Skip to content

soc: apple: sep: remove the xart_writes and provision_keybag parameters (SEP writes always on) - #2591

Merged
DjDeveloperr merged 7 commits into
integration/aurorafrom
djdev/sep-writes-always-on
Nov 20, 2026
Merged

DjDeveloperr merged 7 commits into
integration/aurorafrom
djdev/sep-writes-always-on

Conversation

@DjDeveloperr

@DjDeveloperr DjDeveloperr commented Nov 20, 2026 •

Copy link
Copy Markdown
Contributor

What

This removes the apple_sep module parameters that turn SEP writes off: xart_writes and provision_keybag. SEP writes (xART writes and identity-keybag provisioning) are now always on, on every Mac and in every distribution. There is no read-only mode left to select.

Owner directive (Dj, 2026-11-20): "No machine or distro ever should have sep writes off. that shit should be removed ENTIRELY."

Evidence: the write path works on the T6050

The J714s Touch ID lab boots of 2026-11-11 (#1587, merged as 554c6d1) ran with xart_writes=1 provision_keybag=1 and the xART guard at level 1.

  • The identity keybag and the machine ref-key were provisioned.
  • The enrolment survived a reboot.
  • The enrolled finger matched, and a wrong finger was rejected.
  • The guard refused 0 writes.

The earlier SoCs ran with writes on by default before this change.

Driver changes (drivers/soc/apple)

  • The parameters are deleted, not just defaulted.
    • The store always opens for writing.
    • The bring-up always provisions a missing identity keybag. An existing one returns before any write (sks_provision_identity_keybag).
  • The EINVAL goes too. "provision_keybag=1 requires xart_writes=1" was checked in probe, in complete_attach and in the bring-up; all three are removed.
  • A stale value no longer stops the load. apple_sep.xart_writes=… or apple_sep.provision_keybag=… on the kernel command line, or in any modprobe.d file, used to fail the load when the value was given twice: both were Rust SetOnce parameters, so the second value returned EEXIST and load_module failed. Now the module loader logs apple_sep: unknown parameter 'xart_writes' ignored and the driver loads. That is unknown_module_param_cb in kernel/module/main.c.
  • diag/policy keeps its keys for ABI 3 readers. xart_writes is still the effective writability, and provision_keybag always reads 1.
  • The Touch ID refusal messages now name the read-only store, not a parameter.
  • Docs. The Kconfig help and Documentation/ABI no longer describe a read-only install.

Kept, because it doesn't turn writes off

  • The fail-closed checks. APFS must prove a single unsnapshotted .gl extent, and the store and roots must validate; otherwise the probe fails, as before.
  • The key-store-only initrd phase (keystore_only=1/2). xART stays read-only there until complete_attach reopens it writable. With keystore_only=2 and no guard it is writable in the initrd, as before. initrd_xart_writes() loses its parameter input.
  • The read-only fallback after a failed writable reopen in complete_attach. It still returns the error.
  • The xART guard (xart_guard, T6050 level 1). It refuses rewrites of session records that this Linux install did not create, which protects macOS's records. It is not a writes-off switch.
  • xart_start_sector, trusted_keys, filevault_api and the trace and test parameters.

Behaviour changes to know about

  • J613 26.6.2 lane. A writable attach at probe already required a DT identity and the xART guard. Every normal attach is now writable, so those preconditions now apply to every normal attach. A J613 26.6.2 boot without /chosen/apfs-preboot-uuid, with an explicit UUID override, or with the guard off now fails the probe; before, it could attach read-only with xart_writes=0. The keystore_only=1 path is unchanged and still skips the check: probe opens read-only, and complete_attach reopens writable without it, exactly as with the old defaults. The ABI text that suggested provision_keybag=0 for an existing bag is reworded: an existing bag is kept without a write anyway.
  • Owners who set xart_writes=0. An owner who followed the old docs and put options apple_sep xart_writes=0 provision_keybag=0 in /etc/modprobe.d/ now gets a writable attach, plus the kernel's "unknown parameter ... ignored" lines. That is the point of this change.

tools/aurora-sep (the aurora-sep source the distributions vendor)

  • load-driver: runs modprobe apple_sep with no options. Passing xart_writes=1 provision_keybag=1 again next to a stale command-line value gave each parameter twice, and kernels before this change refused to load.
  • install-aurora-sep.sh --read-only is removed. It wrote options apple_sep xart_writes=0 provision_keybag=0 to /etc/modprobe.d/aurora-sep.conf, and re-runs kept it.
    • --read-only now exits with "--read-only was removed: SEP writes are always on".
    • An earlier --read-only file is deleted on install and on --reset-touchid. It is recognised by its header and its exact contents; any other file is kept, with a warning that kernels from this change on ignore it.
    • The safety notice and the agent prompt no longer offer a read-only install.
    • An owner's own blacklist/install apple_sep line still keeps the driver off: that is the admin not loading the driver at all, not a write switch.
  • aurora-sep status and aurora-sep-ready: the messages name the read-only store (key-store-only phase, or a failed reopen), not a parameter.
  • Docs and packaging: README, XART-SAFETY, M1-SUPPORT and DESKTOP-FIXES lose their read-only text. The packaging release numbers are bumped.
  • Tests:
    • tests/test_sep_writes_always_on.py (15 tests) covers the refusal, the cleanup, the kept-file warning, the owner's block, reset-touchid, load-driver, and that no packaged modprobe.d file sets a write option.
    • .github/workflows/sep-source.yml runs it.
    • Seven installer tests drop the option and keep their intent.

Review follow-ups

Second round:

  • An owner's blacklist/install apple_sep line in /etc/modprobe.d/aurora-sep.conf counts again. The installer no longer writes that file, so whatever is in it is the owner's.
  • --uninstall removes only the file an earlier --read-only install wrote.
  • --reset-touchid refuses when the owner's own file there sets xart_writes=0 or provision_keybag=0 and the installed kernel still has those parameters. Tests use the real sep_block_line, pointed at the test's modprobe.d.
  • The ABI diag/xart text is corrected: complete_attach updates it.
  • The test follow-up is folded into the driver commit, so every commit builds and passes.

First round:

complete_attach falls back to a read-only store when the writable reopen fails. diag/xart now reads disabled there. Before, with keystore_only=2 and no guard, the store was writable at probe and the flag stayed enabled.

Tests

  • Build (rebuilt after the review follow-up). apple-sep.ko was built as an external module against a 7.3 output tree whose SEP sources are identical to integration/aurora acfcd08 (same rustc 1.99.0 and clang 23.1.1). It builds. The 5 dead-code warnings are already there in key.rs and control.rs. modinfo -F parm no longer lists xart_writes or provision_keybag.
  • The tools/aurora-sep driver-source tests. test_diag_policy now follows the fixed provision_keybag=1. test_diag_sensor, test_enrol_admission, test_host_entropy, test_identity_preservation, test_log_ratelimit, test_orphan_delete, test_policy_window, test_aurora_sep_status and test_sep_ready pass.
    • test_keystore_only and test_sbio_exchange fail the same way on acfcd08 without this change: a host rustc compile error in their extracted test.
  • The whole tools/aurora-sep suite. 1093 tests run; the 16 that fail or error fail identically on the base.
    • 4 Rust compile errors in extracted code.
    • 8 need files outside this sparse checkout.
    • 1 runs git show of an old blob, which a blob:none clone can't serve.
    • 1 m3 uninstall test touches a root-owned /etc path.
    • 1 m3_pro_perf assertion fails.

Related

  • The distribution PRs drop the packaged options apple_sep xart_writes=1 provision_keybag=1 file and the read-only override docs (links in the comments below).
  • The M5 rig's default boot drops apple_sep.xart_writes=0 apple_sep.provision_keybag=0 in its own switch.

🤖 Generated with Claude Code

DjDeveloperr added a commit to aurora-silicon/aurora-xart that referenced this pull request Nov 20, 2026
…eybag as obsolete

aurora-silicon/linux#2591 removes the apple_sep xart_writes and
provision_keybag module parameters: the driver always opens the xART
store for writing and always provisions a missing identity keybag, and
a stale value on the command line or in modprobe.d is ignored.

- system.py: PARAMS keeps only xart_guard. The read-only-install reader
  is gone; any xart_writes or provision_keybag setting in modprobe.d or
  on the kernel command line is collected as obsolete, with its file and
  line or the command line.
- state.py: xART writes count as disabled only when the running driver
  says so (diag/xart disabled, or diag/policy xart_writes=0 from an older
  kernel), never from a configured value. status warns when they are,
  and names each obsolete setting to remove (a warning for a 0, which
  turned writes off under an older kernel, otherwise a note). The
  provision_keybag=0 note now comes only from an old kernel's
  diag/policy and says so.
- reset-touchid no longer refuses on an old --read-only modprobe file.
  It refuses only when the running driver reports xART writes disabled,
  because a new keybag needs them and the next boot could not create one.
- status --json: read_only is replaced by xart_writes {disabled, why}
  and obsolete_params; params lists only xart_guard.
- Docs (README, FORMATS.md, aurora-xart.8) and tests follow; version
  0.2.0.

Signed-off-by: Diljit Singh <hey@djdev.me>
DjDeveloperr added a commit to aurora-silicon/aurora-sep-userspace that referenced this pull request Nov 20, 2026
aurora-silicon/linux#2591 removes the apple_sep xart_writes and
provision_keybag parameters: the driver always opens the xART store
for writing and always provisions a missing identity keybag, and a
stale value is ignored.

The initrd design no longer names those parameters. In the
key-store-only phase the xART store is read-only until complete_attach
reopens it writable, and provisioning waits for complete_attach; that
is part of the phase, and there are no write parameters to pass.
keystore_only=2 with no xART guard is writable in the initrd, as
before.

The initrd loader already passes no write options, and test-initrd.sh
keeps enforcing that. Its PARMS_OLD fixture still models a kernel
without keystore_only; a comment now says why it lists xart_writes.

Signed-off-by: Diljit Singh <hey@djdev.me>
DjDeveloperr added a commit to aurora-silicon/aurora-xart that referenced this pull request Nov 20, 2026
…on a keybag

On a kernel from before aurora-silicon/linux#2591 that runs with
provision_keybag=0 (diag/policy), a reset would move the stuck keybag
aside and the next boot would not create a new one, leaving Touch ID
down. Refuse that case, and say why.

Also:
- status: on such a kernel, Touch ID failed with no keybag gets the
  not-provisioned note again;
- an obsolete setting in a packaged file (/usr/lib, /run) is advice to
  update the package, not to edit a file the package owns;
- JSON xart_writes.why is null without the driver, as disabled is;
- the FILES lists name every modprobe.d directory that is read.

Signed-off-by: Diljit Singh <hey@djdev.me>
DjDeveloperr added a commit that referenced this pull request Nov 20, 2026
SEP writes are always on: apple_sep no longer has the xart_writes or
provision_keybag parameters (the previous commits of #2591), so there is
nothing for the loader to set. load-driver now runs a plain
"modprobe apple_sep". Passing the two values on its command line also
gave each of them twice next to a stale apple_sep.xart_writes= on the
kernel command line, and kernels from before #2591 then refused to load
the driver.

The aurora-sep status check and aurora-sep-ready no longer name
xart_writes=0 as a cause of disabled xART writes. They name only the
fail-closed causes: the key-store-only initrd phase or a failed writable
reopen, and a gigalocker that did not validate for writing.
aurora-sep-ready keeps the "xART writes are disabled" prefix that the
NixOS lifecycle check looks for.

README.md, XART-SAFETY.md and M1-SUPPORT.md no longer offer a read-only
setting. XART-SAFETY.md's "Current default" section becomes "SEP writes
are always on", and the m2fix5 opt-in is described as history.

The Fedora and Arch packages ship these files, so they get a new release
(1-13 and pkgrel 18).

Owner directive (2026-11-20): no machine or distro should have SEP
writes off. The write path was proven on T6050 in the 11-11 J714s
Touch ID lab boots (#1587, 554c6d1).

Signed-off-by: Diljit Singh <hey@djdev.me>
DjDeveloperr added a commit that referenced this pull request Nov 20, 2026
install-aurora-sep.sh --read-only wrote /etc/modprobe.d/aurora-sep.conf
with "options apple_sep xart_writes=0 provision_keybag=0". The driver no
longer has either parameter (#2591), and SEP writes are always on, so
the mode is removed: the READ_ONLY setting, the file it wrote, the
"still read-only" and "Done, read-only" messages, --reset-touchid's
refusal on such a Mac, the read-only alternative in the first-boot
write notice and in the agent prompt, and --read-only in the option
combinations and the usage.

--read-only now exits with "--read-only was removed: SEP writes are
always on (#2591)", before anything runs.

An install, a re-run and --reset-touchid (after it is confirmed) delete
the file an earlier --read-only install wrote. That file is recognised
by its "# aurora-sep --read-only:" first line, with nothing else in it
but comments and the options line. On a kernel from before #2591 that
file still kept the driver from writing. Any other file at that path is
kept, with a warning that kernels from #2591 on ignore xart_writes and
provision_keybag. --uninstall still removes the path, and an owner's
own blacklist or install line for apple_sep still keeps the driver off.

The tests that used --read-only now check the same option rules
without it. The new test_sep_writes_always_on.py checks the refusal,
the cleanup and the kept file, --reset-touchid, and that load-driver
and the packaged modprobe.d files pass no write option. The SEP source
workflow runs it.

Signed-off-by: Diljit Singh <hey@djdev.me>
DjDeveloperr added a commit that referenced this pull request Nov 20, 2026
SEP writes are always on: apple_sep no longer has the xart_writes or
provision_keybag parameters (the previous commits of #2591), so there is
nothing for the loader to set. load-driver now runs a plain
"modprobe apple_sep". Passing the two values on its command line also
gave each of them twice next to a stale apple_sep.xart_writes= on the
kernel command line, and kernels from before #2591 then refused to load
the driver.

The aurora-sep status check and aurora-sep-ready no longer name
xart_writes=0 as a cause of disabled xART writes. They name only the
fail-closed causes: the key-store-only initrd phase or a failed writable
reopen, and a gigalocker that did not validate for writing.
aurora-sep-ready keeps the "xART writes are disabled" prefix that the
NixOS lifecycle check looks for.

README.md, XART-SAFETY.md and M1-SUPPORT.md no longer offer a read-only
setting. XART-SAFETY.md's "Current default" section becomes "SEP writes
are always on", and the m2fix5 opt-in is described as history.

The Fedora and Arch packages ship these files, so they get a new release
(1-13 and pkgrel 18).

Owner directive (2026-11-20): no machine or distro should have SEP
writes off. The write path was proven on T6050 in the 11-11 J714s
Touch ID lab boots (#1587, 554c6d1).

Signed-off-by: Diljit Singh <hey@djdev.me>
@DjDeveloperr
DjDeveloperr force-pushed the djdev/sep-writes-always-on branch from 54eb866 to 761b1f8 Compare November 20, 2026 18:29
DjDeveloperr added a commit that referenced this pull request Nov 20, 2026
install-aurora-sep.sh --read-only wrote /etc/modprobe.d/aurora-sep.conf
with "options apple_sep xart_writes=0 provision_keybag=0". The driver no
longer has either parameter (#2591), and SEP writes are always on, so
the mode is removed: the READ_ONLY setting, the file it wrote, the
"still read-only" and "Done, read-only" messages, --reset-touchid's
refusal on such a Mac, the read-only alternative in the first-boot
write notice and in the agent prompt, and --read-only in the option
combinations and the usage.

--read-only now exits with "--read-only was removed: SEP writes are
always on (#2591)", before anything runs.

An install, a re-run and --reset-touchid (after it is confirmed) delete
the file an earlier --read-only install wrote. That file is recognised
by its "# aurora-sep --read-only:" first line, with nothing else in it
but comments and the options line. On a kernel from before #2591 that
file still kept the driver from writing. Any other file at that path is
kept, with a warning that kernels from #2591 on ignore xart_writes and
provision_keybag. --uninstall still removes the path, and an owner's
own blacklist or install line for apple_sep still keeps the driver off.

The tests that used --read-only now check the same option rules
without it. The new test_sep_writes_always_on.py checks the refusal,
the cleanup and the kept file, --reset-touchid, and that load-driver
and the packaged modprobe.d files pass no write option. The SEP source
workflow runs it.

Signed-off-by: Diljit Singh <hey@djdev.me>
@DjDeveloperr
DjDeveloperr marked this pull request as ready for review November 20, 2026 18:29
DjDeveloperr added a commit that referenced this pull request Nov 20, 2026
Distribution kernels from before #2591 still have the
xart_writes and provision_keybag parameters. They default to 1, so writes
stay on, but a leftover 0 from an earlier read-only install or old docs
is still honoured there. XART-SAFETY.md says to remove it, and the
aurora-sep status and aurora-sep-ready messages name it as one cause of
a read-only store, without spelling the setting.

Signed-off-by: Diljit Singh <hey@djdev.me>
DjDeveloperr added a commit to aurora-silicon/fedora-aurora that referenced this pull request Nov 20, 2026
Distribution kernels from before aurora-silicon/linux#2591 still have the
xart_writes and provision_keybag parameters. They default to 1, so writes
stay on, but a leftover 0 from an earlier read-only install or old docs
is still honoured there. XART-SAFETY.md says to remove it, and the
aurora-sep status and aurora-sep-ready messages name it as one cause of
a read-only store, without spelling the setting.

Signed-off-by: Diljit Singh <hey@djdev.me>
DjDeveloperr added a commit to aurora-silicon/alarm-aurora that referenced this pull request Nov 20, 2026
Distribution kernels from before aurora-silicon/linux#2591 still have the
xart_writes and provision_keybag parameters. They default to 1, so writes
stay on, but a leftover 0 from an earlier read-only install or old docs
is still honoured there. XART-SAFETY.md says to remove it, and the
aurora-sep status and aurora-sep-ready messages name it as one cause of
a read-only store, without spelling the setting.

Signed-off-by: Diljit Singh <hey@djdev.me>
DjDeveloperr added a commit to aurora-silicon/nixos-aurora that referenced this pull request Nov 20, 2026
Distribution kernels from before aurora-silicon/linux#2591 still have the
xart_writes and provision_keybag parameters. They default to 1, so writes
stay on, but a leftover 0 from an earlier read-only install or old docs
is still honoured there. XART-SAFETY.md says to remove it, and the
aurora-sep status and aurora-sep-ready messages name it as one cause of
a read-only store, without spelling the setting.

Signed-off-by: Diljit Singh <hey@djdev.me>
SEP writes are always on, on every Mac and in every distribution. The
owner's directive: no machine or distribution should ever run with SEP
writes off, and the switch is removed entirely rather than defaulted.

The write path is proven on the T6050 as on the earlier SoCs: the
J714s lab boots of 2026-11-11 (linux#1587) ran with xART writes and
keybag provisioning on and the guard at level 1. The keybag was
provisioned, the enrolment survived a reboot, the enrolled finger
matched, a wrong finger was rejected, and the guard refused nothing.

So:

- xart_writes and provision_keybag are gone from the module parameters.
  The store is always opened for writing, and the bring-up always
  provisions a missing identity keybag (an existing one returns before
  any write, as before).
- The "provision_keybag=1 requires xart_writes=1" EINVAL in probe,
  complete_attach and the bring-up goes with them.
- A stale apple_sep.xart_writes= or apple_sep.provision_keybag= on the
  command line, or in a modprobe.d file, no longer stops the driver
  loading. The module loader logs "unknown parameter ... ignored" and
  goes on. Before, both were SetOnce Rust parameters, so a second value
  (a command-line value next to a modprobe.d one) failed the load.

What stays, because none of it turns writes off:

- The fail-closed checks: APFS must prove a single unsnapshotted .gl
  extent and the records must validate, or the probe fails as before.
- The key-store-only initrd phase (keystore_only=1/2) still keeps xART
  read-only until complete_attach, which reopens it writable. With
  keystore_only=2 and no guard it is writable in the initrd, as before.
- A failed writable reopen in complete_attach still leaves a read-only
  store behind and returns the error.
- The xART guard, which refuses rewrites of session records this Linux
  install did not create (other OSes' records).

diag/policy keeps its keys for ABI 3 readers: xart_writes is still the
effective writability, and provision_keybag now always reads 1. The
refusal messages name the read-only store instead of a parameter.

Built as an external module against a 7.3 output tree with the same
SEP sources: modinfo lists neither parameter.

Signed-off-by: Diljit Singh <hey@djdev.me>
SEP writes are always on: apple_sep no longer has the xart_writes or
provision_keybag parameters (the previous commits of #2591), so there is
nothing for the loader to set. load-driver now runs a plain
"modprobe apple_sep". Passing the two values on its command line also
gave each of them twice next to a stale apple_sep.xart_writes= on the
kernel command line, and kernels from before #2591 then refused to load
the driver.

The aurora-sep status check and aurora-sep-ready no longer name
xart_writes=0 as a cause of disabled xART writes. They name only the
fail-closed causes: the key-store-only initrd phase or a failed writable
reopen, and a gigalocker that did not validate for writing.
aurora-sep-ready keeps the "xART writes are disabled" prefix that the
NixOS lifecycle check looks for.

README.md, XART-SAFETY.md and M1-SUPPORT.md no longer offer a read-only
setting. XART-SAFETY.md's "Current default" section becomes "SEP writes
are always on", and the m2fix5 opt-in is described as history.

The Fedora and Arch packages ship these files, so they get a new release
(1-13 and pkgrel 18).

Owner directive (2026-11-20): no machine or distro should have SEP
writes off. The write path was proven on T6050 in the 11-11 J714s
Touch ID lab boots (#1587, 554c6d1).

Signed-off-by: Diljit Singh <hey@djdev.me>
install-aurora-sep.sh --read-only wrote /etc/modprobe.d/aurora-sep.conf
with "options apple_sep xart_writes=0 provision_keybag=0". The driver no
longer has either parameter (#2591), and SEP writes are always on, so
the mode is removed: the READ_ONLY setting, the file it wrote, the
"still read-only" and "Done, read-only" messages, --reset-touchid's
refusal on such a Mac, the read-only alternative in the first-boot
write notice and in the agent prompt, and --read-only in the option
combinations and the usage.

--read-only now exits with "--read-only was removed: SEP writes are
always on (#2591)", before anything runs.

An install, a re-run and --reset-touchid (after it is confirmed) delete
the file an earlier --read-only install wrote. That file is recognised
by its "# aurora-sep --read-only:" first line, with nothing else in it
but comments and the options line. On a kernel from before #2591 that
file still kept the driver from writing. Any other file at that path is
kept, with a warning that kernels from #2591 on ignore xart_writes and
provision_keybag. --uninstall still removes the path, and an owner's
own blacklist or install line for apple_sep still keeps the driver off.

The tests that used --read-only now check the same option rules
without it. The new test_sep_writes_always_on.py checks the refusal,
the cleanup and the kept file, --reset-touchid, and that load-driver
and the packaged modprobe.d files pass no write option. The SEP source
workflow runs it.

Signed-off-by: Diljit Singh <hey@djdev.me>
complete_attach falls back to a read-only store when the writable
reopen fails. With keystore_only=2 and no guard the store was already
writable at probe, so diag/xart kept reading enabled. Publish the
read-only state there, as the ABI text says.

Signed-off-by: Diljit Singh <hey@djdev.me>
…etting

Distribution kernels from before #2591 still have the
xart_writes and provision_keybag parameters. They default to 1, so writes
stay on, but a leftover 0 from an earlier read-only install or old docs
is still honoured there. XART-SAFETY.md says to remove it, and the
aurora-sep status and aurora-sep-ready messages name it as one cause of
a read-only store, without spelling the setting.

Signed-off-by: Diljit Singh <hey@djdev.me>
…ners' files

Review follow-up:
- sep_block_line no longer skips /etc/modprobe.d/aurora-sep.conf. The
  installer no longer writes that file, so a blacklist or install line
  there is the owner's and keeps the driver off, as anywhere else. The
  file an earlier --read-only install wrote had no such line.
- --uninstall removes only that earlier --read-only file (through
  read_only_conf_cleanup); an owner's own file at the path stays.
- --reset-touchid refuses when the owner's file there sets xart_writes=0
  or provision_keybag=0 and the installed apple_sep still has those
  parameters (a kernel from before linux#2591): the next boot could not
  create a new keybag. The kept-file warning says older kernels still
  honour a 0.
- A failed removal of the old file warns instead of dying late in an
  install.
- Tests for each, with the real sep_block_line pointed at the test's own
  modprobe.d.

Also the ABI text: diag/xart is not fixed at probe (complete_attach and
a failed writable reopen update it), and the key-store-only read-only
wording separates keystore_only=1 from 2. M1-SUPPORT.md re-wrapped.

Signed-off-by: Diljit Singh <hey@djdev.me>
@DjDeveloperr
DjDeveloperr force-pushed the djdev/sep-writes-always-on branch from a469e1a to 18acebd Compare November 20, 2026 19:19
…tays

read_only_conf_cleanup only warns when it cannot remove the file an
earlier --read-only install wrote. --reset-touchid now stops there,
before anything is moved, because a kernel from before linux#2591 would
still load that file and create no keybag at the next boot.

The modprobe.d directory override sep_block_line reads is now the
test-only AURORA_SEP_TEST_MODPROBE_DIRS, and two ABI paragraphs are
re-wrapped.

Signed-off-by: Diljit Singh <hey@djdev.me>
@DjDeveloperr
DjDeveloperr merged commit ea45e68 into integration/aurora Nov 20, 2026
7 checks passed
AxxzyWasTaken pushed a commit to AxxzyWasTaken/alarm-aurora that referenced this pull request Nov 21, 2026
… the read-only override

Owner directive (2026-11-20): no machine or distribution has SEP writes
off. The 11-11 J714s (M5 Pro, T6050) Touch ID boots ran with writes and
keybag provisioning on and the guard at level 1: the keybag was
provisioned, the enrolment survived a reboot, the enrolled finger
matched, a wrong finger was rejected, and nothing was refused
(aurora-silicon/linux#1587, merged as 554c6d14).

aurora-silicon/linux#2591 removes apple_sep's xart_writes and
provision_keybag parameters, so a stale value on the kernel command line
or in modprobe.d is ignored.

- Remove /usr/lib/modprobe.d/aurora-sep.conf (aurora-sep-modprobe.conf,
  "options apple_sep xart_writes=1 provision_keybag=1"): its PKGBUILD
  source, checksum and package() lines. It existed only so that
  /etc/modprobe.d/aurora-sep.conf could replace it with =0. On kernels
  before #2591 both parameters default to 1, so writes stay on.
- load-driver still runs "modprobe apple_sep" with no options; its comment
  no longer describes an override.
- aurora-sep status: "xART writes are disabled: the store is read-only this
  boot (key-store-only initrd phase, or a failed writable reopen): enrol
  and verify are refused", still a WARN. aurora-sep-ready names only the
  gigalocker that did not validate for writing, under the same "xART
  writes are disabled" prefix.
- XART-SAFETY.md: "SEP writes are always on" replaces "Current default";
  the earlier deployment policy reads as history. README.md no longer
  says aurora-xart reset-touchid refuses on a read-only install. The
  provenance adaptations follow.
- tests/test_sep_modprobe.py, which modelled the read-only override, now
  checks that nothing tracked sets xart_writes or provision_keybag (a
  modprobe.d options line, a modprobe command line or a kernel command
  line), that no PKGBUILD installs a modprobe.d file that sets one, that
  aurora-sep ships only the load-ordering blacklist in modprobe.d (kmod
  agrees), that load-driver passes no options, and that the messages name
  no off switch.

Kept: xart_guard and its xart_guard=0 opt-out, keystore_only, and the
packaged "blacklist apple_sep" in aurora-sep-load.conf (load ordering:
aurora-sep.service always loads the driver).

Signed-off-by: Diljit Singh <hey@djdev.me>
sioodmy pushed a commit to sioodmy/nixos-aurora that referenced this pull request Nov 22, 2026
…he read-only override

Owner directive (2026-11-20): no machine or distribution has SEP writes
off. The 11-11 J714s (M5 Pro, T6050) Touch ID boots ran with writes and
keybag provisioning on and the guard at level 1: the keybag was
provisioned, the enrolment survived a reboot, the enrolled finger
matched, a wrong finger was rejected, and nothing was refused
(aurora-silicon/linux#1587, merged as 554c6d14).

aurora-silicon/linux#2591 removes apple_sep's xart_writes and
provision_keybag parameters, so a stale value on the kernel command line
or in modprobe.d is ignored.

- modules/sep.nix no longer copies sep/aurora-sep-modprobe.conf ("options
  apple_sep xart_writes=1 provision_keybag=1") to
  /run/modprobe.d/aurora-sep.conf. That copy (nixos-aurora#354) existed so
  that a read-only environment.etc."modprobe.d/aurora-sep.conf" could
  replace it, which is no longer offered. The file is removed. With no
  packaged value, nothing gives a parameter twice, which stopped kernels
  before #2591 loading the driver: aurora-silicon#354's problem goes away. On those
  kernels both parameters default to 1, so writes stay on.
- CI: the step that grepped the defaults file now evaluates every profile
  (kde, gnome, minimal) and fails if a modprobe.d entry in environment.etc,
  boot.extraModprobeConfig, boot.initrd.extraModprobeConfig, a tmpfiles
  rule or a kernel parameter names xart_writes or provision_keybag.
- load-driver still runs "modprobe apple_sep" with no options; its comment
  no longer describes an override.
- aurora-sep status: "xART writes are disabled: the store is read-only this
  boot (key-store-only initrd phase, or a failed writable reopen): enrol
  and verify are refused", still a WARN. aurora-sep-ready names only the
  gigalocker that did not validate for writing, under the same "xART
  writes are disabled" prefix that checks/sep-lifecycle.py looks for.
- XART-SAFETY.md: "SEP writes are always on" replaces "Current default";
  the earlier deployment policy reads as history. README.md no longer
  says aurora-xart reset-touchid refuses on a read-only install. The
  provenance adaptations follow.
- tests/test_sep_modprobe.py, which modelled the read-only override, now
  checks that nothing tracked sets xart_writes or provision_keybag (a
  modprobe.d options line, a modprobe command line or a kernel command
  line), that no module mentions either, that the SEP module configures no
  modprobe.d file, that CI checks the evaluated profiles, that load-driver
  passes no options, and that the messages name no off switch.

Kept: xart_guard and its xart_guard=0 opt-out, keystore_only, and
boot.blacklistedKernelModules = [ "apple_sep" ] (load ordering:
aurora-sep.service always loads the driver).

Signed-off-by: Diljit Singh <hey@djdev.me>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant