Compare-and-swap 2026-10-31
An atomic read-modify-write that replaces a location only if its current value matches an expected value. An update based on an old version can retry after a failed compare-and-swap instead of blindly overwriting a competing writer's change.
Exactly-once pin cleanup 2026-11-05
Exactly-once pin cleanup gives one owner responsibility for ending a particular pinning lifetime, despite several possible completion paths. The resource must remain pinned until its consumers finish, and its cleanup must occur once after that condition is met. This combines the safety requirement of asynchronous pin release with a liveness requirement: a pin left permanently owned by nobody can obstruct later memory compaction.
A useful state machine has states “submission in progress”, “completion pending” and “released”. A submission error that guarantees no future consumer can access the address may take the first state directly to released. A pending result transfers cleanup responsibility to a completion path. Immediate completion must follow the API's stated rule about whether a callback will also arrive; it must not independently trigger two owners to dispose of the same handle. Microsoft's native-call example explicitly handles submission exceptions and funnels immediate or deferred completion into cleanup.[ref]
Memory and Span usage guidelines by Microsoft. Rule10, exception branch, non-pending result branch and completion routine
For an original race example, suppose a completion callback and an error-reporting path can both observe an operation as finished. A plain sequence “test flag; if clear, set flag; release pin” can race: both callers read clear before either writes set, and both release. A synchronized ownership claim, using a mutex or an appropriate compare-and-swap transition, can ensure that only one caller changes pending to released. The losing caller performs no cleanup. This is a schematic ownership protocol, not a claim that every pin API tolerates arbitrary racing callbacks.
For a count of releases on one acquired handle, the at-most-once condition is . The requirement that cleanup eventually occurs after all promised access has ended supplies at final completion, so together they give . A one-time flag alone establishes neither the end of access nor eventual execution of the winning path.
The operation's state must also stay alive while any callback can inspect it. Reference counting can protect that state independently of the pin being released. Marking a handle released does not permit another caller to access freed callback state. Keeping resource ownership, callback-state lifetime and data synchronization separate makes the protocol easier to verify. The primary task-based memory rule similarly requires all consumer access to finish by task termination.[ref]
Memory and Span usage guidelines by Microsoft. Rule4, consumer lifetime through terminal task state
O1 supplies mutual exclusion over each complete operation. A reader sees the protected state from a serial order of readers and writers; a writer can safely mutate the existing object while holding the common mutex. If an update reads the old state and modifies it, the entire read-modify-write can be protected against a lost update.
O2 has different semantics. Assume an indivisible aligned access to
p, correct publication ordering, and no later mutation of a published version. A reader chooses one version when it loads p; the subsequent load instructions use the same saved pointer, so an overlapping writer need not block that reader. The reader may continue observing the old complete version after the writer has published the new one. A sufficiently specified snapshot read can still be linearizable, with the load of p as its linearization point: a read overlapping a write is allowed to come before that write in the abstract order. Thus reading an old version during overlap does not by itself prove failure of linearizability.Multiple writers can each construct a valid replacement and race to publish it. If the operation is simply “replace the whole object”, atomic pointer stores give a last-publication order. If each writer instead computes a change from the previously published state, O2 alone does not preserve both changes. Two writers can both read a counter value zero, independently build a version containing one, and publish their respective versions; the final counter is one. A writer mutex or compare-and-swap with retry is needed when the required operation must avoid that lost update. Memory barriers alone do not provide this exclusion.
Releasing the old object's computer memory immediately after publishing a replacement is unsafe. For example:The final operation is a use-after-free: it may crash, see allocator metadata, or read a different object that reuses the same memory address. An atomic update of
reader: q := LOAD [p] // gets the old object
writer: STORE [p], new_object
writer: free(old_object)
reader: LOAD [q + field] // accesses freed storagep does not update the reader's already copied pointer. Freeing the new object immediately would of course also invalidate p itself.Publication and reclamation are separate: publish a fully initialized version, and retain every old version until all readers that might hold it have finished. A read-copy-update grace period or a suitable hazard pointer protocol can provide that lifetime guarantee; naive reference-count acquisition after loading an unprotected pointer can itself race with freeing. Correct memory barriers in part (iii) ensure visibility, not object lifetime.