[renovate]

This PR contains the following updates:

Package	Change	Age	Confidence
multidict	6.7.1 → 6.9.1

---

Warning

Some dependencies could not be looked up. Check the Dependency Dashboard for more information.

---

Multidict: Reference leak in CIMultiDict/MultiDict items-view union and subtraction

CVE-2026-104874 / GHSA-54p9-h82j-f925

More information

Details

Description

A reference leak in the items-view union and subtraction operators of aio-libs/multidict 6.7.0 through 6.9.0 (C extension) lets a remote client drive unbounded, unreclaimable memory growth by having each operand element leak one key-identity object and one value object. The reflected-union path (operand | d.items(), multidict_itemsview_or2_impl) and the subtraction path (d.items() - operand, multidict_itemsview_sub1_impl) parse each element into new strong references but release only the tuple wrapper, never the identity and value. Servers in the aio-libs stack build these views over attacker-supplied HTTP headers and query strings, so the operand size is under remote control. Forced garbage collection does not recover the leaked objects, so resident memory rises monotonically until the process is killed.

---

Root Cause

_multidict_itemsview_parse_item() returns new references: a fresh identity via md_calc_identity() and a fresh value via Py_NewRef(). The or2_impl first parse loop requests both but clears only arg:

// views.h:565 — or2_impl first parse loop
while ((st = PyIter_NextItem(iter, &arg)) > 0) {
    int tmp = _multidict_itemsview_parse_item(
        self, arg, &identity, NULL, &value);   // identity + value: new refs
    if (tmp < 0) goto fail;
    else if (tmp > 0) {
        if (_set_add(tmp_set, identity, value) < 0) goto fail;
    }
    Py_CLEAR(arg);                             // :575 clears arg ONLY
}

_set_add() builds its own tuple with PyTuple_Pack and Py_DECREFs it, so it never borrows the loop's identity/value. On the next iteration those variables are overwritten, so the previous references are lost permanently. The sub1_impl first loop (:686-696) has the identical omission.

This is an editing slip, not an ownership contract: every sibling path clears all three references per iteration, and1_impl (:304-306), and2_impl (:389-391), or1_impl (:506-508), and the second (md_next) loops of both functions (:605-607, :726-728). PR #​1413 added a Py_DECREF(tpl) to the second loop of or2/sub1, fixing a different temporary-tuple leak; it never touched the first parse loop, so this leak remains on master.

views.h is byte-identical between v6.6.0 and v6.7.0, yet 6.6.4 does not leak and 6.7.0 does. The regression is behavioural, introduced by a key-identity ownership change in hashtable.h across that boundary that made md_calc_identity() return a fresh strong reference the unchanged parse loop never releases.

---

Reproduction Environment

Item	Value
Runtime	CPython 3.14.6
multidict	6.9.0 (PyPI binary wheel, C extension)
OS	macOS (darwin 25.6.0, arm64)

---

Proof of Concept

POC Source Code

poc_refcount.py — leak proof with intersection control

import sys, gc
from multidict import CIMultiDict

def probe(name, run_op):
    d = CIMultiDict(); d["seed"] = "x"
    value = object()                 # unique sentinel
    N = 100000
    operand = [("k%d" % i, value) for i in range(N)]
    gc.collect(); before = sys.getrefcount(value)
    run_op(d, operand)
    gc.collect(); after = sys.getrefcount(value)
    print(f"[{name}] leaked strong refs = {after - before}")

probe("or2  operand | items", lambda d, o: o | d.items())    # reflected union
probe("sub1 items - operand", lambda d, o: d.items() - o)    # subtraction
probe("and2 items & operand", lambda d, o: d.items() & o)    # CONTROL: clears, expect 0

poc_dispatch.py — maps which paths leak

import sys, gc
from multidict import CIMultiDict

def probe(name, fn):
    d = CIMultiDict(); d["seed"] = "x"
    v = object(); N = 50000
    operand = [("k%d" % i, v) for i in range(N)]
    gc.collect(); before = sys.getrefcount(v)
    fn(d, operand); gc.collect(); after = sys.getrefcount(v)
    print(f"{name:35s} leaked={after - before}")

probe("operand | d.items()  (or2)",  lambda d, o: o | d.items())
probe("d.items() | operand  (or1)",  lambda d, o: d.items() | o)
probe("d.items() - operand  (sub1)", lambda d, o: d.items() - o)
probe("operand - d.items()  (rsub)", lambda d, o: set(o) - d.items())
probe("d.items() & operand  (and)",  lambda d, o: d.items() & o)

poc_rss.py — availability consequence

import sys, gc, resource
from multidict import CIMultiDict

def rss_mb():
    r = resource.getrusage(resource.RUSAGE_SELF).ru_maxrss
    return r / (1024 * 1024) if sys.platform == "darwin" else r / 1024

d = CIMultiDict(); d["seed"] = "x"
gc.collect(); print(f"start RSS = {rss_mb():.1f} MB")
for i in range(200):
    operand = [(f"k{i}_{j}", f"v{i}_{j}") for j in range(50000)]
    _ = operand | d.items()
    del operand
    gc.collect()                     # prove GC cannot reclaim the leak
    if (i + 1) % 40 == 0:
        print(f"after {(i+1)*50000:>9,} elements: RSS = {rss_mb():.1f} MB")

Execution Steps

1. python3 -m venv venv (CPython 3.10+).
2. ./venv/bin/pip install "multidict==6.9.0" (installs the C-extension wheel).
3. From a directory that is not a multidict checkout, run each script with ./venv/bin/python.

Actual Execution Evidence

[or2  operand | items] leaked strong refs = 100000
[sub1 items - operand] leaked strong refs = 100000
[and2 items & operand] leaked strong refs = 0        <- control

operand | d.items()  (or2)          leaked=50000
d.items() | operand  (or1)          leaked=0
d.items() - operand  (sub1)         leaked=50000
operand - d.items()  (rsub)         leaked=0
d.items() & operand  (and)          leaked=0

start RSS = 15.6 MB
after 2,000,000 elements: RSS = 288.1 MB
after 6,000,000 elements: RSS = 776.8 MB
after 10,000,000 elements: RSS = 1266.8 MB   (gc.collect() every iteration)

Version-boundary evidence

Same poc_dispatch.py, PyPI wheels, identical environment. Last-clean is 6.6.4; first-affected is 6.7.0:

6.6.4 : or2=0      sub1=0      others=0   (clean, last 6.6.x)
6.7.0 : or2=50000  sub1=50000  others=0   (AFFECTED, first)
6.9.0 : or2=50000  sub1=50000  others=0   (affected)

The clean 6.6.x releases return correct set-operation results, so the zero leak reflects correct memory management, not a broken path.

Analysis of Results

The operand holds exactly N references to one sentinel value; after the union its refcount rises by another N and stays there post-GC, so each element leaked one strong reference. Subtraction shows the identical delta. The intersection control, which clears identity and value, leaks zero; the only code difference is the two missing Py_CLEAR calls, so the leak is caused by that omission and not the harness. RSS climbs from 15.6 MB to 1266.8 MB (~250 MB per 2M elements) despite per-iteration GC, confirming the objects are unreachable by the cyclic collector.

---

Impact

A process that evaluates items-view unions or subtractions over remote-influenced operands leaks one identity plus one value object per element, permanently. In the aio-libs stack multidict backs HTTP headers and query strings, so an attacker who enlarges the operand (for example, many repeated header items compared against a fixed allow/deny set) forces steady, unrecoverable heap growth and can eventually exhaust memory in a long-lived server. This is an availability defect only; results stay correct and no data is exposed.

Reachability depends on the application evaluating operand | view.items() (reflected union) or view.items() - operand (subtraction) over a sequence of 2-tuples whose count is remote-influenced. The forward union view.items() | operand routes to or1_impl, which clears correctly and does not leak; a non-tuple operand element takes the parse_item early-return and does not leak. Set algebra over items views is not the most common multidict usage, which bounds exposure and is why this is Medium, not High. This is distinct from PR #​1413, which fixed a temporary-tuple leak in the second loop and did not release these per-element objects.

---

Remediation

Recommended fix. Clear the two per-element references at the end of the first parse loop in both functions, exactly as every sibling path does. In or2_impl (views.h:575) and sub1_impl (views.h:696), replace the lone Py_CLEAR(arg); with:

    Py_CLEAR(arg);
    Py_CLEAR(identity);
    Py_CLEAR(value);

This is an in-repo change to a static internal function; it alters no public API, type, or signature, and asks callers to change nothing. A leak test in the existing tests/test_leaks.py style would lock it in.

Workaround. Avoid operand | view.items() and view.items() - operand on attacker-influenced operands, or install with MULTIDICT_NO_EXTENSIONS=1 to use the unaffected pure-Python build.

---

Severity

• CVSS Score: 5.3 / 10 (Medium)
• Vector String: CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:N/I:N/A:L

References

• https://github.com/aio-libs/multidict/security/advisories/GHSA-54p9-h82j-f925
• https://nvd.nist.gov/vuln/detail/CVE-2026-104874
• https://github.com/aio-libs/multidict/commit/350b4a07bf8ff851b6d7544e81b1c748bdc74c40
• https://github.com/aio-libs/multidict/releases/tag/v6.9.1
• https://github.com/advisories/GHSA-54p9-h82j-f925

This data is provided by the GitHub Advisory Database (CC-BY 4.0).

---

Release Notes

aio-libs/multidict (multidict)

v6.9.1

Compare Source

=====

(2026-09-21)

Bug fixes

• Fixed the C extension reading freed memory on free-threaded builds when a
  list handed to :py:meth:~multidict.MultiDict.update,
  :py:meth:~multidict.MultiDict.extend, :py:meth:~multidict.MultiDict.merge
  or the :py:class:~multidict.MultiDict and :py:class:~multidict.CIMultiDict
  constructors, a [key, value] item inside any iterable handed to them, or a
  list tested with in against :py:meth:~multidict.MultiDict.items, is
  changed by another thread; a call that catches the list shrinking under it
  now raises :py:exc:RuntimeError -- by :user:rodrigobnogueira.

  Related issues and pull requests on GitHub:
  :issue:1437.

• Fixed a data race on the free-threaded build where a retired hash table's
  reader count used relaxed atomics, letting a lock-free get()/getone()/
  __getitem__() read race a concurrent free of that table. The reader-exit
  decrement and the drain's free check now use release/acquire ordering
  instead -- by :user:asvetlov.

  Related issues and pull requests on GitHub:
  :issue:1481.

• Fixed a free-threaded build bug where two threads calling update(),
  merge(), or __setitem__() on the same key at the same time could lose
  the key entirely instead of just racing on which value wins. A decref of the
  replaced value could transiently suspend the writer's critical section,
  letting a second writer for the same key observe the first writer's
  in-progress entry as absent and, once both settled, mistake it for a stale
  duplicate and delete it. Every such decref is now deferred until the writer
  has released its critical section, so the window can no longer open.
  setdefault() had an unrelated instance of the same blind spot (it could
  insert a duplicate rather than recognizing an in-flight key), fixed alongside
  it -- by :user:asvetlov.

  Related issues and pull requests on GitHub:
  :issue:1483.

• Fixed a free-threaded build bug where getall() and the items()/
  keys()/values() equality path could raise KeyError or report a
  present, never-deleted key as missing. A concurrent update()/extend()/
  __setitem__() call can have its critical section transiently suspended
  (a decref triggering a blocking allocator call) while an entry is marked as
  part of its own bookkeeping; a reader landing in that window used to treat
  the mark as "not found" instead of "still there, in flight" -- by
  :user:asvetlov.

  Related issues and pull requests on GitHub:
  :issue:1484.

• Fixed a reference leak in the C extension where operand | md.items()
  and md.items() - operand leaked one key and one value reference per
  element of operand, letting a large operand grow memory without bound
  (:gh:GHSA-54p9-h82j-f925 <aio-libs/multidict/security/advisories/GHSA-54p9-h82j-f925>)
  -- by :user:asvetlov.

  The issue was reported by :user:waydeshi.

  Related commits on GitHub:
  :commit:350b4a0.

• Fixed a segmentation fault on the standard (non-free-threaded) C extension
  build when a value type's __del__ released the GIL (for example by
  calling time.sleep()) while update(), merge(), __setitem__(),
  __delitem__(), pop(), popone(), or popall() was dropping a
  replaced or removed value. Py_BEGIN_CRITICAL_SECTION compiles to a no-op
  on this build, so nothing else was stopping a second thread from mutating the
  very same MultiDict concurrently once the GIL was released mid-mutation.
  Every such decref is now deferred until the mutation has fully finished, the
  same technique already used to close the analogous free-threaded-build race
  -- by :user:asvetlov.

  Related issues and pull requests on GitHub:
  :issue:1489.

• Fixed the C extension reading freed memory while iterating a
  :py:class:~multidict.CIMultiDict whose keys are plain :py:class:str.
  Converting such a key to :py:class:~multidict.istr could run Python code
  (a :py:class:str subclass's __str__ or __del__) or, on free-threaded
  builds, suspend the iterator's critical section, after which the iterator read
  the entry again even though a concurrent mutation could already have freed it.
  As part of the fix, :py:meth:~multidict.MultiDict.copy and re-initializing
  from another multidict now assign a new version in the C extension instead of
  reusing the source's, matching the pure Python implementation
  -- by :user:asvetlov.

  Related issues and pull requests on GitHub:
  :issue:1496.

• Fixed a use-after-free on the free-threaded build where a lock-free
  get(), [] or in could read a hash table that a concurrent
  resize had just retired and another reader was freeing
  -- by :user:asvetlov.

  Related issues and pull requests on GitHub:
  :issue:1497.

Contributor-facing changes

• Removed a redundant include and a duplicated exclude line from
  MANIFEST.in; sdist contents are unchanged -- by :user:asvetlov.

  Related issues and pull requests on GitHub:
  :issue:1478.

• Added .claudeignore file -- by :user:asvetlov

  Related issues and pull requests on GitHub:
  :issue:1479.

• Scaled up the pure-Python pop(), popitem(), __delitem__(),
  add() and item-insertion benchmarks to do more work per measurement.
  Repeated CodSpeed runs on the same commit showed these particular
  benchmarks flagged as dominated by syscalls, understating their real
  cost and adding noise to the reported values; a larger working set
  amortizes that overhead -- by :user:asvetlov.

  Related issues and pull requests on GitHub:
  :issue:1485.

• Replaced deprecated instrumentation codspeed mode with simulation -- by :user:asvetlov

  Related issues and pull requests on GitHub:
  :issue:1493.

• Reorganized the mutating benchmarks (item insertion, update(),
  add() of the same key, pop(), popitem(), clear(),
  __delitem__() and __setitem__()) to copy a fresh multidict and
  apply the operation in a loop, like the add() and extend()
  benchmarks already do. The insertion, update() and clear()
  benchmarks previously mutated a single multidict shared across
  rounds, so only the first round measured the intended operation; the
  rest did a single copy per round, letting per-round overhead dominate
  -- by :user:asvetlov.

  Related issues and pull requests on GitHub:
  :issue:1494.

• The repr() and view inequality benchmarks were updated to repeat their
  operation in a loop, like the other benchmarks, and the CodSpeed benchmark
  job was moved to Python 3.14 -- by :user:asvetlov.

  Related issues and pull requests on GitHub:
  :issue:1498.

• Dropped -I from the AddressSanitizer test command in AGENTS.md
  and in the CI job. It implies -E, which made Python ignore
  PYTHONMALLOC=malloc, so small hash tables were still served from
  pymalloc arenas where use-after-free went undetected
  -- by :user:asvetlov.

  Related issues and pull requests on GitHub:
  :issue:1499.

• The CI/CD workflow was updated to stop superseded runs of the same pull request
  when a new commit is pushed; runs on master, release branches, tags,
  the merge queue, and the daily schedule are never interrupted
  -- by :user:asvetlov.

  Related issues and pull requests on GitHub:
  :issue:1500.

• The release job was changed to upload distributions and their signatures
  to the GitHub Release one file at a time, skipping assets that were
  already attached and retrying after a pause, so that a parallel upload
  burst no longer tripped the GitHub secondary rate limit
  -- by :user:asvetlov.

  Related issues and pull requests on GitHub:
  :issue:1503.

Miscellaneous internal changes

• Corrected several comments in the free-threaded C extension that
  attributed critical-section suspension to a blocking PyMem_Malloc()
  call; allocation alone never suspends an acquired critical section, and
  the real risk at those sites is a decref running
  a finalizer or weakref callback. Also dropped a retry loop in
  md_clone_from_ht() that guarded against the same, non-existent
  allocation-triggered suspension -- by :user:asvetlov.

  Related issues and pull requests on GitHub:
  :issue:1486.

• Changed the free-threaded build's deferred decref buffer, used by
  update() and __setitem__(), to a chain of fixed-size blocks
  with a large inline first block instead of a small inline array that
  was reallocated on growth -- by :user:asvetlov.

  Related issues and pull requests on GitHub:
  :issue:1501.

---

v6.9.0

Compare Source

==========

(2026-09-18)

Bug fixes

• Protected repr() of MultiDict, MultiDictProxy, and their views
  in the C extension with a critical section, avoiding data races on the
  free-threaded build of CPython -- by :user:asvetlov.

  Related issues and pull requests on GitHub:
  :issue:1431.

• Guarded repr() of MultiDictProxy in the C extension and of
  KeysView in both the C extension and the pure-Python
  implementation against infinite recursion on self-referential
  containers, matching the existing guard on MultiDict,
  ItemsView, and ValuesView -- by :user:asvetlov.

  Related issues and pull requests on GitHub:
  :issue:1431.

• Protected MultiDict.update(), .extend(), .merge(), .clear(),
  .copy(), and the MultiDict/CIMultiDict constructors in the C
  extension with a critical section, using the two-object form when a
  second multidict, multidict proxy, or plain dict instance is
  involved, avoiding data races and a segmentation fault on the
  free-threaded build of CPython. .clear() now also publishes the
  empty table before releasing any entry's references, so a concurrent
  caller can never observe a partially-cleared multidict even if releasing
  a value runs arbitrary Python code that suspends the held critical
  section -- by :user:asvetlov.

  Related issues and pull requests on GitHub:
  :issue:1433.

• Protected MultiDict.add(), __setitem__/__delitem__,
  get()/getone()/__getitem__, __contains__, getall(),
  setdefault(), pop()/popone()/popall()/popitem(),
  __eq__, iteration, and the &/|/-/^/in/
  isdisjoint() operations on .keys() and .items() views in the
  C extension with a critical section, avoiding data races and
  use-after-free crashes on the free-threaded build of CPython
  -- by :user:asvetlov.

  Related issues and pull requests on GitHub:
  :issue:1438.

• Fixed a data race on the free-threaded build where MultiDictObject.used
  was written non-atomically while len() read it with a relaxed atomic
  load -- by :user:asvetlov.

  Related issues and pull requests on GitHub:
  :issue:1441.

• Fixed several pre-existing use-after-free races on the free-threaded build: _md_resize()
  and md_clone_from_ht() could allocate a new hash table, have their
  critical section transiently suspended during that allocation, and then use
  a keys-table pointer or size a concurrent operation had already invalidated;
  _md_del_at() (plain del/pop()) could likewise leave self in
  an inconsistent state across a suspending decref -- by :user:asvetlov.

  Related issues and pull requests on GitHub:
  :issue:1441.

• Fixed a data race on the free-threaded build where md->keys was
  published with a plain store while lock-free readers loaded it
  atomically, and a false-negative race in get()/__contains__
  where a hash temporarily marked by a concurrent replace could make a
  present key look absent -- by :user:asvetlov.

  Related issues and pull requests on GitHub:
  :issue:1441.

• Fixed several more pre-existing use-after-free and data-corruption races on
  the free-threaded build, this time in the __setitem__()/update()/
  extend()/merge() replace path: a decref suspending the critical
  section mid-replace could leave a stale table pointer in use, misplace a
  temporarily-marked entry during a concurrent resize, or let one thread's
  duplicate-tracking mark get overwritten by an unrelated key's entry sharing
  the same hash bucket -- by :user:asvetlov.

  Related issues and pull requests on GitHub:
  :issue:1441.

• Fixed an unbounded memory leak on the free-threaded build where a retired
  hash table could sit on md->retired for the rest of the object's
  lifetime under sustained concurrent read traffic instead of being freed
  -- by :user:asvetlov.

  Related issues and pull requests on GitHub:
  :issue:1443.

• Fixed a data race on the free-threaded build of the C extension where
  state->global_version, the version counter shared by every
  multidict instance and used to derive getversion(), was bumped with
  a plain increment instead of an atomic op. Two threads mutating
  different instances at the same time could step on each other's update
  and hand out a duplicate (or non-monotonic) version number
  -- by :user:asvetlov.

  Related issues and pull requests on GitHub:
  :issue:1444.

• Fixed a data race on the free-threaded build of the C extension where
  getall()/popall() and update()/extend()/merge()
  temporarily marked and unmarked a matching entry's hash with a plain,
  non-atomic store, while a lock-free __contains__/get() on another
  thread could load that same field with an atomic operation and no lock at
  all -- by :user:asvetlov.

  Related issues and pull requests on GitHub:
  :issue:1448.

• Fixed a data race in the pure-Python MultiDict/CIMultiDict
  fallback where concurrent mutating calls (add(), __setitem__,
  __delitem__, setdefault(), pop()/popone(),
  popall(), popitem(), update(), extend(), merge(),
  clear(), and re-__init__) could corrupt the shared hash table,
  or silently drop a concurrent write, on a regular, GIL-enabled
  interpreter. Pure-Python bytecode is not atomic under the GIL, so two
  threads could interleave mid hash-table insert or deletion, leaving
  the index table pointing at stale or out-of-range entries and causing
  an infinite probe loop, an AttributeError, or a lost mutation.
  Each MultiDict/CIMultiDict instance now holds its own lock for
  the duration of these operations, with two-object operations locking
  both instances in a fixed order to avoid deadlock. The C extension
  already serializes these operations with a critical section and was
  not affected
  -- by :user:asvetlov.

  Related issues and pull requests on GitHub:
  :issue:1449.

• Fixed a data race on the free-threaded build of the C extension where
  to_dict() cleared every entry's temporary "seen" mark with a plain,
  non-atomic store, while a lock-free __contains__/get() on another
  thread could load that same field with an atomic operation and no lock at
  all -- by :user:asvetlov.

  Related issues and pull requests on GitHub:
  :issue:1450.

• Fixed a data race on the free-threaded build of the C extension where
  update()/extend()/merge() read self's module state without
  holding self's lock, while a concurrent __init__() on the same,
  already-published multidict could rewrite that same field under lock.
  Rather than routing every read through a lock-free lookup, the field is
  now written only once, at object construction, and never touched again by
  __init__(), so it is safe to read unlocked anywhere -- by
  :user:asvetlov.

  Related issues and pull requests on GitHub:
  :issue:1452.

• Fixed items() set algebra (&, |, -, in, isdisjoint()) so it no
  longer corrupts or hangs when a compared value's __eq__() re-enters the
  same :class:~multidict.MultiDict, for example by calling getall() on
  it. The comparison now runs only after the internal hash-chain walk has
  been fully materialized and restored, so the callback can no longer
  observe entries the walk still has marked -- by :user:asvetlov.

  Related issues and pull requests on GitHub:
  :issue:1453.

• Fixed a data race on the free-threaded build of the C extension where
  popall()/popone()/__delitem__/add() rewrote a hash table
  index slot with a plain, non-atomic store, while a lock-free
  get()/getone()/__getitem__/__contains__ on another thread
  walked that same index array with a plain, non-atomic load and no lock at
  all -- by :user:asvetlov.

  Related issues and pull requests on GitHub:
  :issue:1454.

• Stopped folding the case of :class:~multidict.istr keys in the
  case-sensitive :class:~multidict.MultiDict, so both the C and the pure
  Python implementations now keep such a key exactly as given
  -- by :user:youdie006.

  Related issues and pull requests on GitHub:
  :issue:1457.

• Fixed the C extension's update() (and merge(), which shares the same
  code path) silently failing to invalidate an in-progress keys()/
  items()/values() iterator when the call only overwrote an
  already-present key's value in place. Adding a new key already invalidated
  iterators correctly; only the in-place overwrite branch was missing the
  version bump -- by :user:asvetlov.

  Related issues and pull requests on GitHub:
  :issue:1458.

• Fixed the pure-Python MultiDict/CIMultiDict iterator guard so that
  mutating the mapping before the first next() call on an items(),
  keys(), or values() iterator (or on iter(md)) now reliably raises
  RuntimeError, matching the C extension. Previously a mutation that
  happened before the iterator was ever advanced, such as clear() or a
  del/popone() that removed the only remaining entry, could make the
  iterator silently raise StopIteration instead -- by :user:asvetlov.

  Related issues and pull requests on GitHub:
  :issue:1459.

• Fixed a use-after-free race on the free-threaded build of the C extension
  where a retired hash table could be freed while a lock-free
  get()/getone()/__getitem__/__contains__ reader on another
  thread was still walking it: the coarse "no reader in flight" gate
  reaching zero did not reliably mean every such reader had also reached
  its own per-table exit, so a table whose own reader count is still
  nonzero is now deferred for a later retry instead of freed
  -- by :user:asvetlov.

  Related issues and pull requests on GitHub:
  :issue:1461.

Features

• Added a :meth:~multidict.MultiDict.to_dict method returning a plain
  :class:dict that maps every key to the list of all its values. Unlike
  dict(md), which keeps only the first value per key, and unlike a
  {k: md.getall(k) for k in md} comprehension, which emits one entry per
  spelling of a case-insensitive key, it groups by key identity
  -- by :user:rodrigobnogueira.

  Related issues and pull requests on GitHub:
  :issue:783.

• MultiDict, CIMultiDict, MultiDictProxy and CIMultiDictProxy in the
  C extension implemented the vectorcall calling convention for construction.
  This sped up MultiDict(...)-style calls by avoiding an intermediate
  positional-arguments tuple and keyword-arguments dictionary in the common case; the gained burst is ~10.
  -- by :user:asvetlov.

  Related issues and pull requests on GitHub:
  :issue:1436.

• Made MultiDict.__contains__(), .get(), .getone() and
  __getitem__() lock-free on CPython 3.14+'s free-threaded build, instead
  of taking a critical section like the rest of the C extension's API (on
  3.13, which lacks the public API these need to safely read an entry
  concurrently, they still take the critical section, exactly as before). A
  resize/shrink/clear no longer frees the outgoing hash table immediately; it
  is deferred until no lock-free reader could still be walking it -- by
  :user:asvetlov.

  Related issues and pull requests on GitHub:
  :issue:1441.

• Made the pure-Python MultiDict/CIMultiDict implementation safe to
  use from multiple threads under a free-threaded (no-GIL) build. Every
  method that touches an instance's hash table now takes that instance's own
  lock, two-object operations (update(), extend(), merge(),
  __eq__(), copying) lock both objects in a fixed, deadlock-safe order,
  and the version counter shared by every instance is guarded separately. On
  a regular (GIL-enabled) interpreter this adds no overhead: the locked
  methods are the exact same function objects as before -- by
  :user:asvetlov.

  Related issues and pull requests on GitHub:
  :issue:1447.

• Sped up adding many values for the same key to a
  :class:~multidict.MultiDict or :class:~multidict.CIMultiDict, each
  added value no longer gets slower than the one before it
  -- by :user:bdraco.

  Related issues and pull requests on GitHub:
  :issue:1466.

Improved documentation

• Added the plural form "fallbacks" to the docs spell checker's allowed
  word list so :file:CHANGES.rst builds cleanly -- by :user:asvetlov.

  Related issues and pull requests on GitHub:
  :issue:1439.

Contributor-facing changes

• Added race tests that iterate and extend a :class:~multidict.MultiDict
  from several threads on a free-threaded build, so a regression in the C
  extension's locking fails a free-threaded CI leg instead of going unnoticed
  -- by :user:rodrigobnogueira.

  Related issues and pull requests on GitHub:
  :issue:1317.

• Added AddressSanitizer/UndefinedBehaviorSanitizer and
  ThreadSanitizer CI jobs. The former runs the suite under a
  regular CPython with MULTIDICT_ASAN_BUILD=1; the latter builds a
  free-threaded CPython instrumented with --with-thread-sanitizer
  and runs the suite against it via MULTIDICT_TSAN_BUILD=1 -- by
  :user:asvetlov.

  Related issues and pull requests on GitHub:
  :issue:1455.

• Added Hypothesis-based property and stateful fuzz tests covering
  :class:~multidict.MultiDict/:class:~multidict.CIMultiDict semantics,
  views, iterators, and threaded stress scenarios, run against both the
  C-extension and pure-Python backends -- by :user:asvetlov.

  Related issues and pull requests on GitHub:
  :issue:1460.

• Added core dump collection to the AddressSanitizer and ThreadSanitizer
  CI jobs: a crash now uploads the core file alongside the crashing
  interpreter binary and the compiled extension as a downloadable
  artifact, for offline debugging with a matching symbol-carrying binary
  instead of raw hex offsets -- by :user:asvetlov.

  Related issues and pull requests on GitHub:
  :issue:1462.

• Added benchmarks for adding many values for the same key and for
  creating a multidict with many items -- by :user:bdraco.

  Related issues and pull requests on GitHub:
  :issue:1465.

• Skipped abstractmethods from coverage leaks report, they are empty placeholders that are never executed -- by :user:asvetlov

  Related issues and pull requests on GitHub:
  :issue:1470.

• Add benchmarks for __setitem__ -- by :user:asvetlov

  Related issues and pull requests on GitHub:
  :issue:1471.

• The AddressSanitizer and ThreadSanitizer CI jobs were changed to run pytest
  with --capture=no. Previously a sanitizer abort could exit the process
  before pytest flushed its per-test output buffer, silently dropping the
  sanitizer report from the job log -- by :user:asvetlov.

  Related issues and pull requests on GitHub:
  :issue:1472.

• Moved the hypothesis pin out of requirements/pytest.txt into its own
  requirements/pytest-hypothesis.txt, installed only by the dedicated
  Hypothesis CI jobs. hypothesis now ships a Rust extension with no
  prebuilt wheel for some emulated architectures (e.g. musllinux i686),
  which broke wheel-testing jobs that never run the Hypothesis-marked tests
  in the first place -- by :user:asvetlov.

  Related issues and pull requests on GitHub:
  :issue:1475.

Miscellaneous internal changes

• Replaced PyIter_Next() calls with the newer PyIter_NextItem() API in
  the C extension, so the iterator-exhausted and error cases are told apart by
  the return code instead of an ambiguous NULL result -- by :user:asvetlov.

  Related issues and pull requests on GitHub:
  :issue:1440.

---

v6.8.0

Compare Source

=====

(2026-09-09)

Bug fixes

• A segmentation fault that could be triggered when getting an item is now fixed
  -- by :user:Vizonex.

  Related issues and pull requests on GitHub:
  :issue:1310.

• Fixed reference leak in iterators, views and istr
  -- by :user:Vizonex.

  Related issues and pull requests on GitHub:
  :issue:1311.

• Fixed the pure-Python :class:~multidict.MultiDict constructor and
  :py:meth:~multidict.MultiDict.extend,
  :py:meth:~multidict.MultiDict.update, and
  :py:meth:~multidict.MultiDict.merge methods over-allocating their
  internal hash table when called with both a positional argument and
  keyword arguments, because keyword arguments were counted twice in the
  size estimate -- by :user:aiolibsbot.

  Related issues and pull requests on GitHub:
  :issue:1338.

• Fixed __repr__ of :class:~multidict.MultiDict,
  :class:~multidict.CIMultiDict, their proxies, and the keys/items views
  producing invalid output when keys contained quote characters --
  keys are now formatted with :func:repr so the result is a valid Python
  string literal -- by :user:aiolibsbot.

  Related issues and pull requests on GitHub:
  :issue:1342.

• Fixed a segfault when calling :py:meth:~multidict.MultiDict.add with only one of its two required arguments supplied by keyword, e.g. d.add(key="k"). Extra keyword arguments passed to the lookup and removal methods are also now rejected with :exc:TypeError instead of being silently ignored.

  -- by :user:devdanzin

  Related issues and pull requests on GitHub:
  :issue:1376.

• Fixed a segfault when constructing a multidict iterator type directly, e.g. type(iter(md.keys())).__new__(...). Such an iterator had a NULL internal pointer that next() dereferenced. The iterator types now forbid direct instantiation, the same way the view types were fixed in :issue:1163.

  -- by :user:devdanzin

  Related issues and pull requests on GitHub:
  :issue:1377.

• Fixed a segfault when using a :py:class:~multidict.MultiDict or :py:class:~multidict.CIMultiDict created via __new__ without calling __init__ (for example a subclass that does not call super().__init__()). The internal state was left as NULL pointers that the first method call dereferenced. tp_new now initializes the object to a valid empty mapping.

  -- by :user:devdanzin

  Related issues and pull requests on GitHub:
  :issue:1378.

• Fixed two crashes in the C extension caused by holding a raw pointer into a hash table across an operation that could reshape it. Updating a multidict from itself (e.g. d.extend(d)) freed the very table being iterated -- a use-after-free; extend(self) now doubles the contents and update(self)/merge(self) are no-ops. Building a multidict from a list of pairs whose case-insensitive key .lower() shrinks that list read past the end of the list; the length is now re-checked on every iteration.

  -- by :user:devdanzin

  Related issues and pull requests on GitHub:
  :issue:1379.

• Fixed three reference/resource leaks on error paths in the C extension: the items-view __contains__ leaked the first element of a candidate pair when reading the second one raised; __repr__ leaked its PyUnicodeWriter when the multidict was mutated during iteration; and the internal iteration helper leaked the identity reference when key materialization failed under low memory.

  -- by :user:devdanzin

  Related issues and pull requests on GitHub:
  :issue:1381.

• Stopped several feature-detection fallbacks in the C extension from swallowing every exception. When probing an argument (arg.items()/arg.keys()) or measuring it (len(other)) fails, the code now clears only the expected :exc:TypeError/:exc:AttributeError and lets everything else -- notably :exc:MemoryError and :exc:KeyboardInterrupt -- propagate, matching the already-correct sites elsewhere in the module.

  -- by :user:devdanzin

  Related issues and pull requests on GitHub:
  :issue:1382.

• Fixed a segmentation fault when extending a multidict with itself
  -- by :user:cananoo and :user:asvetlov.

  Related issues and pull requests on GitHub:
  :issue:1398.

• Fixed a memory leak when MultiDict and CIMultiDict instances are
  initialized more than once -- by :user:asvetlov.

  Related issues and pull requests on GitHub:
  :issue:1412.

• Fixed a reference leak in items-view set operations
  -- by :user:asvetlov.

  Related issues and pull requests on GitHub:
  :issue:1413.

• Fixed a memory leak when destroying C-extension multidict instances
  and proxies -- by :user:asvetlov.

  Related issues and pull requests on GitHub:
  :issue:1415.

• Fixed a reference leak from repeated __init__ calls on
  MultiDictProxy and CIMultiDictProxy instances -- by :user:asvetlov.

  Related issues and pull requests on GitHub:
  :issue:1419.

• Fixed missing decref for :data:None on error in :meth:~multidict.MultiDict.setdefault
  -- by :user:asvetlov

  Related issues and pull requests on GitHub:
  :issue:1421.

Features

• Added :py:func:reversed support to the keys, values, and items views of
  :class:~multidict.MultiDict, :class:~multidict.CIMultiDict, and their
  proxies in both the C-extension and pure-Python implementations
  -- by :user:aiolibsbot.

  Related issues and pull requests on GitHub:
  :issue:448.

• Optimized key identity comparison for C Extension.

  Now it uses fast path that is equal to :c:func:PyUnicode_Equal from
  Python 3.14+ but without redundant type checks. It gives ~15% speed-up on benchmarks with many key comparisons.

  -- by :user:asvetlov

  Related issues and pull requests on GitHub:
  :issue:1406.

• Changed both the pure-Python and C implementations to mark a temporarily
  removed hash table entry by setting the high bit of its hash instead of
  overwriting it with a sentinel value, so restoring the entry no longer
  recomputes the hash; :meth:~multidict.MultiDict.getall is faster ~10% now in C version
  -- by :user:asvetlov

  Related issues and pull requests on GitHub:
  :issue:1426.

Removals and backward incompatible breaking changes

• Dropped support for Python 3.9 as it has reached end of life.

  Related issues and pull requests on GitHub:
  :issue:1316.

• Dropped support for free-threaded Python 3.13 -- by :user:ngoldbaum.

  Related issues and pull requests on GitHub:
  :issue:1326.

Improved documentation

• Fixed CIMultiDictProxy documentation to state it inherits from
  MultiDictProxy (not MultiDict), and fixed missing word in
  istr section -- by :user:veeceey.

  Related issues and pull requests on GitHub:
  :issue:1298.

• Clarified that istr preserves the original casing and
  compares as a regular str; case-insensitive matching is
  handled by CIMultiDict -- by :user:gyanu2507.

  Related issues and pull requests on GitHub:
  :issue:1397.

• Fixed broken RST markup in items() docstrings
  -- by :user:veeceey.

  Related issues and pull requests on GitHub:
  :issue:1299.

Packaging updates and notes for downstreams

• Dropped support for free-threaded Python 3.13 -- by :user:ngoldbaum.

  Related issues and pull requests on GitHub:
  :issue:1326.

• Wheels for iOS and Android (CPython 3.13+) are now published on
  release-tag builds, so downstreams on those platforms no longer need a
  local compiler to install multidict; these wheels are build-verified
  only, since running the test suite on mobile needs a simulator or an
  emulator -- by :user:aiolibsbot.

  Related issues and pull requests on GitHub:
  :issue:1337.

• Added support for building and shipping riscv64 wheels
  -- by :user:justeph.

  Related issues and pull requests on GitHub:
  :issue:1293.

• The setuptools build dependency lower bound has been restored to be
  >= 47 after an incorrect automated increase in :pr:1315
  -- by :user:aiolibsbot.

  Related issues and pull requests on GitHub:
  :issue:1400.

Contributor-facing changes

• Added support for collecting code coverage of isolated multidict tests
  -- by :user:Vizonex.

  Related issues and pull requests on GitHub:
  :issue:1314.

• Switched to mirrors-clang-format and enabled clang-format in pre-commit.ci
  -- by :user:bdraco.

  Related issues and pull requests on GitHub:
  :issue:1318.

• Added a release-tag-gated CI job that cross-compiles the C extension
  for iOS and Android through cibuildwheel -- by :user:aiolibsbot.

  Related issues and pull requests on GitHub:
  :issue:1337.

• Added an AGENTS.md orientation file at the repository root, covering the
  pull request template, the CHANGES/ news fragment conventions, the
  draft-PR / human-review workflow, and the dual pure-Python and C-extension
  parity rule, so LLM contributors land changes that match project style
  -- by :user:bdraco.

  Related issues and pull requests on GitHub:
  :issue:1339.

• Added init, installable, instantiation, parametrization,
  parametrized, postfix, and unparseable to the docs spelling list
  so news fragments and docs can use these words without failing the spell
  check build -- by :user:pctablet505.

  Related issues and pull requests on GitHub:
  :issue:1343.

• Documented three common agent failure modes in
  :file:AGENTS.md: the docs spell check
  (make doc-spelling) catches unknown words in news fragments
  before CI does; coverage runs over the test tree too, so
  unreachable defensive raise guards and one-sided cleanup
  branches in tests will surface as uncovered on the codecov
  patch report; and branch creation is restricted on
  aio-libs/multidict, so PRs must be pushed from a fork
  -- by :user:bdraco.

  Related issues and pull requests on GitHub:
  :issue:1343, :issue:1345.

• Trimmed the CI test matrix to drop redundant Py_DEBUG jobs on
  macOS and Windows and to gate the windows-11-arm wheel build to
  release-time only, cutting CI wall-clock by roughly a third without
  reducing code coverage -- by :user:aiolibsbot.

  Related issues and pull requests on GitHub:
  :issue:1346.

• Added a CLAUDE.md at the repository root that imports
  :file:AGENTS.md via Claude Code's @-syntax, so the
  project's LLM contributor rules load automatically when
  working in Claude Code
  -- by :user:aiolibsbot.

  Related issues and pull requests on GitHub:
  :issue:1347.

• Enabled the I (import sorting) rule in
  [tool.ruff.lint] so ruff check covers import order on top
  of the existing UP (pyupgrade) group, and updated the
  ruff-check pre-commit hook to run with
  --fix --exit-non-zero-on-fix --show-fixes so import-order
  fixes apply on commit. Sorted imports in four tests/ files
  to match the new rule -- by :user:bdraco.

  Related issues and pull requests on GitHub:
  :issue:1348.

• Enabled the PLC0415 (import-outside-top-level) rule in
  [tool.ruff.lint]. Function-scoped imports must now opt in
  with # noqa: PLC0415 plus a comment explaining the reason
  (typically avoiding a heavy optional dependency at import time).
  This catches a pattern that LLM contributors frequently introduce.
  There are currently no opt-outs in the tree, so the rule is
  enforced everywhere ruff check runs
  -- by :user:bdraco.

  Related issues and pull requests on GitHub:
  :issue:1349.

• Switched the cibuildwheel build frontend to build[uv] so
  that uv provisions every build and test virtual environment
  in the wheel matrix. Test-dependency installation in particular
  drops from a multi-second pip install per ABI to a roughly
  sub-second uv resolve
  -- by :user:bdraco.

  Related issues and pull requests on GitHub:
  :issue:1350.

• Overrode CIBW_BUILD_FRONTEND=build for the odd-arch wheel
  matrix; the upstream manylinux/musllinux images for
  those arches do not ship uv
  -- by :user:bdraco.

  Related issues and pull requests on GitHub:
  :issue:1352.

• Allowed re-running the deploy job after a partial release failure: the
  Make Release step now skips when the GitHub Release already exists,
  and the PyPI publish step uses skip-existing so dists that were
  already uploaded on a prior attempt do not break the retry
  -- by :user:bdraco.

  Related issues and pull requests on GitHub:
  :issue:1353.

• Switched the aarch64 and armv7l wheel builds to GitHub's native ARM
  runners. The aarch64 wheels now build without QEMU emulation, and
  armv7l runs on aarch64 hosts so its 32-bit ARM execution is far
  cheaper than the previous aarch64-on-x86_64 path
  -- by :user:bdraco.

  Related issues and pull requests on GitHub:
  :issue:1354.

• Documented design principles for pure-Python :class:~multidict.istr implementation
  -- by :user:asvetlov

  Related issues and pull requests on GitHub:
  :issue:1360.

• Pinned coverage to the ctrace measurement core so the test suite runs
  on Python 3.14. Coverage picks sysmon there, which cannot record the
  dynamic contexts pytest-cov switches at runtime, and the resulting warning
  became an error under the suite's filterwarnings setting
  -- by :user:rodrigobnogueira.

  Related issues and pull requests on GitHub:
  :issue:1393.

• Dependabot has been restricted to the requirements subdirectory to
  avoid unintended updates outside dependency requirement files
  -- by :user:aiolibsbot.

  Related issues and pull requests on GitHub:
  :issue:1400.

• Enabled automatic upgrades from ruff after each python version that is dropped
  -- by :user:Vizonex.

  Related issues and pull requests on GitHub:
  :issue:1325.

Miscellaneous internal changes

• Explicitly marked empty_htkeys as const.

  Moved the structure to .rodata linker section.

  Unexpected modification of the structure will lead to segfault instead
  of corrupting the data and fail fast.

  .rodata is mount as read-only-mapped, it is shared well across
  multiple threads without cache misses.

  The change is pretty trivial, it is made for the sake of correctness,
  not for speedup.

  -- by :user:asvetlov.

  Related issues and pull requests on GitHub:
  :issue:1405.

• Slightly reorganized multidict creating process.

  1. Removed unnecessary calculations when md_init() creates an empty multidict.
  2. Now tp_alloc slot is used in object cloning instead of bare :c:func:PyType_GenericNew call.

  -- by :user:asvetlov

  Related issues and pull requests on GitHub:
  :issue:1420.

• Dropped dead code, _md_del_at() and _md_del_at_for_upd() never fails
  -- by :user:asvetlov

  Related issues and pull requests on GitHub:
  :issue:1423.

• Fixed a missing space in the :exc:ValueError message
  raised when constructing a multidict from a sequence
  -- by :user:veeceey.

  Related issues and pull requests on GitHub:
  :issue:1295.

• Renamed the benchmark script from benchmarks/becnhmark.py to
  benchmarks/benchmark.py so it matches the invocation documented in
  :doc:benchmark -- by :user:aiolibsbot.

  Related issues and pull requests on GitHub:
  :issue:1335.

---

---

Configuration

📅 Schedule: (UTC)

• Branch creation
  • At any time (no schedule defined)
• Automerge
  • At any time (no schedule defined)

🚦 Automerge: Disabled by config. Please merge this manually once you are satisfied.

♻ Rebasing: Whenever PR becomes conflicted, or you tick the rebase/retry checkbox.

🔕 Ignore: Close this PR and you won't be reminded about this update again.

---

• If you want to rebase/retry this PR, check this box

---

This PR was generated by Mend Renovate. View the repository job log.

[devin-ai-integration]

🔍 Devin Review: 1 flag

Not posted on this PR by your GitHub settings — view it in Devin Review. (Configure)