fix(stage1): give stage-1 an entry point, and refuse to verify an image that is not there

.github/workflows/ci.yml

@@ -250,7 +250,7 @@ jobs:
   ci-gate:
     name: CI Gate
     runs-on: ubuntu-22.04
-    needs: [test, build-arm, static-analysis]
+    needs: [test, build-arm, fuzz-build, static-analysis]
     if: always()
     steps:
       - uses: actions/checkout@v4

.github/workflows/eosim-sanity.yml

@@ -23,6 +23,14 @@ jobs:
   install-validate:
     name: Install & Validate (${{ matrix.os }}, Python ${{ matrix.python-version }})
     runs-on: ${{ matrix.os }}
+    # Every step here is written in bash: `$(...)`, `|| { ... }`, a heredoc,
+    # and /tmp. Without this the Windows legs ran them under PowerShell, where
+    # `SITE_PACKAGES=$(...)` is an unknown command and the job went red, while
+    # the `|| { exit 1 }` guard in "Verify installation" parsed as an unexecuted
+    # script block and could never fail.
+    defaults:
+      run:
+        shell: bash
     strategy:
       fail-fast: false
       matrix:

.github/workflows/scorecard.yml

@@ -24,7 +24,7 @@ jobs:
       - uses: actions/checkout@v4
         with:
           persist-credentials: false
-      - uses: ossf/scorecard-action@v2.4.0
+      - uses: ossf/scorecard-action@v2.4.3
         with:
           results_file: results.sarif
           results_format: sarif

CHANGELOG.md

@@ -13,6 +13,15 @@
 - **`image_verify.c`:** `eos_image_verify_integrity` rejects a zero `image_size`, and an `addr + hdr_size` that wraps `uint32_t`, instead of computing a payload address that is not the payload.
 
 ### Fixed
+- **Stage-0 refuses a stage-1 image it could not read.** The stage-1 hash loop in `stage0/jump_stage1.c` discarded `eos_hal_flash_read()`'s result, so a failed read hashed the previous chunk (or the stack) as if it were stage-1; the mismatch that followed reached recovery with the wrong reason (`0xBAD1`, hash mismatch), and only by luck. A failed read now enters recovery with its own reason (`EBLDR_FAIL_STAGE1_READ`, `0xBAD2`) and never falls through to the jump, matching what `eos_crypto_verify_image()` in `core/` already does. The hash-mismatch path (`EBLDR_FAIL_STAGE1_HASH`, `0xBAD1`) and the recovery-trigger path had the same fall-through -- `eos_recovery_enter()` with nothing after it, so a failed verification reached the `IMAGE_VALID` record and the jump, held back only by recovery never returning -- and both now return. The two detail codes are named in `include/eos_types.h` beside the `EOS_LOG_*` events. `tests/unit/test_stage0_hal_results.py` is a source-level guard that every HAL read or write in `stage0/` has its result tested (not merely assigned; preprocessor lines are ignored), and that every entry into recovery in `ebldr_stage0_main()` is followed by a return.
+- **The stage-1 image was empty, and stage-0 verified it anyway.** On every board build with a stage-1 linker script, `eboot_firmware.bin` was 0 bytes: `stage1/` exported `eboot_main()` only, the script had no `ENTRY()` and kept no section, and with `-nostartfiles` and `--gc-sections` the linker had nothing to anchor. `tools/embed_stage1_hash.py` took the size verbatim and emitted `stage1_expected_size = 0u` with the SHA-256 of the empty string; `stage0/jump_stage1.c` ran its hash loop zero times, matched nothing against nothing, recorded `IMAGE_VALID` and jumped. Three independent fixes: `stage1/reset_entry.c` now owns the vector table and the reset handler (stack pointer, `Reset_Handler` into `eboot_main()`, faults and the system exceptions through SysTick) and `boards/stm32f4/stm32f4_stage1.ld` has `ENTRY(Reset_Handler)` and keeps `.isr_vector`, so the stm32f4 stage-1 links to 13,416 bytes and the embedded digest is that image's; the embed tool refuses an empty image outright and one below a 4096-byte floor (`--min-size`); and stage-0 refuses `stage1_expected_size == 0` before the loop (`EBLDR_FAIL_STAGE1_NO_IMAGE`, `0xBAD3`), because a bootloader does not trust its own build. Stage-0's own vector table stopped at UsageFault while `board_early_init()` enables the SysTick interrupt, so the first tick fetched a handler address from the code after the table; both tables now run through SysTick. `cortex_r5` is left as it was: both of its linker scripts keep `.vectors`, which nothing in the tree emits, so both of its images link to zero bytes and the embed tool now stops that board's build at the empty stage-1 instead of describing it; the port needs an entry shape of its own (#137), and the guard names it as a strict expected failure until then. The guard also checks that every section a stage script keeps is one some source file emits, which is the mismatch that produced both empty images. `tests/unit/test_stage1_image_is_verified.py` pins all of it.
+- **The tree did not configure, compile or link after the 09-07 batch merge.** `tests/CMakeLists.txt` registered `eboot_test_fdt_loader` twice; `core/sha512.c` had been replaced by a version predating the `bitlen[2]`/`buffer_len` context; `core/boot_log.c`, `core/secure_boot.c` and `core/fdt_loader.c` had been dropped from `eboot_core`; `scalarbase()` and `k_low_order[]` were defined twice; and the `eos_boot_log_get_head()` declaration was lost. All restored.
+- **Install-path verification order settled: signature before anti-rollback.** `eos_fw_update_finalize()` verifies the Ed25519 signature over the signed header prefix first and reads the TLV security counter only after the prefix that binds it is authenticated (see `docs/adr/ADR-020`). The `fw_update` and `fw_transport` suites now stream genuinely signed images; `tools/gen_fw_update_test_sigs.py` emits their signatures as `tests/vectors/fw_update_test_sigs.h`, and `tests/unit/test_fw_update_test_sigs.py` pins the committed header to the generator's output.
+- **`tests/CMakeLists.txt`:** the Valgrind list is derived from the registered suites again; a hand-written copy had replaced it, eleven registered suites were missing from `EBLDR_UNIT_TESTS`, and seven of those (`test_eos_sign_boot_path`, `test_fdt_loader`, `test_fw_decrypt`, `test_fw_update_sig`, `test_jump_app_bounds`, `test_qemu_arm64_timer`, `test_secure_boot_policy`) had no Valgrind run at all; the other four were only in the hand-written list. All eleven are appended.
+- **`.github/workflows/ci.yml`:** `fuzz-build` is in the CI gate. It was added after the gate job and the gate never waited for it.
+- **Unit suites count `tests_run`** as each test executes instead of assigning it a literal that the summary line then trusted.
+- **`.github/workflows/eosim-sanity.yml`:** the install-validate job's steps are bash and now run under `shell: bash` on the Windows legs, where PowerShell rejected `SITE_PACKAGES=$(...)` and parsed `|| { exit 1 }` as an unexecuted script block.
+- **`.github/workflows/scorecard.yml`:** `ossf/scorecard-action` moved to v2.4.3, the release hosted on ghcr.io; v2.4.0 pulls from gcr.io, which now requires GCP billing.
 - **The tree did not compile.** `include/eos_image.h` declared `eos_crc32()` as `int eos_crc32(uint32_t, size_t, uint32_t *)` while `core/image_verify.c` defined it as `uint32_t eos_crc32(uint32_t, size_t)` -- a conflicting-types error that stopped the build at the first core source file. The declaration now matches the definition and the documented behaviour.
 - **`ed25519_verify.c`:** `eos_ed25519_verify()` never performed the verification. Two merged copies of the challenge-hash step had been left in the function, the second referring to identifiers that do not exist (`sha512_ctx_t`, `sc_reduce`), and RFC 8032 step 4 -- the `[S]B == R + [k]A` check -- was absent entirely, leaving the function returning an undeclared `diff`. The duplicate is removed and the group-equation check restored; the function now passes the RFC 8032 test vectors and rejects tampered messages, every single-bit signature flip, wrong keys and malleated signatures.
 - **`recovery.c`:** `recovery_handle_write()` declared `slot_size` twice, which does not compile. The bounds check now calls `eos_recovery_write_in_range()` -- the helper the unit tests already exercise -- so the wire-input rule has one definition, and an unmapped slot (`base == 0`) is rejected too.

CMakeLists.txt

@@ -80,6 +80,7 @@ target_include_directories(eboot_hal PUBLIC ${EBLDR_INCLUDE_DIR})
 # ---- Core boot logic ----
 add_library(eboot_core STATIC
     core/bootctl.c
+    core/boot_log.c
     core/image_verify.c
     core/slot_manager.c
     core/boot_policy.c
@@ -106,8 +107,10 @@ add_library(eboot_core STATIC
     core/os_adapter.c
     core/ed25519_verify.c
     core/sha512.c
-    core/keystore.c
+    core/secure_boot.c
+    core/fdt_loader.c
     core/rollback.c
+    core/keystore.c
     core/debug_lock.c
     core/fw_decrypt.c
     core/image_tlv.c
@@ -381,7 +384,10 @@ if(NOT EBLDR_BOARD STREQUAL "none")
         endif()
 
         if(EXISTS "${_LD_STAGE1}")
-            add_executable(eboot_firmware stage1/main.c)
+            # stage1/reset_entry.c owns the vector table the linker script keeps;
+            # without it (and the script's ENTRY) the link had nothing to anchor
+            # and the image was empty. See tools/embed_stage1_hash.py.
+            add_executable(eboot_firmware stage1/main.c stage1/reset_entry.c)
             target_include_directories(eboot_firmware PRIVATE
                 ${EBLDR_INCLUDE_DIR}
                 ${CMAKE_CURRENT_SOURCE_DIR}/boards/${EBLDR_BOARD}

boards/stm32f4/stm32f4_stage1.ld

@@ -13,8 +13,21 @@ MEMORY
 
 _estack = ORIGIN(RAM) + LENGTH(RAM);
 
+/* Stage-0 loads SP from the first word of this image and branches to the
+ * second, so the vector table has to be the first thing in FLASH and the
+ * link has to be anchored on it: with -nostartfiles and --gc-sections and
+ * neither of these lines, nothing was reachable and the image was empty. */
+ENTRY(Reset_Handler)
+
 SECTIONS
 {
+    .isr_vector :
+    {
+        . = ALIGN(4);
+        KEEP(*(.isr_vector))
+        . = ALIGN(4);
+    } > FLASH
+
     .text :
     {
         . = ALIGN(4);

core/ed25519_verify.c

@@ -300,16 +300,6 @@ static int point_is_identity(gf p[4])
     return diff == 0;
 }
 
-static void scalarbase(gf r[4], const uint8_t *s)
-{
-    gf q[4];
-    fe_copy16(q[0], BX);
-    fe_copy16(q[1], BY);
-    fe_copy16(q[2], gf1);
-    fe_mul(q[3], BX, BY);
-    scalarmult(r, q, s);
-}
-
 /* Reject a public key outside the prime-order subgroup.
  *
  * Decoding a point is not enough. Ed25519 has eight points of low order, and