Hi Claes,
Am I reading correct that package-private
ClassLoader.loadLibrary(Class<?> fromClass, String name, boolean
isAbsolute) wouldn't need to consult SecurityManager wasn't it for
accessing system properties in private ClassLoader.initializePath(String
propName). So perhaps if the two calls to initializePath() in the if
block of ClassLoader.loadLibrary() are wrapped with a single
doPrivileged(), you wouldn't have to wrap the call to
JavaLangAccess.loadLibrary in BootLoader. And since initializePath()
would only be called once and then the results cached, you end up with
less doPrivileged() wrappings when SecurityManager is present as soon as
BootLoader.loadLibrary is called the 2nd time.
BTW, there is a data race when accessing usr_paths and sys_paths static
fields in multi-threaded scenario. Their type is String[] which means
that the contents of the arrays could be observed uninitialized
(elements being null) in some thread. Both fields should be made
volatile (not only sys_paths) as assignment to them can be performed
more than once in idempotent racy initialization.
Regards, Peter
On 7/16/19 3:48 PM, Claes Redestad wrote:
Hi,
refactored to use a BootLoader::loadLibrary API that is only visible
within the java.base module:
http://cr.openjdk.java.net/~redestad/8227587/open.03/
I've retained the bridge to ClassLoader::loadLibrary for performance,
but without any magic or privilege escalation involved. Moving sys_path
/ systemNativeLibraries out of ClassLoader seems quite complicated, so
I've not attempted that refactoring for this RFE.
/Claes
On 2019-07-12 20:03, Mandy Chung wrote:
Just to recap what we discussed offline:
We could introduce BootLoader::loadLibrary to load a system native
library for boot loader. sys_path and systemNativeLibraries in
ClassLoader implementation are solely for boot loader and it seems
cleaner for BootLoader to handle loading of system native libraries
where it can be optimized for boot loader. This would require
some refactoring.
One possible solution is to add a shared secret to simply call
package-private ClassLoader::loadLibrary(Class<?> fromClass, String
name)
method and have BootLoader::loadLibrary to check if SM is present
and wrap the call with doPriv; otherwise, just call the shared
secret loadLibrary method.
Then we can investigate in the future to refactor the native library
loading implementation to jdk.internal.loader.
what do you think?
Mandy
On 7/12/19 8:25 AM, Mandy Chung wrote:
Claes,
Thanks for exploring this. I would like to see if we can avoid
introducing
an internal @CS method (keep @CSM only for public APIs will help
security
analysis).
There are other alternatives to improve the footprint performance.
One idea is java.base and java.desktop each has its own utility method
to load library. No change is needed for java.net since only one
loadLibrary
call.
Another idea is to add BootLoader.loadLibrary that skips the security
permission overhead and only for boot loader use. It would require
refactoring.
I think java.base and java.desktop having its own utility method is
a simpler solution.
Mandy
On 7/12/19 7:55 AM, Claes Redestad wrote:
Hi,
I'm dropping the java.desktop changes, and Mandy has asked me to
explore
options that does not add a new @CallerSensitive entry point, even
to a
strongly encapsulated API like JavaLangAccess
Easiest of these is to explicitly provide the class context we're
calling from - with proper assertions. This is marginally more
performant due avoiding the Reflection.getCallerClass, but slightly
less
convenient.
http://cr.openjdk.java.net/~redestad/8227587/open.01/
/Claes
On 2019-07-12 15:27, Roger Riggs wrote:
Hi Claes,
Looks fine.
Thanks for the updates, Roger
On 7/11/19 10:39 AM, Claes Redestad wrote:
Hi Roger,
On 2019-07-11 16:10, Roger Riggs wrote:
Hi Claes,
JavaLangAccess.java:
316: Add @param tag
done.
System.java: 2282, 2287
Runtime.loadLibrary0 makes a second check for a security manager.
Is there any potential gain by calling ClassLoader.loadLibrary
directly?
None of the internal uses would have a separatorChar.
Most of the gain comes from not having to load one-off PA classes or
lambdas, but of course avoiding the indexOf and a few
indirections are
marginally helpful to startup. We should at least assert that the
library names are sane, though.
I expect most of the files need a copyright update.
The script in <repo>/make/scripts/update_copyright_year.sh
should do the work for modified files.
Fixed copyrights and updated in place:
http://cr.openjdk.java.net/~redestad/8227587/open.00
Thanks!
/Claes
