All we need is to store two variables (base and size) and use them if
the define a suitable range. So the code is completely platform agnostic.
We found this useful for Linux only for the standard JDK platforms, but
maybe when people port to other platforms they would find this useful as
well?
Thanks
- Ioi
On 5/7/14, 2:55 AM, David Holmes wrote:
Hi Ioi,
Thanks for the background. I understand this only applies to Linux but
that is the reason it should be confined to Linux-specific files as
much as possible. This wouldn't be the first time we introduced an os
method that only had a non-trivial implementation on one platform.
Thanks,
David
On 7/05/2014 4:14 AM, Ioi Lam wrote:
So backgrounds:
The C code in HotSpot is platform-independent, and will take no effect
if jdk/src/share/bin/main.c does not call JLI_SetStaticSharedSpace().
Since the code inside HotSpot is pretty small, we could always leave it
in (compiled unconditionally for both 32-bit and 64-bit) and push the
platform-dependent decision to the JDK layer.
The need for this fix is only for 32-bit VMs, where ASLR would more
easily cause address conflicts with the predefined SharedBaseAddress
(see globals.hpp). On 64-bit VM, mapping failure due to ASLR is very
rare.
The fix is based on the idea that a variable defined in the BSS section
of the main executable is not randomized on some platforms:
static char staticSharedSpace[32*1024*1024];
so far, we found this is true only in Linux. (And perhaps Solaris as
well, but since we only supported 64-bit on Solaris, I didn't
investigate further).
On MacOS and Windows, the BSS section of the main executable is also
randomized, so this fix will not reduce mapping failures.
- Ioi
On 5/5/14, 7:59 PM, David Holmes wrote:
Hi Yumin,
For something that only addresses one platform this adds a lot of
platform specific code to a bunch of shared files. Are there plans to
extend this to other platforms? Otherwise can we make this more
platform agnostic? Perhaps move the os specific code to the os class?
Seems the changes in filemap.cpp and metaspace.cpp could easily be
factored into an os method that returns a boolean to indicate whether
there is a fixed mapping.
And perhaps define set_static_shared_space in the os class and always
expose it via jni.cpp. Actually, why are we exposing this via jni.cpp
rather than jvm.cpp? This is not part of the JNI interface to the VM.
Thanks,
David
On 6/05/2014 9:15 AM, Yumin Qi wrote:
Hi, Please have codereview for
8042243: Map shared archive to preallocated static address
webrev: http://cr.openjdk.java.net/~minqi/8042243/webrev00/
bug: https://bugs.openjdk.java.net/browse/JDK-8042243
Summary: Mapping shared archive (jsa) some time fail due to ASLR
(Address space layout randomization) on 32bit Linux platforms. To
solve
we come up with mapping shared archive to preallocated static space
since with Linux and the GCC compiler, the BSS section in the main
executable is not randomized, instead, the variable address in the BSS
section is fixed at build time. The tests also showed the extra 32M
space will NOT affect java memory setting though it seems 32M virtual
address space always allocated no matter it is used or not (If
-XX:ShareBaseAddress supply alternative address for dumping, the
static
address will be ignored.) Note the changes is for 32 bit Linux only.
Tests: JPRT, manual test for archive sharing.
Thanks
Yumin