xtc_chash.h

/*-
 * Copyright (c) 2026, The XTC Project
 * Use of this source code is governed by the ISC License.
 *
 * src/inc/xtc_chash.h
 *      RCU-protected concurrent hash table.  M13a's primary xtc_rcu
 *      consumer: chained buckets, an array of per-bucket mutexes for
 *      writers (insert/remove/put serialize only within their own
 *      bucket -- disjoint buckets proceed fully in parallel), and
 *      xtc_rcu for readers, so xtc_chash_get takes NO lock at all --
 *      it is wait-free, bounded only by the chain length it walks.
 *
 *      Storage convention: caller-owned `void *` key and value,
 *      same as xtc_pdict / xtc_chan (the store keeps the pointers
 *      verbatim, no deep copy).  A caller-supplied comparator and
 *      hash function make the table key-type-agnostic (the qsort(3)
 *      convention: cmp returns <0/0/>0, but chash only tests for
 *      equality via == 0).
 *
 *      Concurrency model:
 *        - Bucket array is RCU-protected: a resize (grow always-on;
 *          shrink opt-in) allocates a new array of the new size,
 *          rehashes every live node into it, publishes the new array
 *          pointer, and retires the old array via xtc_rcu_retire.  A
 *          reader that loaded the old array pointer before the swap
 *          keeps working against a fully valid table (larger or smaller
 *          than the new one) until it leaves its read-side; nothing
 *          under it moves.
 *        - Node removal: unlinked from its bucket chain under that
 *          bucket's mutex, then handed to xtc_rcu_retire -- NEVER
 *          freed directly.  A concurrent reader that already read a
 *          pointer to the node (chased into the chain before the
 *          unlink) keeps a valid node until it leaves its read-side;
 *          RCU reclaims the node only after every such reader has
 *          drained.
 *        - Node insertion: publish-then-link.  The new node is fully
 *          initialized, then linked with a single release-ordered
 *          store of its bucket's head pointer (or the previous node's
 *          `next`) -- a concurrent reader either sees the old chain or
 *          the new one with the new node fully formed, never a torn
 *          write.
 *
 *      xtc_chash_get returns a value pointer valid only inside the
 *      caller's OWN xtc_rcu_read_lock/xtc_rcu_read_unlock bracket,
 *      exactly like every other xtc_rcu consumer: the table does not
 *      take the read-side for you (nesting is supported, see
 *      xtc_rcu.h, so a caller already inside a read-side may call get
 *      freely) because a caller usually wants to do more than one
 *      lookup, or read the value's contents, inside one critical
 *      section.  xtc_chash_insert / _remove take the read-side
 *      internally for their own traversal and enter/leave it around
 *      the mutable part; callers do not need an outer read-side for
 *      them.
 *
 *      Resize: xtc_chash grows automatically past a 75% load factor
 *      and, IF auto-shrink is enabled (xtc_chash_set_auto_shrink,
 *      default OFF), shrinks automatically once the load factor drops
 *      below 18.75% -- never below the initial/minimum bucket count.
 *      Grow is always-on because memory then tracks the true working-
 *      set peak monotonically; shrink is opt-in because it adds a
 *      rehash to the remove path and reclaims memory a user may have
 *      wanted stable.  Enable it if your workload has large delete
 *      phases and you want memory reclaimed.  The wide gap between the
 *      grow (0.75) and shrink (0.1875) triggers is deliberate
 *      hysteresis: a table hovering near a boundary cannot thrash
 *      grow<->shrink (a shrink halves the array, doubling the load
 *      factor to <0.375, so re-growing needs a 2x count increase).
 *      Both directions use the identical RCU-swap machinery.
 */
 
#ifndef XTC_CHASH_H
#define XTC_CHASH_H
 
#include "xtc_export.h"
 
#include <stddef.h>
#include <stdint.h>
 
#include "xtc.h"
 
typedef struct xtc_chash xtc_chash_t;
 
/* Key comparator: qsort(3) convention, but chash only ever tests the
 * sign of the result for == 0 (equal) vs. != 0 (not equal) -- it does
 * not order buckets, so a cmp that only distinguishes equal/unequal
 * (e.g. always returns 0 or 1) is fine too. */
typedef int      (*xtc_chash_cmp_fn)(const void *a, const void *b);
 
/* Key hash: any distribution is correct; a poor one just chains more.
 * Same key must always hash the same for the table's lifetime. */
typedef uint64_t (*xtc_chash_hash_fn)(const void *key);
 
/*
 * PUBLIC: int    xtc_chash_create __P((xtc_chash_cmp_fn, xtc_chash_hash_fn, size_t, xtc_chash_t **));
 * PUBLIC: void   xtc_chash_destroy __P((xtc_chash_t *));
 * PUBLIC: int    xtc_chash_get __P((xtc_chash_t *, const void *, void **));
 * PUBLIC: int    xtc_chash_insert __P((xtc_chash_t *, void *, void *, void **));
 * PUBLIC: int    xtc_chash_remove __P((xtc_chash_t *, const void *, void **));
 * PUBLIC: size_t xtc_chash_size __P((const xtc_chash_t *));
 * PUBLIC: void   xtc_chash_set_auto_shrink __P((xtc_chash_t *, int));
 * PUBLIC: int    xtc_chash_get_auto_shrink __P((const xtc_chash_t *));
 */
 
/*
 * Create a table with `initial_capacity` buckets (rounded up to the
 * next power of two; 0 defaults to 16).  cmp and hash are required
 * (non-NULL).  Returns XTC_OK and writes *out, or XTC_E_INVAL /
 * XTC_E_NOMEM.
 */
XTC_API int    xtc_chash_create(xtc_chash_cmp_fn cmp, xtc_chash_hash_fn hash,
                                size_t initial_capacity, xtc_chash_t **out);
 
/*
 * Destroy the table.  NOT concurrency-safe with any other call on
 * this table -- the caller must ensure no reader or writer is active
 * (same contract as xtc_rcu_fini: quiesce first).  Frees every
 * remaining node's bookkeeping but NOT the caller-owned key/value
 * pointers (caller-owned lifetime, same as xtc_chan/xtc_pdict).
 */
XTC_API void   xtc_chash_destroy(xtc_chash_t *h);
 
/*
 * Look up `key`.  On a hit, returns XTC_OK and writes the value
 * pointer to *out_value.  On a miss, returns XTC_E_NOTFOUND and does
 * not touch *out_value.  MUST be called inside the caller's own
 * xtc_rcu_read_lock/_read_unlock bracket (nesting is fine); the value
 * pointer written to *out_value is valid only until the caller's
 * read-side ends -- do not stash it past xtc_rcu_read_unlock unless
 * you know the value's lifetime is otherwise pinned.
 */
XTC_API int    xtc_chash_get(xtc_chash_t *h, const void *key, void **out_value);
 
/*
 * Insert (key, value), replacing any existing entry for an equal key.
 * Takes the read-side internally for its traversal; safe to call
 * without an outer read-side.  On replace, the OLD value pointer is
 * written to *out_old_value (out_old_value may be NULL if the caller
 * does not care) and the old key/value pointers are the caller's
 * responsibility to free (this table never frees caller payloads).
 * On a fresh insert, *out_old_value (if non-NULL) is set to NULL.
 * May trigger a resize (a grow).  Returns XTC_OK or XTC_E_NOMEM.
 */
XTC_API int    xtc_chash_insert(xtc_chash_t *h, void *key, void *value,
                                void **out_old_value);
 
/*
 * Remove the entry for `key`.  On a hit, returns XTC_OK and writes
 * the removed value pointer to *out_value (may be NULL).  The node's
 * memory is retired via xtc_rcu_retire, never freed synchronously --
 * a concurrent reader that is already inside the chain sees a fully
 * valid node until it leaves its read-side.  On a miss, returns
 * XTC_E_NOTFOUND.
 */
XTC_API int    xtc_chash_remove(xtc_chash_t *h, const void *key, void **out_value);
 
/* Approximate live entry count (an _Atomic counter; exact at any
 * instant with no concurrent writers, a fresh-enough estimate
 * otherwise). */
XTC_API size_t xtc_chash_size(const xtc_chash_t *h);
 
/*
 * Enable (on != 0) or disable (on == 0) automatic shrink; default OFF.
 * When enabled, xtc_chash_remove may halve the bucket array once the
 * load factor drops below 18.75%, never below the initial capacity
 * passed to xtc_chash_create (minimum 16).  Shrink uses the same
 * RCU-swap machinery as grow: it takes every stripe lock, rehashes
 * every live node into the smaller array, publishes it with a release
 * store, and retires the old array via xtc_rcu_retire, so a concurrent
 * reader mid-lookup on the old array stays correct until its read-side
 * ends.  The 0.75-grow / 0.1875-shrink gap is deliberate hysteresis --
 * a table near a boundary cannot thrash grow<->shrink.  Enable it if
 * your workload has large delete phases and you want the memory
 * reclaimed; leave it OFF (the default) if you want stable memory and
 * a lighter remove path.  Takes effect immediately and is safe to call
 * from any thread (a single atomic flag).
 */
XTC_API void   xtc_chash_set_auto_shrink(xtc_chash_t *h, int on);
XTC_API int    xtc_chash_get_auto_shrink(const xtc_chash_t *h);
 
#endif /* XTC_CHASH_H */