Messages in this thread | | | From | Thomas Gleixner <> | Subject | Re: [PATCH] genirq: fix reference leaks on irq affinity notifiers | Date | Tue, 17 Mar 2020 20:25:35 +0100 |
| |
Ben Hutchings <ben@decadent.org.uk> writes: > On Tue, 2020-03-17 at 10:58 +0000, Edward Cree wrote: >> On 15/03/2020 20:29, Ben Hutchings wrote: >> > ...since the pending work item holds a reference to the notification >> > state, it's still not clear to me why or whether "genirq: Prevent use- >> > after-free and work list corruption" was needed. >> Yeah, I think that commit was bogus. The email thread[1] doesn't >> exactly inspire confidence either. I think the submitter just didn't >> realise that there was a ref corresponding to the work; AFAICT there's >> no way the alleged "work list corruption" could happen. >> >> > If it's reasonable to cancel_work_sync() when removing a notifier, I >> > think we can remove the kref and call the release function directly. >> I'd prefer to stick to the smaller fix for -rc and stable. But if you >> want to remove the kref for -next, I'd be happy to Ack that patch. > > OK, then you can add: > > Acked-by: Ben Hutchings <ben@decadent.org.uk> > > to this one. > >> Btw, we (sfc linux team) think there's still a use-after-free issue in >> the cpu_rmap lib, as follows: >> 1) irq_cpu_rmap_add creates a glue and notifier, adds glue to rmap->obj >> 2) someone else does irq_set_affinity_notifier. >> This causes cpu_rmap's notifier (old_notify) to get released, and so >> irq_cpu_rmap_release kfrees glue. But it's still in rmap->obj >> 3) free_irq_cpu_rmap loops over obj, finds the glue, tries to clear its >> notifier. >> Now one could say that this UAF is academic, since having two bits of >> code trying to register notifiers for the same IRQ is broken anyway >> (in this case, the rmap would stop getting updated, because the >> "someone else" stole the notifier). > > So far as I can remember, my thinking was that only non-shared IRQs > will have notifiers and only the current user of the IRQ will set the > notifier. The doc comment for irq_set_affinity_notifier() implies the > latter restriction, but it might be worth spelling this out explicitly.
Bah. I so wish these notifiers would have never been introduced at all.
| |