RFR(XS)(10): 8175342: assert(InstanceKlass::cast(k)->is_initialized()) failed: need to increase java_thread_min_stack_allowed calculation
david.holmes at oracle.com
Thu Mar 16 05:14:11 UTC 2017
On 16/03/2017 2:57 PM, Chris Plummer wrote:
> Please review the following:
I think you want to disable the guardpage always, not just when a
specific stack size is requested. You might not miss 64KB in 8MB but
logically the guard page is never needed.
> The assert is somewhat misleading, although it did properly detect a
> "too small" stack issue. The test was run with -Xss256k on a system with
> 64k pages. On this system 256k is suppose to be the smallest allowed
> stack size, so -Xss256k should work. The thread that the assert happens
> on is the main thread created by ContinueInNewThread0(). By default
> pthread gives new threads a guard page the size of an OS page. pthreads
> is suppose to add additional stack space for the guard page, but it
> doesn't. Later we call current_stack_region(), which among other things,
> computes the size of the stack. It has the following code to deal with
> the pthread guard page bug:
> // Work around NPTL stack guard error.
> size_t guard_size = 0;
> rslt = pthread_attr_getguardsize(&attr, &guard_size);
> *bottom += guard_size;
> *size -= guard_size;
> So the net effect is hotspot treats the stack as only being 192k, not
> 256k. However, in terms of usable stack space, hotspot then also
> subtracts the red, yellow, and shadow zones. Each of these is one OS
> page. So that subtracts another 192k, leaving us with 0k. The assert is
> a by product of this.
> It turns out this pthread guard page problem is already fixed for all
> java threads except the main thread. We do the following in
> For java threads, os::Linux::default_guard_size() returns 0, so the
> above code removes the guard page for java threads. The fix I'm
> proposing for the main thread does the same.
> Tested by running the test in question dozens of times on all supported
> platforms. Also ran most tests we do for nightlies except for longer
> running ones.
More information about the core-libs-dev