Skip to content

v1.17

Choose a tag to compare

@github-actions github-actions released this 31 Oct 23:03
· 20 commits to master since this release

Based on upstream DFRoot 4.6 - 4.7, merged into this fork. versionCode 34, versionName 1.17.

D2 - the download-mode authorisation byte, read and written from the phone

Screenshot_20261101_000536(1)
  • The DL/Recovery row is now read from the vault instead of assumed. The record is
    [lock, maintenance, at_command], and the row says D2 Blocked on exactly [1, 0, 0] - the one
    triple the bootloader refuses Odin on - and Allowed otherwise. Unknown until the vault answers,
    Not gated here where the firmware never gated download mode.
  • Byte 2 can be authorised or cleared from the app, behind Expert Mode. Every call is one
    read-modify-verify: read the DMC vault, refuse a record whose first three bytes are not each 0 or 1,
    change only byte 2, write it back, re-read it, and report written only when it stuck.
    already-authorized, write-failed, verify-failed and unsupported stay separate outcomes, and
    nothing retries a blind second time.
  • The writer is a one-shot process, not a daemon: a 5.5 kB launcher execves app_process64 with this
    app's own installed base.apk on the classpath, and the harness drops itself to uid 1000. Nothing is
    copied to disk, nothing is injected into system_server, no sepolicy rule is added, and nothing stays
    resident - ps -A after a run holds none of it. Measured on the SM-S931B. Not watched end to end:
    booting patched firmware into Odin and seeing it accepted. The full write-up, including that gap, is
    docs/d2-download-mode-writer.md.

Autorun Writer honours its word while the app is running

  • A reboot clears the vendor byte, and DirtyFrag's root is a page-cache patch, so with the switch on and
    Expert Mode armed the app re-reads the vault and writes the byte again at each reroot - once per root,
    because a refused su is a state and not a retry cue. The boot path stays, but on this vendor it
    cannot be the only one.
  • A write the switch makes on its own is reported in the run log, not only in logcat.
  • The recovery reading now follows root: the card repaints when su appears or leaves the path, instead
    of going on showing the answer it computed before a reboot.

UI

  • The app now carries no androidx at all - Total dependency-less rewrite in Java

Engine - upstream 4.6 - 4.7

  • The instruction patch stopped guessing prologues: it writes BTI C, MOV x0, #0, RET at the
    function entry instead of decoding preambles and picking retaa/retab per SoC.
  • On kernels built without CONFIG_SECURITY_SELINUX_DEVELOP, the module reads selinux_state back and
    patches avc_has_perm and avc_has_perm_noaudit to return 0 (upstream #172).
  • A Log buffer size of Off in Developer Options is now reported up front instead of turning up later as
    an unexplained timeout (upstream #173).

Fixed - the Android version that picks the LKM (issue #73)

  • The kernel's GKI level picks the module, read from uname -r and ro.vendor.build.version.release
    rather than the marketing version - which on an android15-6.6 phone reports 17. Where a level has
    no exact row the nearest one is taken, ties broken downward, and the run log says which module was
    guessed. Kernels with no row are still refused, not guessed at.

Fixed - opening a saved log crashed below Android 14

  • readLastLog() used Matcher overloads that only exist from API 34 while minSdk is 32, so Android
    12L and 13 threw NoSuchMethodError as soon as a log file was read. The overloads that have existed
    since the first API level do the same job.