Is multidict safe?

GHSA-54p9-h82j-f925

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(), multidictitemsviewor2impl) and the subtraction path (d.items() - operand, multidictitemsviewsub1impl) 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.

---

multidictitemsviewparseitem() returns new references: a fresh identity via mdcalcidentity() and a fresh value via PyNewRef(). The or2impl 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) {
        if (_set_add(tmp_set, identity, value) < 0) goto fail;
    }
    Py_CLEAR(arg);                             // :575 clears arg ONLY
}

setadd() builds its own tuple with PyTuplePack and PyDECREFs 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 sub1impl 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, and1impl (:304-306), and2impl (:389-391), or1impl (:506-508), and the second (mdnext) loops of both functions (:605-607, :726-728). PR #1413 added a PyDECREF(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 mdcalcidentity() return a fresh strong reference the unchanged parse loop never releases.

---

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

---

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

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)

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")

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

[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)

Same pocdispatch.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.

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 PyCLEAR 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.

---

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 or1impl, which clears correctly and does not leak; a non-tuple operand element takes the parseitem 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.

---

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 or2impl (views.h:575) and sub1impl (views.h:696), replace the lone PyCLEAR(arg); with: