On Fri, Feb 5, 2021 at 6:24 AM Josh Poimboeuf jpoimboe@redhat.com wrote:
KASAN reserves "redzone" areas between stack frames in order to detect stack overruns. A read or write to such an area triggers a KASAN "stack-out-of-bounds" BUG.
Normally, the ORC unwinder stays in-bounds and doesn't access the redzone. But sometimes it can't find ORC metadata for a given instruction. This can happen for code which is missing ORC metadata, or for generated code. In such cases, the unwinder attempts to fall back to frame pointers, as a best-effort type thing.
This fallback often works, but when it doesn't, the unwinder can get confused and go off into the weeds into the KASAN redzone, triggering the aforementioned KASAN BUG.
But in this case, the unwinder's confusion is actually harmless and working as designed. It already has checks in place to prevent off-stack accesses, but those checks get short-circuited by the KASAN BUG. And a BUG is a lot more disruptive than a harmless unwinder warning.
Disable the KASAN checks by using READ_ONCE_NOCHECK() for all stack accesses. This finishes the job started by commit 881125bfe65b ("x86/unwind: Disable KASAN checking in the ORC unwinder"), which only partially fixed the issue.
Fixes: ee9f8fce9964 ("x86/unwind: Add the ORC unwinder") Reported-by: Ivan Babrou ivan@cloudflare.com Cc: stable@vger.kernel.org Signed-off-by: Josh Poimboeuf jpoimboe@redhat.com
I haven't seen any previously observed issues with this after a day of uptime.
Tested-by: Ivan Babrou ivan@cloudflare.com