On 09/09/22(Fri) 14:41, Martin Pieuchot wrote:
> On 09/09/22(Fri) 12:25, Theo Buehler wrote:
> > > Yesterday gnezdo@ fixed a race in uvn_attach() that lead to the same
> > > assert. Here's an rebased diff for the bug discussed in this thread,
> > > could you try again and let us know? Thanks!
> >
> > This seems to be stable now. It's been running for nearly 5 days.
> > Without gnezdo's fix it would blow up within at most 2 days.
>
> Thanks! I'm looking for oks then.
Here's an alternative possible fix. The previous one got reverted
because it exposes a bug on arm64 machines with Cortex-A72 CPUs.
The idea of the diff below is to flush data to physical pages that we keep
around when munmap(2) is called. I hope that the page daemon does the right
thing and don't try to grab a reference to the vnode if all pages are PG_CLEAN.
Could you try that and tell me if this prevents the panic you're seeing?
Index: uvm/uvm_vnode.c
===================================================================
RCS file: /cvs/src/sys/uvm/uvm_vnode.c,v
retrieving revision 1.130
diff -u -p -r1.130 uvm_vnode.c
--- uvm/uvm_vnode.c 20 Oct 2022 13:31:52 -0000 1.130
+++ uvm/uvm_vnode.c 9 Nov 2022 16:08:57 -0000
@@ -329,7 +329,7 @@ uvn_detach(struct uvm_object *uobj)
*/
if (uvn->u_flags & UVM_VNODE_CANPERSIST) {
/* won't block */
- uvn_flush(uobj, 0, 0, PGO_DEACTIVATE|PGO_ALLPAGES);
+ uvn_flush(uobj, 0, 0, PGO_CLEANIT|PGO_DEACTIVATE|PGO_ALLPAGES);
goto out;
}
