3DS System Flaws

From 3dbrew
Revision as of 11:38, 30 December 2016 by Mtgxyz (talk | contribs)
Jump to navigation Jump to search

Exploits are used to execute unofficial code (homebrew) on the Nintendo 3DS. This page is a list of publicly known system flaws, for userland applications/applets flaws see here.

Stale / Rejected Efforts

  • Neimod has been working on a RAM dumping setup for a little while now. He's de-soldered the 3DS's RAM chip and hooked it and the RAM pinouts on the 3DS' PCB up to a custom RAM dumping setup. A while ago he published photos showing his setup to be working quite well, with the 3DS successfully booting up. However, his flickr stream is now private along with most of his work.
  • Someone (who will remain unnamed) has released CFW and CIA installers, all of which is copied from the work of others, or copyrighted material.

Tips and info

The 3DS uses the XN feature of the ARM11 processor. There's no official way from applications to enable executable permission for memory containing arbitrary unsigned code(there's a SVC for this, but only RO-module has access to it). A usable userland exploit would still be useful: you could only do return-oriented-programming with it initially. From ROP one could then exploit system flaw(s), see below.

SD card extdata and SD savegames can be attacked, for consoles where the console-unique movable.sed was dumped(accessing SD data is far easier by running code on the target 3DS however).

System flaws

Hardware

Summary Description Fixed with hardware model/revision Newest hardware model/revision this flaw was checked for Timeframe this was discovered Discovered by
ARM9/ARM11 bootrom vectors point at uninitialized RAM ARM9's and ARM11's exception vectors are hardcoded to point at the CPU's internal memory (0x08000000 region for ARM9, AXIWRAM for ARM11). While the bootrom does set them up to point to an endless loop at some point during boot, it does not do so immediately. As such, a carefully-timed fault injection (via hardware) to trigger an exception (such as an invalid instruction) will cause execution to fall into ARM9 RAM.

Since RAM isn't cleared on boot (see below), one can immediately start execution of their own code here to dump bootrom, OTP, etc. The ARM9 bootrom does the following at reset: reset vector branches to another instruction, then branches to bootrom+0x8000. Hence, there's no way to know for certain when exactly the ARM9 exception-vector data stored in memory gets initialized.

This requires *very* *precise* timing for triggering the hardware fault.

It has been exploited by derrek (others?) to dump the ARM9 bootrom as of Summer 2015.

None: all available 3DS models at the time of writing have the exact same ARM9/ARM11 bootrom for the unprotected areas. New3DS End of February 2014 derrek, WulfyStylez (May 2015) independently
Missing AES key clearing The hardware AES engine does not clear keys when doing a hard reset/reboot. None New3DS August 2014 Mathieulh/Others
No RAM clearing on reboots On an MCU-triggered reboot all RAM including FCRAM/ARM9 memory/AXIWRAM/VRAM keeps its contents. None New3DS March 2014 derrek
32bits of actual console-unique TWLNAND keydata On retail the 8-bytes at ARM9 address 0x01FFB808 are XORed with hard-coded data, to generate the TWL console-unique keys, including TWLNAND. On Old3DS the high u32 is always 0x0, while on New3DS that u32 is always 0x2. On top of this, the lower u32's highest bit is always ORed. only 31 bits of the TWL console-unique keydata / TWL consoleID are actually console-unique.

This allows one to easily bruteforce the TWL console-unique keydata with *just* data from TWLNAND. On DSi the actual console-unique data for key generation is 8-bytes(all bytes actually set).

None New3DS 2012? Yellows8
DSi / 3DS-TWL key-generator After using the key generator to generate the normal-key, you could overwrite parts of the normal-key with your own data and then recover the key-generator output by comparing the new crypto output with the original crypto output. From the normal-key outputs, you could deduce the TWL key-generator function.

This applies to the keyX/keyY too.

This attack does not work for the 3DS key-generator because keyslots 0-3 are only for TWL keys.

None New3DS 2011 Yellows8
3DS key-generator The algorithm for generating the normal-keys for keyslots is cryptographically weak. As a result, it is easily susceptible to differential cryptanalysis if the normal-key corresponding to any scrambler-generated keyslot is discovered.

Several such pairs of matching normal-keys and KeyY values were found, leading to deducing the key-generator function.

None New3DS February 2015 Yellows8, plutoo
FIRM partitions known-plaintext The FIRM partitions are encrypted with AES-CTR without a MAC. Since this works by XOR'ing data with a static (per-console in this case) keystream, one can deduce the keystream of a portion of each FIRM partition if they have the actual FIRM binary stored in it.

This can be paired with many exploits. For example, it allows minor FIRM downgrades (i.e. 10.4 to 9.6 or 9.5 to 9.4, but not 9.6 to 9.5).

This can be somewhat addressed by having a FIRM header skip over previously used section offsets, but this would just air-gap newer FIRMs without fixing the core bug. This can also only be done a limited number of times due to the size of FIRM versus the size of the partitions.

None New3DS Everyone

ARM9 software

boot9

Summary Description Fixed with hardware model/revision Newest hardware model/revision this flaw was checked for Timeframe this was discovered Discovered by
Incorrect padding check The FIRM signature is using PKCS #1 padding, which mandates that the padding is all 0xFF bytes. This is not checked correctly, because it only checks whether none of the bytes in the padding are 0x00. This allows a signature to be crafted more easily (sighax) N/A New3DS Summer 2015 derrek and/or nedwill
No bound checks inside of ASN.1 parser The hash inside of the signature is stored in an ASN.1 structure. However the length fields are not bounds-checked, allowing one to point the header hash to the hash the 3DS calculated before verification. This and because of the aforementioned bug, you can brute-force a signature that will always work easily, as essentially only a few bytes need to be valid. N/A New3DS Summer 2015 derrek and/or nedwill

arm9loader

Summary Description Successful exploitation result Fixed in FIRM system version Last FIRM system version this flaw was checked for Timeframe this was discovered Public disclosure timeframe Discovered by
Rearrangable keys in the NAND keystore Due to the keystore being encrypted with AES-ECB, one can rearrange blocks and still have the NAND keystore decrypt in a deterministic way.

Using 10.0 FIRM it is possible to rearrange keys such that ARM9 memory is executed. As such using existing ARM9 execution 10.0 FIRM can be written to NAND and a payload written to memory, with the payload to be executed post-K9L using an MCU reboot.

arm9loaderhax given existing ARM9 code execution None 11.2.0-X Early 2016 27 September 2016 Myria, dark_samus; mathieulh (independently); plutoo (independently) + others
Uncleared OTP hash keydata in console-unique 0x11 key-generation Kernel9Loader does not clear the SHA_HASH register after use. As a result, the data stored here as K9L hands over to Kernel9 is the hash of OTP data used to seed the console-unique NAND keystore decryption key set on keyslot 0x11.

Retrieving this keydata and the NAND keystore of the same device allows calculating the decrypted New3DS NAND keystore (non-unique, common to all New3DS units), which contains AES normal keys, also set on keyslot 0x11, which are then used to derive all current New3DS-only AES keyXs including the newer batch introduced in 9.6.0-X. From there, it is trivial to perform the same key derivation in order to initialize those keys on any system version, and even on Old3DS.

This can be performed by exploiting the "arm9loaderhax" vulnerability to obtain post-K9L code execution after an MCU reboot (the bootrom section-loading fail is not relevant here, this attack was performed without OTP data by brute-forcing keys), and using this to dump the SHA_HASH register. This attack works on any FIRM version shipping a vulnerable version of K9L, whereas OTP dumping required a boot of <3.0.0-X.

This attack results in obtaining the entire (0x200-bytes) NAND keystore - it was confirmed at a later date that this keystore is encrypted with the same key (by comparing the decrypted data from multiple units), and therefore using another key in this store will not remedy the issue as all keys are known (i.e. later, unused keys decrypt to the same 0x200-bytes constant with the same OTP hash). Later keys could have been encrypted differently but this is not the case. As a result of this, it is not possible for Nintendo to use K9L again in its current format for its intended purpose, though this was not news from the moment people dumped a New3DS OTP.

Derivation of all New3DS keys generated via the NAND keystore (0x1B "Secure4" etc.) None 11.2.0-X ~April 2015, implemented in May 2015 13 January 2016 WulfyStylez, Dazzozo, shinyquagsire23 (complimentary + implemented), plutoo, Normmatt (discovered independently)
enhanced-arm9loaderhax See the 32c3 3ds talk.

Since this is a combination of a trick with the arm9-bootrom + arm9loaderhax, and since you have to manually write FIRM to the firm0/firm1 NAND partitions, this can't be completely fixed. Any system with existing ARM9 code execution and an OTP/OTP hash dump can exploit this. Additionally, by using the FIRM partition known-plaintext bug and bruteforcing the second entry in the keystore, this can currently be exploited on all New3DS systems without any other prerequisite hacks.

arm9loaderhax which automatically occurs at hard-boot. See arm9loaderhax / description. See arm9loaderhax / description. Theorized around mid July, 2015. Later implemented+tested by plutoo and derrek. 32c3 3ds talk (December 27, 2015) Yellows8
Missing verification-block for the 9.6 keys (arm9loaderhax) Starting with 9.6.0-X a new set of NAND-based keys were introduced. However, no verification block was added to verify that the new key read from NAND is correct. This was technically an issue from 9.5.0-X with the original sector+0 keydata, however the below is only possible with 9.6.0-X since keyslots 0x15 and 0x16 are generated from different 0x11 keyXs.

Writing an incorrect key to NAND will cause arm9loader to decrypt the ARM9 kernel as garbage and then jump to it.

This allows an hardware-based attack where you can boot into an older exploited firmware, fill all memory with NOP sleds/jump-instructions, and then reboot into executing garbage. By automating this process with various input keydata, eventually you'll find some garbage that jumps to your code.

This gives very early ARM9 code execution (pre-ARM9 kernel). As such, it is possible to dump RSA keyslots with this and calculate the 6.x save, and 7.x NCCH keys. This cannot be used to recover keys initialized by arm9loader itself. This is due to it wiping the area used for its stack during NAND sector decryption and keyslot init.

Due to FIRMs on both Old and New 3DS using the same RSA data, this can be exploited on Old3DS as well, but only if one already has the actual plaintext normalkey from New3DS NAND sector 0x96 offset-0 and has dumped the OTP area of the Old3DS.

Recovery of 6.x save key/7.x NCCH key, access to uncleared OTP hash keydata None 11.2.0-X March, 2015 plutoo
Uncleared New3DS keyslot 0x11 Originally the New3DS FIRM arm9bin loader only cleared keyslot 0x11 when it gets executed at firmlaunch. This was fixed with 9.5.0-X by completely clearing keyslot 0x11 immediately after the loader finishes using keyslot 0x11.

This means that any ARM9 code that can execute before the loader clears the keyslot at firmlaunch(including firmlaunch-hax) can get access to the uncleared keyslot 0x11, which then allows one to generate all <=v9.5 New3DS keyXs which are generated by keyslot 0x11.

Therefore, to completely fix this the loader would have to generate more keys using different keyslot 0x11 keydata. This was done with 9.6.0-X.

New3DS keyXs generation Mostly fixed with 9.5.0-X, completely fixed with new keys with 9.6.0-X. February 3, 2015 (one day after 9.5.0-X release) Yellows8

Process9

Summary Description Successful exploitation result Fixed in FIRM system version Last FIRM system version this flaw was checked for Timeframe this was discovered Public disclosure timeframe Discovered by
Leak of normal-key matching a key-scrambler key New 3DS firmware versions 8.1.0 through 9.2.0 set the encryption key for Amiibo data using a hardcoded normal-key in Process9. In firmware 9.3.0, Nintendo "fixed" this by using the key scrambler instead, by calculating the keyY value for keyslot 0x39 that results in the same normal-key, then hardcoding that keyY into Process9.

Nintendo's fix is actually the problem: Nintendo revealed the normal-key matching an unknown keyX and a known keyY. Combined with the key scrambler using an insecure scrambling algorithm (see "Hardware" above), the key scrambler function could be deduced.

Deducing the keyX for keyslot 0x39 and the key scrambler algorithm New 3DS 9.3.0-X, sort of 10.0.0-X Sometime in 2015 after the hardware key-generator was broken. 32c3 3ds talk (December 27, 2015) Yellows8
Leak of normal-key matching a key-generator key During the 3DS' development (June/July 2010) Nintendo added support installing encrypted content (CIA). Common-key index1 was intended to be a hardware generated key. However while they added code to generate the key in hardware, they forgot to remove the normal-key for index1 (used elsewhere, likely old debug code). Nintendo later removed the normal key sometime before the first non-prototype firmware release.


Knowing the keyY and the normal-key for common-key index1, the devkit key-generator algorithm can be deduced (see "Hardware" above). Additionally the remaining devkit common-keys can be generated once the common-key keyX is recovered.

Note the devkit key-generator was discovered to be the same as the retail key-generator.

Deducing the keyX for keyslot 0x3D and hardware key-generator algorithm. Generate remaining devkit common-keys. pre-1.0.0-X Shortly after the key-generator was revealed to be flawed at the 32c3 3ds talk January 20, 2016 jakcron
ntrcardhax ARM9 code execution 10.4.0-29 March 2015 32c3 3ds talk (December 27, 2015) plutoo
Title downgrading via AM(PXI) When a title is *already* installed, Process9 will compare the installed title-version with the title-version being installed. When the one being installed is older, Process9 would return an error.

However, this can be bypassed by just deleting the title first via the service command(s) for that: with the title removed from the Title_Database, Process9 can't compare the input title-version with anything. Hence, titles can be downgraded this way.

11.0.0-X fixed this for key system titles (MSET, Home Menu, spider, ErrDisp, SKATER, NATIVE_FIRM, and every retail system module), by checking the version of the title to install against a hard-coded list of (titleID, minimumVersionRequired) pairs.

Bypassing title version check at installation, which then allows downgrading any title. 11.0.0-X, for key system titles. NATIVE_FIRM / AM-sysmodule 11.0.0-X ? ?
FAT FS code null-deref When FSFile:Read is used with a file which is corrupted on a FAT filesystem(in particular SD), Process9 can crash. This particular crash is caused by a function returning NULL instead of an actual ptr due to an error. The caller of that function doesn't check for NULL which then triggers a read based at NULL.

Sample "fsck.vfat -n -v -V <fat image backup>" output for the above crash: 

...
Starting check/repair pass.
<FilePath0> and
<FilePath1>
 share clusters.
 Truncating second to 3375104 bytes.
<FilePath1>
 File size is 2787392 bytes, cluster chain length is 16384 bytes.
 Truncating file to 16384 bytes.
Checking for unused clusters.
Reclaimed 1 unused cluster (16384 bytes).
Checking free cluster summary.
Free cluster summary wrong (1404490 vs. really 1404491)
 Auto-correcting.
Starting verification pass.
Checking for unused clusters.
Leaving filesystem unchanged.
 Useless null-based-read None 9.6.0-X July 8-9, 2015 Yellows8
RSA signature padding checks The TWL_FIRM RSA sig padding check code used for all TWL RSA sig-checks has issues, see here.

The main 3DS RSA padding check code(non-certificate, including NATIVE_FIRM) uses the function used with the above to extract more padding + the actual hash from the additional padding. This isn't really a problem here because there's proper padding check code which is executed prior to this.

None 9.5.0-X March 2015 Yellows8
AMPXI:ValidateDSiWareSectionMAC AES keyslot reuse When the input DSiWare section index is higher than <max number of DSiWare sections supported by this FIRM>, Process9 uses keyid 0x40 for calculating the AESMAC, which translates to keyslot 0x40. The result is that the keyslot is left at whatever was already selected before, since the AES selectkeyslot code will immediately return when keyslot is >=0x40. However, actually exploiting this is difficult: the calculated AESMAC is never returned, this command just compares the calculated AESMAC with the input AESMAC(result-code depends on whether the AESMACs match). It's unknown whether a timing attack would work with this.

This is basically a different form of the pxips9 keyslot vuln, except with AESMAC etc.

See description. None 11.2.0-X March 15, 2015 December 29, 2015 Yellows8
pxips9 AES keyslot reuse This requires access to the ps:ps/pxi:ps9 services. One way to get access to this would be snshax on system-version <=10.1.0-X(see 32c3 3ds talk).

When an invalid key-type value is passed to any of the PS commands, Process9 will try to select keyslot 0x40. That aesengine_setkeyslot() code will then immediately return due to the invalid keyslot value. Since that function doesn't return any errors, Process9 will just continue to do crypto with whatever AES keyslot was selected before the PS command was sent.

Reusing the previously used keyslot, for crypto with PS. None 11.2.0-X Roughly the same time(same day?) as firmlaunch-hax. December 29, 2015 Yellows8
firmlaunch-hax: FIRM header ToCToU This can't be exploited from ARM11 userland.

During FIRM launch, the only FIRM header the ARM9 uses at all is stored in FCRAM, this is 0x200-bytes(the actual used FIRM RSA signature is read to the Process9 stack however). The ARM9 doesn't expect "anything" besides the ARM9 to access this data. With 9.5.0-22 the address of this FIRM header was changed from a FCRAM address, to ARM9-only address 0x01fffc00.

ARM9 code execution 9.5.0-22 2012, 3 days after Yellows8 started Process9 code RE. Yellows8
Uninitialized data output for (PXI) command replies PXI commands for various services(including some here and many others) can write uninitialized data (like from ARM registers) to the command reply. This happens with stubbed commands, but this can also occur with certain commands when returning an error.

Certain ARM11 service commands have this same issue as well.

None 9.3.0-X ? Yellows8
FSPXI OpenArchive SD permissions Process9 does not use the exheader ARM9 access-mount permission flag for SD at all.

This would mean ARM11-kernelmode code / fs-module itself could directly use FSPXI to access SD card without ARM9 checking for SD access, but this is rather useless since a process is usually running with SD access(Home Menu for example) anyway.

None 9.3.0-X 2012 Yellows8
AMPXI:ExportDSiWare export path Process9 allocates memory on Process9 heap for the export path then verifies that the actual allocated size matches the input size. Then Process9 copies the input path from FCRAM to this buffer, and uses it with the Process9 FS openfile code, which use paths in the form of "<mountpoint>:/<path>".

Process9 does not check the contents of this path at all before passing it to the FS code, besides writing a NUL-terminator to the end of the buffer.

Exporting of DSiWare to arbitrary Process9 file-paths, such as "nand:/<path>" etc. This isn't really useful since the data which gets written can't be controlled. None 9.5.0-22 April 2013 Yellows8
DSiWare_Exports CTCert verification Just like DSi originally did, 3DS verifies the APCert for DSiWare on SD with the CTCert also in the DSiWare .bin. On DSi this was fixed with with system-version 1.4.2 by verifying with the actual console-unique cert instead(stored in NAND), while on 3DS it's still not(?) fixed.

On 3DS however this is rather useless, due to the entire DSiWare .bin being encrypted with the console-unique movable.sed keyY.

When the movable.sed keyY for the target 3DS is known and the target 3DS CTCert private-key is unknown, importing of modified DSiWare SD .bin files. Unknown, probably none. ? April 2013 Yellows8
Gamecard_Services_PXI unchecked REG_CTRCARDCNT transfer-size The u8 REG_CTRCARDCNT transfer-size parameter for the Gamecard_Services_PXI read/write CTRCARD commands is used as an index for an array of u16 values. Before 5.0.0-X this u8 value wasn't checked, thus out-of-bounds reads could be triggered(which is rather useless in this case). Out-of-bounds read for a value which gets written to a register. 5.0.0-X 2013? Yellows8
PXI cmdbuf buffer overrun The Process9 code responsible PXI communications didn't verify the size of the incoming command before writing it to a C++ member variable. Probably ARM9 code execution 5.0.0-11 March 2015, original timeframe if any unknown plutoo/Yellows8/maybe others(?)
PXIAM command 0x003D0108(See also this) When handling this command, Process9 allocates a 0x2800-byte heap buffer, then copies the 4 FCRAM input buffers to this heap buffer without checking the sizes at all(only the buffers with non-zero sizes are copied). Starting with 5.0.0-X, the total combined size of the input data must be <=0x2800. ARM9 code execution 5.0.0-X May 2013 Yellows8
PS RSA commands buffer overflows pxips9 cmd1(not accessible via ps:ps) and VerifyRsaSha256: unchecked copy to a buffer in Process9's .bss, from the input FCRAM buffer. The buffer is located before the pxi cmdhandler threads' stacks. SignRsaSha256 also has a buf overflow, but this isn't exploitable.

The buffer for this is the buffer for the signature data. With v5.0, the signature buffer was moved to stack, with a check for the signature data size. When the signature data size is too large, Process9 uses svcBreak.

ARM9 code execution 5.0.0-X 2012 Yellows8
PXI pxi_id bad check The Process9 code responsible for PXI communications read pxi_id as a signed char. There were two flaws:
  • They used it as index to a lookup-table without checking the value at all.
  • Another function verified that pxi_id < 7, allowing negative values to pass the check. This would also cause an out-of-range table-lookup.
 Maybe ARM9 code execution 3.0.0-5 March 2015, originally 2012 for the first issue at least plutoo, Yellows8, maybe others(?)

Kernel9

Summary Description Successful exploitation result Fixed in FIRM system version Last FIRM system version this flaw was checked for Timeframe this was discovered Discovered by
CFG_SYSPROT9 bit1 not set by Kernel9 Old versions of Kernel9 never set bit1 of CFG_SYSPROT9. This leaves the 0x10012000-region unprotected (this region should be locked early during boot!). Since it's never locked, you can dump it once you get ARM9 code execution.

From 3.0.0-X this was fixed by setting the bit in Kernel9 after poking some registers in that region. On New3DS arm9loader sets this bit instead of Kernel9, which is exploitable through a hardware + software vulnerability (see arm9loaderhax / description).

This flaw resurged when it gained a new practical use: retrieving the OTP data for a New3DS console in order to decrypt the key data used in arm9loader (see enhanced-arm9loaderhax / description). This was performed by downgrading to a vulnerable system version. By accounting for differences in CTR-NAND crypto (0x05 -> 0x04, see partition encryption types here), it is possible to boot a New3DS using Old3DS firmware 1.0-2.X and an Old3DS NCSD Header to retrieve the required OTP data using this flaw.

Dumping of the OTP area 3.0.0-X February 2015 plutoo, Normmatt independently

ARM11 software

Kernel11

Summary Description Successful exploitation result Fixed in FIRM system version Last FIRM system version this flaw was checked for Timeframe this was discovered Discovered by
SVC table too small The table of function pointers for SVC's only contains entries up to 0x7D, but the biggest allowed SVC for the table is 0x7F. Thus, executing SVC7E or SVC7F would make the SVC-handler read after the buffer, and interpret some ARM instructions as function pointers.

However, this would require patching the kernel .text or modifying SVC-access-control. Even if you could get these to execute, they would still jump to memory that isn't mapped as executable.

None 11.2.0-X 2012 Everyone
svcBackdoor (0x7B) This backdoor allows executing SVC-mode code at the user-specified code-address. This is used by Process9, using this on the ARM11 (with NATIVE_FIRM) required patching the kernel .text or modifying SVC-access-control. See description 11.0.0-X (deleted) Everyone
veryslowpidhax This is completely different from the kernelmode-code-execution vuln described in the below separate entry.

When updating the kernel global PID counter under svcCreateProcess the kernel does not check for wraparound to 0x0(the PID for the very first process). This only matters because SM-module allows processes with PID value less than <total ARM11 FIRM modules> to access all services, without checking exheader service-access-control; and because Kernel11 checks for the PID to be 1 (loader) to use the input mem-region value on ControlMemory. This alone does not affect access the SVCs access table at all.

Inlined ldrex+strex code is used for updating the above counter. 11.2.0-X had changes for similar code, but it was only for dedicated ldrex+strex functions(mainly for kernel objects) and hence this PID code was not affected.

With launching+terminating a sysmodule repeatedly with this via ns:s, it would take weeks to finish(if not at least about a month?).

Access to all services, ControlMemory on any given mem-region. None 11.2.0-X 2012 maybe?
slowhax/waithax svcWaitSynchronizationN does not decrement the references to valid handles in an array before returning an error when it encounters an invalid handle. This allows one to (slowly) overflow the reference count for a handle object to zero. ARM11 kernel-mode code execution 11.2.0-X 11.2.0-X 2016 nedwill, derrek, others?
0xEFF00000 / 0xDFF00000 ARM11 kernel virtual-memory The ARM11 kernel-mode 0xEFF00000/0xDFF00000 virtual-memory(size 0x100000) is mapped to phys-mem 0x1FF00000(entire DSP-mem + entire AXIWRAM), with permissions RW-. This is used during ARM11 kernel startup for loading the FIRM-modules from the FIRM section located in DSP-mem, this never seems to be used after that, however. This is never unmapped either. None 11.2.0-X
memchunkhax2.1 Nintendo's fix for memchunkhax2 in 10.4.0-X did not fix the GPU case: one may cause the requisite ToCToU race using gspwn, bypassing the new validation.

derrek's original 32c3 presentation for memchunkhax2 commented that a GPU-based attack was possible, but would be difficult. However, memchunkhax2.1 showed that it was possible to do fairly reliably.

ARM11 kernel code execution 11.0.0-X, via the new memchunkhdr MAC which prevents modifying memchunkhdr data with DMA. 11.0.0-X derrek, aliaspider
memchunkhax2 When allocating a block of memory, the "next" pointer of the memchunkhdr is accessed without being checked after being mapped to userland.

This allows a race condition, where the process can change the next pointer just before it's accessed. By pointing the next pointer to a crafted memchunckhdr in the kernel SlabHeap, some of the SlabHeap is allocated to the calling process, allowing to change vtables of kernel objects.

ARM11 kernel code execution 10.4.0-X (partially, see memchunkhax2.1) 10.4.0-X derrek
heaphax Can change the size of free memchunk structures stored in FCRAM using DMA, which leads to the ability to allocate memory chunks over already-allocated memory. This can be used in the SYSTEM region to allocate RW memory over any part of the NS system module, which is enough to take it over. Code execution with access to all of NS's privileges. (including downgrading) Code execution within any applet. 11.0.0-X, via the new memchunkhdr MAC which prevents modifying memchunkhdr data with DMA. 11.0.0-X April 2015 ? smea
snshax Can force creation of Safe NS process into gspwn-able memory, allowing for takeover. Code execution with access to all of NS's privileges. (including downgrading) 10.1.0-X 10.1.0-X April 2015 ? smea
AffinityMask/processorid validation With 10.0.0-X the following functions were updated: svcGetThreadAffinityMask, svcGetProcessAffinityMask, svcSetProcessAffinityMask, and svcCreateThread. The code changes for all but svcCreateThread are identical.

The original code with the first 3 did the following:

  • if(u32_processorcount > ~0x80000001)return 0xe0e01bfd;
  • if(s32_processorcount > <total_cores>)return 0xd8e007fd;

The following code replaced the above:

  • if(u32_processorcount >= <total_cores+1>)return 0xd8e007fd;

In theory the latter should catch everything that the former did, so it's unknown if this was really a security issue.

The svcCreateThread changes with 10.0.0-X definitely did fix a security issue.

  • Original code: "if(s32_processorid > <total_cores>)return 0xd8e007fd;"
  • New code: "if(s32_processorid >= <total_cores> || s32_processorid <= -4)return 0xd8e007fd;"

This fixed an off-by-one issue: if one would use processorid=total_cores, which isn't actually a valid value, svcCreateThread would accept that value on <10.0.0-X. This results in data being written out-of-bounds(baseaddr = arrayaddr + entrysize*processorid), which has the following result:

  • Old3DS: Useless kernel-mode crash due to accessing unmapped memory.
  • New3DS: uncontrolled data write into a kernel-mode L1 MMU-table. This isn't really useful: the data can't be controlled, and the data which gets overwritten is all-zero anyway(this isn't anywhere near MMU L1 entries for actually mapped memory).

The previous version also allowed large negative s32_processorid values(negative processorid values are special values not actual procids), but it appears using values like that won't actually do anything(meaning no crash) besides the thread not running / thread not running for a while(besides triggering a kernelpanic with certain s32_processorid value(s)).

Nothing useful 10.0.0-X 10.0.0-X svcCreateThread issue: May 31, 2015. The rest: September 8, 2015, via v9.6->v10.0 ARM11-kernel code-diff. Yellows8
memchunkhax The kernel originally did not validate the data stored in the FCRAM kernel heap memchunk-headers for free-memory at all. Exploiting this requires raw R/W access to these memchunk-headers, like physical-memory access with gspwn.

There are multiple ways to exploit this, but the end-result for most of these is the same: overwrite code in AXIWRAM via the 0xEFF00000/0xDFF00000 kernel virtual-memory mapping.

This was fixed in 9.3.0-X by checking that the memchunk(including size, next, and prev ptrs) is located within the currently used heap memory. The kernel may also check that the next/prev ptrs are valid compared to other memchunk-headers basically. When any of these checks fail, kernelpanic() is called.

When combined with other flaws: ARM11-kernelmode code execution 9.3.0-21 February 2014 Yellows8
Multiple KLinkedListNode SlabHeap use after free bugs The ARM11-kernel did access the 'key' field of KLinkedListNode objects, which are located on the SlabHeap, after freeing them. Thus, triggering an allocation of a new KLinkedListNode object at the right time could result in a type-confusion. Pseudo-code:

SlabHeap_free(KLinkedListNode); KObject *obj = KLinkedListNode->key; // the object there might have changed! This bug appeared all over the place.

ARM11-kernelmode code exec maybe 8.0.0-18 April 2015 derrek
PXI Command input/output buffer permissions Originally the ARM11-kernel didn't check permissions for PXI input/output buffers for commands. Starting with 6.0.0 PXI input/output buffers must have RW permissions, otherwise kernelpanic is triggered. 6.0.0-11 2012 Yellows8
svcStartInterProcessDma For svcStartInterProcessDma, the kernel code had the following flaws:
  • Originally the ARM11-kernel read the input DmaConfig structure directly in kernel-mode(ldr(b/h) instructions), without checking whether the DmaConfig address is readable under userland. This was fixed by copying that structure to the SVC-mode stack, using the ldrbt instruction.
  • Integer overflows for srcaddr+size and dstaddr+size are now checked(with 6.0.0-11), which were not checked before.
  • The kernel now also checks whether the srcaddr/dstaddr (+size) is within userland memory (0x20000000), the kernel now (with 6.0.0-11) returns an error when the address is beyond userland memory. Using an address >=0x20000000 would result in the kernel reading from the process L1 MMU table, beyond the memory allocated for that MMU table(for vaddr->physaddr conversion).
 6.0.0-11 DmaConfig issue: unknown. The rest: 2014 plutoo, Yellows8 independently
svcControlMemory Parameter checks For svcControlMemory the parameter check had these two flaws:
  • The allowed range for addr0, addr1, size parameters depends on which MemoryOperation is being specified. The limitation for GSP heap was only checked if op=(u32)0x10003. By setting a random bit in op that has no meaning (like bit17?), op would instead be (u32)0x30003, and the range-check would be less strict and not accurate. However, the kernel doesn't actually use the input address for LINEAR memory-mapping at all besides the range-checks, so this isn't actually useful. This was fixed in the kernel by just checking for the LINEAR bit, instead of comparing the entire MemoryOperation value with 0x10003.
  • Integer overflows on (addr0+size) are now checked that previously weren't (this also applies to most other address checks elsewhere in the kernel).
 5.0.0-11 plutoo
Command request/response buffer overflow Originally the kernel did not check the word-values from the command-header. Starting with 5.0.0-11, the kernel will trigger a kernelpanic() when the total word-size of the entire command(including the cmd-header) is larger than 0x40-words (0x100-bytes). This allows overwriting threadlocalstorage+0x180 in the destination thread. However, since the data written there would be translate parameters (such as header-words + buffer addresses), exploiting this would likely be very difficult, if possible at all.

If the two words at threadlocalstorage+0x180 could be overwritten with controlled data this way, one could then use a command with a buffer-header of ((size<<14) | 2) to write arbitrary memory to any RW userland memory in the destination process.

5.0.0-11 v4.1 FIRM -> v5.0 code diff Yellows8
SVC stack allocation overflows
  • Syscalls that allocate a variable-length array on stack, only checked bit31 before multiplying by 4/16 (when calculating how much memory to allocate). If a large integer was passed as input to one of these syscalls, an integer overflow would occur, and too little memory would have been allocated on stack resulting in a buffer overrun.
  • The alignment (size+7)&~7 calculation before allocation was not checked for integer overflow.

This might allow for ARM11 kernel code-execution.

(Applies to svcSetResourceLimitValues, svcGetThreadList, svcGetProcessList, svcReplyAndReceive, svcWaitSynchronizationN.)

5.0.0-11 v4.1 FIRM -> v5.0 code diff plutoo, Yellows8 complementary
svcControlMemory MemoryOperation MAP memory-permissions svcControlMemory with MemoryOperation=MAP allows mapping the already-mapped process virtual-mem at addr1, to addr0. The lowest address permitted for addr1 is 0x00100000. Originally the ARM11 kernel didn't check memory permissions for addr1. Therefore .text as addr1 could be mapped elsewhere as RW- memory, which allowed ARM11 userland code-execution. 4.1.0-8 2012 Yellows8
Command input/output buffer permissions Originally the ARM11 kernel didn't check memory permissions for the input/output buffers for commands. Starting with 4.0.0-7 the ARM11 kernel will trigger a kernelpanic() if the input/output buffers don't have the required memory permissions. For example, this allowed a FSUSER file-read to .text, which therefore allowed ARM11-userland code execution. 4.0.0-7 2012 Yellows8
svcReadProcessMemory/svcWriteProcessMemory memory permissions Originally the kernel only checked the first page(0x1000-bytes) of the src/dst buffers, for svcReadProcessMemory and svcWriteProcessMemory. There is no known retail processes which have access to these SVCs. 4.0.0-7 2012? Yellows8

FIRM Sysmodules

Summary Description Successful exploitation result Fixed in FIRM system version Last FIRM system version this flaw was checked for Timeframe this was discovered Discovered by
"srv:pm" process registration Originally any process had access to the port "srv:pm". The PID's used for the (un)registration commands are not checked either. This allowed any process to re-register itself with "srv:pm", and therefore allowed the process to give itself access to any service, bypassing the exheader service-access-control list.

This was fixed in 7.0.0-13: starting with 7.0.0-13 "srv:pm" is now a service instead of a globally accessible port. Only processes with PID's less than 6 (in other words: fs, ldr, sm, pm, pxi modules) have access to it. With 7.0.0-13 there can only be one session for "srv:pm" open at a time(this is used by pm module), svcBreak will be executed if more sessions are opened by the processes which can access this.

This flaw was needed for exploiting the <=v4.x Process9 PXI vulnerabilities from ARM11 userland ROP, since most applications don't have access to those service(s).

Access to arbitrary services 7.0.0-13 2012 Yellows8
FSDIR null-deref FS-module may crash in some cases when handling directory reading. The trigger seems to be due to using FSDir:Close without closing the dir-handle afterwards?(Perhaps this is caused by out-of-memory?) This seems to be useless since it's just a null-deref. None 9.6.0-X May 19(?)-20, 2015 Yellows8

Standalone Sysmodules

Summary Description Successful exploitation result Fixed in system-module system-version Last system-module system-version this flaw was checked for Timeframe this was discovered Timeframe this was added to wiki Discovered by
AM stack/.bss infoleak via AM:ReadTwlBackupInfo(Ex) After writing the output-info structure to stack, it then copies that structure to the output buffer ptr using the size from the command. The size is not checked. This could be used to read data from the AM-service-thread stack handling the command + .bss.

This was not tested on hardware.

Stack/.bss reading None 10.0.0-27(AM v9217) Roughly October 17, 2016 October 25, 2016 Yellows8
MVD: Stack buffer overflow with MVDSTD:SetupOutputBuffers. The input total_entries is not validated when initially processing the input entry-list. This fixed-size input entry-list is copied to stack from the command request. The loop for processing this initializes a global table, the converted linearmem->physaddrs used there are also copied to stack(0x8-bytes of physaddrs per entry).

If total_entries is too large, MVD-sysmodule will crash due to reading unmapped memory following the stack(0x10000000). Afterwards if the out-of-bounds total_entries is smaller than that, it will crash due accessing address 0x0, hence this useless.

MVD-sysmodule crash. None 9.0.0-20 April 22, 2016 (Tested on the 25th) April 25, 2016 Yellows8
NWM: Using CTRSDK heap with UDS sharedmem from the user-process. See the HTTP-sysmodule section below.

CTRSDK heap is used with the sharedmem from NWMUDS:InitializeWithVersion. Buffers are allocated/freed under this heap using NWMUDS:Bind and NWMUDS:Unbind.

Hence, overwriting sharedmem with gspwn then using NWMUDS:Unbind results in the usual controlled CTRSDK memchunk-header write, similar to HTTP-sysmodule.

This could be done by creating an UDS network, without any other nodes on the network.

Besides CTRSDK memchunk-headers, there are no addresses stored under this sharedmem.

ROP under NWM-module. None 9.0.0-X April 10, 2016 April 16, 2016 Yellows8
DLP: Out-of-bounds memory access during spectator data-frame checksum calculation DLP doesn't validate the frame_size when receiving spectator data-frames at all, unlike non-spectator data-frames. The actual spectator data-frame parsing code doesn't use that field either. However, the data-frame checksum calculation code called during checksum verification does use the frame_size for loading the size of the framebuf.

Hence, using a large frame_size like 0xFFFF will result in the checksum calculation code reading data out-of-bounds. This isn't really useful, you could trigger a remote local-WLAN DLP-sysmodule crash while a 3DS system is scanning for DLP networks(due to accessing unmapped memory), but that's about all(trying to infoleak with this likely isn't useful either).

DLP-sysmodule crash, handled by dlplay system-application by a "connection interrupted" error eventually then a fatal-error via ErrDisp. None 10.0.0-X April 8, 2016 (Tested on the 10th) April 10, 2016 Yellows8
DLP: Out-of-bounds output data writing during spectator sysupdate titlelist data-frame handling The total_entries and out_entryindex fields for the titlelist DLP spectator data-frames are not validated. This is parsed during DLP network scanning. Hence, the specified titlelist data can be written out-of-bounds using the specified out_entryindex and total_entries. A crash will occur while reading the input data-frame titlelist if total_entries is larger than 0x27A, due to accessing unmapped memory.

There's not much non-zero data to overwrite following the output buffer(located in sharedmem), any ptrs are located in sharedmem. Overwriting certain ptr(s) are only known to cause a crash when attempting to use the DLP-client shutdown service-command.

There's no known way to exploit the above crash, since the linked-list code involves writes zeros(with a controlled start ptr).

None 10.0.0-X April 8-9, 2016 April 10, 2016 Yellows8
IR: Stack buffer overflow with custom hardware Originally IR sysmodule used the read value from the I2C-IR registers TXLVL and RXLVL without validating them at all. See here for the fix. This is the size used for reading the data-recv FIFO, etc. The output buffer for reading is located on the stack.

This should be exploitable if one could successfully setup the custom hardware for this and if the entire intended sizes actually get read from I2C.

ROP under IR sysmodule. 10.6.0-31 February 23, 2016 (Unknown if it was noticed before then) February 23, 2016 Yellows8
HTTP: Using CTRSDK heap with sharedmem from the user-process. The data from httpcAddPostDataAscii and other commands is stored under a CTRSDK heap. That heap is the sharedmem specified by the user-process via the HTTPC Initialize command.

Normally this sharedmem isn't accessible to the user-process once the sysmodule maps it, hence using it is supposed to be "safe".

This isn't the case due to gspwn however. Since CTRSDK heap code is so insecure in general, one can use gspwn to locate the HTTPC sharedmem + read/write it, then trigger a mem-write under the sysmodule. This can then be used to get ROP going under HTTP-sysmodule.

This is exploited by ctr-httpwn.

ROP under HTTP sysmdule. None 9.6.0-X (Latest sysmodule version as of 10.7.0-32) Late 2015 March 22, 2016 Yellows8
NIM: Downloading old title-versions from eShop Multiple NIM service commands(such as NIMS:StartDownload) use a title-version value specified by the user-process, NIM does not validate that this input version matches the latest version available via SOAP. Therefore, when combined with AM(PXI) title-downgrading via deleting the target eShop title with System Settings Data Management(if the title was already installed), this allows downloading+installing any title-version from eShop if it's still available from CDN.

The easiest way to exploit this is to just patch the eShop system-application code using these NIM commands(ideally the code which loads the title-version).

Originally this was tested with a debugging-system via modded-FIRM, eventually smea implemented it in HANS for the 32c3 release.

Downloading old title-versions from eShop None 10.0.0-X October 24, 2015 (Unknown when exactly the first eShop title downgrade was actually tested, maybe November) January 7, 2016 (Same day Ironfall v1.0 was removed from CDN via the main-CXI files) Yellows8
SPI service out-of-bounds write cmd1 has out-of-bounds write allowing overwrite of some static variables in .data. None 9.5.0-22 March 2015 plutoo
NFC module service command buf-overflows NFC module copies data with certain commands, from command input buffers to stack without checking the size. These commands include the following, it's unknown if there's more commands with similar issues: "nfc:dev" <0x000C....> and "nfc:s" <0x0037....>.

Since both of these commands are stubbed in the Old3DS NFC module from the very first version(those just return an error), these issues only affect the New3DS NFC module.

There's no known retail titles which have access to either of these services.

ROP under NFC module. New3DS: None New3DS: 9.5.0-22 December 2014? Yellows8
NEWSS service command notificationID validation failure This module does not validate the input notificationID for "news:s" service commands. This is an out-of-bounds array index bug. For example, NEWSS:SetNotificationHeader could be used to exploit news module: this copies the input data(size is properly checked) to: out = newsdb_savedata+0x10 + (someu32array[notificationID]*0x70). ROP under news module. None 9.7.0-X December 2014 Yellows8
NWMUDS:DecryptBeaconData heap buffer overflow input_size = 0x1E * <value the u8 from input_networkstruct+0x1D>. Then input_tag0 is copied to a heap buffer. When input_size is larger than 0xFA-bytes, it will then copy input_tag1 to <end_address_of_previous_outbuf>, with size=input_size-0xFA.

This can be triggered by either using this command directly, or by boadcasting a wifi beacon which triggers it while a 3DS system running the target process is in range, when the process is scanning for hosts to connect to. Processes will only pass tag data to this command when the wlancommID and other thing(s) match the values for the process.

There's no known way to actually exploit this for getting ROP under NWM-module, at the time of originally adding this to the wiki. This is because the data which gets copied out-of-bounds *and* actually causes crash(es), can't be controlled it seems(with just broadcasting a beacon at least). It's unknown whether this could be exploited from just using NWMUDS service-cmd(s) directly.

Without any actual way to exploit this: NWM-module DoS, resulting in process termination(process crash). This breaks *everything* involving wifi comms, a reboot is required to recover from this. None 9.0.0-20 ~September 23, 2014(see the NWMUDS:DecryptBeaconData page history) August 3, 2015 Yellows8
HID module shared-mem HID module does not validate the index values in sharedmem(just changes index to 0 when index == maxval when updating), therefore large values will result in HID module writing HID data to arbitrary addresses. ROP under HID module, but this is *very* unlikely to be exploitable since the data written is HID data. None 9.3.0-21 2014? Yellows8
gspwn GSP module does not validate addresses given to the GPU. This allows a user-mode application/applet to read/write to a large part of physical FCRAM using GPU DMA. From this, you can overwrite the .text segment of the application you're running under, and gain real code-execution from a ROP-chain. Normally applets' .text(Home Menu, Internet Browser, etc) is located beyond the area accessible by the GPU, except for CROs used by applets(Internet Browser for example).

FCRAM is gpu-accessible up to physaddr 0x26800000 on Old3DS, and 0x2DC00000 on New3DS. This is BASE_memregion_start(aka SYSTEM_memregion_end)-0x400000 with the default memory-layout on Old3DS/New3DS.

User-mode code execution. None 9.6.0-X Early 2014 smea, Yellows8/others before then
rohax Using gspwn, it is possible to overwrite a loaded CRO0/CRR0 after its RSA-signature has been validated. Badly validated CRO0 header leads to arbitrary read/write of memory in the ro-process. This gives code-execution in the ro module, who has access to syscalls 0x70-0x72, 0x7D.

This was fixed after ninjhax release by adding checks on CRO0-based pointers before writing to them.

Memory-mapping syscalls. 9.3.0-21 9.4.0-21 smea, plutoo joint effort
Region free Only Home Menu itself checks gamecards' region when launching them. Therefore, any application launch that is done directly with NS without signaling Home Menu to launch the app, will result in region checks being bypassed.

This essentially means launching the gamecard with the "ns:s" service. The main way to exploit this is to trigger a FIRM launch with an application specified, either with a normal FIRM launch or a hardware reboot.

Launching gamecards from any region + bypassing Home Menu gamecard-sysupdate installation None Last tested with 10.1.0-X. June(?) 2014 Yellows8
NWM service-cmd state null-ptr deref The NWMUDS service command code loads a ptr from .data, adds an offset to that, then passes that as the state address for the actual command-handler function. The value of the ptr loaded from .data is not checked, therefore this will cause crashes due to that being 0x0 when NWMUDS was not properly initialized.

It's unknown whether any NWM services besides NWMUDS have this issue.

This is rather useless since it's only a crash caused by a state ptr based at 0x0. None 9.0.0-20 2013? Yellows8

General/CTRSDK

Summary Description Successful exploitation result Fixed in version Last version this flaw was checked for Timeframe this was discovered Discovered by
UDS beacon additional-data buffer overflow Originally CTRSDK did not validate the UDS additional-data size before using that size to copy the additional-data to a networkstruct. This was eventually fixed.

This was discovered while doing code RE with an old dlp-module version. It's unknown in what specific CTRSDK version this was fixed, or even what system-version updated titles with a fixed version.

It's unknown if there's any titles using a vulnerable CTRSDK version which are also exploitable with this(dlp module can't be exploited with this).

The maximum number of bytes that can be written beyond the end of the outbuf is 0x37-bytes, with additionaldata_size=0xFF.

Perhaps ROP, very difficult if possible with anything at all ? September(?) 2014 Yellows8
CTPK buffer overflow At offset 0x20 in CTPK is an array for each texture, each entry is 0x20-bytes. This contains a wordindex(entry+0x18) for some srcdata relative to CTPK+0, and an u8 wordsize(entry+0x14) for this data. The CTRSDK function handling this doesn't validate the size, when copying srcdata using this size to the output buffer. Applications usually have the output buffer on the stack, hence stack buffer overflow.

While CTPK(*.ctpk) are normally only loaded from RomFS, some application(s) load from elsewhere too.

ROP under the target application. None? "[SDK+NINTENDO:CTR_SDK-11_4_0_200_none]" November 14, 2016 Yellows8
Retrieved from "https://www.3dbrew.org/w/index.php?title=3DS_System_Flaws&oldid=19080"