# CVE-2026-109618 / CVE-2026-109619: UAF via OOM Bailout

Use-after-free in PHP's session upload progress handler, exploitable
over HTTP against Apache mod_php with **default PHP configuration**.
No LFI, no framework dependency, no non-default settings.

## Affected versions

PHP 5.4.0 through 8.3.35, 8.4.0 through 8.4.26, 8.5.0 through 8.5.11,
running under Apache mod_php with prefork MPM.

Patched in PHP 8.3.36, 8.4.27, 8.5.12.

PHP-FPM is **not vulnerable**: FPM workers exit on bailout (code 70),
so stale state never persists across requests.

## Prerequisites

- Apache mod_php with prefork MPM (default on Debian/Ubuntu)
- `session.upload_progress.enabled = 1` (default)
- `memory_limit` reachable via session data decode (~4MB for 33000 entries)
- Any PHP-handled URL (even a 404 page)
- A method to write a session file with many entries

No LFI. No specific framework or library. No non-default PHP settings.

## How it works

### CVE-2026-109618: OOM triggers longjmp past cleanup

When `memory_limit` is exceeded during `session_decode()` inside the
RFC1867 progress callback, PHP calls `zend_bailout()` which executes a
`longjmp` back to `php_request_startup()`. This skips
`php_var_unserialize_destroy()`, leaving `BG(unserialize).level = 1`
and `BG(unserialize).data` pointing at a freed `php_unserialize_data`
struct (8216 bytes, 3 ZendMM pages).

Apache prefork workers survive `zend_bailout()` and accept new
connections with the stale state intact.

### CVE-2026-109619: Stale state reused on next request

On the next request, `sapi_activate()` runs the RFC1867 callback
BEFORE `PHP_RINIT_FUNCTION(basic)` has a chance to zero `BG` state.
The callback calls `php_session_decode()`, which calls
`php_var_unserialize_init()`. That function sees
`BG(unserialize).level > 0` and returns the stale `.data` pointer
without allocating a new struct.

This is a use-after-free: the pointer was freed when the previous
request's heap was destroyed during `zend_deactivate()`.

### Phase 1: Plant session file

A session file with 33000 entries (~682KB) is planted. When decoded,
this forces `nTableSize = 65536`, requiring an `arData` allocation of
65536 * 36 = 2,359,296 bytes (~2.25MB). With `memory_limit = 4MB`,
this exceeds the limit during `HashTable` resize.

### Phase 2: OOM trigger

A multipart POST with `PHP_SESSION_UPLOAD_PROGRESS` and a file field
triggers the RFC1867 progress callback. The callback reads the planted
session file and calls `php_session_decode()`. The decode triggers OOM
during `HashTable` resize. `zend_bailout()` longjmps past
`php_var_unserialize_destroy()`.

After bailout:
- `BG(unserialize).level = 1` (should be 0)
- `BG(unserialize).data` = freed heap pointer
- `php_session_cancel_decode()` deletes the session file
- Worker survives, serves new connections

### Phase 3: UAF crash

Another multipart POST (no intermediate GET requests, which would zero
`BG` state via `PHP_RINIT_FUNCTION`) includes 100 filler form fields
of 8200 bytes each. Each field becomes a 3-page `zend_string`
allocation. The allocations consume ZendMM pages sequentially through
the chunk, eventually reclaiming the 3 pages where the stale struct
resided (typically page ~288 of 512).

The `zend_string` GC header (8 bytes: `refcount` + `type_info`) at
offset 0 of the allocation overwrites the struct's `last` pointer
(a `var_entries*`). When the session decode calls `var_push()`, it
reads `last` and dereferences the GC header value as a pointer:

```
var_entries *var_hash = (*var_hashx)->last;  // GC header, not a valid address
var_hash->data[var_hash->used_slots++] = rval;  // SIGSEGV
```

The worker segfaults, confirming the use-after-free.

## Exploitation status

**Confirmed primitives:**

| Primitive | Status |
|---|---|
| OOM during RFC1867 session decode | Confirmed |
| Apache prefork worker survival after OOM | Confirmed |
| Stale `BG(unserialize)` state across requests | Confirmed (level=1, data=freed ptr) |
| Page reclamation via filler field spray | Confirmed (27+ fields crashes worker) |
| UAF crash (SIGSEGV in `var_tmp_var`) | Confirmed (reproducible) |
| Type confusion via `var_access` | Confirmed (stale zval, type matches OOM session) |

**Not achieved: full RCE.**

Two UAF primitives are confirmed, but they are mutually exclusive:

**1. Crash (filler spray path):** Filler fields reclaim the stale
struct pages. `var_tmp_var` reads `last_dtor` at struct offset 16
(now overwritten by the `zend_string` `len` field, e.g. 0x2008).
Dereferences unmapped address. SIGSEGV.

**2. Type confusion (no spray path):** Session data with `R:N;`
backreferences resolves through `var_access`, which reads inline
entries at struct offset 56. Without spray, entries.data[0] holds
a stale pointer to the attacker's own data from request 1. The
cached ZendMM chunk retains page data (only the header is zeroed).
PHP reads the stale zval without crashing: type matches whatever
the OOM session's first entry was (string, integer, etc.).

The two primitives cannot coexist: `var_tmp_var` reads offset 16
BEFORE `var_access` reads offset 56. Filler spray corrupts offset
16, crashing before the type confusion can execute.

Full RCE is blocked by:
1. **No info leak**: type confusion reads own data, not heap metadata
2. **Mutual exclusion**: spray crashes `var_tmp_var` before `var_access`
3. **GC header overlap**: uncontrolled bytes at offset 0-23 crash var_push
4. **No usable classes**: SplFileObject/PDO have function pointer calls
   in `free_obj`, but fake objects need libphp.so base (ASLR)
5. **Session key validation**: php serializer rejects keys containing `|`

## Usage

```
python3 exploit.py --target http://host:port/index.php
python3 exploit.py --target http://host:port/index.php --oom-only
```

### Options

| Flag | Default | Description |
|---|---|---|
| `--target` | (required) | Target URL (any PHP endpoint) |
| `--session-path` | `/var/www/uaf-test/sessions` | `session.save_path` on the target |
| `--entries` | 33000 | Entries in OOM session file |
| `--spray-fields` | 100 | Filler fields for crash trigger |
| `--oom-only` | off | Stop after Phase 2 |

## The patch

The fix (PHP 8.3.36, 8.4.27, 8.5.12) adds three independent defense
layers. Any one of them blocks the attack:

1. **`php_basic_early_rinit()`**: Called from `php_request_startup()`
   AFTER `zend_activate()` but BEFORE `sapi_activate()`. Zeros
   `BG(serialize_lock)`, `BG(serialize)`, and `BG(unserialize)` with
   `memset`. Eliminates the stale state window entirely.

2. **Delayed OOM**: `zend_bailout()` on OOM is removed. The allocator
   either returns NULL (fallible mode) or succeeds beyond the soft limit
   (non-fallible mode) and queues a delayed error. No longjmp, so all
   cleanup code runs normally. A hard limit calls `_exit(137)`.

3. **Session handler hardening**: RFC1867 session decode runs in
   `incomplete_classes_only` mode (no `__wakeup`/`__unserialize`).
   `php_rinit_session_globals()` at `MULTIPART_EVENT_START`.
   `$_SESSION` cleanup at `MULTIPART_EVENT_END`.

## Test environment

Tested on PHP 8.3.35 (built from source as Apache module), Apache
2.4.58 (prefork MPM, MaxRequestWorkers=1), Ubuntu 24.04, x86_64.

```
php_admin_value memory_limit 4M
php_admin_value session.upload_progress.enabled 1
php_admin_value session.upload_progress.cleanup 0
```

## References

- CVE-2026-109618: OOM during multipart parsing triggers longjmp
- CVE-2026-109619: Session globals survive across requests
- CVSS: 9.2 (both CVEs, per Rapid7)