[11u] RFR(S): 8223266: PPC64: Check for branch to illegal address before checking for mem serialization

[Gustavo Romero]

On 05/03/2019 06:30 AM, Lindenmaier, Goetz wrote:
> thanks for fixing this issue this quickly.
> I put the patch into our testing.

Thanks a lot for reporting and testing it, Goetz.


> You please need to flag the bug as jdk11u-fix-request and
> add the correspornding 'Fix Request' comment referring to
> this review thread.

oh, sure. I'll do.


> Looking at the issue:
> actually the current VM would think the signal is a
> memory serialization and return true, while it is
> a real crash? (Or a simulated one as in the test?)

is_memory_serialization() never returns, so the VM never knows if it's
indeed a memory serialization case or not.

Because the issue on some old kernels forbids us to rely on 'si_addr' we
are inspecting the instruction at 'pc' to extract the registers used in
the instruction to determine the actual faulty address before calling
os::is_memory_serialize_page(), similarly to what we do to dertermine the
faulty address in get_stack_bang_address().

In doing it's assumed that SIGSEGV can only be caused due to a load/store
(Data Storage Interrupt), because it assumes it's  always possible to read
the instruction at 'pc'. This is not true, particularly for a SIGSEGV due
to a branch to an invalid address (e.g. not mapped / no permission to
read/exec address), which is generated due to Instruction Storage Interrupt
(ISI). Hence we can use the interruption type (passed in 'trap' member to
the signal handler) to discern when it's possible to inspect the 'pc' (DSI)
and when it's not possible (ISI).

The test case 13 precisely stresses that case of branching to an invalid
address by forcing a function bad pointer = 0xf and calling it (on PPC64
that crash shows pc=0xc in the hs_err log due to the code alignment
requirements for execution).

So the additional code block checking for SIGSEGV related to UseMemBar
feature on 11u needs to be adapted like the  previous block using
get_stack_bang_address(), which is already fixed on jdk/jdk tip.

Kind regards,
Gustavo