Add a docs/ tree - flash map, boot chain, root filesystem, GPON/OMCI, recovery and artifact inventory - plus tools/ that carve a dump by the published offsets, verify every published partition hash, and refuse to publish unit identity. Every figure was re-derived from the 16 MB dump rather than carried over. Corrections against the artifacts: the 512-byte vendor header is little-endian u32 throughout (0x008 is a 2022-11-12 build stamp, and 0x014/0x018 are the rootfs extent and offset); the U-Boot and kernel image headers both use ih_comp=3; the earlier ih_type values were being read as a compression byte; the B bank's trailing fill starts at 0x521000, not 0x710000; the Config partition's JFFS2 inventory is counted rather than described; the rootfs mkfs time is 13:53:01Z while 13:52:58Z is /version's build time; /www holds 169 files; S90mknode's Realtek node lines are commented out and S03egis cannot load the egis.ko it names, since only egisdev.ko ships. Unreproducible teardown claims are recorded under "Reported, not verified" instead of being asserted, and the recovery procedure now orders the ImageB write so the bank being written is never the one being booted. |
||
|---|---|---|
| docs | ||
| tools | ||
| .gitignore | ||
| README.md | ||
Adtran SDX 611 — firmware and flash analysis
Analysis of the Adtran SDX 611 (Adtran-611-V2 REV1.0) GPON ONT from a full-chip
read of its 16 MB SPI NOR (Winbond W25Q128JVSQ), taken with a CH341A and a SOIC-8
clip with the ONT unpowered. What makes a read trustworthy is the double-read
comparison in docs/recovery.md; the programmer's own verify pass
is reported but unverifiable, because no log was kept — see Reported, not
re-verified below.
Firmware analysed: GRO100_ADTRAN611V2 V2.2.9.11 (V1.0.0(V2.2.9.11)), rootfs
built 2022-11-12 13:52:58 UTC.
The dump itself is not published. Every figure below was re-derived from it, and
the hashes in docs/artifacts.md identify the dump and its
partitions without carrying any unit identity. The scripts in tools/ reproduce
every slice and re-check every published partition hash.
Hardware
| Field | Value |
|---|---|
| Model | Adtran SDX 611, board Adtran-611-V2 REV1.0 |
| Chipset / SoC | Realtek RTL9601D (Realtek "Luna" SDK) |
| Flash | 16 MB SPI NOR (Winbond W25Q128JVSQ) |
| Bootloader | U-Boot 2011.12.NA-g61AC973 (Mar 10 2022 - 19:23:10) |
| Kernel | 3.18.24 (uImage, LZMA, load address 0x80000000) |
| Root filesystem | squashfs 4.0, XZ compressed, 128 KiB blocks, 988 inodes |
| Userland | uClibc 0.9.33 (NPTL) + BusyBox 1.34.1 |
| Optics | SC/APC (U&T UTG24P05) — see Reported, not re-verified |
Firmware identity
Read out of /version in the rootfs and the 512-byte vendor header at the head of
each image bank:
| Field | Value |
|---|---|
| Project | GRO100_ADTRAN611V2 |
| Version | V2.2.9.11 |
| ONU version string | V1.0.0(V2.2.9.11) |
| Vendor id | ADTRAN |
| Factory version | ADTN-R01B010De25f3ad0 |
| Build user | mayinrong |
| GIT revision | e25f3ad0 (openreach_rtl9601d_gpon_20210805) |
| Build time | 2022-11-12 21:52:58 +0800 (13:52:58 UTC) |
| squashfs mkfs time | 2022-11-12 13:52:58 UTC |
Key findings
- The flash layout is not in the environment.
mtdpartsis a compiled-in default of the U-Boot uImage at0x15F40; it was recovered from the 186,808-byte decompressed stage, where the string sits at0x27E9C, and the stored environment has nomtdpartsat all. Seedocs/flash-map.md. - ImageB is a byte-identical copy of ImageA, not a second firmware: both banks hash to the same value, so the second is a recovery copy rather than an A/B pair.
- The device boots to a root shell on serial, and that shell runs the init
scripts.
/etc/inittabrespawns-/bin/sh;rcSonly doesmount -a; theS*scripts are launched by thefor i in /etc/init.d/S??*loop in/etc/profile, guarded by/var/telnetflag. Seedocs/boot-chain.md. - Telnet is disabled in two independent places:
rcShastelnetdcommented out, andS50telnetdis a 161-byte script whosestart)branch is empty. BusyBox still carries thetelnetdapplet. Seedocs/rootfs.md. - Unit identity is not in the environment.
ethaddris the Realtek placeholder00:11:22:33:44:66; the GPON serial and MAC live inSECTION_EGIS(tag at0xFE0024) and are mirrored intoomcimib.confon the Config partition. Seedocs/gpon-omci.md. - The image ships a fixed Dropbear RSA/DSS key pair at
/etc/default/, butS95startsspnever installs it: it copies a per-unit pair from/var/configand generates one if absent. This unit's keys differ from the shipped pair. - The Config partition is JFFS2 with 4 KiB erase blocks, holding the OMCI MIB
store and the per-unit SSH keys;
S02mountconfigerases and remounts it when the mount fails.
Repository layout
| Path | Contents |
|---|---|
docs/flash-map.md |
Partition table, recovered mtdparts, tiling arithmetic, image-bank internals |
docs/boot-chain.md |
Preloader → U-Boot → tw_boot → bank select → kernel, plus the environment dump |
docs/rootfs.md |
squashfs parameters, init topology, SSH keys, port maps, userland |
docs/gpon-omci.md |
GPON/OMCI library inventory and the omcimib.conf record structure |
docs/recovery.md |
Re-reading and re-writing the chip, serial console commands, brick recovery |
docs/artifacts.md |
Hash inventory, extraction methods, and what is deliberately not published |
tools/ |
Carve a dump by the published offsets, verify hashes, run the redaction check |
Reproducing this
The published offsets and hashes are checkable against any 16 MB dump of this build:
python tools/carve_partitions.py --dump SDX611.bin --out parts/
sh tools/verify_hashes.sh --dump SDX611.bin
SDX611.bin stands for wherever the dump is kept. On the reference dump both
commands exit 0 and print all twelve slices ok; parts/ is gitignored. On another
unit, run sh tools/verify_hashes.sh --dump other-unit.bin --ignore-unit-parts, which
treats Env, Config and SECTION_EGIS as expected to differ.
carve_partitions.py slices Boot/Env/ImageA/ImageB/Config/SECTION_EGIS plus the
kernel and rootfs payloads and compares each slice against
tools/expected.sha256, which was generated from the dump described in
docs/artifacts.md. A dump from a different unit matches on every partition
except Env, Config and SECTION_EGIS, which carry unit state.
Extraction used no vendor tooling: the LZMA uImages and the squashfs were unpacked
with the Python lzma module and a hand-written squashfs 4.0 reader, the JFFS2
partitions with a node walker. docs/artifacts.md records which method produced
which figure.
Redaction policy
The raw 16 MB dump is withheld because it carries one subscriber's line identity:
- the GPON serial (8 bytes at
0xFE0028) and the unit MAC (6 bytes at0xFE0030) inSECTION_EGIS, mirrored inomcimib.conf; - the
omcimib.confMIB store, which is ONT provisioning state; - the per-unit Dropbear host keys on the Config partition.
This repository publishes field layout, offsets and partition hashes only. Hashes of the whole dump and of individual partitions are published because a hash of a withheld file does not disclose it.
tools/redact_check.py makes that mechanical: it derives the identity fields from
a dump at run time and refuses if any of them appear in the tree in raw, hex or
MAC-string form. It is run over the staged tree before every commit.
Related
- hack-gpon.org device page: https://hack-gpon.org/ont-adtran-sdx-611/
- PacketForge device entry:
devices/adtran/sdx611.json
Reported, not re-verified
Claims carried over from the teardown that this analysis cannot re-derive from the flash image. They are kept here instead of being presented as measured:
- Optics are SC/APC with a U&T UTG24P05 module. The PacketForge device entry for the same unit says LC/UPC; the two records disagree and neither is a firmware fact.
- The CH341A programmer reported a verify pass on the read. No programmer log was
retained, so that claim cannot be re-checked; only the dump itself is re-verifiable,
and
docs/recovery.mdmakes the double-read comparison the check that matters. - The
mpU-Boot command does not exist in this build: there is no such entry in the command table, and the only literalmpbytes in the stage sit inside MIPS instructions.docs/recovery.mdtherefore lists only commands that are present —xf [offset](help textxmodem to flash),loadb,loady,sfand theupimgtar/upvmimgupdate commands. - The kernel command line template has three further
%sslots betweenACTIVE=%candchiptype=RTL9601D; whattw_bootsubstitutes into them was not recovered. - The
omcimib.confpayload never matched a documented MIB encoding; only the 128-byte header and the identity block are described here.
Licence
No licence is granted. This is a reverse-engineering write-up, not a code release; add a licence before reusing any of it.