RFR(S): 8170548: VM may crash at startup because StdoutLog/StderrLog logging stream can be badly aligned

[Thomas Stüfe]

On Thu, Dec 1, 2016 at 2:32 PM, Volker Simonis <volker.simonis at gmail.com>
wrote:

> On Thu, Dec 1, 2016 at 2:23 PM, David Holmes <david.holmes at oracle.com>
> wrote:
> > Hi Volker,
> >
> >
> > On 1/12/2016 8:35 PM, Volker Simonis wrote:
> >>
> >> Hi,
> >>
> >> can I please have a review and sponsor for the following fix:
> >>
> >> http://cr.openjdk.java.net/~simonis/webrevs/2016/8170548/
> >> https://bugs.openjdk.java.net/browse/JDK-8170548
> >>
> >> Change "8146009: "pure virtual method called" with using new GC
> >> logging mechanism" introduced a sophisticated initialization mechanism
> >> for the logging stream. In order to avoid deconstruction of the
> >> streams before the VM exits, it creates them with a placement new into
> >> statically allocated memory:
> >>
> >> static bool initialized;
> >> static char stdoutmem[sizeof(LogStdoutOutput)];
> >> static char stderrmem[sizeof(LogStderrOutput)];
> >>
> >> LogStdoutOutput &StdoutLog =
> >> reinterpret_cast<LogStdoutOutput&>(stdoutmem);
> >> LogStderrOutput &StderrLog =
> >> reinterpret_cast<LogStderrOutput&>(stderrmem);
> >>
> >> LogFileStreamInitializer::LogFileStreamInitializer() {
> >>   if (!initialized) {
> >>     ::new (&StdoutLog) LogStdoutOutput();
> >>     ::new (&StderrLog) LogStderrOutput();
> >>     initialized = true;
> >>   }
> >> }
> >>
> >> Unfortunately it is not guaranteed, that the static memory (which is a
> >> char array) is well-aligned for the stream objects. Actually, the C++
> >> standard only defines that it has to be at least 'char' aligned which
> >> is obviously not enough for a stream object.
> >> When building 'slowdebug' on Solaris with SS12u4 we indeed observed
> >> reproducible crashes during VM initialization because of this issue.
> >>
> >> The fix is easy - just wrap the character arrays into unions to align
> >> them appropriately.
> >
> >
> > Seems reasonable - though I don't know the C++ alignment rules for static
> > unions to know whether this is guaranteed to be correct, or just very
> likely
> > to be correct.
> >
>
> The rule is that the union is aligned such that all of its members are
> correctly aligned. With the fix, the union should be correctly aligned
> for jlong which should be 'good enough' for
> LogStdoutOutput/LogStderrOutput.
>
> > I'm also wondering if anyone knows exactly why only the Solaris slowdebug
> > build detected the problem?
> >
>
> Probably because the Sun Studio compiler takes -O0 very seriously and
> places the character arrays at the next available address while other
> compilers still aligned at least 'good enough' for the object
> placement.
>
>
Also, do not some platform just swallow unaligned accesses quietly with
just a speed penalty? I think this is the case for x86.


> > Thanks,
> > David
> >
> >
> >> Thank you and best regards,
> >> Volker
> >>
> >
>