Hi Mike,
On 29/02/2012 5:23 AM, Mike Duigou wrote:
On Feb 27 2012, at 20:50 , David Holmes wrote:
Hi Mike,
The problem with changing initialization order is that you can never enumerate
all the possible initialization paths. I don't see anything problematic with
the move in relation to Thread.currentThread() or ThreadGroup.add(). But the
call to setJavaLangAccess requires initialization of SharedSecrets which in
turn requires initialization of Unsafe which in turn initializes
sun.reflect.Reflection which in turn initializes a bunch of Collection classes
- and suddenly we may have a whole swag of classes now being initialized before
the VM appears to be booted.
The class initialization I'm interested in does occur during the SharedSecrets
initialization path. Because, in the current implementation, booted() hasn't
been called the only way I can determine if the class is being used before
boot() is to check if the SharedSecrets have been initialized. This ends up
being cumbersome and permanently leaves some checking logic in the execution
path for what is a very narrow window. I've looked at where isBooted is
currently being used and none of the uses would seem to be sensitive to being
called
Could Win32ErrorMode.initialize() now be called too late in some case?
What about sun.nio.cs.ext.ExtendedCharsets.init()?
...
If you then throw into the mix the possibility of different system properties
affecting initialization actions, or the existence of an agent, then who knows
what might actually happen.
Do you believe that the change is as a result too risky to consider? Are there
validations that could/should be done to reduce or eliminate the risk?
I think this change may well have unanticipated consequences that might
take a long time to surface. If I was confident that regular testing
might expose these then I'd say go ahead, but I think these will be
obscure changes in very specific contexts that probably are not tested.
Given it is only a convenience fix is it worth any risk? You could
always add a new field to reflect the end of initializeSystemClass. Or
make booted an integer:
private static volatile int booted = 0;
public static void booted() {
booted++;
}
public static boolean isBooted() {
return booted > 0;
}
public static void fullyBooted() {
booted++;
}
public static boolean isFullyBooted() {
return booted > 1;
}
Cheers,
David
Cheers,
David
-----
On 28/02/2012 1:57 PM, Mike Duigou wrote:
Hello all;
This issue is a patch for review that I have split out of a larger issue I'll
be submitting later.
WEBREV @ http://cr.openjdk.java.net/~mduigou/7149320/0/webrev/
sun.misc.VM.booted() is currently called before setJavaLangAccess(). For code
which uses the JavaLangAccess shared secrets it's convenient to be able to
check whether the secrets are initialized without having to call
sun.misc.SharedSecrets.getJavaLangAccess() every time the secrets are used and
checking for null. In particular with the static inner class holder idiom it
would be desirable to do :
static class Holder {
final sun.misc.JavaLangAccess ACCESS =
sun.misc.SharedSecrets.getJavaLangAccess();
}
...
if(sun.misc.VM.isBooted()&& Holder.ACCESS...
In my research on this issue I was unable to determine why sun.misc.VM.booted()
isn't the currently the last activity in System.initSystemClass(). Neither of
the two tasks which currently follow it depend upon booted state in any way
that I could determine. I am tempted, thinking about it, to add a comment to
System.initSystemClass before the call to sun.misc.VM.booted() asking future
maintainers to leave boot() as the last activity or explain why not.
I have done JTReg runs on linux x86, linux x64 and Solaris 64 which
incorporated this change without any problems encountered. It's rather
difficult to write a unit test for this issue as it requires a class that is
initialized after java.lang.System but before
java.lang.System.initSystemClass() and uses JavaLangAccess.
Thanks,
Mike
