(9) RFR: 8064815 Zero+PPC64: Stack overflow when running Maven

[Volker Simonis]

On Mon, Nov 17, 2014 at 2:54 PM, Severin Gehwolf <sgehwolf at redhat.com> wrote:
> Hi Volker,
>
> On Fri, 2014-11-14 at 14:46 +0100, Volker Simonis wrote:
>> Hi Severin,
>>
>> I can sponsor this change if we get one more review.
>
> Thanks!
>
>> The only comment I have is that in ZeroStack::suggest_size() there
>> doesn't seem to be a handling for the potentially negative values
>> returned by ZeroStack::abi_stack_available().
>
> OK. I could do something like this.
>
> diff --git a/src/cpu/zero/vm/stack_zero.cpp
> b/src/cpu/zero/vm/stack_zero.cpp
> --- a/src/cpu/zero/vm/stack_zero.cpp
> +++ b/src/cpu/zero/vm/stack_zero.cpp
> @@ -30,7 +30,9 @@
>
>  int ZeroStack::suggest_size(Thread *thread) const {
>    assert(needs_setup(), "already set up");
> -  return align_size_down(abi_stack_available(thread) / 2, wordSize);
> +  ssize_t abi_available = abi_stack_available(thread);
> +  assert(abi_available >= 0, "available abi stack must be >= 0");
> +  return align_size_down(abi_available / 2, wordSize);
>  }
>
> Did you have something else in mind? I'm not sure if it's necessary
> since suggest_size is only being called when the zero stack is set up
> (src/cpu/zero/vm/stubGenerator_zero.cpp). Thoughts?
>

I think an assertion would be good (but 'abi_available' should be an
'int' because 'ssize_t' is only specified to hold values  from -1 to
SSIZE_MAX).

As far as I can see, the possibly negative result of
ZeroStack::suggest_size() is casted into a size_t in
StubGenerator::call_stub() before doing an alloca() and just a few
lines later we call EntryFrame::build() which will do an overflow
check. Nevertheless its better to fail early in the case of a problem.

Thank,
Volker

> Thanks,
> Severin
>
>> On Fri, Nov 14, 2014 at 2:18 PM, Severin Gehwolf <sgehwolf at redhat.com> wrote:
>> > Hi,
>> >
>> > Could I please get a review and sponsor for the following fix:
>> >
>> > bug: https://bugs.openjdk.java.net/browse/JDK-8064815
>> > webrev:
>> > https://jerboaa.fedorapeople.org/bugs/openjdk/JDK-8064815/webrev.0/
>> >
>> > When running Maven ("mvn") on a Zero variant build on PPC/PPC64 hardware
>> > it throws a StackOverflowError. This is because the stack bound
>> > calculation does not account for red and yellow pages.
>> >
>> > The bug has a slightly different patch attached. The changes to
>> > hotspot/src/os/linux/vm/os_linux.cpp aren't needed for this bug.
>> >
>> > Testing done: A Zero variant build of OpenJDK 9 on PPC/PPC64 throws
>> > StackOverflowError without this fix and works fine with this fix
>> > applied.
>> >
>> > Note that this problem seems to surface on architectures where pages are
>> > large. PPC is one such instance. Page size there is 64KB and Zero
>> > initially sets its minimal stack allowance to 64KB (one page),
>> > src/os_cpu/linux_zero/vm/os_linux_zero.cpp. In os::init_2 this gets
>> > potentially increased if min_stack_allowed is small. The case on PPC
>> > Zero.
>> >
>> > However, then later at runtime the calculation of available stack is
>> > wrong since it does not account for red and yellow pages. Thus it thinks
>> > there is too little stack available where in fact more stack is
>> > available.
>> >
>> > Thanks,
>> > Severin
>> >
>
>
>