| RFC 9267 | Vulnerabilities in DNS RR Processing | July 2022 |
| Dashevskyi, et al. | Informational | [Page] |
This memo describes common vulnerabilities related to Domain Name System (DNS) resource record (RR) processing as seen in several DNS client implementations. These vulnerabilities may lead to successful Denial-of-Service and Remote Code Execution attacks against the affected software. Where applicable, violations of RFC 1035 are mentioned.¶
This document is not an Internet Standards Track specification; it is published for informational purposes.¶
This is a contribution to the RFC Series, independently of any other RFC stream. The RFC Editor has chosen to publish this document at its discretion and makes no statement about its value for implementation or deployment. Documents approved for publication by the RFC Editor are not candidates for any level of Internet Standard; see Section 2 of RFC 7841.¶
Information about the current status of this document, any errata, and how to provide feedback on it may be obtained at https://www.rfc-editor.org/info/rfc9267.¶
Copyright (c) 2022 IETF Trust and the persons identified as the document authors. All rights reserved.¶
This document is subject to BCP 78 and the IETF Trust's Legal Provisions Relating to IETF Documents (https://trustee.ietf.org/license-info) in effect on the date of publication of this document. Please review these documents carefully, as they describe your rights and restrictions with respect to this document.¶
Major vulnerabilities in DNS implementations recently became evident and raised attention to this protocol as an important attack vector, as discussed in [SIGRED], [SADDNS], and [DNSPOOQ], the latter being a set of 7 critical issues affecting the DNS forwarder "dnsmasq".¶
The authors of this memo have analyzed the DNS client implementations of several major TCP/IP protocol stacks and found a set of vulnerabilities that share common implementation flaws (anti-patterns). These flaws are related to processing DNS resource records (RRs) (discussed in [RFC1035]) and may lead to critical security vulnerabilities.¶
While implementation flaws may differ from one software project to another, these anti-patterns are highly likely to span multiple implementations. In fact, one of the first "Common Vulnerabilities and Exposures" (CVE) documents related to one of the anti-patterns [CVE-2000-0333] dates back to the year 2000. The observations are not limited to DNS client implementations. Any software that processes DNS RRs may be affected, such as firewalls, intrusion detection systems, or general-purpose DNS packet dissectors (e.g., the DNS dissector in Wireshark; see [CVE-2017-9345]). Similar issues may also occur in DNS-over-HTTPS [RFC8484] and DNS-over-TLS [RFC7858] implementations. However, any implementation that deals with the DNS wire format is subject to the considerations discussed in this document.¶
[DNS-COMPRESSION] and [RFC5625] briefly mention some of these anti-patterns, but the main purpose of this memo is to provide technical details behind these anti-patterns, so that the common mistakes can be eradicated.¶
We provide general recommendations on mitigating the anti-patterns. We also suggest that all implementations should drop malicious/malformed DNS replies and (optionally) log them.¶
[RFC1035] defines the DNS message compression scheme that can be used to reduce the size of messages. When it is used, an entire domain name or several name labels are replaced with a (compression) pointer to a prior occurrence of the same name.¶
The compression pointer is a combination of two octets: the two most significant bits are set to 1, and the remaining 14 bits are the OFFSET field. This field specifies the offset from the beginning of the DNS header, at which another domain name or label is located:¶
+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+ | 1 1| OFFSET | +--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+--+¶
The message compression scheme explicitly allows a domain name to be represented as one of the following: (1) a sequence of unpacked labels ending with a zero octet, (2) a pointer, or (3) a sequence of labels ending with a pointer.¶
However, [RFC1035] does not explicitly state that blindly following compression pointers of any kind can be harmful [DNS-COMPRESSION], as we could not have had any assumptions about various implementations that would follow.¶
Yet, any DNS packet parser that attempts to decompress domain names without validating the value of OFFSET is likely susceptible to memory corruption bugs and buffer overruns. These bugs make it easier to perform Denial-of-Service attacks and may result in successful Remote Code Execution attacks.¶
Pseudocode that illustrates a typical example of a broken domain name parsing implementation is shown below (Figure 1):¶
1: decompress_domain_name(*name, *dns_payload) {
2:
3: name_buffer[255];
4: copy_offset = 0;
5:
6: label_len_octet = name;
7: dest_octet = name_buffer;
8:
9: while (*label_len_octet != 0x00) {
10:
11: if (is_compression_pointer(*label_len_octet)) {
12: ptr_offset = get_offset(label_len_octet,
label_len_octet+1);
13: label_len_octet = dns_payload + ptr_offset + 1;
14: }
15:
16: else {
17: length = *label_len_octet;
18: copy(dest_octet + copy_offset,
label_len_octet+1, *length);
19:
20: copy_offset += length;
21: label_len_octet += length + 1;
22: }
23:
24: }
25: }
Such implementations typically have a dedicated function for decompressing domain names (for example, see [CVE-2020-24338] and [CVE-2020-27738]). Among other parameters, these functions may accept a pointer to the beginning of the first name label within an RR ("name") and a pointer to the beginning of the DNS payload to be used as a starting point for the compression pointer ("dns_payload"). The destination buffer for the domain name ("name_buffer") is typically limited to 255 bytes as per [RFC1035] and can be allocated either in the stack or in the heap memory region.¶
The code of the function in Figure 1 reads the domain name label by label from an RR until it reaches the NUL octet ("0x00") that signifies the end of a domain name. If the current label length octet ("label_len_octet") is a compression pointer, the code extracts the value of the compression offset and uses it to "jump" to another label length octet. If the current label length octet is not a compression pointer, the label bytes will be copied into the name buffer, and the number of bytes copied will correspond to the value of the current label length octet. After the copy operation, the code will move on to the next label length octet.¶
The first issue with this implementation is due to unchecked compression offset values. The second issue is due to the absence of checks that ensure that a pointer will eventually arrive at a decompressed domain label. We describe these issues in more detail below.¶
[RFC1035] states that a compression pointer is "a pointer to a prior occurance [sic] of the same name." Also, according to [RFC1035], the maximum size of DNS packets that can be sent over UDP is limited to 512 octets.¶
The pseudocode in Figure 1 violates these constraints, as it will accept a compression pointer that forces the code to read outside the bounds of a DNS packet. For instance, a compression pointer set to "0xffff" will produce an offset of 16383 octets, which is most definitely pointing to a label length octet somewhere past the bounds of the original DNS packet. Supplying such offset values will most likely cause memory corruption issues and may lead to Denial-of-Service conditions (e.g., a Null pointer dereference after "label_len_octet" is set to an invalid address in memory). For additional examples, see [CVE-2020-25767], [CVE-2020-24339], and [CVE-2020-24335].¶
The pseudocode in Figure 1 allows jumping from a compression pointer to another compression pointer and does not restrict the number of such jumps. That is, if a label length octet that is currently being parsed is a compression pointer, the code will perform a jump to another label, and if that other label is a compression pointer as well, the code will perform another jump, and so forth until it reaches a decompressed label. This may lead to unforeseen side effects that result in security issues.¶
Consider the DNS packet excerpt illustrated below:¶
+----+----+----+----+----+----+----+----+----+----+----+----+
+0x00 | ID | FLAGS | QDCOUNT | ANCOUNT | NSCOUNT | ARCOUNT |
+----+----+----+----+----+----+----+----+----+----+----+----+
->+0x0c |0xc0|0x0c| TYPE | CLASS |0x04| t | e | s | t |0x03|
| +----+--|-+----+----+----+----+----+----+----+----+----+----+
| +0x18 | c | o| | m |0x00| TYPE | CLASS | ................ |
| +----+--|-+----+----+----+----+----+----+----+----+----+----+
| |
-----------------
¶
The packet begins with a DNS header at offset +0x00, and its DNS payload contains several RRs. The first RR begins at an offset of 12 octets (+0x0c); its first label length octet is set to the value "0xc0", which indicates that it is a compression pointer. The compression pointer offset is computed from the two octets "0xc00c" and is equal to 12. Since the broken implementation in Figure 1 follows this offset value blindly, the pointer will jump back to the first octet of the first RR (+0x0c) over and over again. The code in Figure 1 will enter an infinite-loop state, since it will never leave the "TRUE" branch of the "while" loop.¶
Apart from achieving infinite loops, the implementation flaws in Figure 1 make it possible to achieve various pointer loops that have other undesirable effects. For instance, consider the DNS packet excerpt shown below:¶
+----+----+----+----+----+----+----+----+----+----+----+----+
+0x00 | ID | FLAGS | QDCOUNT | ANCOUNT | NSCOUNT | ARCOUNT |
+----+----+----+----+----+----+----+----+----+----+----+----+
->+0x0c |0x04| t | e | s | t |0xc0|0x0c| ...................... |
| +----+----+----+----+----+----+--|-+----+----+----+----+----+
| |
------------------------------------------
¶
With such a domain name, the implementation in Figure 1 will first copy the domain label at offset "0xc0" ("test"); it will then fetch the next label length octet, which happens to be a compression pointer ("0xc0"). The compression pointer offset is computed from the two octets "0xc00c" and is equal to 12 octets. The code will jump back to offset "0xc0" where the first label "test" is located. The code will again copy the "test" label and then jump back to it, following the compression pointer, over and over again.¶
Figure 1 does not contain any logic that restricts multiple jumps from the same compression pointer and does not ensure that no more than 255 octets are copied into the name buffer ("name_buffer"). In fact,¶
Therefore, blindly following compression pointers may lead not only to Denial-of-Service conditions, as pointed out by [DNS-COMPRESSION], but also to successful Remote Code Execution attacks, as there may be other implementation issues present within the corresponding code.¶
Some implementations may not follow [RFC1035], which states:¶
The first two bits are ones. This allows a pointer to be distinguished from a label, since the label must begin with two zero bits because labels are restricted to 63 octets or less. (The 10 and 01 combinations are reserved for future use.)¶
Figures 2 and 3 show pseudocode that implements two functions that check whether a given octet is a compression pointer; Figure 2 shows a correct implementation, and Figure 3 shows an incorrect (broken) implementation.¶
1: unsigned char is_compression_pointer(*octet) {
2: if ((*octet & 0xc0) == 0xc0)
3: return true;
4: } else {
5: return false;
6: }
7: }
1: unsigned char is_compression_pointer(*octet) {
2: if (*octet & 0xc0) {
3: return true;
4: } else {
5: return false;
6: }
7: }
The correct implementation (Figure 2) ensures that the two most significant bits of an octet are both set, while the broken implementation (Figure 3) would consider an octet with only one of the two bits set to be a compression pointer. This is likely an implementation mistake rather than an intended violation of [RFC1035], because there are no benefits in supporting such compression pointer values. The implementations related to [CVE-2020-24338] and [CVE-2020-24335] had a broken compression pointer check, similar to the code shown in Figure 3.¶
While incorrect implementations alone do not lead to vulnerabilities, they may have unforeseen side effects when combined with other vulnerabilities. For instance, the first octet of the value "0x4130" may be incorrectly interpreted as a label length by a broken implementation. Such a label length (65) is invalid and is larger than 63 (as per [RFC1035]); a packet that has this value should be discarded. However, the function shown in Figure 3 will consider "0x41" to be a valid compression pointer, and the packet may pass the validation steps.¶
This might give attackers additional leverage for constructing payloads and circumventing the existing DNS packet validation mechanisms.¶
The first occurrence of a compression pointer in an RR (an octet with the two highest bits set to 1) must resolve to an octet within a DNS record with a value that is greater than 0 (i.e., it must not be a Null-terminator) and less than 64. The offset at which this octet is located must be smaller than the offset at which the compression pointer is located; once an implementation makes sure of that, compression pointer loops can never occur.¶
In small DNS implementations (e.g., embedded TCP/IP stacks), support for nested compression pointers (pointers that point to a compressed name) should be discouraged: there is very little to be gained in terms of performance versus the high probability of introducing errors such as those discussed above.¶
The code that implements domain name parsing should check the offset with respect to not only the bounds of a packet but also its position with respect to the compression pointer in question. A compression pointer must not be "followed" more than once. We have seen several implementations using a check that ensures that a compression pointer is not followed more than several times. A better alternative may be to ensure that the target of a compression pointer is always located before the location of the pointer in the packet.¶
[RFC1035] restricts the length of name labels to 63 octets and lengths of domain names to 255 octets (i.e., label octets and label length octets). Some implementations do not explicitly enforce these restrictions.¶
Consider the function "copy_domain_name()" shown in Figure 4 below. The function is a variant of the "decompress_domain_name()" function (Figure 1), with the difference that it does not support compressed labels and only copies decompressed labels into the name buffer.¶
1: copy_domain_name(*name, *dns_payload) {
2:
3: name_buffer[255];
4: copy_offset = 0;
5:
6: label_len_octet = name;
7: dest_octet = name_buffer;
8:
9: while (*label_len_octet != 0x00) {
10:
11: if (is_compression_pointer(*label_len_octet)) {
12: length = 2;
13: label_len_octet += length + 1;
14: }
15:
16: else {
17: length = *label_len_octet;
18: copy(dest_octet + copy_offset,
label_len_octet+1, *length);
19:
20: copy_offset += length;
21: label_len_octet += length + 1;
22: }
23:
24: }
25: }
This implementation does not explicitly check for the value of the label length octet: this value can be up to 255 octets, and a single label can fill the name buffer. Depending on the memory layout of the target, how the name buffer is allocated, and the size of the malformed packet, it is possible to trigger various memory corruption issues.¶
Both Figures 1 and 4 restrict the size of the name buffer to 255 octets; however, there are no restrictions on the actual number of octets that will be copied into this buffer. In this particular case, a subsequent copy operation (if another label is present in the packet) will write past the name buffer, allowing heap or stack metadata to be overwritten in a controlled manner.¶
Similar examples of vulnerable implementations can be found in the code relevant to [CVE-2020-25110], [CVE-2020-15795], and [CVE-2020-27009].¶
As a general recommendation, a domain label length octet must have a value of more than 0 and less than 64 [RFC1035]. If this is not the case, an invalid value has been provided within the packet, or a value at an invalid position might be interpreted as a domain name length due to other errors in the packet (e.g., misplaced Null-terminator or invalid compression pointer).¶
The number of domain label characters must correspond to the value of the domain label octet. To avoid possible errors when interpreting the characters of a domain label, developers may consider recommendations for the preferred domain name syntax outlined in [RFC1035].¶
The domain name length must not be more than 255 octets, including the size of decompressed domain names. The NUL octet ("0x00") must be present at the end of the domain name and must be within the maximum name length (255 octets).¶
A domain name must end with a NUL ("0x00") octet, as per [RFC1035]. The implementations shown in Figures 1 and 4 assume that this is the case for the RRs that they process; however, names that do not have a NUL octet placed at the proper position within an RR are not discarded.¶
This issue is closely related to the absence of label and name length checks. For example, the logic behind Figures 1 and 4 will continue to copy octets into the name buffer until a NUL octet is encountered. This octet can be placed at an arbitrary position within an RR or not placed at all.¶
Consider the pseudocode function shown in Figure 5. The function returns the length of a domain name ("name") in octets to be used elsewhere (e.g., to allocate a name buffer of a certain size): for compressed domain names, the function returns 2; for decompressed names, it returns their true length using the "strlen(3)" function.¶
1: get_name_length(*name) {
2:
3: if (is_compression_pointer(name))
4: return 2;
5:
6: name_len = strlen(name) + 1;
7: return name_len;
8: }
"strlen(3)" is a standard C library function that returns the length of a given sequence of characters terminated by the NUL ("0x00") octet. Since this function also expects names to be explicitly Null-terminated, the return value "strlen(3)" may also be controlled by attackers. Through the value of "name_len", attackers may control the allocation of internal buffers or specify the number by octets copied into these buffers, or they may perform other operations, depending on the implementation specifics.¶
The absence of explicit checks for placement of the NUL octet may also facilitate controlled memory reads and writes. An example of vulnerable implementations can be found in the code relevant to [CVE-2020-25107], [CVE-2020-17440], [CVE-2020-24383], and [CVE-2020-27736].¶
As a general recommendation for mitigating such issues, developers should never trust user data to be Null-terminated. For example, to fix/mitigate the issue shown in the code in Figure 5, developers should use the function "strnlen(3)", which reads at most X characters (the second argument of the function), and ensure that X is not larger than the buffer allocated for the name.¶
As stated in [RFC1035], every RR contains a variable-length string of octets that contains the retrieved resource data (RDATA) (e.g., an IP address that corresponds to a domain name in question). The length of the RDATA field is regulated by the resource data length field (RDLENGTH), which is also present in an RR.¶
Implementations that process RRs may not check for the validity of the RDLENGTH field value when retrieving RDATA. Failing to do so may lead to out-of-bound read issues, whose impact may vary significantly, depending on the implementation specifics. We have observed instances of Denial-of-Service conditions and information leaks.¶
Therefore, the value of the data length byte in response DNS records (RDLENGTH) must reflect the number of bytes available in the field that describes the resource (RDATA). The format of RDATA must conform to the TYPE and CLASS fields of the RR.¶
Examples of vulnerable implementations can be found in the code relevant to [CVE-2020-25108], [CVE-2020-24336], and [CVE-2020-27009].¶
According to [RFC1035], the DNS header contains four two-octet fields that specify the amount of question records (QDCOUNT), answer records (ANCOUNT), authority records (NSCOUNT), and additional records (ARCOUNT).¶
Figure 6 illustrates a recurring implementation anti-pattern for a function that processes DNS RRs. The function "process_dns_records()" extracts the value of ANCOUNT ("num_answers") and the pointer to the DNS data payload ("data_ptr"). The function processes answer records in a loop, decrementing the "num_answers" value after processing each record until the value of "num_answers" becomes zero. For simplicity, we assume that there is only one domain name per answer. Inside the loop, the code calculates the domain name length ("name_length") and adjusts the data payload pointer ("data_ptr") by the offset that corresponds to "name_length + 1", so that the pointer lands on the first answer record. Next, the answer record is retrieved and processed, and the "num_answers" value is decremented.¶
1: process_dns_records(dns_header, ...) {
// ...
2: num_answers = dns_header->ancount
3: data_ptr = dns_header->data
4:
5: while (num_answers > 0) {
6: name_length = get_name_length(data_ptr);
7: data_ptr += name_length + 1;
8:
9: answer = (struct dns_answer_record *)data_ptr;
10:
11: // process the answer record
12:
13: --num_answers;
14: }
// ...
15: }
If the ANCOUNT number retrieved from the header ("dns_header->ancount") is not checked against the amount of data available in the packet and it is, for example, larger than the number of answer records available, the data pointer ("data_ptr") will read outside the bounds of the packet. This may result in Denial-of-Service conditions.¶
In this section, we used an example of processing answer records. However, the same logic is often reused for implementing the processing of other types of records, e.g., the number of question (QDCOUNT), authority (NSCOUNT), and additional (ARCOUNT) records. The specified numbers of these records must correspond to the actual data present within the packet. Therefore, all record count fields must be checked before fully parsing the contents of a packet. Specifically, Section 6.3 of [RFC5625] recommends that such malformed DNS packets should be dropped and (optionally) logged.¶
Examples of vulnerable implementations can be found in the code relevant to [CVE-2020-25109], [CVE-2020-24340], [CVE-2020-24334], and [CVE-2020-27737].¶
Security issues are discussed throughout this memo; it discusses implementation flaws (anti-patterns) that affect the functionality of processing DNS RRs. The presence of such anti-patterns leads to bugs that cause buffer overflows, read-out-of-bounds, and infinite-loop issues. These issues have the following security impacts: information leaks, Denial-of-Service attacks, and Remote Code Execution attacks.¶
This document lists general recommendations for the developers of DNS record parsing functionality that allow those developers to prevent such implementation flaws, e.g., by rigorously checking the data received over the wire before processing it.¶
This document has no IANA actions. Please see [RFC6895] for a complete review of the IANA considerations introduced by DNS.¶
We would like to thank Shlomi Oberman, who has greatly contributed to the research that led to the creation of this document.¶