists at that point. Unless I'm missing something, that's an unlocked window where chunk can be destroyed by the workfn, as the report suggests? -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>