A little background

When we talk about UEFI Secure Boot, a lot of times we talk about what we call the whitelist and the blacklist. These databases go by many names—formally EFI_IMAGE_SECURITY_DATABASE and EFI_IMAGE_SECURITY_DATABASE1, sometimes d719b2cb-3d3a-4596-a3bc-dad00e67656f:db and d719b2cb-3d3a-4596-a3bc-dad00e67656f:dbx, but most often just db and dbx. These two[^1] databases, stored in UEFI Authenticated Variables[^2], constitute lists of which binaries can and cannot be executed within a UEFI environment when Secure Boot is enabled.

When a UEFI binary is loaded, the system first checks to see if any revocations in dbx are applicable to that binary. A dbx entry may apply in three ways—it may contain the hash of a specific binary, an X.509 certificate, or the hash of a certificate[^3]. If any of these matches the binary in question, UEFI raises the error EFI_SECURITY_VIOLATION, and your binary will not go to space today.

If a binary successfully passes that hurdle, then the same verification method is processed with db, the whitelist, but with the opposite policy: if any entry describes the binary in question, verification has succeeded. If no entry does, EFI_SECURITY_VIOLATION is raised.

The need for updates

When a UEFI binary is discovered to have a serious security flaw which would allow a malicious user to circumvent Secure Boot, it becomes necessary to prevent it from running on machines. When this happens, the UEFI CA issues a dbx update. The mechanism for the update is a file that’s structured as a UEFI Authenticated Variable append. This file gets distributed as part of an OS update, and when the update is applied, the UEFI variable is updated.

One mechanism for doing this in Linux is via dbxtool. In Fedora, the dbxtool package includes the current dbx updates in /usr/share/dbxtool/DBXUpdate-2014-04-13-22-14-00.bin, as well as a systemd service to apply them during boot[^4]. In the special case of Fedora’s shim being added to a blacklist, we would include a dependency on the fixed version of shim in the dbxtool package so that systems would remain bootable.

The structure of a `dbx` update

The dbx variable is composed of an array of EFI_SIGNATURE_LIST structures, each themselves containing an array of EFI_SIGNATURE_DATA entries. A dbx update is that same structure, but wrapped in a EFI_VARIABLE_AUTHENTICATION_2 structure to authenticate it. The definitions look like this[^5]:

typedef struct {
efi_guid_t SignatureOwner; // who owns this entry
uint8_t SignatureData[0]; // the data we want to
// fish out of this thing
typedef struct {
efi_guid_t SignatureType; // type of structure in
uint32_t SignatureListSize; // Total size of the signature
// list, including this header.
uint32_t SignatureHeaderSize; // Size of type-specific header
uint32_t SignatureSize; // The size of each individual
// in this list.
// uint8_t SignatureHeader[SignatureHeaderSize]
// this is a header defined by
// and for each specific
// signature type. Of course
// none of them actually define
// a header.
// EFI_SIGNATURE_DATA[...][SignatureSize] // actual signature data
typedef struct {
efi_guid_t HashType;
uint8_t PublicKey[256];
uint8_t Signature[256];
typedef struct {
uint32_t dwLength; // Length of this structure
uint16_t wRevision; // Revision of this structure (2)
uint16_t wCertificateType; // The kind of signature this is
//uint16_t bCertificate[0]; // The signature data itself. This
// is actually, and not the least
// bit confusingly, the rest of
// structure wrapping this one.
#define WIN_CERT_TYPE_PKCS_SIGNED_DATA 0x0002 </div></div>
#define WIN_CERT_TYPE_EFI_PKCS115 0x0ef0
#define WIN_CERT_TYPE_EFI_GUID 0x0ef1</span>
typedef struct {
WIN_CERTIFICATE Hdr; // Info about which structure this is
efi_guid_t CertType; // Type of certificate in CertData
uint8_t CertData[0]; // A certificate of some kind
typedef struct {
EFI_TIME TimeStamp; // monotonically increasing
// timestamp to prevent replay
// attacks.
WIN_CERTIFICATE_EFI_GUID AuthInfo; // Information about how to
// authenticate this variable
// against some KEK entry
</pre></div></figure> Conceptually, this means the structure we've got is:
[ dbx update file:
[ Authentication structure:
[ monotonic number | timestamp ]
[ Cert Type ]
[ Certificate Data ] ]
... ]
So a full update looks something like this:
``` plain auth2.TimeStamp 00000000 da 07 03 06 13 11 15 00 00 00 00 00 00 00 00 00 |................| ``` 2010-03-06 19:17:21 GMT+0000 ``` plain auth2.AuthInfo.Hdr.DwLength 00000010 bd 0c 00 00 |.... | ``` It is 0x00000cbd bytes long[^6].
``` plain auth2.AuthInfo.Hdr.wRevision 00000010 00 02 | .. | ``` It's revision is 2. It is always revision 2. ``` plain auth2.AuthInfo.Hdr.wCertificateType 00000010 f1 0e | .. | ``` 0x0ef1, which as we see above is `WIN_CERT_TYPE_EFI_GUID`. The interesting bit here is that `.bCertificate` isn't quite a real thing. This is actually describing that `auth2.AuthInfo` is a `WIN_CERTIFICATE_EFI_GUID`, and `.bCertificate` is actually the fields other than `auth2.AuthInfo.Hdr`—`auth2.AuthInfo.CertType` and `auth2.AuthInfo.CertData`. Again, this shouldn't confuse you at all. As a result, next, in the place of `.bCertificate`, we have: ``` plain auth2.AuthInfo.CertType 00000010 9d d2 af 4a df 68 ee 49 | ...J.h.I| 00000020 8a a9 34 7d 37 56 65 a7 |..4}7Ve. | ``` `4aafd29d-68df-49ee-8aa9-347d375665a7`, aka `EFI_GUID_PKCS7_CERT`, which means that `.CertData` is verified against `EFI_SIGNATURE_DATA` entries in the Key Exchange Keys database (`KEK`), but only those entries which are contained in `EFI_SIGNATURE_LIST` structures with `.SignatureType` of `EFI_CERT_X509_GUID`. *Whew*. ``` plain auth2.AuthInfo.CertData 00000020 30 82 0c a1 02 01 01 31 | 0......1| 00000030 0f 30 0d 06 09 60 86 48 01 65 03 04 02 01 05 00 |.0...`.H.e......| [ what an X.509 certificate looks like is left as an exercise for the reader ] 00000cb0 b6 2b 89 02 73 c4 86 57 83 6f 28 57 e0 12 cb 05 |.+..s..W.o(W....| 00000cc0 6d d0 3e 60 8f 85 9f dd fc 46 ac 54 44 |m.>`.....F.TD | ``` Yep, that's an ASN.1 DER encoding of a PKCS-7 Certificate.
That's the end of the `EFI_VARIABLE_AUTHENTICATION_2` structure, and so on to the actual *data*: ``` plain EFI_SIGNATURE_LIST.SignatureType 00000cc0 26 16 c4 | &..| 00000cd0 c1 4c 50 92 40 ac a9 41 f9 36 93 43 28 |.LP.@..A.6.C( | ``` That's c1c41626-504c-4092-aca9-41f936934328, aka `EFI_GUID_SHA256` ``` plain EFI_SIGNATURE_LIST.SignatureListSize 00000cd0 cc 01 00 | ...| 00000ce0 00 |. | ``` The list size is 0x00001cc bytes ``` plain EFI_SIGNATURE_LIST.SignatureHeaderSize: 00000ce0 00 00 00 00 | .... | ``` This is actually always 0. ``` plain EFI_SIGNATURE_LIST.SignatureSize 00000ce0 30 00 00 00 | 0... | ``` 0x00000030 bytes. That's 16 bytes for the `efi_guid_t`, and since `.SignatureType` was `EFI_GUID_SHA256`, 32 bytes of SHA-256 data. ``` plain EFI_SIGNATURE_DATA[0].SignatureOwner 00000ce0 bd 9a fa 77 59 03 32 |.....0......wY.2| 00000cf0 4d bd 60 28 f4 e7 8f 78 4b |M.`(...xK | ``` 77fa9abd-0359-4d32-bd60-28f4e78f784b , which is the GUID Microsoft uses to identify themselves[^7] ``` plain EFI_SIGNATURE_DATA[0].SignatureData 00000cf0 80 b4 d9 69 31 bf 0d | ...i1..| 00000d00 02 fd 91 a6 1e 19 d1 4f 1d a4 52 e6 6d b2 40 8c |.......O..R.m.@.| 00000d10 a8 60 4d 41 1f 92 65 9f 0a |.`MA..e.. | ``` This is the actual SHA-256 data. It goes on like this: ``` plain EFI_SIGNATURE_DATA[1..8] 00000d10 bd 9a fa 77 59 03 32 | ...wY.2| 00000d20 4d bd 60 28 f4 e7 8f 78 4b f5 2f 83 a3 fa 9c fb |M.`(...xK./.....| 00000d30 d6 92 0f 72 28 24 db e4 03 45 34 d2 5b 85 07 24 |...r($...E4.[..$| 00000d40 6b 3b 95 7d ac 6e 1b ce 7a bd 9a fa 77 59 03 32 |k;.}.n..z...wY.2| 00000d50 4d bd 60 28 f4 e7 8f 78 4b c5 d9 d8 a1 86 e2 c8 |M.`(...xK.......| 00000d60 2d 09 af aa 2a 6f 7f 2e 73 87 0d 3e 64 f7 2c 4e |-...*o..s..>d.,N| 00000d70 08 ef 67 79 6a 84 0f 0f bd |..gyj.... | ... 00000e60 bd 9a fa 77 59 03 32 | ...wY.2| 00000e70 4d bd 60 28 f4 e7 8f 78 4b 53 91 c3 a2 fb 11 21 |M.`(...xKS.....!| 00000e80 02 a6 aa 1e dc 25 ae 77 e1 9f 5d 6f 09 cd 09 ee |.....%.w..]o....| 00000e90 b2 50 99 22 bf cd 59 92 ea |.P."..Y..| ```

How the structure is used

When you apply a database update, you're basically doing a `SetVariable()` call to UEFI with a couple of flags set: ``` c rc = efi_set_variable(EFI_IMAGE_SECURITY_DATABASE_GUID, "dbx", data, len, EFI_VARIABLE_TIME_BASED_AUTHENTICATED_WRITE_ACCESS | EFI_VARIABLE_APPEND_WRITE | EFI_VARIABLE_BOOTSERVICE_ACCESS | EFI_VARIABLE_RUNTIME_ACCESS | EFI_VARIABLE_NON_VOLATILE); ``` These flags tell the firmware some crucial things - that this variable is authenticated with the `EFI_VARIABLE_AUTHENTICATION_2` structure, that this is an append, that both *Boot Services* (i.e. the firmware) and *Runtime Services* (i.e. the OS) should have access to it, and that it should persist across a reboot. As a special case in the spec, an append has a special meaning for the UEFI security databases:

For variables with the GUID EFI_IMAGE_SECURITY_DATABASE_GUID (i.e. where the data buffer is formatted as EFI_SIGNATURE_LIST), the driver shallnot perform an append of EFI_SIGNATURE_DATA values that are already part of the existing variable value. Note: This situation is not considered an error, and shall in itself not cause a status code other than EFI_SUCCESS to be returned or the timestamp associated with the variable not to be updated.

UEFI Specification section 7.2.1 revision 2.4
As a result, what happens here is that the first time you write to `dbx`, any `EFI_SIGNATURE_LIST` structures and the `EFI_SIGNATURE_DATA` entries they contain get added to the variable, but not the `EFI_VARIABLE_AUTHENTICATION_2` structure. Then when later `dbx` updates are issued, they contain a superset of the previous ones. When you apply them, the firmware only appends the difference to the variable.


Obviously a system this complex needs some tools. To manage these databases on linux, I've written a tool called `dbxtool`, which may be available in your linux distribution of choice. It can be used to apply `dbx` changes[^8], to list the contents of the UEFI Security Databases, and to list the contents of updates files:
fenchurch:~$ dbxtool -l
1: {microsoft} {sha256} 80b4d96931bf0d02fd91a61e19d14f1da452e66db2408ca8604d411f92659f0a
2: {microsoft} {sha256} f52f83a3fa9cfbd6920f722824dbe4034534d25b8507246b3b957dac6e1bce7a
3: {microsoft} {sha256} c5d9d8a186e2c82d09afaa2a6f7f2e73870d3e64f72c4e08ef67796a840f0fbd
4: {microsoft} {sha256} 363384d14d1f2e0b7815626484c459ad57a318ef4396266048d058c5a19bbf76
5: {microsoft} {sha256} 1aec84b84b6c65a51220a9be7181965230210d62d6d33c48999c6b295a2b0a06
6: {microsoft} {sha256} e6ca68e94146629af03f69c2f86e6bef62f930b37c6fbcc878b78df98c0334e5
7: {microsoft} {sha256} c3a99a460da464a057c3586d83cef5f4ae08b7103979ed8932742df0ed530c66
8: {microsoft} {sha256} 58fb941aef95a25943b3fb5f2510a0df3fe44c58c95e0ab80487297568ab9771
9: {microsoft} {sha256} 5391c3a2fb112102a6aa1edc25ae77e19f5d6f09cd09eeb2509922bfcd5992ea
10: {microsoft} {sha256} d626157e1d6a718bc124ab8da27cbb65072ca03a7b6b257dbdcbbd60f65ef3d1
11: {microsoft} {sha256} d063ec28f67eba53f1642dbf7dff33c6a32add869f6013fe162e2c32f1cbe56d
12: {microsoft} {sha256} 29c6eb52b43c3aa18b2cd8ed6ea8607cef3cfae1bafe1165755cf2e614844a44
13: {microsoft} {sha256} 90fbe70e69d633408d3e170c6832dbb2d209e0272527dfb63d49d29572a6f44c
As usual, there are a couple of places where vendors have not gotten everything quite right, and sometimes things fail to work correctly—but that's for another post. [^1]: There is actually a third in this set, `dbt`, which can be used in revocation processing. [^2]: Strictly speaking these are only an analog to *Authenticated Variables*, which can only be appended to, replaced, or deleted by updates signed with the key that created them. Key database updates are instead controlled by a list of keys stored in another variable called `KEK` - the Key Exchange Keys. Otherwise the mechanism is the same. [^3]: That is, the digest of the certificate's `TBSCertificate` as defined in RFC 5280 section, using the digest specified in the database entry itself. [^4]: This is currently disabled by default in Fedora. I'm looking at enabling this as an F22 feature. Getting these things right is important, and it takes time. [^5]: I have left out the definitions of `EFI_TIME` and `efi_guid_t`; they are quite boring. [^6]: Here `dwLength` includes the size of `Hdr` itself (that is, the size of `dwLength`, `wRevision`, and `wCertificateType`) as well as the data following it (`bCertificate`). Because we live in the best of all possible worlds, Authenticode signatures—the signatures on binaries themselves—use the same structure but only include the size of `Hdr.bCertificate`. This hasn't ever confused *anybody*. [^7]: Big congratulations to Acer, who used 55555555-5555-5555-5555-555555555555 for this in one of their db entries. Not only did they win the random number generator lottery to get that, but also experienced a minimum of 3 single bit errors, since the closest valid GUID to that is 55555555-5555-4555-9555-555555555555. [^8]: Also theoretically `db` updates, `dbt` updates, and `KEK` updates, but those are much more rare.