NAME
    Data::HashMap::Shared - Multiprocess shared-memory hash maps with LRU
    eviction and per-key TTL

SYNOPSIS
        use Data::HashMap::Shared::II;

        # Create or open a shared map (file-backed mmap)
        my $map = Data::HashMap::Shared::II->new('/tmp/mymap.shm', 100000);

        # Keyword API (fastest)
        shm_ii_put $map, 42, 100;
        my $val = shm_ii_get $map, 42;

        # Method API
        $map->put(42, 100);
        my $v = $map->get(42);

        # Atomic counters (under the read lock, without LRU or TTL)
        shm_ii_incr $map, 1;            # 1
        shm_ii_incr_by $map, 1, 10;     # 11
        shm_ii_max $map, 1, 50;         # monotonic: store max(current, 50) -> 50

        # Compare-and-swap (all variants; byte-compare for string values)
        shm_ii_cas $map, 1, 50, 42;     # swap to 42 only if current == 50

        # LRU cache (evicts least-recently-used when full)
        my $cache = Data::HashMap::Shared::II->new('/tmp/cache.shm', 100000, 1000);
        shm_ii_put $cache, 42, 100;    # auto-evicts LRU entry if size > 1000

        # TTL (entries expire after N seconds)
        my $ttl_map = Data::HashMap::Shared::II->new('/tmp/ttl.shm', 100000, 0, 60);
        shm_ii_put $ttl_map, 1, 10;          # expires in 60s
        shm_ii_put_ttl $ttl_map, 2, 20, 5;   # per-key: expires in 5s

        # Multiprocess
        if (fork() == 0) {
            my $child = Data::HashMap::Shared::II->new('/tmp/mymap.shm', 100000);
            shm_ii_incr $child, 1;   # atomic increment visible to parent
            exit;
        }
        wait;

DESCRIPTION
    Data::HashMap::Shared provides type-specialized hash maps stored in
    file-backed shared memory (mmap(MAP_SHARED)), enabling efficient
    multiprocess data sharing on Linux. With opt-in LRU eviction and per-key
    TTL it doubles as a fast cross-process cache; lookups take a lock-free
    seqlock fast path.

    Linux-only. Requires 64-bit Perl.

  Features
    *   File-backed mmap for cross-process sharing

    *   Futex-based read-write lock (fast userspace path)

    *   Atomic counters (incr/decr under the read lock on maps without LRU or
        TTL)

    *   Elastic capacity (starts small, grows/shrinks automatically)

    *   Arena allocator for string storage in shared memory

    *   Keyword API via XS::Parse::Keyword for maximum speed

    *   Opt-in LRU eviction -- clock/second-chance algorithm; reads stay
        lock-free

    *   Opt-in per-key TTL expiry -- lazy removal on access; monotonic clock

    *   Stale lock recovery for both writers and readers (dead PIDs detected
        and drained automatically)

  Variants
    Data::HashMap::Shared::I16 - int16 to int16
    Data::HashMap::Shared::I32 - int32 to int32
    Data::HashMap::Shared::II - int64 to int64
    Data::HashMap::Shared::I16S - int16 to string
    Data::HashMap::Shared::I32S - int32 to string
    Data::HashMap::Shared::IS - int64 to string
    Data::HashMap::Shared::SI16 - string to int16
    Data::HashMap::Shared::SI32 - string to int32
    Data::HashMap::Shared::SI - string to int64
    Data::HashMap::Shared::SS - string to string

  Integer Range and Wrapping
    Integer keys and values are stored as fixed-width two's-complement
    integers: "I16"/"SI16"/"I16S" use a signed 16-bit range (-32768 .. 32767),
    "I32"/"SI32"/"I32S" a signed 32-bit range, and "II"/"IS"/"SI" a signed
    64-bit range. A key or value outside the variant's range is silently
    truncated to the low bits (two's complement), with no warning: on an "I16"
    map, "$map->put(70000, ...)" stores under key 4464 ("70000 & 0xFFFF"), so
    get(70000) and get(4464) address the same entry. "incr"/"decr" wrap the
    same way ("32767 + 1" becomes -32768). Pick a variant wide enough for your
    data.

  Constructor
        my $map = Data::HashMap::Shared::II->new($path, $max_entries);
        my $map = Data::HashMap::Shared::II->new(undef, $max_entries);    # anonymous
        my $map = Data::HashMap::Shared::II->new($path, $max_entries, $max_size);
        my $map = Data::HashMap::Shared::II->new($path, $max_entries, $max_size, $ttl);
        my $map = Data::HashMap::Shared::II->new($path, $max_entries, $max_size, $ttl, $lru_skip);
        my $map = Data::HashMap::Shared::SS->new($path, $max_entries, 0, 0, 0, $arena_cap); # explicit arena bytes
        my $map = Data::HashMap::Shared::II->new($path, $max_entries, $max_size, $ttl, $lru_skip, $arena_cap, $file_mode);
        my $map = Data::HashMap::Shared::II->new_sharded($prefix, $shards, $max_entries, $max_size, $ttl, $lru_skip, $arena_cap, $file_mode);
        my $map = Data::HashMap::Shared::II->new_memfd($name, $max_entries, ...); # memfd-backed
        my $map = Data::HashMap::Shared::II->new_from_fd($fd);            # reopen memfd
        my $fd  = $map->memfd;                                            # -1 if not memfd

    Creates or opens a shared hash map backed by file $path. Passing "undef"
    as the path creates an anonymous "MAP_SHARED|MAP_ANONYMOUS" mapping that
    is inherited across "fork" but has no filesystem presence.

    "new_memfd" creates an unlinked memfd-backed map whose file descriptor can
    be passed to another process (via "SCM_RIGHTS", "fork"+"exec", or
    duped+open). "new_from_fd" reopens such a descriptor. The descriptor you
    pass is duplicated ("F_DUPFD_CLOEXEC"), so it stays yours to close and
    closing it does not disturb the handle. Both require a 64-bit Perl on
    Linux (memfd_create(2)).

    "$map->memfd" goes the other way: it returns the handle's own descriptor,
    not a copy. Pass it, but do not close it: the handle closes it when the
    map goes away, and in between the number is handed out again to the next
    file the process opens, so closing it early makes the map's eventual close
    hit an unrelated file. Dup it first if it has to outlive the map.

    $max_entries, $max_size, $ttl, $lru_skip, $arena_cap, and $file_mode are
    used only when creating a new file; when opening an existing one, all
    parameters are read from the stored header and the constructor arguments
    are ignored -- but still range-checked, so a value the module would never
    accept at creation is rejected either way.

    $shards is the exception: it is not recorded in any file, so every process
    opening a sharded set must pass the same count. A different count routes
    keys to the wrong shard and creates the missing shard files, which loses
    access to roughly the data that hashes elsewhere -- silently. Treat it as
    part of the path. Multiple processes can open the same file
    simultaneously. Dies if the file exists but was created by a different
    variant or is corrupt.

    Optional $max_size enables LRU eviction: when the map reaches $max_size
    entries, the least-recently-used entry is evicted on insert. Set to 0
    (default) to disable. LRU uses a clock/second-chance algorithm: "get" sets
    an accessed bit (lock-free, no write lock), and eviction gives a second
    chance to recently accessed entries before evicting.

    A $max_size at or above the slot count -- 2048 for a map created with
    1000, not the 1536 "max_entries" reports -- can never drive eviction:
    above it the bound is never reached, and at exactly it the insert fails on
    the full table first. The map then fills up and refuses further inserts,
    keeping its oldest keys. Every constructor that opens a map for writing
    warns about this (category "misc"), reading the bound off the map rather
    than its arguments, so attaching to a sound file never warns however its
    arguments are written. It warns rather than dies: the bound is merely
    unreachable, and LRU ordering still drives "pop"/"shift". Silence it with
    "no warnings 'misc'".

    Eviction counts entries, not bytes, so on a string variant the arena can
    run out while the count is still below $max_size: nothing is evicted and
    every insert needing arena space fails, leaving the cache stuck with its
    oldest entries. A single key or value may be at most 1 GB; longer ones
    croak. Keys and values of 7 bytes or fewer are stored inline and keep
    succeeding, which is how an exhausted arena can look like a working cache
    -- so size $arena_cap for what you store, and check what the insert
    returned. An insert at $max_size evicts before it stores, so the first
    failure costs one entry and later ones cost nothing until an insert
    succeeds again.

    Optional $ttl sets a default time-to-live in seconds for all entries.
    Expired entries are reclaimed lazily, by the next mutating access to that
    key. "remove", "update", "take", "cas", "touch", "persist" and "set_ttl"
    free the slot; "incr", "add" and "get_or_set" free it and insert afresh,
    so the key is live again when they return; "put" overwrites it in place. A
    read -- "get", "exists", "get_with_ttl", "get_multi", "ttl_remaining" --
    reports it absent but leaves it there, so "size" still counts it. Set to 0
    (default) to disable.

    TTLs have whole-second granularity and the deadline is truncated, so an
    entry given $n seconds expires somewhere between "$n-1" and $n seconds
    later: a TTL of 1 can expire almost immediately. Leave at least a second
    of margin over any refresh interval that has to beat it.

    Any operation that stores a value resets the entry's TTL to the map
    default: "put", "update", "cas", "incr"/"decr"/"incr_by", "max"/"min",
    "get_or_set" on a hit, and "set_multi", as well as the documented "touch"
    and "swap". A permanent entry (TTL 0) stays permanent. To carry a per-key
    TTL across a value change, write it with "put_ttl"/"update_ttl" or restore
    it afterwards with "set_ttl".

    An expired entry still occupies its slot until something reclaims it, so a
    TTL map without LRU eviction can fill up with entries that every read
    reports as absent: once no slot is free, inserts fail and "incr" dies
    while "keys" returns nothing. Call "flush_expired" (or
    "flush_expired_partial") on a timer if keys are not revisited; $max_size
    avoids the problem entirely because eviction reclaims slots. When TTL is
    active, "get" and "exists" check expiry. Expiry is measured against a
    monotonic clock ("CLOCK_MONOTONIC_COARSE"): TTLs track elapsed running
    time and do not advance while the system is suspended or hibernating.

    That clock is local to the current boot: it restarts at zero on reboot and
    is unrelated between machines, while the expiry timestamps live in the
    file. A map that outlives the boot which wrote it -- or is copied to
    another host, a frozen map shipped elsewhere included -- therefore carries
    deadlines on a timeline that no longer exists, and the error runs
    whichever way the destination clock does. Where its uptime has not yet
    reached the stored values, entries live longer than their TTL and correct
    themselves as it passes them; where its uptime already exceeds them -- a
    map built on a freshly booted host and served on a long-running one --
    every TTL'd entry arrives expired while "size" still counts it. Nothing
    crashes, but do not rely on TTL across a reboot or a host move:
    "flush_expired", "persist" what you ship, or rebuild.

    Optional $lru_skip (0-99, default 0; 100 or more disables skipping, a
    negative value is rejected like any out-of-range size) reduces how often
    LRU promotion reorders the recency list -- higher values skip more.
    Promotion runs only where an operation updates an existing entry under the
    write lock ("put"/"incr"/"get_or_set" on a hit, the update family); "get"
    and "get_with_ttl" never promote, they set the lock-free accessed bit that
    clock eviction consumes. Skipping cuts write-lock churn on Zipfian
    workloads where a few hot keys dominate. The eviction victim itself is
    never skipped, so eviction stays correct. Set to 0 for strict LRU
    ordering.

    Optional $arena_cap (bytes) sizes the string arena explicitly instead of
    deriving it from $max_entries. The default is roughly 128 bytes per entry
    (4096 minimum), which a few large strings can exhaust while the table is
    nearly empty. Clamped to "[4096, 0xFFFFFFFF]" (arena offsets are 32-bit);
    integer-only variants ("II"/"I16"/"I32") have none and ignore it; for
    sharded maps it is per shard, like $max_entries.

    Size it from the rounded lengths, not the byte totals: blocks are powers
    of two with a 16-byte minimum, so a 100-byte value takes 128 and a
    1100-byte value 2048, and a total can be short by up to half. Blocks are
    recycled by exact size class, so a freed large one never serves a smaller
    request -- a workload alternating value sizes needs room for one block of
    each size it uses, not just the largest.

    Optional $file_mode (octal, default 0600) sets the permission bits used
    when the backing file is created; the exact mode is applied via "fchmod",
    so the process umask does not narrow it. It is ignored when attaching an
    existing file and for anonymous or memfd-backed maps. The default is
    owner-only; pass a wider mode such as 0666 to opt in to cross-user
    sharing. Before version 0.14 the default was 0666.

    Zero-cost when disabled: with both "$max_size=0" and "$ttl=0", the fast
    lock-free read path is used. The only overhead is a branch (predicted
    away).

  String Keys/Values and UTF-8
    String-key variants ("SS", "SI", "SI16", "SI32") compare keys as raw
    bytes: two keys are the same entry if and only if they contain the same
    byte sequence. The SV UTF-8 flag is stored alongside the key so retrieval
    round-trips it to the returned SV, but it is not part of key identity.
    Consequences:

    *   ASCII keys with a toggled UTF-8 flag hash and match the same entry
        ("use utf8", "utf8::upgrade", and "utf8::downgrade" on ASCII are all
        equivalent from the map's point of view).

    *   Non-ASCII keys with different byte encodings are distinct. "caf\xe9"
        (latin-1, 4 bytes) and the same character sequence under "use utf8"
        ("caf\xc3\xa9", 5 UTF-8 bytes) are two different keys. If your input
        comes in mixed encodings, normalize with "Encode::encode_utf8" before
        use.

    Because the flag is not part of key identity, the stored key keeps the
    flag it had when the entry was first inserted: a later "put" with the same
    bytes and the opposite flag replaces the value but not the key, so "keys"
    reports the original flag.

    String-value variants ("SS", "IS", "I16S", "I32S") store the SV UTF-8 flag
    alongside each value and round-trip it on retrieval. The "cas" comparison
    of $expected against the stored value is byte-only -- the UTF-8 flag on
    $expected is ignored (same rationale as string-key equality).

  Sharding
        my $map = Data::HashMap::Shared::II->new_sharded($path_prefix, $shards, $max_entries, ...);

    Creates $shards independent maps (files "$path_prefix.0",
    "$path_prefix.1", ...) behind a single handle, each with up to
    $max_entries entries (each sized as if it were a map of its own -- see
    "max_entries" below for what the total comes to). Per-key operations
    automatically route to the correct shard via hash dispatch. Writes to
    different shards proceed in parallel with independent locks. "new_sharded"
    requires a filesystem $path_prefix; anonymous ("undef"-path) sharded maps
    are not supported.

    The batch ops ("set_multi", "get_multi", "remove_multi") dispatch each key
    to its shard independently rather than holding one lock for the whole
    call, so on a sharded map a batch is not atomic across shards (the "single
    lock" note in the API below applies to non-sharded maps).

    All operations work transparently on sharded maps: "put", "get", "remove",
    "exists", "add", "update", "swap", "take", "incr", "max", "min", "cas",
    "cas_take", "get_or_set", "put_ttl", "add_ttl", "update_ttl", "touch",
    "persist", "set_ttl", "keys", "values", "items", "to_hash", "set_multi"
    (method only), "remove_multi" (method only), "get_multi" (method only),
    "get_with_ttl" (method only), "each", "pop", "shift", "drain", "clear",
    "flush_expired", "flush_expired_partial", "size", "stats" (method only),
    "reserve", and all diagnostic keywords.

    Diagnostic counters and capacities reported for a sharded handle are
    aggregate totals across all shards: "size", "capacity", "max_entries",
    "max_size", "tombstones", "mmap_size", "arena_used", "arena_cap", and the
    "stats" eviction/expiry/recovery counts all sum over the shards. ("ttl" is
    the shared per-entry default, so it reports a single shard's value.)
    "reserve $n" pre-grows each shard to $n entries (not $n in total).

    Every shard file must come from the same configuration. Nothing checks
    this: opening a set whose files disagree -- one left behind by an earlier
    run with a different $ttl or $max_size -- succeeds, and each key then
    behaves according to the shard it routes to. The accessors do not reveal
    it either: "ttl" and "frozen" report shard 0's setting, while the summed
    ones add the disagreeing values into a figure belonging to no shard at all
    (two shards capped at 50 and 20 report "max_size" 70).

    Shard routing and slot placement both come from the same hash, so within a
    shard only one home slot in $shards is reachable -- that shard's own table
    capacity divided by $shards, and a single slot once $shards reaches it.
    Probe runs lengthen with the shard count and with the entries each shard
    holds: not measurably at a handful of shards, around 2x at 4096 shards
    holding a few entries each, and 10x or more at 4096 shards holding a
    thousand each. Use the smallest shard count that relieves your lock
    contention, not the largest you can afford. "max_entries" reports the
    entry count at the table's 75% design load, which is three quarters of the
    maximum slot count and so is neither the constructor argument nor the slot
    count -- a map created with 1000 reports 1536, over 2048 slots. It is not
    a hard ceiling either: the table itself holds "capacity" slots (the next
    power of two above the requested load factor), and inserts keep succeeding
    until every slot is occupied. Probe length grows sharply over the last few
    percent, though: a miss on a table at 99% costs roughly an order of
    magnitude more than at 95%, and one on a completely full table has to walk
    every slot before it can report absence. Treat "max_entries" as the size
    to run at, not the size to reach. On a TTL map "size" counts entries that
    have expired but not yet been reclaimed, which "keys" and the iterators
    skip; "flush_expired" reconciles the two.

    Cursors chain across shards automatically. "cursor_seek" routes to the
    correct shard based on key hash. $shards is rounded up to the next power
    of 2.

  API
    Replace "xx" with variant prefix: "i16", "i32", "ii", "i16s", "i32s",
    "is", "si16", "si32", "si", "ss".

        my $ok = shm_xx_put $map, $key, $value;   # insert or overwrite
        my $ok = shm_xx_add $map, $key, $value;   # insert only if key absent
        my $ok = shm_xx_update $map, $key, $value; # overwrite only if key exists
        my $old = shm_xx_swap $map, $key, $value; # put + return old value (undef if new)
        my $ok = shm_xx_cas $map, $key, $expected, $desired; # compare-and-swap
        my $v  = shm_xx_cas_take $map, $key, $expected; # compare-and-remove; returns value on match, undef otherwise
        my $n  = $map->set_multi($k, $v, ...);   # batch put under single lock, returns count
        my $n  = $map->remove_multi(@keys);      # batch remove under single lock, returns count
        my @v  = $map->get_multi($k1, $k2, ...); # batch get under single lock with prefetch pipeline
        my ($v, $ttl) = $map->get_with_ttl($key); # atomic snapshot; () if missing, $ttl is undef on non-TTL map, 0 = permanent; sets LRU clock bit
        my $v  = shm_xx_get $map, $key;           # returns undef if not found
        my $ok = shm_xx_remove $map, $key;        # returns false if not found
        my $ok = shm_xx_exists $map, $key;        # returns boolean
        my $s  = shm_xx_size $map;
        my $m  = shm_xx_max_entries $map;
        my @k  = shm_xx_keys $map;
        my @v  = shm_xx_values $map;
        my @items = shm_xx_items $map;            # flat (k, v, k, v, ...)
        while (my ($k, $v) = shm_xx_each $map) { ... }  # auto-resets at end
        shm_xx_iter_reset $map;
        shm_xx_clear $map;
        my $href = shm_xx_to_hash $map;
        my $v  = shm_xx_get_or_set $map, $key, $default;  # returns value

    Several calls below fail for want of room. No room means the table is full
    (every slot occupied -- see "capacity") or, on a variant with string keys
    or values, the arena is.

    "get_or_set" returns the existing value, or stores and returns $default
    when the key is absent; "undef" only when the key is absent and there is
    no room.

    "cas", available for all variants, returns true when the stored value
    matched $expected and was atomically replaced with $desired; false if the
    key is missing or expired, the value did not match, or there is no room.
    See "String Keys/Values and UTF-8" for the byte-only comparison rule.

    "swap" returns the previous value, or "undef" when the key did not exist
    -- and also "undef" when there is no room, in which case an existing key
    keeps its old value. It therefore cannot by itself tell a fresh insert
    from a failure; check "exists" or "size" first if that matters. On a TTL
    map it refreshes an existing entry's TTL to the default and assigns the
    default on insert, leaving a permanent entry (TTL 0) permanent.

    Integer-value variants also have:

        my $n = shm_xx_incr $map, $key;           # returns new value
        my $n = shm_xx_decr $map, $key;           # returns new value
        my $n = shm_xx_incr_by $map, $key, $delta;
        my $n = shm_xx_max $map, $key, $desired;  # store max(current, desired), return it
        my $n = shm_xx_min $map, $key, $desired;  # store min(current, desired), return it

    A missing key is created starting from zero (Redis-style): the first
    "incr" returns 1, "decr" returns -1, and "incr_by" returns $delta. These
    die only when the key is new and there is no room for it. The result wraps
    at the variant's integer width (see "Integer Range and Wrapping").

    "max"/"min" atomically store "max($current, $desired)" / "min($current,
    $desired)" and return the resulting value; a missing key is inserted as
    $desired. The compare-and-set runs under a single lock acquisition, so
    there is no read-modify-write gap for a concurrent
    "incr_by"/"cas"/"max"/"min" on the same key: the result is monotonic
    ("max" never lowers, "min" never raises) and never clobbers a concurrent
    increment. On a map with neither LRU nor TTL the common "value already
    past the bound" case completes under a shared read lock, writing nothing;
    with either enabled every call takes the write lock, promotes the entry in
    the LRU order and refreshes its TTL even when it stores nothing. Like
    "incr_by", they die only when the key is new and there is no room for it,
    and the result wraps at the variant's integer width.

    LRU/TTL operations ("put_ttl", "add_ttl", and "update_ttl" require a
    TTL-enabled map):

        my $ok = shm_xx_put_ttl $map, $key, $value, $ttl_sec;  # per-key TTL (0 = permanent); requires TTL-enabled map
        my $ok = shm_xx_add_ttl $map, $key, $value, $ttl_sec;  # insert-if-absent with per-key TTL (0 = permanent)
        my $ok = shm_xx_update_ttl $map, $key, $value, $ttl_sec; # overwrite-only with per-key TTL (0 = permanent)
        my $ms = shm_xx_max_size $map;            # LRU capacity (0 = disabled)
        my $t  = shm_xx_ttl $map;                 # default TTL in seconds
        my $r  = shm_xx_ttl_remaining $map, $key; # seconds left (0 = permanent, undef if missing/expired/no TTL)
        my $ok = shm_xx_touch $map, $key;         # refresh TTL to default (permanent entries stay permanent); promotes in LRU; false if no TTL/LRU
        my $ok = shm_xx_persist $map, $key;       # remove TTL, make key permanent; false on non-TTL maps
        my $ok = shm_xx_set_ttl $map, $key, $sec; # change TTL without changing value (0 = permanent); false on non-TTL maps
        my $n  = shm_xx_flush_expired $map;       # proactively expire all stale entries, returns count
        my ($n, $done) = shm_xx_flush_expired_partial $map, $limit;  # gradual: scan $limit slots ($limit per shard on sharded maps; $done true once every shard completes a cycle)

    Atomic remove-and-return:

        my $v = shm_xx_take $map, $key;           # remove key and return value (undef if missing)
        my ($k, $v) = shm_xx_pop $map;            # remove+return from LRU tail / scan forward
        my ($k, $v) = shm_xx_shift $map;          # remove+return from LRU head / scan backward
        my @kv = shm_xx_drain $map, $n;           # remove+return up to N entries as flat (k,v,...) list

    "pop" and "shift" remove from opposite ends: "pop" takes the LRU tail
    (oldest / least recently used) while "shift" takes the LRU head (newest /
    most recently used). On a sharded map they walk the shards in turn and
    take from each shard's own end, so a sequence of "pop"s is not in global
    recency order. On non-LRU maps, "pop" scans forward and "shift" scans
    backward. "drain" removes in "pop" order (tail-first). Useful for
    work-queue patterns and batch processing.

    Cursors (independent iterators, allow nesting and removal during
    iteration):

        my $cur = shm_xx_cursor $map;             # create cursor
        while (my ($k, $v) = shm_xx_cursor_next $cur) { ... }
        shm_xx_cursor_reset $cur;                 # restart from beginning
        my $ok = shm_xx_cursor_seek $cur, $key;   # position at key (best-effort across resize); true if found, false if missing/expired
        # cursor auto-destroyed when out of scope

    "shm_xx_each" is also safe to use with "remove" during iteration.

    Leaving an "each" loop early -- "last", "return", an exception -- leaves
    the built-in iterator open on that handle, and tombstone compaction and
    shrink stay deferred for as long as it is: a long-lived handle that keeps
    removing and re-inserting keys then grows its table instead of compacting
    it, all the way to its maximum slot count. Removals on their own leave
    tombstones without growing it. Unlike Perl's "each", "keys" does not reset
    it. Call "iter_reset" when you abandon a pass, or run it to completion.
    The deferral is also per handle, not per map, so another process can
    compact or shrink the table underneath your iteration; that restarts it,
    and an abandoned pass can then yield keys it has already returned.
    Tombstone compaction and shrink are deferred until iteration ends; growth
    is not -- a load-driven insert still resizes, and an iteration in progress
    restarts on the resulting table-generation bump. Compaction is likewise
    not deferred once the table has reached its maximum capacity, can only be
    compacted in place, and its load (live entries plus tombstones) has passed
    75% of the slots -- since deferring it there would wedge a full table: an
    insert during an iteration can restart it, and keys already visited are
    visited again. On a sharded map a cursor restarts only within the shard it
    has reached, so shards it already passed are not revisited; take a fresh
    cursor after a "clear" if you need a complete pass.

    Diagnostics:

        my $cap = shm_xx_capacity $map;           # current table capacity (slots)
        my $tb  = shm_xx_tombstones $map;         # tombstone count
        my $au  = shm_xx_arena_used $map;         # arena high-water mark (0 for int-only)
        my $ac  = shm_xx_arena_cap $map;          # arena total capacity (0 for int-only)
        my $sz  = shm_xx_mmap_size $map;          # backing file size in bytes
        my $ok  = shm_xx_reserve $map, $n;        # pre-grow (false if exceeds max)
        my $ev  = shm_xx_stat_evictions $map;     # cumulative LRU eviction count
        my $ex  = shm_xx_stat_expired $map;       # cumulative TTL expiration count
        my $rc  = shm_xx_stat_recoveries $map;    # cumulative stale lock recovery count
        my $p   = $map->path;                    # backing file path (method only)
        my $s   = $map->stats;                   # hashref with all diagnostics in one call
        # stats keys: size, capacity, max_entries, tombstones, mmap_size,
        #   arena_used, arena_cap, evictions, expired, recoveries, max_size, ttl,
        #   frozen, readonly

    "set_multi", "get_multi", "remove_multi", "get_with_ttl", "stats", "path",
    "sync", and "unlink" are method-only (no keyword form).

    Keywords take their arguments as a list, so a keyword that takes more than
    one argument must be written without parentheses around them:

        shm_ii_put $map, $key, $value;            # correct
        shm_ii_put($map, $key, $value);           # error, usually at compile time

    A single-argument keyword accepts either form. The method call
    "$map->put($key, $value)" is always available if you prefer parentheses.

    "keys", "values", "items", "each", "get_multi", "get_with_ttl", "pop",
    "shift", "drain", "flush_expired_partial" and the cursor's "next" return
    lists. Like any Perl sub returning a list, in scalar context they yield
    their last element -- not a count, and not the first -- so call them in
    list context and use "size" when you want a count.

    Calling "no Data::HashMap::Shared::II;" disables that variant's keywords
    for the rest of the enclosing lexical scope.

    File management:

        $map->sync;                               # flush the mmap to the backing file (msync MS_SYNC)
        $map->unlink;                             # remove backing file (mmap stays valid)
        Data::HashMap::Shared::II->unlink($path); # class method form (single file)

    "sync" issues a synchronous msync(2) over the whole mapping (every shard,
    for sharded maps) and dies on error. Use it to force durability of a
    file-backed map; it is a no-op for anonymous mappings, which have no
    backing file. Changes are visible to other processes sharing the mapping
    without "sync" -- it only affects on-disk persistence.

    "unlink" reports through its return value rather than by dying: it returns
    true when the file (every shard, for sharded maps) was removed and false
    otherwise, including when the file was already gone and when removal was
    refused -- a read-only directory, for instance. Check it if the removal
    mattered; an ignored false is how a file survives a cleanup unnoticed.

  Frozen (Read-Only) Mode
        $map->freeze;                                       # seal the file immutable (durable)
        my $ro = Data::HashMap::Shared::II->new_readonly($path);
        my $v  = $ro->get($key);                            # lock-free query; writes nothing
        my $is_frozen   = $map->frozen;                     # true once sealed
        my $is_readonly = $ro->readonly;                    # true for a read-only handle

    "freeze" permanently seals a map's contents so it can be shipped and
    served read-only (it works on anonymous and memfd maps too, though only a
    file can be shipped). It takes the write lock, sets a seal flag in the
    header, and flushes it to disk ("msync"). Afterwards every mutator on that
    handle croaks and the handle itself becomes read-only. A sharded map seals
    every shard file. Freezing is one-way; there is no unfreeze.

    Quiesce your writers first. A mutator tests the seal on entry and takes
    the write lock afterwards, so another process already inside a mutating
    call when "freeze" runs completes its write after the seal: the sealed
    file changes once more, and a "new_readonly" reader -- which takes no lock
    precisely because a sealed map cannot change -- can observe it. Seal a map
    only when nothing else is writing to it. Stopping the writers, or a
    handshake that leaves them all idle, is the caller's responsibility;
    "freeze" cannot detect a writer that has passed the check but not yet
    reached the lock.

    On a sharded map this is not one straggling write. Whole-map and batch
    operations -- "set_multi", "remove_multi", "clear", "drain", "pop",
    "shift", "flush_expired", "flush_expired_partial", "reserve" -- test the
    seal once and then take each shard's lock in turn, so one that is under
    way when "freeze" lands keeps writing for the whole remainder of the call,
    across every shard it has not reached yet. Quiescing matters more here,
    not less.

    "new_readonly" opens an already-frozen file with "O_RDONLY" and maps it
    "PROT_READ". Because a sealed map's contents cannot change and it can have
    no new writers, queries take no lock at all -- no reader-slot bookkeeping,
    no LRU clock bit, no lazy TTL cleanup -- and never write the mapping. A
    read-only view therefore works from a read-only file or filesystem, and
    any number of processes can share one frozen file at once. All queries and
    full iteration are supported: "get", "exists", "get_with_ttl",
    "get_multi", "keys", "values", "items", "to_hash", "each", and cursors
    ("cursor" and "cursor_next"). Every mutator croaks (including the integer
    counters "incr"/"decr"/"max"/"min", which on a writable map update in
    place under the read lock). "sync" is a silent no-op. "frozen" and
    "readonly" report the state, and "stats" gains matching "frozen" and
    "readonly" keys.

    The seal reuses a previously reserved header byte, so the on-disk format
    and version are unchanged: a file written by an older release is simply
    not frozen (the byte reads 0) and opens read-write exactly as before.

    The two modes never mix: opening a frozen file read-write ("new",
    "new_from_fd") is refused -- open it with "new_readonly" instead -- and
    "new_readonly" refuses a file that has not been frozen (its lock-free
    readers must never race a live writer).

    "new_readonly" is for a single backing file, and there is no read-only
    sharded constructor, so "freeze" on a sharded map seals a set that no
    constructor will reopen: "new_sharded" refuses the frozen shards, and
    reading it back means opening each shard file by name and probing them.
    Freeze single-file maps.

    Portability: a frozen file is a raw memory image. Read it back on the same
    architecture that wrote it (same word size and endianness; the native
    magic and variant id reject a mismatched or wrong-variant file at attach
    time). Ship it by copying the file; do not serve it over NFS or another
    network filesystem while another host has it mapped.

  Crash Safety
    If a process dies (e.g., SIGKILL, OOM kill) while holding the write lock,
    other processes detect the stale lock within 2 seconds and automatically
    recover. The writer's PID is encoded in the rwlock word itself (single
    atomic CAS, no crash window). On "FUTEX_WAIT" timeout, waiters "kill($pid,
    0)" the holder and CAS-release the lock if it's dead.

    Reader-side recovery uses a 1024-slot table in the shared mmap (one slot
    per handle, claimed lazily on first lock -- a process holding several
    handles on one map uses a slot for each; fork()'d children claim a fresh
    slot via "pthread_atfork"). A reader's entire contribution to the lock is
    the "rdepth" word in its own slot -- there is no shared reader counter --
    so a dead reader is neutralised by clearing that one slot's PID, and a
    draining writer does so unconditionally as it scans. Because no orphaned
    count can exist, there is no quiescent force-reset: a worker killed
    mid-"incr_by" cannot pin the lock at all. An occupancy bitmap (one bit per
    slot, published before the slot can hold a lock) lets the writer visit
    only occupied slots rather than all 1024. Beyond 1024 simultaneous handles
    per map, a handle that cannot claim a slot proceeds "slotless"; see
    "Reader-slot exhaustion" for the one case that recovery cannot cover.

    The same path validates and rebuilds the LRU doubly-linked list if a dead
    writer left it inconsistent. "stat_recoveries" in "stats" counts stale
    write-lock recoveries; a dead reader drained by a writer is not counted,
    so the counter staying at zero does not mean nothing has been recovered.

    Recovery uses "kill($pid, 0)" for liveness, which cannot tell a reused PID
    from the original -- and the lock word lives in the file, so it lasts as
    long as the file does. Within one running system the risk is small: the
    holder must die in the window it holds the lock and the kernel must
    reissue that exact PID to a long-lived process before the next waiter
    looks.

    It is not small once the file outlives the PID space that wrote it. A
    reboot, a container restart against a persisted volume, or a copy taken
    while a writer held the lock leaves a lock word naming a PID the new
    system may already have reissued -- and for a low PID certainly has, those
    being init and kernel threads, which never exit. If it went to a
    long-lived process, every writer waits on a holder that will never
    release, unbounded and silent: no error, no warning, no timeout; readers
    wait too when the crash was mid-publish (an odd sequence counter). If it
    went to a short-lived one they stall until it exits, and if it was never
    reissued the first writer recovers at once. Nothing in the API can break
    such a lock -- the file has to be recreated. A killed reader strands its
    slot the same way, since that records a PID too, and the read lock is held
    by "each", "keys", "values", "items" and, without LRU or TTL, by "incr",
    "max" and "min". So carry a map across a reboot or a container restart
    only if every process that used it exited cleanly, and copy one only while
    nothing is using it.

    Limitation: PID-based recovery assumes all processes share the same PID
    namespace. Cross-container sharing (different PID namespaces) is not
    supported.

    A full filesystem arrives as SIGBUS, not as an error. The backing file is
    sized once, at creation, for the map's maximum geometry, and its table and
    arena are otherwise sparse: growing the table writes into pages that were
    never allocated rather than extending the file. So "mmap_size" is the
    space the file will need once every page has been touched, not what it
    occupies now -- "du" reports less, and far less for a map with neither LRU
    nor TTL, since those two side arrays are the part written in full at
    creation: a fresh LRU+TTL map is already around 40% allocated where a
    plain one is well under 1% -- and if the filesystem fills while a page is
    first written, the kernel raises SIGBUS in the writing process instead of
    returning an error. One raised in the middle of a table resize takes with
    it the entries not yet re-inserted, exactly as a SIGKILL there would.
    Leave "mmap_size" bytes of headroom on the filesystem, or "fallocate -l"
    the file after creating it to take the allocation failure up front rather
    than at an arbitrary later insert. Use exactly "mmap_size" bytes: a file
    longer than the size recorded in its header is refused as corrupt, and for
    a sharded set "mmap_size" is the total across shards, so use each shard
    file's own size rather than the aggregate.

    After recovery from a mid-mutation crash, the map data may be partially
    inconsistent (e.g., one entry was being updated when the writer died).
    Locks, the LRU chain and the entry counters are restored. The arena free
    lists are not rebuilt, so blocks in flight at the crash may leak; the
    specific entry being mutated may have stale or partial bytes; and a crash
    part-way through a table resize permanently drops the entries that had not
    yet been re-inserted, because they existed only in the dead writer's
    private buffer. Calling "clear" after detecting a stale lock recovery is
    recommended for safety-critical applications.

    An interrupted create is recovered too. A creator killed after the file is
    sized but before its header is committed leaves a full-size, all-zero
    file, which "new" re-initializes -- but only when the file is exactly the
    size the requested geometry needs, is owned by your effective uid, and is
    still entirely zero. A file holding data is never re-initialized. If the
    creator got as far as writing part of the header the file cannot be told
    from a corrupt one, and "new" croaks with "incomplete map file left by an
    interrupted create; remove it and retry". An abandoned create never held
    data, so removing it is safe -- but a header corrupted after the fact
    reaches the same croak, so check before deleting anything you care about.

    Recovery is run by whichever process next takes a lock, readers included,
    and comes from the code that touches the file -- so a map shared with a
    process running anything older than 0.18 keeps that release's crash
    windows. Upgrade every process sharing a map together.

  Reader-slot exhaustion
    Dead-process recovery attributes a crashed lock holder's contribution
    through its reader-slot. The slot table holds 1024 entries, one per open
    handle rather than per process. If more handles than that are open on one
    mapping at once, a reader that cannot claim a slot proceeds "slotless" --
    it still takes the read lock but leaves no per-process record. If such a
    slotless reader is then killed while holding the read lock, its share of
    the lock cannot be attributed to a dead process, so writer recovery cannot
    reclaim it and writers may block until the mapping is recreated. Reaching
    this needs more than 1024 handles open on one mapping at once plus a crash
    in the brief read-lock window; the dead-process slot reclaim keeps the
    table from filling with stale entries, so in practice it is very unlikely.

BENCHMARKS
    Throughput versus other shared-memory / on-disk solutions, 25K entries,
    single process, Linux x86_64. Each benchmarked sub runs over all 25,000
    entries, so the figures below are "Benchmark" rates -- whole passes per
    second, higher is better -- and not operations per second. Multiply by
    25,000 for the rate of the operation named: "Shared::II" LOOKUP at 383 is
    about 9.6 M lookups/s. (A pass can do more than the operation it is named
    for: DELETE refills the map first, so its 25,000 removals come with 25,000
    inserts.) The cross-process table further down is already in operations
    per second. Run "perl -Mblib bench/vs.pl 25000" to reproduce.

    Integer key - integer value> (Shared::II):

                  BerkeleyDB   LMDB   Shared::II
        INSERT          31       46         184
        LOOKUP          35       40         383
        INCREMENT       16       18         165

    String key - string value, short> (inline <= 7B, Shared::SS):

                  FastMmap   BerkeleyDB   LMDB   SharedMem   Shared::SS
        INSERT        11          26       40        62          130
        LOOKUP        10          32       34       146          213
        DELETE        14          18       --        32           68

    String key - string value, long> (~50-100B, Shared::SS):

                  BerkeleyDB   LMDB   SharedMem   Shared::SS
        INSERT        25         37        61          133
        LOOKUP        30         33       125          229

    LRU cache lookup (25K entries, lock-free clock eviction):

                  plain   LRU
        II         350    373   (lock-free, ~6% faster via clock)
        SS         159    159

    Cross-process (25K SS entries, 2 processes, ops/s):

                      Shared::SS   SharedMem       LMDB
        READS        3,250,000    1,986,000     728,000
        WRITES       2,801,000      826,000      95,000
        MIXED 50/50  3,691,000    1,963,000     211,000

    LMDB benchmarked with
    MDB_WRITEMAP|MDB_NOSYNC|MDB_NOMETASYNC|MDB_NORDAHEAD. BerkeleyDB with
    DB_PRIVATE|128MB cache.

    Key takeaways:

    *   10x faster lookups than LMDB for integer keys (lock-free seqlock path)

    *   1.5x faster than Hash::SharedMem for short string lookups (inline
        strings, no arena overhead)

    *   1.8x faster than Hash::SharedMem for long string lookups

    *   4.5x faster cross-process reads than LMDB; 3.4x faster writes than
        SharedMem

    *   LRU reads are lock-free (clock eviction) -- no overhead vs plain maps

    *   Atomic "incr" is 9x faster than get+put on competitors

    *   Strings <= 7 bytes stored inline in node (zero arena overhead)

SEE ALSO
    Data::Buffer::Shared - typed shared array

    Data::Queue::Shared - FIFO queue

    Data::PubSub::Shared - publish-subscribe ring

    Data::ReqRep::Shared - request-reply

    Data::Sync::Shared - synchronization primitives

    Data::Pool::Shared - fixed-size object pool

    Data::Stack::Shared - LIFO stack

    Data::Deque::Shared - double-ended queue

    Data::Log::Shared - append-only log (WAL)

    Data::Heap::Shared - priority queue

    Data::Graph::Shared - directed weighted graph

    Data::BitSet::Shared - shared bitset (lock-free per-bit ops)

    Data::RingBuffer::Shared - fixed-size overwriting ring buffer

SECURITY
    Backing files are created with mode 0600 (owner-only) by default, so only
    the creating user can open and attach them. To share a backing file across
    users, pass an explicit octal file mode such as 0660 as the last argument
    to "new"; the mode is applied when the file is created, and when a file
    left behind by an interrupted create is re-initialized (see "Crash
    Safety"); a file already in use keeps its own permissions. The file is
    opened with "O_NOFOLLOW", so a symlink planted at the path is refused, and
    created with "O_EXCL"; the on-disk header is validated when the file is
    attached. Any process you grant write access to a shared mapping is
    trusted not to corrupt its contents while other processes are using it.

    A backing file written before 0.16 uses the previous on-disk format and is
    rejected with a version-mismatch error when attached; recreate the map
    from its source data. Anonymous and memfd maps are process-local and
    unaffected.

AUTHOR
    vividsnow

LICENSE
    This is free software; you can redistribute it and/or modify it under the
    same terms as Perl itself.

    It bundles xxHash by Yann Collet, used under the BSD 2-Clause licence; see
    LICENSE.xxhash in the distribution.

