On 10/27/2011 1:17 PM, Peter wrote:
That's exactly what the original implementers needed to do, make those fields
volatile.
They're private implementation fields though.
Okay, but are there any usage patterns where we could add the use of a volatile?
The JMM says that multiple non-volatile fields can be made visible by a single
volatile write and then another thread making a volatile read. So if we are
adding properties, we should do a volatile write after that, and then if there
is a place before the use of permissioncollection, by another thread, that we
can force a volatile read of the same volatile field, then that should fix the
visibility. It's not pretty...
Should we create an issue in the JDK bugzilla?
Gregg
Trouble is, none of these old jvm homogenous PermissionCollection's have been
exposed to any more than single threads before and the last thing I want to do
is reimplement them. They're supposed to be thread safe but many have
visibility issues.
Considering java security policy is a occassional write, multi read, it should
be simple to make it scale very well, using immutability and concurrency utils.
There's just some legacy cruft that spoils it a little.
I guess I could make a wrapper class that uses volatile and write replace, but
then if it changes you still have to replace the underlying
PermissionCollection, and still wear the synchronisation cost.
Cheers,
Peter.
Cheers,
Peter.
----- Original message -----
What about a volatile as the visibility control? Write after update, read
before access? It would at least expose the changes to other threads, not be a
lock, and represent a fairly limited overhead on most hardware.
Gregg
On 10/27/2011 8:55 AM, Peter wrote:
The problem:
Stale references allowed and noted in comments:
java.security.Permissions
java.security.BasicPermissions.BasicPermissionCollection
The stale reference in Permissions is an AllPermission object - an
optimisation. If a thread doesn't see the current value, it just checks the
internal Map, which is synchronised, no biggy.
Problem is, Permissions is a heterogenous PermissionCollection, it contains a
Map, synchronzed thread access, which prevents a similar optimisation in the
homogenous BasicPermissionCollection from being seen in the stale state.
Every ProtectionDomain has its own Permissions and each Permission class type
has it's own unique PermissionCollection shared with all others with the same
type for a ProtectionDomain.
I replaced Permissions with a class called ConcurrentPermissions that uses a
ConcurrentMap
Trouble is BasicPermissionCollection is no longer protected by synchronization
in Permissions. BasicPermissionCollection now exposed to multiple threads has
a stale reference optimisation for wildcard * permissions.
What happens in my concurrent policy implementation is the Permission isn't
necessarily found in the BasicPermissionCollection by a second thread, so it
checks the PermissionGrants (immutable objects that contain data from policy
files or dynamic grants) again and adds all the permissions to
BasicPermissionCollection again. So it doesn't fail, but it doesn't scale
well with contention, because you've still got the synchronisation bottleneck,
can't see the Permission and have to process again, wasting resources, on the
second occassion.
Problem is, BasicPermissionCollection is the bread and butter
PermissionCollection implementation many Permission classes use.
Now you have to remember, these classes were designed well before concurrency
was ever a consideration. Nowadays these classes would be immutable, since
policy's don't change much, they're mostly read access.
But I can't change it because many are part of the decision process.
Now I could put a synchronized wrapper PermissionCollection class around these
things, which fixes the bug, creating long lived objects that live on the heap
and will likely cause L2 cache misses or contended locks.
How about something different?
Create the PermissionCollection's on demand, then discard immediately after
use. The Permission objects themselves are long lived immutable objects.
Why?
It'll be used only by one thread, so the jvm will optimise out the
synchronised locks.
The object will be created on the threads local memory stack, instead of the
heap and die in the young generation, so it doesn't incur gc heap generation
movements or memory heap copy to cpu cache stalls.
But what about single thread applications or those with few threads and little
contention? They would run slower, although object allocation costs aren't as
bad as people think, say 10 to 20 cpu cycles compared to 200 for a cache miss,
or worse for a contended lock.
Pattern matching of strings is the most expensive computation of most
permission decisions and has to be repeated for every ProtectionDomain on the
call stack for each thread, the impact on single core machines won't be much.
I can test for that, but not the high end stuff.
Arrghh decisions! Not enough test hardware.
Cheers,
Peter.
