P4
Amongst the changes in 7039089 was a newly added assert that checks that
when a thread issues a younger_refs_iterate() call, the observed used_region()
is a superset of use_region_at_save_marks(). The intention of the assertion
was to ensure that a save_marks() call had been done by the GC coordinator
thread (aka the VM thread) prior to the younger_refs_iterate() call.
The save_marks() mechanism itself may be considered a vestige of the
single-threaded Cheney scan scheme where one could keep track of grey
objects as they were copied into a contiguous space without the
need for an explicit stack. Given the use of promotion LABs and such
in the parallel collectors, save_marks() is really obsolete for its
original purpose of avoiding stacks to keep track of grey objects.
Rather its sole use in parallel collector code is to record, in the
case of contiguous or pseduo-contiguous spaces, the highest address
to which it is reasonable to look for refs (in other words to avoid
looking at newly promoted objects in the case of contiguous spaces).
In any case, it is reasonable to expect that parallel worker threads
should always find that the used_region() is a strict superset of
used_region_at_save_marks(), with (in the case of parallel collectors)
the latter being a constant during any single collection.
It appears as though that is not the case with ParNew+CMS, and we
need to fix the code so the invariant holds, otherwise there is
some chance that a flickering used_region_at_save_marks() could
miss refs that needed to be looked at or attempt to access
cards that do not have objects on them. The former would be
a more insidious problem than the latter, of course.
If investigation reveals that this is a more serious problem,
we may revise the target and priority of this CR.
Since the issue appears to be a long-standing one that just happened
to be exposed on account of the newly-added assert, we have set this
at priority P4 and targeted it to post-7 given the late stage of
JDK 7 development and the presence of potentially bigger fish (or bugs)
to fry (or respectively catch) at the moment.
when a thread issues a younger_refs_iterate() call, the observed used_region()
is a superset of use_region_at_save_marks(). The intention of the assertion
was to ensure that a save_marks() call had been done by the GC coordinator
thread (aka the VM thread) prior to the younger_refs_iterate() call.
The save_marks() mechanism itself may be considered a vestige of the
single-threaded Cheney scan scheme where one could keep track of grey
objects as they were copied into a contiguous space without the
need for an explicit stack. Given the use of promotion LABs and such
in the parallel collectors, save_marks() is really obsolete for its
original purpose of avoiding stacks to keep track of grey objects.
Rather its sole use in parallel collector code is to record, in the
case of contiguous or pseduo-contiguous spaces, the highest address
to which it is reasonable to look for refs (in other words to avoid
looking at newly promoted objects in the case of contiguous spaces).
In any case, it is reasonable to expect that parallel worker threads
should always find that the used_region() is a strict superset of
used_region_at_save_marks(), with (in the case of parallel collectors)
the latter being a constant during any single collection.
It appears as though that is not the case with ParNew+CMS, and we
need to fix the code so the invariant holds, otherwise there is
some chance that a flickering used_region_at_save_marks() could
miss refs that needed to be looked at or attempt to access
cards that do not have objects on them. The former would be
a more insidious problem than the latter, of course.
If investigation reveals that this is a more serious problem,
we may revise the target and priority of this CR.
Since the issue appears to be a long-standing one that just happened
to be exposed on account of the newly-added assert, we have set this
at priority P4 and targeted it to post-7 given the late stage of
JDK 7 development and the presence of potentially bigger fish (or bugs)
to fry (or respectively catch) at the moment.