State Machine

PTP uses fine-grained atomic state machine transitions to avoid locks. This document helps developers understand how it works, in order to reduce development difficulty.
Most logic is optimistic concurrency control based on CAS: once CAS fails, it is directly skipped / recovered, while a small part uses spin-waiting. Logic that is spin-waited on must be instantaneous operations.
PTP state machines basically (in fact, all of them) only care about their final states, so the ABA problem does not need to be considered.

PowerPool State Machines

State Machine	Description	Values
PoolState	Running state of the thread pool itself	NotRunning (Default), IdleChecked, Running
CanCreateNewWorker	Throttling when creating threads, to prevent creating too many threads under concurrency	Allowed (Default), NotAllowed
CanDeleteRedundantWorker	Throttling when threads need to be destroyed after exceeding limits, to prevent destroying unnecessary threads under concurrency	Allowed (Default), NotAllowed

PoolState

Trigger Event	State Change	If CAS Fails	Prerequisite
When PoolIdled event and other cleanup work are completed	NotRunning
When it is detected during running that there is no running, waiting, or async task	IdleChecked (CAS)	Directly skip
When attempting to enqueue a task while the current state is NotRunning	Running (CAS)	Directly skip

CanCreateNewWorker

Trigger Event	State Change	If CAS Fails	Prerequisite
When preparing to create a thread	NotAllowed (CAS)	Directly skip
When thread creation is completed	Allowed

CanDeleteRedundantWorker

Trigger Event	State Change	If CAS Fails	Prerequisite
When preparing to destroy a thread	NotAllowed (CAS)	Directly skip
When thread destruction is completed	Allowed
When it is double-confirmed that the thread does not need to be destroyed	Allowed

Worker State Machines

State Machine	Description	Values
WorkerState	Running state of the Worker	Idle (Default), Running, ToBeDisposed
CanDispose	Used to prevent multiple Dispose calls	Allowed (Default), NotAllowed
CanForceStop	Used to prevent multiple ForceStop calls	Allowed (Default), NotAllowed
CanGetWork	Indicates whether this Worker can accept new work	Allowed, NotAllowed (Default), ToBeDisabled, Disabled
WorkStealability	Indicates whether this Worker can have its tasks stolen	Allowed (Default), NotAllowed
WorkHeldState	Indicates whether the current work is forcibly kept held	NotHeld (Default), Held

WorkerState

Trigger Event	State Change	If CAS Fails	Prerequisite
When a new task is enqueued	Running (CAS)	CAS is used to check the previous state; failure has no impact. If the state before CAS is Idle, initialize; otherwise just enqueue or put into DAC Inbox.
When the thread needs to be naturally disposed	ToBeDisposed (CAS)	Restore CanGetWork state and skip	CanGetWork=Disabled
When the thread needs to be destroyed due to exceeding the thread count limit	ToBeDisposed		CanDeleteRedundantWorker=NotAllowed & CanGetWork=Disabled
When actually entering the Idle state	Idle		CanGetWork=ToBeDisabled
When a temporary helper Worker is created	Running
When the temporary helper Worker finishes	Idle
When forced stop begins	ToBeDisposed

CanDispose

Trigger Event	State Change	If CAS Fails	Prerequisite
On Dispose	NotAllowed (CAS)	Directly skip

CanForceStop

Trigger Event	State Change	If CAS Fails	Prerequisite
When preparing to force stop	NotAllowed (CAS)	Directly skip; if there is an associated return value, return the value representing force-stop failure
When force stop is requested while WorkerState is already in a non-Running state	Allowed

CanGetWork

Trigger Event	State Change	If CAS Fails	Prerequisite
When a Worker is newly created due to insufficient count	Allowed
When a Worker is attempted to be woken from Idle	NotAllowed (CAS)	Skip and put it back into the Idle set
When a running Worker is attempted to be obtained to enqueue a new task	NotAllowed (CAS)	Directly skip
When a Worker better suited for enqueuing the new task is found	Allowed
When being force-stopped	Disabled
When thread count exceeds the limit and the Worker is attempted to be destroyed	Disabled (CAS)	Directly skip	CanDeleteRedundantWorker=NotAllowed
When it is double-confirmed that the thread does not need to be destroyed	Allowed
When a new task has been successfully enqueued (excluding stealing)	Allowed
When attempting to enter Idle state	ToBeDisabled (CAS)	Directly skip
When attempting to enter Idle state but a new task is obtained within this time window	Allowed (CAS)	CAS is only for verification during development
When attempting to enter Idle state succeeds and there is no keep-alive setting	Disabled (CAS)	CAS is only for verification during development
When attempting to enter Idle state succeeds, waiting to be woken up or destroyed after keep-alive timeout	Allowed (CAS)	CAS is only for verification during development
When the thread needs to be naturally disposed	Disabled (CAS)	Use spinning; must exit spinning when it has succeeded or is already in Disabled state (failed but the state requirement is satisfied)
When natural disposal of the thread fails (due to race; the current Worker has resumed running)	Allowed (CAS)	CAS is only for verification during development

WorkStealability

Trigger Event	State Change	If CAS Fails	Prerequisite
When obtained in the help-execution logic in order to steal tasks	NotAllowed (CAS)	Directly skip
When finishing stealing tasks in the help-execution logic	Allowed
When obtained in the work-stealing logic in order to steal tasks	NotAllowed (CAS)	Directly skip
When, in the work-stealing logic, a Worker better suited for enqueuing new tasks is found	Allowed
When tasks are stolen from this Worker and the steal succeeds	Allowed
When the relationship between the Worker and its currently held work needs to be frozen (during force stop or cancellation)	NotAllowed (CAS)	Spin until the state changes. Both force stop and cancellation are non-blocking operations; freezing will always complete immediately, so the spinning will certainly be short-spin
When the relationship between the Worker and its currently held work needs to be unfrozen	Allowed

WorkHeldState

Trigger Event	State Change	If CAS Fails	Prerequisite
When the relationship between the Worker and its currently held work needs to be frozen (during force stop or cancellation)	Held (CAS)	Spin until the state changes. Both force stop and cancellation are non-blocking operations; freezing will always complete immediately, so the spinning will certainly be short-spin	WorkStealability=NotAllowed
When the relationship between the Worker and its currently held work needs to be unfrozen	NotHeld