Difference between revisions of "Savegames"

From 3dbrew
Jump to navigation Jump to search
(→‎Partitions: additional info on the partition's DIFI/IVFC/DPFS structures...)
m (fix typos)
 
(134 intermediate revisions by 5 users not shown)
Line 1: Line 1:
This page describes the format, de/encryption, etc. of savegames found in 3DS game cartridges/gamecards. You can find savegames from various 3DS games on the [[Games]] page.
+
This page describes the format and encryption of savegames contained in gamecards, SD and NAND. You can find savegames from various 3DS games on the [[Games]] page.
  
 +
== Overview ==
 +
Savegames are stored in [[DISA and DIFF|DISA container format]]. Inside the DISA container, it forms a [[Inner FAT|FAT filesystem]]. '''Please refer to these pages for how to fully extract save files'''. This page only describes additional encryption wear leveling on top of the DISA container. These layers only apply to gamecard save games. SD savegames and NAND savegames are DISA containers in plaintext after decrypting the common SD/NAND encryption layer.
  
=== Encryption ===
+
== Gamecard savegame Encryption ==
  
On the 3DS savegames are stored much like on the DS, that is on a FLASH chip in the gamecart. On the DS these savegames were stored in plain-text but on the 3DS a layer of encryption was added. This is highly likely a streamcipher, as the contents of several savegames exhibit the odd behavior that xor-ing certain parts of the savegame together will result in the plain-text appearing.
+
Gamecard encryption is AES-CTR applied on top of DISA container, but below the wear leveling layer (if exists). The same key Y used for encryption is also used for DISA CMAC signing. Several versions of encryption scheme have been introduced over the time.
  
The reason this works is because the stream cipher used has a period of 512 bytes. That is to say, it will repeat the same keystream after 512 bytes. The way you encrypt with a stream cipher is you XOR your data with the keystream as it is produced. Unfortunately, if your streamcipher repeats and you are encrypting a known plain-text (in our case, zeros) you are basically giving away your valuable keystream.
+
{| class="wikitable" border="1"
 
 
So how do you use this to decrypt a savegame on a 3DS? First off, you chunk up the savegame into 512 byte chunks. Then, you bin these chunks by their contents, discarding any that contain only FF. Now look for the most common chunk. This is your keystream. Now XOR the keystream with your original savegame and you should have a fully decrypted savegame. XOR with the keystream again to produce an encrypted savegame.
 
 
 
=== Wear leveling ===
 
 
 
The 3DS employs a wear leveling scheme on the savegame FLASH chips. This is done through the usage of blockmaps and a journal. The blockmap is located at offset 0 of the flash chip, and is immediately followed by the journal. The initial state is dictated by the blockmap, and the journal is then applied to that.
 
 
 
First, there are 8 bytes whose purposes are currently unknown. Then comes the actual blockmap.
 
The blockmap structure is simple:
 
<pre>
 
struct header_entry {
 
        uint8_t phys_sec; // when bit7 is set, block has checksums, otherwise checksums are all zero
 
        uint8_t alloc_cnt;
 
        uint8_t chksums[8];
 
} __attribute__((__packed__));
 
</pre>
 
 
 
There's one entry per sector, counting from physical sector 1 (sector 0 contains the blockmap/journal).
 
 
 
The 2 bytes that follow the blockmap are the CRC16 (with starting value 0xFFFF (like modbus)) of the first 8 bytes and the blockmap.
 
 
 
Then comes the journal.
 
The journal structure is as follows:
 
<pre>
 
struct sector_entry {
 
        uint8_t virt_sec;      // Mapped to sector
 
        uint8_t prev_virt_sec;  // Physical sector previously mapped to
 
        uint8_t phys_sec;      // Mapped from sector
 
        uint8_t prev_phys_sec;  // Virtual sector previously mapped to
 
        uint8_t phys_realloc_cnt;      // Amount of times physical sector has been remapped
 
        uint8_t virt_realloc_cnt;      // Amount of times virtual sector has been remapped
 
        uint8_t chksums[8];
 
} __attribute__((__packed__));
 
 
 
struct long_sector_entry{
 
        struct sector_entry sector;
 
        struct sector_entry dupe;
 
        uint32_t magic;
 
}__attribute__((__packed__));
 
</pre>
 
 
 
With magic being a constant 0x080d6ce0.
 
 
 
The checksums in the blockmap/journal entries work as follows:
 
* each byte is the checksum of an encrypted 0x200 bytes large block
 
* to calculate the checksum, a CRC16 of the block (with starting value 0xFFFF) is calculated, and the two bytes of the CRC16 are XORed together to produce the 8bit checksum
 
 
 
=== Partitions ===
 
 
 
There can be multiple partitions on the chip.
 
The partitions are represented by tables of DIFI blobs inside a DISA structure.
 
The order of the DIFI blobs is the order of the partitions in the chip.
 
 
 
'''DISA'''
 
 
 
* If the uint32 @ 0x168 into the image in the DISA is a %1=1, then first table is is hashed, otherwise the second DIFI table is hashed.
 
* If the table has more then 1 DIFI then the uint32 @ 0x168 is the offset from the DATA partition to the file base (masked with 0xFFFFFFFE).
 
 
 
{| class="wikitable"
 
|-
 
! Start
 
! Length
 
! Description
 
|-
 
| 0x00
 
| 4
 
| Magic ("DISA")
 
 
|-
 
|-
| 0x04
+
!  FW Introduced
| 4
+
!  Old3DS
| Unknown
+
!  [[AES#Keyslot|AES Keyslots]] (Encryption / CMAC)
 +
!  KeyY generation method
 +
!  Repeating CTR
 
|-
 
|-
| 0x08
+
| The initial version
| 8
+
| style="background: #ccffbb" | Yes
| Unknown
+
| 0x37 / 0x33
 +
| v1
 +
| style="background: #ccffbb" | Yes
 
|-
 
|-
| 0x10
+
| [[2.0.0-2]]
| 8
+
| style="background: #ccffbb" | Yes
| Offset to first "DIFI" blob table in DISA (usually 0x0200)
+
| 0x37 / 0x33
 +
| v2
 +
| style="background: #ccffbb" | Yes
 
|-
 
|-
| 0x18
+
| [[2.2.0-4]]
| 8
+
| style="background: #ccffbb" | Yes
| Offset to second "DIFI" blob table in DISA (usually 0x0330)
+
| 0x37 / 0x33
 +
| v2
 +
| style="background: #ffccbb" | No
 
|-
 
|-
| 0x20
+
| [[6.0.0-11]]
| 8
+
| style="background: #ccffbb" | Yes
| Size of the "DIFI" blobs table
+
| 0x37 / 0x33
 +
| v3
 +
| style="background: #ffccbb" | No
 
|-
 
|-
| 0x28
+
| [[9.6.0-24|9.6.0-X]]
| 8
+
| style="background: #ffccbb" | No
| Unknown (padding?)
+
| 0x1A / 0x19
|-
+
| v2?
| 0x30
+
| style="background: #ffccbb" | No
| 8
 
| Size of data to be hashed
 
|-
 
| 0x38
 
| 8*6
 
| Unknown
 
|-
 
| 0x68
 
| 4
 
| Active table (and the offset to the filebase)
 
|-
 
| 0x6C
 
| 0x20
 
| Hashes
 
|-
 
| 0x8C
 
| 4*29
 
| Unknown
 
 
|}
 
|}
  
 +
=== Repeating CTR Fail ===
 +
On the 3DS savegames are stored much like on the DS, that is on a FLASH chip in the gamecart. On the DS these savegames were stored in plain-text but on the 3DS a layer of encryption was added. This is AES-CTR, as the contents of several savegames exhibit the odd behavior that xor-ing certain parts of the savegame together will result in the plain-text appearing.
  
<pre>
+
The reason this works is because the stream cipher used has a period of 512 bytes. That is to say, it will repeat the same keystream after 512 bytes. The way you encrypt with a stream cipher is you XOR your data with the keystream as it is produced. Unfortunately, if your streamcipher repeats and you are encrypting a known plain-text (in our case, zeros) you are basically giving away your valuable keystream.
struct disa_header {
 
        char magic[4];
 
        u8 unknown0[0xC];
 
        u64 first_difi_table_offset;
 
        u64 second_difi_table_offset;
 
        u64 table_size;
 
        u64 Padding;
 
        u64 hash_size;
 
        u8 unknown1[0x30];
 
        u32 active_table;              // and the offset to the filebase
 
        u8 hash[0x20];
 
        u8 unknown2[0x74];
 
} __attribute__((__packed__));
 
</pre>
 
* The hash in the DISA blob hashes 'hash_size' bytes of the 'active_table'.
 
  
The DIFIs table @ 0x200 (into the image) is written twice, (Meaning, if there's 4 DIFI blobs then the table is 2 DIFIs long).
+
So how do you use this to decrypt a savegame on a 3DS? First off, you chunk up the savegame into 512 byte chunks. Then, you bin these chunks by their contents, discarding any that contain only FF. Now look for the most common chunk. This is your keystream. Now XOR the keystream with your original savegame and you should have a fully decrypted savegame. XOR with the keystream again to produce an encrypted savegame.
  
The second table is for backup. The active table is mentioned at 0x13C into the image (1=First table, other=Second Table)
 
  
'''DIFI'''
+
=== KeyY Generation method ===
  
These 0x130 large blobs describe the partitions. Every DIFI blob describes a partition. In order to find the partitions, you will need the uint32_t at 0x9C into the DIFI block, and the uint32_t at 0xA4. The uint32_t at 0x9C describes the length of the hash table at the start of the partition, the uint32_t at 0xA4 is the length of the filesystem. Partitions are catted together, so after the end of one partition is the beginning of the next (on the next 0x1000 block).
+
The [[NCSD]] partition flags determine the method used to generate this keyY.
  
Actually DIFI blobs are 0x12C large because the last 4 are not used and appear 0xFFFFFFFF at the encrypted image.
+
==== v1 ====
  
For most games there's only 1 partition (The SAVE partition) and some (like Asphalt 3D and Steel Diver) has 2 partitions.
+
When all of the flags checked by the running NATIVE_FIRM are clear, the keyY is the following:
 
+
{| class="wikitable" border="1"
* 2 Partitions means that the files inside the SAVE partition is on the other partition (we would call it DATA partition).
 
 
 
* No more than 2 partitions have been seen yet.
 
 
 
{| class="wikitable"
 
 
|-
 
|-
! Start
+
! Offset
! Length
+
! Size
! Description
+
! Description
 
|-
 
|-
| 0x00
+
| 0x0
| 4
+
| 0x8
| Magic ("DIFI")
+
| First 8-bytes from the plaintext [[NCCH#CXI|CXI]] accessdesc signature.
 
|-
 
|-
| 0x04
+
| 0x8
| 4
+
| 0x4
| Unknown
+
| u32 CardID0 from [[Gamecards|gamecard]] plaintext-mode command 0x90, Process9 reads this with the [[NTRCARD]] hw. The actual cmdID used by Process9 is different since Process9 reads it with the gamecard in encrypted-mode.
 
|-
 
|-
| 0x08
+
| 0xC
| 8
+
| 0x4
| Offset to "IVFC" blob in DIFI (usually 0x44)
+
| u32 CardID1 from [[Gamecards|gamecard]] plaintext-mode command 0xA0, Process9 reads this with the [[NTRCARD]] hw. The actual cmdID used by Process9 is different since Process9 reads it with the gamecard in encrypted-mode.
 +
|}
 +
 
 +
==== v2 ====
 +
 
 +
Key Y is the first 0x10 bytes of SHA-256 calculated over the following data
 +
 
 +
{| class="wikitable" border="1"
 
|-
 
|-
| 0x10
+
!  Offset
| 8
+
!  Size
| Size of "IVFC" blob
+
!  Description
 
|-
 
|-
| 0x18
+
| 0x0
| 8
+
| 0x8
| Offset to "DPFS" blob in DIFI (usually 0xBC)
+
| First 8-bytes from the plaintext [[NCCH#CXI|CXI]] accessdesc signature.
|-
 
| 0x20
 
| 8
 
| Size of "DPFS" blob
 
|-
 
| 0x28
 
| 8
 
| Offset to the hashes in DIFI (usually 0x010C)
 
|-
 
| 0x30
 
| 8
 
| Size of this hashes
 
|-
 
| 0x38
 
| 8
 
| Unknown
 
 
|-
 
|-
 +
| 0x8
 
| 0x40
 
| 0x40
| 4
+
| read from a gamecard command(this 0x40-byte data is also read by [[Process_Services_PXI|GetRomId]], which is the gamecard-uniqueID)
| Unknown (0x00 filled)
 
 
|}
 
|}
  
'''IVFC'''
+
This keyY generation method was implemented with [[2.0.0-2]] via NCSD partition flag[3], however the proper CTR wasn't implemented for flag[7] until [[2.2.0-4]]. The hashed keyY flag[3] implemented with [[2.0.0-2]] was likely never used with retail gamecards.
 +
 
 +
==== v3 ====
  
{| class="wikitable"
+
[[6.0.0-11]] implemented support for generating the savegame keyY with a new method, this method is much more complex than previous keyY methods. This is enabled via new [[NCSD]] partition flags, all retail games which have the NCSD image finalized after the [[6.0.0-11]] release(and [[6.0.0-11]]+ in the system update partition) will have these flags set for using this new method.
|-
 
! Start
 
! Length
 
! Description
 
|-
 
| 0x00
 
| 4
 
| Magic ("IVFC")
 
|-
 
| 0x04
 
| 0x54
 
| Unknown
 
|-
 
| 0x58
 
| 8
 
| Hash Table Length
 
|-
 
| 0x60
 
| 8
 
| Partition Length
 
|-
 
| 0x68
 
| 8
 
| Hashed Block Length (1 << Value)
 
|-
 
| 0x70
 
| 8
 
| Unknown (usually 0x78=120)
 
|-
 
|}
 
  
'''DPFS'''
+
First, a SHA-256 hash is calculated over the following data
  
{| class="wikitable"
+
{| class="wikitable" border="1"
 
|-
 
|-
! Start
+
! Offset
! Length
+
! Size
! Description
+
! Description
 
|-
 
|-
| 0x00
+
| 0x0
| 4
+
| 0x8
| Magic ("DPFS")
+
| First 8-bytes from the plaintext [[NCCH#CXI|CXI]] accessdesc signature.
|-
 
| 0x04
 
| 0x3C
 
| Unknown
 
 
|-
 
|-
 +
| 0x8
 
| 0x40
 
| 0x40
| 8
+
| Same ID as [[Process_Services_PXI|GetRomId]]
| Offset To Next Partition (or to the end of this partiton)
 
 
|-
 
|-
 
| 0x48
 
| 0x48
| 8
+
| 0x8
| Unknown
+
| CXI Program ID
 
|-
 
|-
 +
| 0x50
 +
| 0x20
 +
| ExeFS:/.code hash from the decrypted [[ExeFS]] header
 
|}
 
|}
  
* The u64 at 0x40 is an offset because if you summarize the hashtable's length and the partition's length you'll get sometimes less then this value. so it isn't the length of the total partition but an offset.
+
Then an [[AES]]-CMAC is calculated over this hash. The output CMAC is used for keyY. The key slot for this CMAC is 0x2F.
 +
 
 +
The 0x2F keyY used for calculating this AES-CMAC (not to be confused with the final keyY for decrypting/signing savegames) is initialized while NATIVE_FIRM is loading, this keyY is generated via the [[RSA]] engine. The RSA slot used here is slot0(key-data for slot0 is initialized by bootrom), this RSA slot0 key-data is overwritten during system boot. This RSA slot0 key-data gets overwritten with the RSA key-data used for verifying RSA signatures, every time Process9 verifies any RSA signatures except for [[NCCH|NCCH]] accessdesc signatures. Starting with [[7.0.0-13]] this key-init function used at boot is also used to initialize a separate keyslot used for the new [[NCCH]] encryption method.
 +
 
 +
This [[FIRM|Process9]] key-init function first checks if a certain 0x10-byte block in the 0x01FF8000 region is all-zero. When all-zero it immediately returns, otherwise it clears that block then continues to do the key generation. This is likely for supporting launching a v6.0+ NATIVE_FIRM under this FIRM.
 +
 
 +
== Gamecard wear leveling ==
  
==== The SAVE partition ====
+
The 3DS employs a wear leveling scheme on the savegame FLASH chips(only used for CARD1 gamecards). This is done through the usage of blockmaps and a journal. The blockmap is located at offset 0 of the flash chip, and is immediately followed by the journal. The initial state is dictated by the blockmap, and the journal is then applied to that.
  
Savefiles are stored on the FLASH in a custom filesystem called SAVE. SAVE has a header which describes where the various bits of the filesystem live.
+
There are two versions of wear leveling have been observed. V1 is used for 128KB and 512 KB CARD1 flash chips. V2 is used for 1MB CARD1 flash chips (uncommon. Pokemon Sun/Moon is an example).
  
The first partition starts at 0x2000. The hashtable at the start of the partition describe each in-use block in the partition with a SHA256 of the 0x200 sized block.
+
First, there are two 32-bit integers whose purposes are currently unknown. They generally increase the value as the savegame is written more times, so probably counter for how many times the journal became full and got flushed into the block map, and/or how many times <code>alloc_cnt</code> has wrapped around.  
  
* The first two hashes don't seem to be associated with any 0x200 block.
+
Then comes the actual blockmap. The block map contains entries of 10 bytes (V1) or 2 bytes (V2) with total number of <code>(flash_size / 0x1000 - 1)</code>.
 +
The blockmap entry is simple:
 +
<pre>
 +
struct blockmap_entry_v1 {
 +
        uint8_t phys_sec; // when bit7 is set, block is initialized and has checksums, otherwise checksums are all zero
 +
        uint8_t alloc_cnt;
 +
        uint8_t chksums[8];
 +
} __attribute__((__packed__));
  
Finding the file system file base:
+
struct blockmap_entry_v2 {
* If there's no DATA partiton the file base will be at the uint32 @ 0x58 into the SAVE header.
+
        // Note that the phys_sec and alloc_cnt field are swapped in v2,
* Otherwise the file base will be at ('active_table' & 0xFFFFFFFE) into the DATA partition.
+
        // but the initialized bit is still on the first byte
 +
        uint8_t alloc_cnt; // when bit7 is set, block is initialized
 +
        uint8_t phys_sec;
 +
        // v2 has no chksums
 +
} __attribute__((__packed__));
 +
</pre>
  
Finiding the file system table:
+
There's one entry per 0x1000-byte sector, counting from physical sector 1 (sector 0 contains the blockmap/journal).
* If the uint32 @ 0x58 and uint32 @ 0x6C into the SAVE header aren't 0 then the table is at [@0x58] + [@0x6C]*0x200.
 
* Otherwise, (meaning the files are at another partition) the exact offset into the SAVE is @ 0x68.
 
  
Once you've found the FST, parsing it is fairly straightforward.
+
A 2-byte CRC16 follows the block map. For V1 it immediately follows the last block map entry. For V2 it is located at 0x3FE, and bytes before the CRC is padded with zero. The CRC16 checks all the bytes before it, including the two unknown integers, the block map, and the padding bytes for V2. The CRC standard used looks like CRC-16-IBM (modbus). Here is the code in Rust for it
  
 
<pre>
 
<pre>
struct fs_entry {
+
fn crc16(data: &[u8]) -> u16 {
    u32 node_cnt;
+
    let poly = 0xA001;
    u8  filename[0x10];
+
    let mut crc = 0xFFFFu16;
    u32 index;
+
    for byte in data {
    u32 unk1; // magic?
+
        crc ^= <u16>::from(*byte);
    u32 block_offset;
+
        for _ in 0..8 {
    u32 file_size;
+
            let b = crc & 1 != 0;
    u32 unk2;
+
            crc >>= 1;
    u32 unk3; // flags and/or date?
+
            if b {
    u32 unk4;
+
                crc ^= poly;
}
+
            }
 +
        }
 +
    }
 +
    crc
 +
}
 
</pre>
 
</pre>
  
The first entry is the root directory, easily identifiable by the node_cnt being larger than 1. The node_cnt includes the root directory itself, so there are node_cnt - 1 files in the root directory. The entries that follow after the root directory are the actual files. Reading them out is as simple as taking the file base offset and adding (block_offset * 0x200) to it.
+
Then comes the journal. The journal contains entries that describes how sectors should be remapped. The rest bytes before 0x1000 after all journal entries are padded with 0xFF
 
+
The journal entry structure is as follows:
Here's a follow-up example from the Legend of Zelda: Ocarina of Time 3D:
 
 
<pre>
 
<pre>
//FST entry = SAVE base + File base + (FST offset * 0x200) + (FST entry # * 0x30)
+
struct journal_entry_half {
//0x2600    = 0x2000    + 0x400     + (0x1        * 0x200) + (0x0        * 0x30)
+
        uint8_t virt_sec;      // Mapped to sector
 +
        uint8_t prev_virt_sec;  // Physical sector previously mapped to
 +
        uint8_t phys_sec;      // Mapped from sector
 +
        uint8_t prev_phys_sec;  // Virtual sector previously mapped to
 +
        uint8_t phys_realloc_cnt;      // Amount of times physical sector has been remapped
 +
        uint8_t virt_realloc_cnt;      // Amount of times virtual sector has been remapped
 +
        uint8_t chksums[8];     // Unused & uninitialized for V2
 +
} __attribute__((__packed__));
  
00002600: 03000000 09000000 00000000 00000000  ................
+
struct journal_entry{
00002610: 00000000 00000000 00000000 00000000  ................
+
        struct journal_entry_half entry;
00002620: 00000000 00000000 00000000 00000000  ................
+
        struct journal_entry_half dupe; // same data as `entry`. No idea what this is used fore
00002630: 01000000 73797374 656D2E64 61740000  ....system.dat..
+
        uint32_t uninitialized;        // 0xFFFFFFFF in newer system
00002640: 00000000 00000000 D57B1100 02000000  ........Õ{......
+
}__attribute__((__packed__));
00002650: 22000000 00000000 E8121500 00000000  ".......è.......
 
00002660: 01000000 73617665 30302E62 696E0000  ....save00.bin..
 
00002670: 00000000 01000000 69921100 03000000  ........i’......
 
00002680: DC140000 00000000 04000000 00000000  Ü...............
 
 
</pre>
 
</pre>
  
Example for a different file with different SAVE structure:
 
  
<pre>
+
The checksums in the blockmap/journal entries work as follows:
00002400 53415645 00000400 20000000 00000000 SAVE.... ....... //Save from Steel Diver
+
* each byte is the checksum of an encrypted 0x200 bytes large block
00002410 28030000 00000000 00020000 00000000 (...............
+
* to calculate the checksum, a CRC16 of the block (same CRC16 algorithm as above) is calculated, and the two bytes of the CRC16 are XORed together to produce the 8bit checksum
00002420 00000000 00020000 88000000 00000000 ................
+
 
00002430 03000000 00020000 94000000 00000000 ................
+
== Initialization ==
00002440 43000000 00020000 A0010000 00000000 C...............
+
 
00002450 28030000 00020000 00000000 00000000 (............... //[0x58] = 0 and  
+
When a save FLASH contains all xFFFF blocks it's assumed uninitialized by the game cartridges and it initializes default data in place, without prompting the user. The 0xFFFFFFFF blocks are uninitialized data. When creating a non-gamecard savegame and other images/files, it's initially all 0xFFFFFFFF until it's formatted where some of the blocks are overwritten with encrypted data.
00002460 28030000 00020000 E81A0000 00000000 (............... //[0x6C] = 0, but
 
00002470 00000000 00020000 381B0000 00000000 ........8....... //On offset 0x68 There's an exact offset to the FST
 
00002480 40000000 00020000 01000000 00000000 @............... //meaning a uint32_t at 0x68 into the SAVE struct
 
00002490 00000000 00000000 00000000 02000000 ................
 
  
00003F30 00000000 00000000 04000000 41000000 ............A... // first fs_entry '04 00 00 00'
+
I got a new game SplinterCell3D-Pal and I downloaded the save and it was 128KB of 0xFF, except the first 0x10 bytes which were the letter 'Z' (uppercase) --[[User:Elisherer|Elisherer]] 22:41, 15 October 2011 (CEST)
00003F40 00000000 00000000 00000000 00000000 ................ //exectly 0x1B38 from the save header struct
 
00003F50 00000000 00000000 00000000 00000000 ................
 
00003F60 00000000 00000000 01000000 67686F73 ............ghos
 
00003F70 742E7374 30370000 00000000 00000000 t.st07..........
 
00003F80 D57B1100 00000000 7E290000 00000000 .{......~)......
 
00003F90 00000000 00000000 01000000 73617665 ............save
 
00003FA0 2E737562 00000000 00000000 01000000 .sub............
 
00003FB0 D57B1100 15000000 9C090000 00000000 .{..............
 
00003FC0 00000000 00000000 01000000 73617665 ............save
 
00003FD0 2E706572 69730000 00000000 02000000 .peris..........
 
00003FE0 D57B1100 1A000000 29070000 00000000 .{......).......
 
00003FF0 00000000 00000000 00000000 00000000 ................
 
</pre>
 
  
known struct until now:
+
== Fun Facts ==
  
<pre>
+
If you have facts that you found out by looking at the binary files please share them here:
struct save_header {
 
        char magic[4];          //'SAVE'
 
        u8 unknown0[0x54];
 
        u32 file_base_offset;
 
        u8 unknown1[0x10];
 
        u32 fst_block_offset;    //FST is in [file_base_offset] * 0x200 + [fst_block_offset]
 
        u8 unknown2[8];          //or (if no filebase)
 
        uint32 fst_exact_offset; //The exact offset from the header start
 
}
 
</pre>
 
  
=== Initialization ===
+
* From one save to another the game backups the last files that were in the partition and the entire image header in "random" locations.. --[[User:Elisherer|Elisherer]] 22:41, 15 October 2011 (CEST)
  
When a save EEPROM contains all xFFFF blocks it's assumed uninitialized by the game cartridges and it initializes default data in place, without prompting the user.
+
== Tools ==
  
 +
* [https://github.com/wwylele/save3ds save3ds] supports reading and modifying savegames, extdata and title database in FUSE filesystem or batch extracting/importing.
 +
* [https://github.com/3dshax/3ds/tree/master/3dsfuse 3dsfuse] supports reading and modifying savegames. In the mounted FUSE filesystem, the /output.sav is the raw FLASH save-image. When the save was modified, a separate tool to update the CMAC must be used with /clean.sav, prior to writing output.sav to a gamecard. (This is an old tool that doesn't handle the savegame format correctly. --[[User:Wwylele|Wwylele]] ([[User talk:Wwylele|talk]]) 16:13, 2 December 2019 (CET))
 +
* [[3DSExplorer]] supports reading of savegames, it doesn't support reading the new encrypted savegames and maybe in the future it will support modifying (some of the modyfing code is already implemented).
 +
* [https://github.com/wwylele/3ds-save-tool wwylele's 3ds-save-tool] supports extracting files from savegames and extdata. It properly reconstructs data from the DPFS tree and extracts files in directories hierarchy.
  
 
[[セーブデータ|Japanese]]
 
[[セーブデータ|Japanese]]

Latest revision as of 15:15, 3 September 2021

This page describes the format and encryption of savegames contained in gamecards, SD and NAND. You can find savegames from various 3DS games on the Games page.

Overview[edit]

Savegames are stored in DISA container format. Inside the DISA container, it forms a FAT filesystem. Please refer to these pages for how to fully extract save files. This page only describes additional encryption wear leveling on top of the DISA container. These layers only apply to gamecard save games. SD savegames and NAND savegames are DISA containers in plaintext after decrypting the common SD/NAND encryption layer.

Gamecard savegame Encryption[edit]

Gamecard encryption is AES-CTR applied on top of DISA container, but below the wear leveling layer (if exists). The same key Y used for encryption is also used for DISA CMAC signing. Several versions of encryption scheme have been introduced over the time.

FW Introduced Old3DS AES Keyslots (Encryption / CMAC) KeyY generation method Repeating CTR
The initial version Yes 0x37 / 0x33 v1 Yes
2.0.0-2 Yes 0x37 / 0x33 v2 Yes
2.2.0-4 Yes 0x37 / 0x33 v2 No
6.0.0-11 Yes 0x37 / 0x33 v3 No
9.6.0-X No 0x1A / 0x19 v2? No

Repeating CTR Fail[edit]

On the 3DS savegames are stored much like on the DS, that is on a FLASH chip in the gamecart. On the DS these savegames were stored in plain-text but on the 3DS a layer of encryption was added. This is AES-CTR, as the contents of several savegames exhibit the odd behavior that xor-ing certain parts of the savegame together will result in the plain-text appearing.

The reason this works is because the stream cipher used has a period of 512 bytes. That is to say, it will repeat the same keystream after 512 bytes. The way you encrypt with a stream cipher is you XOR your data with the keystream as it is produced. Unfortunately, if your streamcipher repeats and you are encrypting a known plain-text (in our case, zeros) you are basically giving away your valuable keystream.

So how do you use this to decrypt a savegame on a 3DS? First off, you chunk up the savegame into 512 byte chunks. Then, you bin these chunks by their contents, discarding any that contain only FF. Now look for the most common chunk. This is your keystream. Now XOR the keystream with your original savegame and you should have a fully decrypted savegame. XOR with the keystream again to produce an encrypted savegame.


KeyY Generation method[edit]

The NCSD partition flags determine the method used to generate this keyY.

v1[edit]

When all of the flags checked by the running NATIVE_FIRM are clear, the keyY is the following:

Offset Size Description
0x0 0x8 First 8-bytes from the plaintext CXI accessdesc signature.
0x8 0x4 u32 CardID0 from gamecard plaintext-mode command 0x90, Process9 reads this with the NTRCARD hw. The actual cmdID used by Process9 is different since Process9 reads it with the gamecard in encrypted-mode.
0xC 0x4 u32 CardID1 from gamecard plaintext-mode command 0xA0, Process9 reads this with the NTRCARD hw. The actual cmdID used by Process9 is different since Process9 reads it with the gamecard in encrypted-mode.

v2[edit]

Key Y is the first 0x10 bytes of SHA-256 calculated over the following data

Offset Size Description
0x0 0x8 First 8-bytes from the plaintext CXI accessdesc signature.
0x8 0x40 read from a gamecard command(this 0x40-byte data is also read by GetRomId, which is the gamecard-uniqueID)

This keyY generation method was implemented with 2.0.0-2 via NCSD partition flag[3], however the proper CTR wasn't implemented for flag[7] until 2.2.0-4. The hashed keyY flag[3] implemented with 2.0.0-2 was likely never used with retail gamecards.

v3[edit]

6.0.0-11 implemented support for generating the savegame keyY with a new method, this method is much more complex than previous keyY methods. This is enabled via new NCSD partition flags, all retail games which have the NCSD image finalized after the 6.0.0-11 release(and 6.0.0-11+ in the system update partition) will have these flags set for using this new method.

First, a SHA-256 hash is calculated over the following data

Offset Size Description
0x0 0x8 First 8-bytes from the plaintext CXI accessdesc signature.
0x8 0x40 Same ID as GetRomId
0x48 0x8 CXI Program ID
0x50 0x20 ExeFS:/.code hash from the decrypted ExeFS header

Then an AES-CMAC is calculated over this hash. The output CMAC is used for keyY. The key slot for this CMAC is 0x2F.

The 0x2F keyY used for calculating this AES-CMAC (not to be confused with the final keyY for decrypting/signing savegames) is initialized while NATIVE_FIRM is loading, this keyY is generated via the RSA engine. The RSA slot used here is slot0(key-data for slot0 is initialized by bootrom), this RSA slot0 key-data is overwritten during system boot. This RSA slot0 key-data gets overwritten with the RSA key-data used for verifying RSA signatures, every time Process9 verifies any RSA signatures except for NCCH accessdesc signatures. Starting with 7.0.0-13 this key-init function used at boot is also used to initialize a separate keyslot used for the new NCCH encryption method.

This Process9 key-init function first checks if a certain 0x10-byte block in the 0x01FF8000 region is all-zero. When all-zero it immediately returns, otherwise it clears that block then continues to do the key generation. This is likely for supporting launching a v6.0+ NATIVE_FIRM under this FIRM.

Gamecard wear leveling[edit]

The 3DS employs a wear leveling scheme on the savegame FLASH chips(only used for CARD1 gamecards). This is done through the usage of blockmaps and a journal. The blockmap is located at offset 0 of the flash chip, and is immediately followed by the journal. The initial state is dictated by the blockmap, and the journal is then applied to that.

There are two versions of wear leveling have been observed. V1 is used for 128KB and 512 KB CARD1 flash chips. V2 is used for 1MB CARD1 flash chips (uncommon. Pokemon Sun/Moon is an example).

First, there are two 32-bit integers whose purposes are currently unknown. They generally increase the value as the savegame is written more times, so probably counter for how many times the journal became full and got flushed into the block map, and/or how many times alloc_cnt has wrapped around.

Then comes the actual blockmap. The block map contains entries of 10 bytes (V1) or 2 bytes (V2) with total number of (flash_size / 0x1000 - 1). The blockmap entry is simple:

struct blockmap_entry_v1 {
        uint8_t phys_sec; // when bit7 is set, block is initialized and has checksums, otherwise checksums are all zero
        uint8_t alloc_cnt;
        uint8_t chksums[8];
} __attribute__((__packed__));

struct blockmap_entry_v2 {
        // Note that the phys_sec and alloc_cnt field are swapped in v2, 
        // but the initialized bit is still on the first byte
        uint8_t alloc_cnt; // when bit7 is set, block is initialized
        uint8_t phys_sec; 
        // v2 has no chksums
} __attribute__((__packed__));

There's one entry per 0x1000-byte sector, counting from physical sector 1 (sector 0 contains the blockmap/journal).

A 2-byte CRC16 follows the block map. For V1 it immediately follows the last block map entry. For V2 it is located at 0x3FE, and bytes before the CRC is padded with zero. The CRC16 checks all the bytes before it, including the two unknown integers, the block map, and the padding bytes for V2. The CRC standard used looks like CRC-16-IBM (modbus). Here is the code in Rust for it

fn crc16(data: &[u8]) -> u16 {
    let poly = 0xA001;
    let mut crc = 0xFFFFu16;
    for byte in data {
        crc ^= <u16>::from(*byte);
        for _ in 0..8 {
            let b = crc & 1 != 0;
            crc >>= 1;
            if b {
                crc ^= poly;
            }
        }
    }
    crc
}

Then comes the journal. The journal contains entries that describes how sectors should be remapped. The rest bytes before 0x1000 after all journal entries are padded with 0xFF The journal entry structure is as follows:

struct journal_entry_half {
        uint8_t virt_sec;       // Mapped to sector
        uint8_t prev_virt_sec;  // Physical sector previously mapped to
        uint8_t phys_sec;       // Mapped from sector
        uint8_t prev_phys_sec;  // Virtual sector previously mapped to
        uint8_t phys_realloc_cnt;       // Amount of times physical sector has been remapped
        uint8_t virt_realloc_cnt;       // Amount of times virtual sector has been remapped
        uint8_t chksums[8];     // Unused & uninitialized for V2
} __attribute__((__packed__));

struct journal_entry{
        struct journal_entry_half entry;
        struct journal_entry_half dupe; // same data as `entry`. No idea what this is used fore
        uint32_t uninitialized;         // 0xFFFFFFFF in newer system
}__attribute__((__packed__));


The checksums in the blockmap/journal entries work as follows:

  • each byte is the checksum of an encrypted 0x200 bytes large block
  • to calculate the checksum, a CRC16 of the block (same CRC16 algorithm as above) is calculated, and the two bytes of the CRC16 are XORed together to produce the 8bit checksum

Initialization[edit]

When a save FLASH contains all xFFFF blocks it's assumed uninitialized by the game cartridges and it initializes default data in place, without prompting the user. The 0xFFFFFFFF blocks are uninitialized data. When creating a non-gamecard savegame and other images/files, it's initially all 0xFFFFFFFF until it's formatted where some of the blocks are overwritten with encrypted data.

I got a new game SplinterCell3D-Pal and I downloaded the save and it was 128KB of 0xFF, except the first 0x10 bytes which were the letter 'Z' (uppercase) --Elisherer 22:41, 15 October 2011 (CEST)

Fun Facts[edit]

If you have facts that you found out by looking at the binary files please share them here:

  • From one save to another the game backups the last files that were in the partition and the entire image header in "random" locations.. --Elisherer 22:41, 15 October 2011 (CEST)

Tools[edit]

  • save3ds supports reading and modifying savegames, extdata and title database in FUSE filesystem or batch extracting/importing.
  • 3dsfuse supports reading and modifying savegames. In the mounted FUSE filesystem, the /output.sav is the raw FLASH save-image. When the save was modified, a separate tool to update the CMAC must be used with /clean.sav, prior to writing output.sav to a gamecard. (This is an old tool that doesn't handle the savegame format correctly. --Wwylele (talk) 16:13, 2 December 2019 (CET))
  • 3DSExplorer supports reading of savegames, it doesn't support reading the new encrypted savegames and maybe in the future it will support modifying (some of the modyfing code is already implemented).
  • wwylele's 3ds-save-tool supports extracting files from savegames and extdata. It properly reconstructs data from the DPFS tree and extracts files in directories hierarchy.

Japanese