BUG #19527: Double-Abort Crash in `ResOwnerReleaseOSSLCipher` via `encrypt_iv` with Oversized Input

Started by PG Bug reporting form5 days ago2 messagesbugs
Jump to latest
#1PG Bug reporting form
noreply@postgresql.org

The following bug has been logged on the website:

Bug reference: 19527
Logged by: Yuelin Wang
Email address: 3020001251@tju.edu.cn
PostgreSQL version: 19beta1
Operating system: Linux (Ubuntu 24.04, x86_64)
Description:

**Component**: `contrib/pgcrypto/openssl.c`, `ResOwnerReleaseOSSLCipher`
(line 833), `free_openssl_cipher` (line 291)

Any role with `EXECUTE` on `encrypt_iv` (granted by default when pgcrypto is
installed) can crash the backend with a single statement:

```sql
CREATE EXTENSION IF NOT EXISTS pgcrypto;
SELECT encrypt_iv(
repeat('A', 1073741308)::bytea,
decode('00112233445566778899aabbccddeeff', 'hex'),
decode('000102030405060708090a0b0c0d0e0f', 'hex'),
'aes'
);
```

Expected vs actual output:

| Step | Expected | Actual |
|---|---|---|
| `encrypt_iv(...)` | `ERROR: invalid memory alloc request size ...` then
clean abort | `WARNING: AbortTransaction while in ABORT state` followed by
`ERROR: ResourceOwnerForget called for pgcrypto OpenSSL cipher handle after
release started` then backend crash |
| Server state after | Normal | `FATAL: the database system is in recovery
mode` (postmaster restarted) |

`1073741308 + 512 + 4 = 1073741824 = 0x40000000 > MaxAllocSize
(0x3FFFFFFF)`, so `palloc` throws `ERROR` and longjmps past `px_combo_free`,
leaving the `OSSLCipher` registered with the current `ResourceOwner`. When
`AbortTransaction` calls `ResourceOwnerRelease`, `ResOwnerReleaseOSSLCipher`
calls `free_openssl_cipher` with `od->owner` still set.
`free_openssl_cipher` then calls `ResourceOwnerForgetOSSLCipher`, which
checks `owner->releasing == true` and throws `elog(ERROR, ...)`. An ERROR
thrown inside `AbortTransaction` re-enters `AbortCurrentTransaction`,
calling `ResourceOwnerRelease` a second time on the already-released owner.
The second call invokes `EVP_CIPHER_CTX_free` on the already-freed EVP
context, producing a null pointer dereference (ASan: `SEGV in
EVP_CIPHER_CTX_reset`).

Server output:
```
WARNING: AbortTransaction while in ABORT state
ERROR: ResourceOwnerForget called for pgcrypto OpenSSL cipher handle after
release started
server closed the connection unexpectedly
```

ASan confirmation:
```
==ERROR: AddressSanitizer: SEGV on unknown address 0x000000000659
#0 EVP_CIPHER_CTX_reset (libcrypto.so.3)
#1 EVP_CIPHER_CTX_free (libcrypto.so.3)
#2 free_openssl_cipher openssl.c:294
#3 ResOwnerReleaseOSSLCipher openssl.c:835
#4 ResourceOwnerReleaseAll resowner.c:395
#5 AbortTransaction xact.c:3016
SUMMARY: AddressSanitizer: SEGV (EVP_CIPHER_CTX_reset)
```

The fix is to clear `od->owner` before calling `free_openssl_cipher` in
`ResOwnerReleaseOSSLCipher`:

```c
static void
ResOwnerReleaseOSSLCipher(Datum res)
{
OSSLCipher *od = (OSSLCipher *) DatumGetPointer(res);
od->owner = NULL;
free_openssl_cipher(od);
}
```

#2Tom Lane
tgl@sss.pgh.pa.us
In reply to: PG Bug reporting form (#1)
Re: BUG #19527: Double-Abort Crash in `ResOwnerReleaseOSSLCipher` via `encrypt_iv` with Oversized Input

PG Bug reporting form <noreply@postgresql.org> writes:

Any role with `EXECUTE` on `encrypt_iv` (granted by default when pgcrypto is
installed) can crash the backend with a single statement:

```sql
CREATE EXTENSION IF NOT EXISTS pgcrypto;
SELECT encrypt_iv(
repeat('A', 1073741308)::bytea,
decode('00112233445566778899aabbccddeeff', 'hex'),
decode('000102030405060708090a0b0c0d0e0f', 'hex'),
'aes'
);
```

The fix is to clear `od->owner` before calling `free_openssl_cipher` in
`ResOwnerReleaseOSSLCipher`:

Good catch, thanks for the report!

regards, tom lane