1. Allocation frame protocol¶
.intro: This document explains the design of the support for allocation frames in MPS.
.readership: This document is intended for any MM developer.
.overview: Allocation frames are used for implementing stack pools; each stack frame corresponds to an allocation frame. Allocation frames may also be suitable for implementing other sub-pool groupings, such as generations and ramp allocation patterns.
.overview.ambition: We now believe this to be a design that loses too many advantages of stack allocation for questionable gains. The requirements are almost entirely based on unanalysed anecdote, instead of actual clients.
We plan to supersede this with a stack pool design at some point in the future. Pekka P. Pirinen, 2000-03-09.
.def.alloc-frame: An allocation frame is a generic name for a device which groups objects together with other objects at allocation time, and which may have a parent/child relationship with other allocation frames.
.purpose.stack-allocation: The allocation frame protocol is intended to support efficient memory management for stack allocation, that is, the allocation of objects which have dynamic extent.
.purpose.general: The allocation frame protocol is intended to be sufficiently general that it will be useful in supporting other types of nested allocation patterns too. For example, it could be used to for EPVM-style save and restore, ramp allocation patterns or generations.
1.4.1. Known requirements¶
.req.stack-alloc: Provide a interface for clients to describe a stack allocation pattern, as an alternative to using the control stack.
.req.efficient: Permit an implementation which is comparable in efficiency to allocating on the control stack.
.req.ap: Support allocation via allocation points (APs).
.req.format: Support the allocation of formatted objects.
.req.scan: Ensure that objects in allocation frames can participate in garbage collection by being scanned.
.req.fix: Ensure that objects in allocation frames can participate in garbage collection by accepting Fix requests.
.req.condemn: Ensure that objects in allocation frames can participate in garbage collection by being condemned.
.attr.locking: Minimize the synchronization cost for the creation and destruction of frames.
.proto-req: The following are possible requirements that might be important in the future. The design does not necessarily meet all these requirements, but it does consider them all. Each requirement either has direct support in the framework, or could be supported with future additions to the framework.
.req.parallels: The allocation frame protocol should provide a framework for exploiting the parallels between stack extents, generations and “ramps”.
.req.pool-destroy: It should be possible to use allocation frames to free all objects in a pool without destroying the pool.
.req.epvm: It should be possible to implement EPVM-style save and restore operations by creating and destroying allocation frames.
.req.subst: It should be possible to substitute a stack pool with a GC-ed pool so that erroneous use of a stack pool can be detected.
.req.format-extensions: It should be possible for stack pools to utilize the same format as any other pool, including debugging formats that include fenceposting, etc.
.req.mis-nest: Should ensure “mis-nested” stacks are safe.
.req.non-top-level: Should support allocation in the non-top stack extent.
.req.copy-if-necessary: Should ensure that stack pools can support “copy-if-necessary” (so that low-level system code can heapify stack objects.)
.req.preserve: When an object is in an allocation frame which is being destroyed, it should be possible to preserve that object in the parent frame.
.req.contained: Should allow clients to ask if an object is “contained” in a frame. The object is contained in a frame if it is affected when the frame is ended.
.req.alloc-with-other: Should allow clients to allocate an object in the same frame as another object.
.frame-classes: The protocol supports different types of allocation frames, which are represented as “frame classes”. It’s up to pools to determine which classes of allocation frames they support. Pools which support more than one frame class rely on the client to indicate which class is currently of interest. The client indicates this by means of an operation which stores the class in the buffer to which the allocation point is attached.
.frame-handles: Allocation frames are described via abstract “frame handles”. Pools may choose what the representation of a frame handle should be. Frame handles are static, and the client need not store them in a GC root.
.lightweight-frames: The design includes an extension to the allocation point protocol, which permits the creation and destruction of allocation frames without the necessity for claiming the arena lock. Such frames are called “lightweight frames”.
.op.intro: Each operation has both an external (client) interface and an internal (MPS) interface. The external function takes an allocation point as a parameter, determines which buffer and pool it belongs to, and calls the internal function with the buffer and pool as parameters.
.op.obligatory: The following operations are supported on any allocation point which supports allocation frames:-
.operation.push: The FramePush operation creates a new allocation frame of the currently chosen frame class, makes this new frame the current frame, and returns a handle for the frame.
.operation.pop: The FramePop operation takes a frame handle as a parameter. Some pool classes might insist or assume that this is the handle for the current frame. It finds the parent of that frame and makes it the current frame. The operation indicates that all children of the new current frame contain objects which are likely to be dead. The reclaim policy is up to the pool; some classes might insist or assume that the objects must be dead, and eagerly free them. Note that this might introduce the possibility of leaving dangling pointers elsewhere in the arena. If so, it’s up to the pool to decide what to do about this.
.op.optional: The following operations are supported for some allocation frames, but not all. Pools may choose to support some or all of these operations for certain frame classes. An unsupported operation will return a failure value:-
.operation.select: The FrameSelect operation takes a frame handle as a parameter and makes that frame the current frame. It does not indicate that any children of the current frame contain objects which are likely to be dead.
.operation.select-addr: The FrameSelectOfAddr operation takes an address as a parameter and makes the frame of that address the current frame. It does not indicate that any children of the current frame contain objects which are likely to be dead.
.operation.in-frame: The FrameHasAddr operation determines whether the supplied address is the address of an object allocated in the supplied frame, or any child of that frame.
.operation.set: The SetFrameClass operation takes a frame class and an allocation point as parameters, and makes that the current frame class for the allocation point. The next FramePush operation will create a new frame of that class.
1.7.1. External types¶
typedef struct mps_frame_class_s *mps_frame_class_t¶
.type.client.frame-class: Frame classes are defined as an abstract type.
.type.client.frame-class.access: Clients access frame classes by means of dedicated functions for each frame class.
1.7.2. External functions¶
.fn.client.select: This following function is used by clients to invoke the FrameSelect operation.
.fn.client.select-addr: This function is used by clients to invoke the FrameSelectOfAddr operation.
mps_res_t mps_ap_addr_in_frame(mps_bool_t *inframe_o, mps_ap_t buf, mps_addr_t *addrref, mps_frame_t frame)¶
.fn.client.in-frame: This function is used by clients to invoke the FrameHasAddr operation.
.fn.client.set: This function is used by clients to invoke the SetFrameClass operation.
.fn.client.stack-frame-class: This function is used by clients to access the frame class used for simple stack allocation.
1.7.3. Internal types¶
typedef struct AllocFrameStruct *AllocFrame¶
.type.frame-handle: Frame handles are defined as an abstract type.
typedef struct AllocFrameClassStruct *AllocFrameClass¶
.type.frame-class: Frame classes are defined as an abstract type.
.fn.push: A pool method of this type is called (if needed) to invoke the FramePush operation.
.fn.pop: A pool method of this type is called (if needed) to invoke the FramePop operation:
.fn.select: A pool method of this type is called to invoke the FrameSelect operation.
.fn.select-addr: A pool method of this type is called to invoke the FrameSelectOfAddr operation.
typedef Res (*PoolFrameHasAddrMethod)(Bool *inframeReturn, Pool pool, Seg seg, Addr *addrref, AllocFrame frame)¶
.fn.in-frame: A pool method of this type is called to invoke the FrameHasAddr operation.
.fn.set: A pool method of this type is called to invoke the SetFrameClass operation.
1.8. Lightweight frames¶
.lw-frame.overview: Allocation points provide direct support for lightweight frames, and are designed to permit FramePush and FramePop operations without the need for locking and delegation to the pool method. The pool method will be called whenever synchronization is required for other reasons (e.g. the buffer is tripped).
.lw-frame.model: Lightweight frames offer direct support for a particular model of allocation frame use, whereby the FramePush operation returns the current allocation pointer as a frame handle, and the FramePop operation causes the allocation pointer to be reset to the address of the frame handle. This model should be suitable for simple stack frames, where more advanced operations like FrameSelect are not supported. It may also be suitable for more advanced allocation frame models when they are being used simply. The use of a complex operation always involves synchronization via locking, and the pool may disable lightweight synchronization temporarily at this time.
.lw-frame.sync: The purpose of the design is that mutator may access the state of an AP without locking with MPS (via the external functions). The design assumes the normal MPS restriction that an operation on an AP may only be performed by a single mutator thread at a time. Each of the operations on allocation frames counts as an operation on an AP.
IF ap->init != ap->alloc FAIL ELSE IF ap->init < ap->limit *frame_o = ap->init; ELSE WITH_ARENA_LOCK PerformInternalFramePushOperation(...) END END
.lw-frame.push.limit: The reason for testing
ap->limit and not
ap->init <= ap->limit is that a frame pointer
at the limit of a buffer (and possibly therefore of a segment) would
be ambiguous: is it at the limit of the segment, or at the base of the
segment that’s adjacent in memory? The internal operation must handle
this case, for example by refilling the buffer and setting the frame
at the beginning.
IF ap->init != ap->alloc FAIL ELSE IF BufferBase(ap) <= frame AND frame < ap->init ap->init = ap->alloc = frame; ELSE WITH_ARENA_LOCK PerformInternalFramePopOperation(...) END END
.lw-frame.pop.buffer: The reason for testing that
frame is in
the buffer is that if it’s not, then we’re popping to an address in
some other segment, and that means that some objects in the other
segment (and all objects in any segments on the stack in between) are
now dead, and the only way for the pool to mark them as being dead is
to do a heavyweight pop.