There is a race in the evac failure handling code that causes the condition the assert checks not to be true. The fix is to replace the too-strong assert with a more targeted one.
Reviewed-by: johnc, ysr, jcoomes
We should only undirty cards after we decide that they are not on a young region, not before. The fix also includes improvements to the verify_dirty_region() method which print out which cards were not found dirty.
Reviewed-by: johnc, brutisso
We were calling STS join and leave during a STW pause and we are not suppoesed to. I now only call those during concurrent phase. I also added stress code in the non-product builds to force an overflows (the condition that ws uncovering the bug) to make sure it does not happen again.
Reviewed-by: johnc, brutisso
In G1 heap verification, we no longer scan perm to G1-collected heap refs as part of process_strong_roots() but rather in a separate explicit oop iteration over the perm gen. This preserves the original perm card-marks. Added a new assertion in younger_refs_iterate() to catch a simple subcase where the user may have forgotten a prior save_marks() call, as happened in the case of G1's attempt to iterate perm to G1 refs when verifying the heap before exit. The assert was deliberately weakened for ParNew+CMS and will be fixed for that combination in a future CR. Also made some (non-G1) cleanups related to code and comments obsoleted by the migration of Symbols to the native heap.
Reviewed-by: iveresov, jmasa, tonyp
It introduces non-product cmd line parameter G1DummyRegionsPerGC which indicates how many "dummy" regions to allocate at the end of each GC. This allows the G1 heap to grow artificially and makes concurrent marking cycles more frequent irrespective of what the application that is running is doing. The dummy regions will be found totally empty during cleanup so this parameter can also be used to stress the concurrent cleanup operation.
Reviewed-by: brutisso, johnc
Missing parentheses around a casted expression and some missing casts were causing build failures with the jdk6 build tools.
Reviewed-by: kvn, brutisso
A referent object that is only weakly reachable at the start of concurrent marking but is re-attached to the strongly reachable object graph during marking may not be marked as live. This can cause the reference object to be processed prematurely and leave dangling pointers to the referent object. Implement a read barrier for the java.lang.ref.Reference::referent field by intrinsifying the Reference.get() method, and intercepting accesses though JNI, reflection, and Unsafe, so that when a non-null referent object is read it is also logged in an SATB buffer.
Reviewed-by: kvn, iveresov, never, tonyp, dholmes
Use _max_num_q = max(discovery_degree, processing_degree), and let balance_queues() redistribute from discovery_degree to processing_degree of queues. This should also allow a more dynamic and flexible parallelism policy in the future.
Reviewed-by: jmasa, johnc
Bulk zeroing reduction didn't work with G1, because arraycopy would call pre-barriers on uninitialized oops. The solution is to have version of arraycopy stubs that don't have pre-barriers. Also refactored arraycopy stubs generation on SPARC to be more readable and reduced the number of stubs necessary in some cases.
Reviewed-by: jrose, kvn, never
Repurpose the existing flag G1PrintRegionLivenessInfo to print out the liveness distribution across the regions in the heap at the end of marking.
Reviewed-by: iveresov, jwilhelm
The epochs on the card cache array are initialized to 0 and our initial epoch also starts at 0. So, until the first GC, it might be possible to successfully "claim" a card which was in fact never initialized.
Reviewed-by: johnc, iveresov, ysr
7023151: G1: refactor the code that operates on _cur_alloc_region to be re-used for allocs by the GC threads
7018286: G1: humongous allocation attempts should take the GC locker into account
First, this change replaces the asymmetric locking scheme in the G1 slow alloc path by a summetric one. Second, it factors out the code that operates on _cur_alloc_region so that it can be re-used for allocations by the GC threads in the future.
Reviewed-by: stefank, brutisso, johnc
Manage allocation/freeing of the card cache counts and epochs arrays directly so that an allocation failure while attempting to expand these arrays does not abort the JVM. Failure to expand these arrays is not fatal.
Reviewed-by: iveresov, tonyp
Changed default value of WorkStealingYieldsBeforeSleep from 1000 to 5000. Added more information to G1 pause logging.
Reviewed-by: jwilhelm, tonyp, jmasa
7007446: G1: expand the heap with a single step, not one region at a time
Changed G1CollectedHeap::expand() to expand the committed space by calling VirtualSpace::expand_by() once rather than for every region in the expansion amount. This allows the success or failure of the expansion to be determined before creating any heap regions. Introduced a develop flag G1ExitOnExpansionFailure (false by default) that, when true, will exit the VM if the expansion of the committed space fails. Finally G1CollectedHeap::expand() returns a status back to it's caller so that the caller knows whether to attempt the allocation.
Reviewed-by: brutisso, tonyp
There's a potential deadlock between the concurrent cleanup thread and the GC workers that are trying to allocate and waiting for more free regions to be made available.
Reviewed-by: iveresov, jcoomes
A race between the concurrent cleanup thread and the VM thread while it is processing the "expanded sparse table list" causes both threads to try to free the same sparse table entry and either causes one of the threads to fail or leaves the entry in an inconsistent state. The solution is purge all entries on the expanded list that correspond go regions that are being cleaned up.
Reviewed-by: brutisso, johnc
Two small fixes to deal with a guarantee failure (the marking thread should join the SuspendibleThreadSet before calling a method that does pause prediction work so that said method is never called during a pause) and an assert failure (an assert is too strong).
Reviewed-by: iveresov, johnc
This changeset removes the zero-filling thread from G1 and collapses the two free region lists we had before (the "free" and "unclean" lists) into one. The new free list uses the new heap region sets / lists abstractions that we'll ultimately use it to keep track of all regions in the heap. A heap region set was also introduced for the humongous regions. Finally, this change increases the concurrency between the thread that completes freeing regions (after a cleanup pause) and the rest of the system (before we'd have to wait for said thread to complete before allocating a new region). The changest also includes a lot of refactoring and code simplification.
Reviewed-by: jcoomes, johnc
This changeset introduces filtering of SATB buffers at the point when they are about to be enqueued. If this filtering clears enough entries on each buffer, the buffer can then be re-used and not enqueued. This cuts down the number of SATB buffers that need to be processed by the concurrent marking threads.
Reviewed-by: johnc, ysr
First attempt to allocate out the current alloc region using a CAS instead of taking the Heap_lock (first level of G1's slow allocation path). Only if that fails and it's necessary to replace the current alloc region take the Heap_lock (that's the second level of G1's slow allocation path).
Reviewed-by: johnc, brutisso, ysr
During evacuation failure handling we refine the BOT to reflect the location of all the objects in the regions we scan. The changeset includes some minor cleanup: a) non-product print_on() method on the G1 BOT class, b) added more complete BOT verification during heap / region verification, c) slight modification to the BOT set up for humongous regions to be more consistent with the BOT set up during evac failure handling, and d) removed a couple of unused methods.
Reviewed-by: johnc, ysr
Age bits need not enter the mark-word preservation calculus; also affected, in addition to CMS, per CR synopsis above, were ParNew (but not DefNew), ParallelScavenge and G1, albeit to a lesser degree than CMS.
Reviewed-by: tonyp, johnc
Initialize the _is_alive_non_header field of G1's reference processor with an instance of the G1CMIsAliveClosure. This will stop adding reference objects with live referents to the discovered reference lists unnecessarily.
Reviewed-by: tonyp, ysr, jwilhelm, brutisso
The concurrent marking thread can complete its operation and increment the full GC counter during a Full GC. This causes the nesting of increments to the start and end of Full GCs that we are expecting to be wrong. the fix is for the marking thread to join the suspendible thread set before incrementing the counter so that it's blocked until the Full GC (or any other safepoint) is finished. The change also includes some minor code cleanup (I renamed a parameter).
Reviewed-by: brutisso, ysr
