pub trait Collection<VM: VMBinding> {
    // Required methods
    fn stop_all_mutators<F>(tls: VMWorkerThread, mutator_visitor: F)
       where F: FnMut(&'static mut Mutator<VM>);
    fn resume_mutators(tls: VMWorkerThread);
    fn block_for_gc(tls: VMMutatorThread);
    fn spawn_gc_thread(tls: VMThread, ctx: GCThreadContext<VM>);

    // Provided methods
    fn out_of_memory(_tls: VMThread, err_kind: AllocationError) { ... }
    fn schedule_finalization(_tls: VMWorkerThread) { ... }
    fn post_forwarding(_tls: VMWorkerThread) { ... }
    fn vm_live_bytes() -> usize { ... }
    fn is_collection_enabled() -> bool { ... }
    fn create_gc_trigger() -> Box<dyn GCTriggerPolicy<VM>> { ... }
}
Expand description

VM-specific methods for garbage collection.

Required Methods§

source

fn stop_all_mutators<F>(tls: VMWorkerThread, mutator_visitor: F)where F: FnMut(&'static mut Mutator<VM>),

Stop all the mutator threads. MMTk calls this method when it requires all the mutator to yield for a GC. This method should not return until all the threads are yielded. The actual thread synchronization mechanism is up to the VM, and MMTk does not make assumptions on that. MMTk provides a callback function and expects the binding to use the callback for each mutator when it is ready for stack scanning. Usually a stack can be scanned as soon as the thread stops in the yieldpoint.

Arguments:

  • tls: The thread pointer for the GC worker.
  • mutator_visitor: A callback. Call it with a mutator as argument to notify MMTk that the mutator is ready to be scanned.
source

fn resume_mutators(tls: VMWorkerThread)

Resume all the mutator threads, the opposite of the above. When a GC is finished, MMTk calls this method.

This method may not be called by the same GC thread that called stop_all_mutators.

Arguments:

  • tls: The thread pointer for the GC worker. Currently it is the tls of the embedded GCWorker instance of the coordinator thread, but it is subject to change, and should not be depended on.
source

fn block_for_gc(tls: VMMutatorThread)

Block the current thread for GC. This is called when an allocation request cannot be fulfilled and a GC is needed. MMTk calls this method to inform the VM that the current thread needs to be blocked as a GC is going to happen. Then MMTk starts a GC. For a stop-the-world GC, MMTk will then call stop_all_mutators() before the GC, and call resume_mutators() after the GC.

Arguments:

  • tls: The current thread pointer that should be blocked. The VM can optionally check if the current thread matches tls.
source

fn spawn_gc_thread(tls: VMThread, ctx: GCThreadContext<VM>)

Ask the VM to spawn a GC thread for MMTk. A GC thread may later call into the VM through these VM traits. Some VMs have assumptions that those calls needs to be within VM internal threads. As a result, MMTk does not spawn GC threads itself to avoid breaking this kind of assumptions. MMTk calls this method to spawn GC threads during initialize_collection().

Arguments:

  • tls: The thread pointer for the parent thread that we spawn new threads from. This is the same tls when the VM calls initialize_collection() and passes as an argument.
  • ctx: The context for the GC thread.
    • If Controller is passed, it means spawning a thread to run as the GC controller. The spawned thread shall call memory_manager::start_control_collector.
    • If Worker is passed, it means spawning a thread to run as a GC worker. The spawned thread shall call memory_manager::start_worker. In either case, the Box inside should be passed back to the called function.

Provided Methods§

source

fn out_of_memory(_tls: VMThread, err_kind: AllocationError)

Inform the VM of an out-of-memory error. The binding should hook into the VM’s error routine for OOM. Note that there are two different categories of OOM:

  • Critical OOM: This is the case where the OS is unable to mmap or acquire more memory. MMTk expects the VM to abort immediately if such an error is thrown.
  • Heap OOM: This is the case where the specified heap size is insufficient to execute the application. MMTk expects the binding to notify the VM about this OOM. MMTk makes no assumptions about whether the VM will continue executing or abort immediately.

See AllocationError for more information.

Arguments:

  • tls: The thread pointer for the mutator which failed the allocation and triggered the OOM.
  • err_kind: The type of OOM error that was encountered.
source

fn schedule_finalization(_tls: VMWorkerThread)

Inform the VM to schedule finalization threads.

Arguments:

  • tls: The thread pointer for the current GC thread.
source

fn post_forwarding(_tls: VMWorkerThread)

A hook for the VM to do work after forwarding objects.

This function is called after all of the following have finished:

  • The life and death of objects are determined. Objects determined to be live will not be reclaimed in this GC.
  • Live objects have been moved to their destinations. (copying GC only)
  • References in objects have been updated to point to new addresses. (copying GC only)

And this function may run concurrently with the release work of GC, i.e. freeing the space occupied by dead objects.

It is safe for the VM to read and write object fields at this time, although GC has not finished yet. GC will be reclaiming spaces of dead objects, but will not damage live objects. However, the VM cannot allocate new objects at this time.

One possible use of this hook is enqueuing {Soft,Weak,Phantom}Reference instances to reference queues (for Java). VMs (including JVM implementations) do not have to handle weak references this way, but mmtk-core provides this opportunity.

Arguments:

  • tls_worker: The thread pointer for the worker thread performing this call.
source

fn vm_live_bytes() -> usize

Return the amount of memory (in bytes) which the VM allocated outside the MMTk heap but wants to include into the current MMTk heap size. MMTk core will consider the reported memory as part of MMTk heap for the purpose of heap size accounting.

This amount should include memory that is kept alive by heap objects and can be released by executing finalizers (or other language-specific cleaning-up routines) that are executed when the heap objects are dead. For example, if a language implementation allocates array headers in the MMTk heap, but allocates their underlying buffers that hold the actual elements using malloc, then those buffers should be included in this amount. When the GC finds such an array dead, its finalizer shall free the buffer and reduce this amount.

If possible, the VM should account off-heap memory in pages. That is, count the number of pages occupied by off-heap objects, and report the number of bytes of those whole pages instead of individual objects. Because the underlying operating system manages memory at page granularity, the occupied pages (instead of individual objects) determine the memory footprint of a process, and how much memory MMTk spaces can obtain from the OS.

However, if the VM is incapable of accounting off-heap memory in pages (for example, if the VM uses malloc and the implementation of malloc is opaque to the VM), the VM binding can simply return the total number of bytes of those off-heap objects as an approximation.

Performance note

This function will be called when MMTk polls for GC. It happens every time the MMTk allocators have allocated a certain amount of memory, usually one or a few blocks. Because this function is called very frequently, its implementation must be efficient. If it is too expensive to compute the exact amount, an approximate value should be sufficient for MMTk to trigger GC promptly in order to release off-heap memory, and keep the memory footprint under control.

source

fn is_collection_enabled() -> bool

Callback function to ask the VM whether GC is enabled or disabled, allowing or disallowing MMTk to trigger garbage collection. When collection is disabled, you can still allocate through MMTk, but MMTk will not trigger a GC even if the heap is full. In such a case, the allocation will exceed MMTk’s heap size (the soft heap limit). However, there is no guarantee that the physical allocation will succeed, and if it succeeds, there is no guarantee that further allocation will keep succeeding. So if a VM disables collection, it needs to allocate with careful consideration to make sure that the physical memory allows the amount of allocation. We highly recommend to have GC always enabled (i.e. that this method always returns true). However, we support this to accomodate some VMs that require this behavior. Note that handle_user_collection_request() calls this function, too. If this function returns false, handle_user_collection_request() will not trigger GC, either. Note also that any synchronization involving enabling and disabling collections by mutator threads should be implemented by the VM.

source

fn create_gc_trigger() -> Box<dyn GCTriggerPolicy<VM>>

Ask the binding to create a GCTriggerPolicy if the option gc_trigger is set to crate::util::options::GCTriggerSelector::Delegated.

Implementors§