THE COLLECTION FRAMEWORK
design.mps.collection
incomplete design
pekka 1998-03-20
INTRODUCTION
.intro: This document describes the Collection Framework. It's a framework for
implementing garbage collection techniques and integrating them into a system
of collectors that all cooperate in recycling garbage.
Document History
.hist.0: Version 0 was a different document.
.hist.1: Version 1 was a different document.
.hist.2: Written in January and February 1998 by Pekka P. Pirinen on the basis
of the current implementation of the MPS, analysis.async-gc, [that note on the
independence of collections] and analysis.tracer.
OVERVIEW
.framework: MPS provides a framework that allows the integration of many
different types of GC strategies and provides many of the basic services that
those strategies use. .framework.cover: The framework subsumes most major GC
strategies and allows many efficient techniques, like in-line allocation or
software barriers.
.framework.overhead: The overhead due to cooperation is low. [But not
non-existent. Can we say something useful about it?]
.framework.benefits: The ability to combine collectors contributes
significantly to the flexibility of the system. The reduction in code
duplication contributes to reliability and integrity. The services of the
framework make it easier to write new MM strategies and collectors.
.framework.mpm: The Collection Framework is merely a part of the structure of
the MPM. See design.mps.architecture and design.mps.arch [Those two documents
should be combined into one. Pekka 1998-01-15] for the big picture. Other
notable components that the MPM manages to integrate into a single framework
are manually-managed memory [another missing document here?] and finalization
services (see design.mps.finalize).
.see-also: This document assumes basic familiarity with the ideas of pool (see
design.mps.arch.pools) and segment (see design.mps.seg.over.*).
COLLECTION ABSTRACTIONS
Colours, scanning and fixing
.state: The framework knows about the three colours of the tri-state
abstraction and free blocks. Recording the state of each object is the
responsibility of the pool, but the framework gets told about changes in the
states and keeps track of colours in each segment. Specifically, it records
whether a segment might contain white, grey and black objects wrt. each active
trace (see .tracer) [black not currently implemented -- Pekka 1998-01-04]. (A
segment might contain objects of all colours at once, or none.) This
information is approximate, because when an object changes colour, or dies, it
usually is too expensive to determine if it was the last object of its former
colour.
.state.transitions: The possible state transitions are as follows:
free ---alloc--> black (or grey) or white or none
none --condemn-> white
none --refine--> grey
grey ---scan---> black
white ----fix---> grey (or black)
black --revert--> grey
white --reclaim-> free
black --reclaim-> none
.none-is-black: Outside of a trace, objects don't really have colour, but
technically, the colour is black. Objects are only allocated grey or white
during a trace, and by the time the trace has finished, they are either dead or
black, like the other surviving objects. We might then reuse the colour field
for another trace, so it's convenient to set the colour to black when
allocating outside a trace. This means that refining the foundation
(analysis.tracer.phase.condemn.refine), actually turns black segments grey,
rather than vice versa, but the principle is the same.
.scan-fix: "Scanning" an object means applying the "fix" function to all
references in that object. Fixing is the generic name for the operation that
takes a reference to a white object and makes it non-white (usually grey, but
black is a possibility, and so is changing the reference as we do for weak
references). Typical examples of fix methods are copying the object into
to-space or setting its mark bit.
.cooperation: The separation of scanning and fixing is what allows different GC
techniques to cooperate. The scanning is done by a method on the pool that the
scanned object resides in, and the fixing is done by a method on the pool that
the reference points to.
.scan-all: Pools provide a method to scan all the grey objects in a segment.
Reference sets
.refsets: The cost of scanning can be significantly reduced by storing
remembered sets. We have chosen a very compact and efficient implementation,
called reference sets, or refsets for short (see idea.remember
[design.mps.refset is empty! Perhaps some of this should go there. -- Pekka
1998-02-19]). This makes the cost of maintaining them low, so we maintain them
for all references out of all scannable segments.
.refsets.approx: You might describe refsets as summaries of all references out
of an area of memory, so they are only approximations of remembered sets. When
a refset indicates that an interesting reference might be present in a segment,
we still have to scan the segment to find it.
.refsets.scan: The refset information is collected during scanning. The scan
state protocol provides a way for the pool and the format scan methods to
cooperate in this, and to pass this information to the tracer module which
checks it and updates the segment (see design.mps.scan [Actually, there's very
little doc there. Pekka 1998-02-17]).
.refsets.maintain: The MPS tries to maintain the refset information when it
moves or changes object.
.refsets.pollution: Ambiguous references and pointers outside the arena will
introduce spurious zones into the refsets. We put up with this to keep the
scanning costs down. Consistency checks on refsets have to take this into
account.
.refsets.write-barrier: A write-barrier are needed to keep the mutator from
invalidating the refsets when writing to a segment. We need one on any
scannable segment whose refset is not a superset of the mutator's (and that the
mutator can see). If we know what the mutator is writing and whether it's a
reference, we can just add that reference to the refset (figuring out whether
anything can be removed from the refset is too expensive). If we don't know or
if we cannot afford to keep the barrier up, the framework can union the
mutator's refset to the segment's refset.
.refset.mutator: The mutator's refset could be computed during root scanning in
the usual way, and then kept up to date by using a read-barrier. It's not a
problem that the mutator can create new pointers out of nothing behind the
read-barrier, as they won't be real references. However, this is probably not
cost-effective, since it would cause lots of barrier hits. We'd need a
read-barrier on every scannable segment whose refset is not a subset of the
mutator's (and that the mutator can see). So instead we approximate the
mutator's refset with the universal refset.
THE TRACER
.tracer: The tracer is an engine for implementing multiple garbage collection
processes. Each process (called a "trace") proceeds independently of the
others through five phases as described in analysis.tracer. The following
sections describe how the action of each phase fits into the framework. See
design.mps.trace for details [No, there's not much there, either. Possibly
some of this section should go there. Pekka 1998-02-18]).
.combine: The tracer can also combine several traces for some actions, like
scanning a segment or a root. The methods the tracer calls to do the work get
an argument that tells them which traces they are expected to act for. [extend
this@@@@]
.trace.begin: Traces are started by external request, usually from a client
function or an action (see design.mps.action).
.trace.progress: The tracer gets time slices from the arena to work on a given
trace [This is just a provisional arrangement, in lieu of real progress
control. Pekka 1998-02-18]. In each slice, it selects a small amount of work
to do, based on the state of the trace, and does it, using facilities provided
by the pools. .trace.scan: A typical unit of work is to scan a single
segment. The tracer can choose to do this for multiple traces at once,
provided the segment is grey for more than one trace.
.trace.barrier: Barrier hits might also cause a need to scan a segment (see
.hw-barriers.hit). Again, the tracer can choose to combine traces, when it
does this.
.mutator-colour: The framework keeps track of the colour of the mutator
separately for each trace.
The Condemn Phase
.phase.condemn: The agent that creates the trace (see .trace.begin) determines
the condemned set and colours it white. The tracer then examines the refsets
on all scannable segments, and if it can deduce some segment cannot refer to
the white set, it's immediately coloured black, otherwise the pool is asked to
grey any objects in the segment that might need to be scanned (in copying
pools, this is typically the whole segment).
.phase.condemn.zones: To get the maximum benefit from the refsets, we try to
arrange that the zones are a minimal superset (e.g., generations uniquely
occupy zones) and a maximal subset (there's nothing else in the zone) of the
condemned set. This needs to be arranged at allocation time (or when copying
during collection, which is much like allocation) [soon, this will be handled
by segment loci, see design.mps.locus].
.phase.condemn.mutator: At this point, the mutator might reference any objects,
i.e., it is grey. Allocation can be in any colour, most commonly white [more
could be said about this].
The Grey Mutator Phase
.phase.grey-mutator: Grey segments are chosen according to some sort of
progress control and scanned by the pool to make them black. Eventually, the
tracer will decide to flip or it runs out of grey segments, and proceeds to the
next phase. [Currently, this phase has not been implemented; all traces flip
immediately after condemn. Pekka 1998-02-18]
.phase.grey-mutator.copy: At this stage, we don't want to copy condemned
objects, because we would need an additional barrier to keep the mutator's view
of the heap consistent (see analysis.async-gc.copied.pointers-and-new-copy).
.phase.grey-mutator.ambig: This is a good time to get all ambiguous scanning
out of the way, because we usually can't do any after the flip [write a
detailed explanation of this some day] and because it doesn't cause any copying.
The Flip Phase
.phase.flip: The roots (see design.mps.root) are scanned. This has to be an
atomic action as far as the mutator is concerned, so all threads are suspended
for the duration.
.phase.flip.mutator: After this, the mutator is black: if we use a strong
barrier (analysis.async-gc.strong), this means it cannot refer to white
objects. Allocation will be in black (could be grey as well, but there's no
point to it).
The Black Mutator Phase
.phase.black-mutator: Grey segments are chosen according to some sort of
progress control and scanned by the pool to make them black. Eventually, the
tracer runs out of segments that are grey for this trace, and proceeds to the
next phase.
.phase.black-mutator.copy: At this stage white objects can be relocated,
because the mutator cannot see them (as long as a strong barrier is used, as we
must do for a copying collection, see analysis.async-gc.copied.pointers).
The Reclaim Phase
.phase.reclaim: The tracer finds the remaining white segments and asks the pool
to reclaim any white objects in them.
.phase.reclaim.barrier: Once a trace has started reclaiming objects, the others
shouldn't try to scan any objects that are white for it, because they might
have dangling pointers in them [xref doc yet to be written]. [Currently, we
reclaim atomically, but it could be incremental, or even overlapped with a new
trace on the same condemned set. Pekka 1997-12-31]
BARRIERS
[An introduction and a discussion of general principles should go here. This
is a completely undesigned area.]
Hardware Barriers
.hw-barriers: Hardware barrier services cannot, by their very nature, be
independently provided to each trace. A segment is either protected or not,
and we have to set the protection on a segment if any trace needs a hardware
barrier on it.
.hw-barriers.supported: The framework currently supports segment-oriented
Appel-Ellis-Li barriers (analysis.async-gc.barrier.appel-ellis-li), and
write-barriers for keeping the refsets up-to-date. It would not be hard to add
Steele barriers (analysis.async-gc.barrier.steele.scalable).
.hw-barriers.hit: When a barrier hit happens, the arena determines which
segment it was on. The segment colour info is used to determine whether it had
trace barriers on it, and if so, the appropriate barrier action is performed,
using the methods of the owning pool. If the segment was write-protected, its
refset is unioned with the refset of the mutator [in practice, RefSetUNIV].
.hw-barriers.hit.multiple: Fortunately, if we get a barrier hit on a segment
with multiple trace barriers on it, we can scan it for all the traces that it
had a barrier for, see .combine.@@@@
Software barriers
[@@@@Have to say something about software barriers]
| 2002-06-07 | RB | Converted from MMInfo database design document. |
This document is copyright © 1995-2002 Ravenbrook Limited. All rights reserved. This is an open source license. Contact Ravenbrook for commercial licensing options.
Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met:
This software is provided by the copyright holders and contributors "as is" and any express or implied warranties, including, but not limited to, the implied warranties of merchantability, fitness for a particular purpose, or non-infringement, are disclaimed. In no event shall the copyright holders and contributors be liable for any direct, indirect, incidental, special, exemplary, or consequential damages (including, but not limited to, procurement of substitute goods or services; loss of use, data, or profits; or business interruption) however caused and on any theory of liability, whether in contract, strict liability, or tort (including negligence or otherwise) arising in any way out of the use of this software, even if advised of the possibility of such damage.
$Id: //info.ravenbrook.com/project/mps/version/1.105/design/collection/index.html#1 $
Ravenbrook / Projects / Memory Pool System / Master Product Sources / Design Documents