Date: Mon, 12 Aug 2019 20:33:05 -0700 From: Greg Lewis <glewis@eyesbeyond.com> To: freebsd-hackers@freebsd.org Subject: Re: Java stack overflow segfaults Message-ID: <20190813033305.GA85090@misty.eyesbeyond.com> In-Reply-To: <20190812161629.GA99971@misty.eyesbeyond.com> References: <20190812161629.GA99971@misty.eyesbeyond.com>
next in thread | previous in thread | raw e-mail | index | archive | help
On Mon, Aug 12, 2019 at 09:16:29AM -0700, Greg Lewis wrote: > I'm investigating an issue where, on FreeBSD, Java will crash rather than > throw a StackOverflowError given a simple test program with a function > that just calls itself over and over. There's an example of such a test > in https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=222146 > > This affects, I suspect, every native version of Java in the ports tree, > although I've only tried openjdk8 and higher. My investigation has mostly > focused on openjdk11. > > To outline the situation, Java uses pthreads internally for threading. It > doesn't use the pthreads own guard page(s), but instead creates it's own > guard area at the bottom of the stack (which grows downward) using > mprotect. It then installs a signal handler and examines any SIGSEGV's > fault address to see if it falls within the guard area, and if so throws a > StackOverflowError. This logic is the same across all of the OSes I've > looked at and works on OpenBSD, Linux, etc. On FreeBSD though, the fault > address lies in the page above the guard zone, rather than in the guard > zone, which results in a crash rather than throwing StackOverflowError. > > An diagram may help here: > > --- <- Stack top > | > | Untouched memory + stack frames + etc. > | > | > | <-- SIGSEGV signal info fault address (< 1 page above guard zone) > --- <- Start of JVM reserved zone / guard zone > | > | JVM Reserved page > | > --- <- Start of JVM yellow zone > | > | JVM Yellow pages > | > --- <- Start of JVM red zone > | > | JVM Red page > | > --- <- Stack bottom > | > | Pthread guard page(s) > | > --- > > On my FreeBSD 11.3/amd64 machine the JVM uses a total of four pages for the > guard zone (1 reserved, 2 yellow, 1 red). The page size is 4K, and I see > the follow mprotect calls with truss: > > mprotect(stack bottom address, 4K, PROT_NONE) (Just the red zone) > mprotect(stack bottom address, 16K, PROT_NONE) (The entire guard zone) > mprotect(top of red zone address, 12K, PROT_READ|PROT_WRITE) (Reserved + yellow) > mprotect(top of red zone address, 12K, PROT_NONE) (Reserved + yellow) > > While I've committed a workaround for openjdk8, which just rounds down the > fault address, it isn't entirely satisfactory (it's a hack) and I wondered > if anyone had any insight into what may be going on. I've done an analysis > of the sizes and addresses being used and used truss to check the parameters > to the mprotect calls, and everything appears to add up. > > The same problem also occurs under FreeBSD 12.0/i386 and on aarch64, so it > doesn't appear to be either version or platform specific. I've simplified > a little here, but am happy to provide additional details and code > references. This appears to be due to security.bsd.stack_guard_page and setting that to different values alters the extra space which may contain the fault address. -- Greg
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20190813033305.GA85090>