Arena

author Pekka P. Pirinen
date 1997-08-11
index terms pair: arena; design
revision //info.ravenbrook.com/project/mps/version/1.114/design/arena.txt#1
status incomplete design
tag design.mps.arena

Introduction

.intro: This is the design of the arena structure.

.readership: MPS developers.

Overview

.overview: The arena serves two purposes. It is a structure that is the top-level state of the MPS, and as such contains a lot of fields which are considered "global". And it provides raw memory to pools.

An arena belongs to a particular arena class. The class is selected when the arena is created. Classes encapsulate both policy (such as how pool placement preferences map into actual placement) and mechanism (such as where the memory originates: operating system virtual memory, client provided, or via malloc). Some behaviour (mostly serving the "top-level datastructure" purpose) is implemented by generic arena code, and some by arena class code.

Definitions

.def.grain: The arena manages memory in units called arena grains, whose size is returned by the macro ArenaGrainSize(). Memory allocated by ArenaAlloc() is a contiguous sequence of arena grains, whose base address and size are multiples of the arena grain size.

.def.tract: A tract is a data structure containing information about a region of address space: which pool it belongs to (if any), for which traces the contents is white, and so on. Tracts are the hook on which the segment module is implemented. Pools which don't use segments may use tracts for associating their own data with ranges of address.

Requirements

Note

Where do these come from? Need to identify and document the sources of requirements so that they are traceable to client requirements. Most of these come from the architectural design (design.mps.architecture) or the fix function design (design.mps.fix). Richard Brooksby, 1995-08-28.

They were copied from design.mps.arena.vm(1) and edited slightly. David Jones, 1999-06-23.

Block management

.req.fun.block.alloc: The arena must provide allocation of contiguous blocks of memory.

.req.fun.block.free: It must also provide freeing of contiguously allocated blocks owned by a pool, whether or not the block was allocated via a single request.

.req.attr.block.size.min: The arena must support management of blocks down to the larger of (i) the grain size of the virtual mapping interface (if a virtual memory interface is being used); and (ii) the grain size of the memory protection interface (if protection is used).

Note

On all the operating systems we support, these grain sizes are the same and are equal to the operating system page size. But we want the MPS to remain flexible enough to be ported to operating systems where these are different.

.req.attr.block.size.max: It must also support management of blocks up to the maximum size allowed by the combination of operating system and architecture. This is derived from req.dylan.attr.obj.max (at least).

.req.attr.block.align.min: The alignment of blocks shall not be less than MPS_PF_ALIGN for the architecture. This is so that pool classes can conveniently guarantee pool allocated blocks are aligned to MPS_PF_ALIGN. (A trivial requirement.)

Address translation

.req.fun.trans: The arena must provide a translation from any address to the following information:

.req.fun.trans.arena: Whether the address is managed by the arena.

.req.fun.trans.pool: Whether the address is managed by a pool within the arena, and if it is, the pool.

.req.fun.trans.arbitrary: If the address is managed by a pool, an arbitrary pointer value that the pool can associate with a group of contiguous addresses at any time.

.req.fun.trans.white: If the address is managed by an automatic pool, the set of traces for which the address is white. This is required so that the second-stage fix protocol can reject non-white addresses quickly. See design.mps.critical-path.

.req.attr.trans.time: The translation shall take no more than @@@@ [something not very large -- drj 1999-06-23]

Arena partition

.req.fun.set: The arena must provide a method for approximating sets of addresses.

.req.fun.set.time: The determination of membership shall take no more than @@@@ [something very small indeed]. (the non-obvious solution is refsets)

Constraints

.req.attr.space.overhead: req.dylan.attr.space.struct implies that the arena must limit the space overhead. The arena is not the only part that introduces an overhead (pool classes being the next most obvious), so multiple parts must cooperate in order to meet the ultimate requirements.

.req.attr.time.overhead: Time overhead constraint?

Note

How can there be a time "overhead" on a necessary component? David Jones, 1999-06-23.

Architecture

Statics

.static: There is no higher-level data structure than a arena, so in order to support several arenas, we have to have some static data in impl.c.arena. See impl.c.arena.static.

.static.init: All the static data items are initialized when the first arena is created.

.static.serial: arenaSerial is a static Serial, containing the serial number of the next arena to be created. The serial of any existing arena is less than this.

.static.ring: arenaRing is the sentinel of the ring of arenas.

.static.ring.init: arenaRingInit is a Bool showing whether the ring of arenas has been initialized.

.static.ring.lock: The ring of arenas has to be locked when traversing the ring, to prevent arenas being added or removed. This is achieved by using the (non-recursive) global lock facility, provided by the lock module.

.static.check: The statics are checked each time any arena is checked.

Arena classes

typedef mps_arena_s *Arena

.class: The Arena data structure is designed to be subclassable (see design.mps.protocol). Clients can select what arena class they'd like when instantiating one with mps_arena_create_k(). The arguments to mps_arena_create_k() are class-dependent.

.class.init: However, the generic ArenaInit() is called from the class-specific method, rather than vice versa, because the method is responsible for allocating the memory for the arena descriptor and the arena lock in the first place. Likewise, ArenaFinish() is called from the finish method.

.class.fields: The grainSize (for allocation and freeing) and zoneShift (for computing zone sizes and what zone an address is in) fields in the arena are the responsibility of the each class, and are initialized by the init method. The responsibility for maintaining the commitLimit, spareCommitted, and spareCommitLimit fields is shared between the (generic) arena and the arena class. commitLimit (see .commit-limit) is changed by the generic arena code, but arena classes are responsible for ensuring the semantics. For spareCommitted and spareCommitLimit see .spare-committed below.

.class.abstract: The basic arena class (AbstractArenaClass) is abstract and must not be instantiated. It provides little useful behaviour, and exists primarily as the root of the tree of arena classes. Each concrete class must specialize each of the class method fields, with the exception of the describe method (which has a trivial implementation) and the extend, retract and spareCommitExceeded methods which have non-callable methods for the benefit of arena classes which don't implement these features.

.class.abstract.null: The abstract class does not provide dummy implementations of those methods which must be overridden. Instead each abstract method is initialized to NULL.

Chunks

.chunk: Each contiguous region of address space managed by the MPS is represented by a chunk.

.chunk.tracts: A chunk contains a table of tracts. See `.tract`_.

.chunk.lookup: Looking of the chunk of an address is the first step in the second-stage fix operation, and so on the critical path. See design.mps.critical-path.

.chunk.tree: For efficient lookup, chunks are stored in a balanced tree; arena->chunkTree points to the root of the tree. Operations on this tree must ensure that the tree remains balanced, otherwise performance degrades badly with many chunks.

.chunk.insert: New chunks are inserted into the tree by calling ArenaChunkInsert(). This calls TreeInsert(), followed by TreeBalance() to ensure that the tree is balanced.

.chunk.delete: There is no corresponding function ArenaChunkDelete(). Instead, deletions from the chunk tree are carried out by calling TreeToVine(), iterating over the vine (where deletion is possible, if care is taken) and then calling TreeBalance() on the remaining tree. The function TreeTraverseAndDelete() implements this.

.chunk.delete.justify: This is because we don't have a function that deletes an item from a balanced tree efficiently, and because all functions that delete chunks do so in a loop over the chunks (so the best we can do is O(n) time in any case).

.chunk.delete.tricky: Deleting chunks from the chunk tree is tricky in the virtual memory arena because vmChunkDestroy() unmaps the memory containing the chunk, which includes the tree node. So the next chunk must be looked up before deleting the current chunk. The function TreeTraverseAndDelete() ensures that this is done.

Tracts

.tract.table: The arena maintains tables of tract structures such that every address managed by the arena belongs to exactly one tract.

.tract.size: Each tract covers exactly one arena grain. This is an implementation detail, not a requirement.

.tract.structure: The tract structure definition looks like this:

typedef struct TractStruct { /* Tract structure */
  PagePoolUnion pool; /* MUST BE FIRST (design.mps.arena.tract.field.pool) */
  void *p;                    /* pointer for use of owning pool */
  Addr base;                  /* Base address of the tract */
  TraceSet white : TRACE_MAX; /* traces for which tract is white */
  BOOLFIELD(hasSeg);          /* does tract have a seg in p? */
} TractStruct;

.tract.field.pool: The pool.pool field indicates to which pool the tract has been allocated (.req.fun.trans.pool). Tracts are only valid when they are allocated to pools. When tracts are not allocated to pools, arena classes are free to reuse tract objects in undefined ways. A standard technique is for arena class implementations to internally describe the objects as a union type of TractStruct and some private representation, and to set the pool field to NULL when the tract is not allocated. The pool field must come first so that the private representation can share a common prefix with TractStruct. This permits arena classes to determine from their private representation whether such an object is allocated or not, without requiring an extra field.

.tract.field.p: The p field is used by pools to associate tracts with other data (.req.fun.trans.arbitrary). It's used by the segment module to indicate which segment a tract belongs to. If a pool doesn't use segments it may use the p field for its own purposes. This field has the non-specific type (void *) so that pools can use it for any purpose.

.tract.field.hasSeg: The hasSeg bit-field is a Boolean which indicates whether the p field is being used by the segment module. If this field is TRUE, then the value of p is a Seg. See design.mps.type.bool.bitfield for why this is declared using the BOOLFIELD macro.

.tract.field.base: The base field contains the base address of the memory represented by the tract.

.tract.field.white: The white bit-field indicates for which traces the tract is white (.req.fun.trans.white). This information is also stored in the segment, but is duplicated here for efficiency during a call to TraceFix() (see design.mps.trace.fix).

.tract.limit: The limit of the tract's memory may be determined by adding the arena grain size to the base address.

.tract.iteration: Iteration over tracts is described in design.mps.arena.tract-iter(0).

Bool TractOfAddr(Tract *tractReturn, Arena arena, Addr addr)

.tract.if.tractofaddr: The function TractOfAddr() finds the tract corresponding to an address in memory. (See .req.fun.trans.) If addr is an address which has been allocated to some pool, then TractOfAddr() returns TRUE, and sets *tractReturn to the tract corresponding to that address. Otherwise, it returns FALSE. This function is similar to TractOfBaseAddr() (see design.mps.arena.tract-iter.if.contig-base) but serves a more general purpose and is less efficient.

Control pool

.pool: Each arena has a "control pool", arena->controlPoolStruct, which is used for allocating MPS control data structures by calling ControlAlloc().

Polling

.poll: ArenaPoll() is called "often" by other code (for instance, on buffer fill or allocation). It is the entry point for doing tracing work. If the polling clock exceeds a set threshold, and we're not already doing some tracing work (that is, insidePoll is not set), it calls TracePoll() on all busy traces.

.poll.size: The actual clock is arena->fillMutatorSize. This is because internal allocation is only significant when copy segments are being allocated, and we don't want to have the pause times to shrink because of that. There is no current requirement for the trace rate to guard against running out of memory.

Note

Clearly it really ought to: we have a requirement to not run out of memory (see req.dylan.prot.fail-alloc, req.dylan.prot.consult), and emergency tracing should not be our only story. David Jones, 1999-06-22.

BufferEmpty() is not taken into account, because the splinter will rarely be useable for allocation and we are wary of the clock running backward.

.poll.clamp: Polling is disabled when the arena is "clamped", in which case arena->clamped is TRUE. Clamping the arena prevents background tracing work, and further new garbage collections from starting. Clamping and releasing are implemented by the ArenaClamp() and ArenaRelease() methods.

.poll.park: The arena is "parked" by clamping it, then polling until there are no active traces. This finishes all the active collections and prevents further collection. Parking is implemented by the ArenaPark() method.

Commit limit

.commit-limit: The arena supports a client configurable "commit limit" which is a limit on the total amount of committed memory. The generic arena structure contains a field to hold the value of the commit limit and the implementation provides two functions for manipulating it: ArenaCommitLimit() to read it, and ArenaSetCommitLimit() to set it. Actually abiding by the contract of not committing more memory than the commit limit is left up to the individual arena classes.

.commit-limit.err: When allocation from the arena would otherwise succeed but cause the MPS to use more committed memory than specified by the commit limit ArenaAlloc() should refuse the request and return ResCOMMIT_LIMIT.

.commit-limit.err.multi: In the case where an ArenaAlloc() request cannot be fulfilled for more than one reason including exceeding the commit limit then class implementations should strive to return a result code other than ResCOMMIT_LIMIT. That is, ResCOMMIT_LIMIT should only be returned if the only reason for failing the ArenaAlloc() request is that the commit limit would be exceeded. The client documentation allows implementations to be ambiguous with respect to which result code in returned in such a situation however.

Spare committed (aka "hysteresis")

.spare-committed: See mps_arena_spare_committed(). The generic arena structure contains two fields for the spare committed memory fund: spareCommitted records the total number of spare committed bytes; spareCommitLimit records the limit (set by the user) on the amount of spare committed memory. spareCommitted is modified by the arena class but its value is used by the generic arena code. There are two uses: a getter function for this value is provided through the MPS interface (mps_arena_spare_commit_limit_set()), and by the SetSpareCommitLimit() function to determine whether the amount of spare committed memory needs to be reduced. spareCommitLimit is manipulated by generic arena code, however the associated semantics are the responsibility of the class. It is the class's responsibility to ensure that it doesn't use more spare committed bytes than the value in spareCommitLimit.

.spare-commit-limit: The function ArenaSetSpareCommitLimit() sets the spareCommitLimit field. If the limit is set to a value lower than the amount of spare committed memory (stored in spareCommitted) then the class specific function spareCommitExceeded is called.

Locks

.lock.ring: ArenaAccess() is called when we fault on a barrier. The first thing it does is claim the non-recursive global lock to protect the arena ring (see design.mps.lock(0)).

.lock.arena: After the arena ring lock is claimed, ArenaEnter() is called on one or more arenas. This claims the lock for that arena. When the correct arena is identified or we run out of arenas, the lock on the ring is released.

.lock.avoid: Deadlocking is avoided as described below:

.lock.avoid.mps: Firstly we require the MPS not to fault (that is, when any of these locks are held by a thread, that thread does not fault).

.lock.avoid.thread: Secondly, we require that in a multi-threaded system, memory fault handlers do not suspend threads (although the faulting thread will, of course, wait for the fault handler to finish).

.lock.avoid.conflict: Thirdly, we avoid conflicting deadlock between the arena and global locks by ensuring we never claim the arena lock when the recursive global lock is already held, and we never claim the binary global lock when the arena lock is held.

Location dependencies

.ld: Location dependencies use fields in the arena to maintain a history of summaries of moved objects, and to keep a notion of time, so that the staleness of location dependency can be determined.

Finalization

.final: There is a pool which is optionally (and dynamically) instantiated to implement finalization. The fields finalPool and isFinalPool are used.

Implementation

Tract cache

.tract.cache: When tracts are allocated to pools by ArenaAlloc(), the first tract of the block and it's base address are cached in arena fields lastTract and lastTractBase. The function TractOfBaseAddr() (see design.mps.arena.tract-iter.if.block-base(0)) checks against these cached values and only calls the class method on a cache miss. This optimizes for the common case where a pool allocates a block and then iterates over all its tracts (for example, to attach them to a segment).

.tract.uncache: When blocks of memory are freed by pools, ArenaFree() checks to see if the cached value for the most recently allocated tract (see .tract.cache) is being freed. If so, the cache is invalid, and must be reset. The lastTract and lastTractBase fields are set to NULL.

Control pool

.pool.init: The control pool is initialized by a call to PoolInit() during ArenaCreate().

.pool.ready: All the other fields in the arena are made checkable before calling PoolInit(), so PoolInit() can call ArenaCheck(arena). The pool itself is, of course, not checkable, so we have a field arena->poolReady, which is false until after the return from PoolInit(). ArenaCheck() only checks the pool if poolReady.

Traces

.trace: arena->trace[ti] is valid if and only if TraceSetIsMember(arena->busyTraces, ti).

.trace.create: Since the arena created by ArenaCreate() has arena->busyTraces = TraceSetEMPTY, none of the traces are meaningful.

.trace.invalid: Invalid traces have signature SigInvalid, which can be checked.

Polling

.poll.fields: There are three fields of a arena used for polling: pollThreshold, insidePoll, and clamped (see above). pollThreshold is the threshold for the next poll: it is set at the end of ArenaPoll() to the current polling time plus ARENA_POLL_MAX.

Location dependencies

.ld.epoch: arena->epoch is the "current epoch". This is the number of 'flips' of traces in the arena since the arena was created. From the mutator's point of view locations change atomically at flip.

.ld.history: arena->history is an array of ARENA_LD_LENGTH elements of type RefSet. These are the summaries of moved objects since the last ARENA_LD_LENGTH epochs. If e is one of these recent epochs, then

arena->history[e % ARENA_LD_LENGTH]

is a summary of (the original locations of) objects moved since epoch e.

.ld.prehistory: arena->prehistory is a RefSet summarizing the original locations of all objects ever moved. When considering whether a really old location dependency is stale, it is compared with this summary.

Roots

.root-ring: The arena holds a member of a ring of roots in the arena. It holds an incremental serial which is the serial of the next root.

Document History

  • 1997-08-11 Pekka P. Pirinen. First draft, based on design.mps.space(0) and mail.richard.1997-04-25.11-52(0).
  • 1999-04-16 Tony Mann. Updated for separation of tracts and segments.
  • 2002-06-07 RB Converted from MMInfo database design document.
  • 2013-03-11 GDR Converted to reStructuredText.
  • 2014-02-17 RB Updated first field of tract structure.

Docutils System Messages

System Message: ERROR/3 (version/1.114/design/arena.txt, line 239); backlink

Unknown target name: ".tract".