dangernoodle-io / breadboard / 30525250435

97%

push

github

fix: switch stack-overflow detection off silently-dropped CHECK_STACKOVERFLOW=2 (#1150)

CONFIG_FREERTOS_CHECK_STACKOVERFLOW is a Kconfig choice symbol in ESP-IDF
5.5.4 (CONFIG_FREERTOS_CHECK_STACKOVERFLOW_NONE / _PTRVAL / _CANARY), not
an int. Nine sdkconfig.defaults files (examples/smoke's esp32, esp32c3,
tdongle, elecrow-p4-hmi7, esp32-autofan, esp32-cache-sweep,
esp32-boot-progress, esp32-boot-status, plus examples/floor) either set
the invalid form `CONFIG_FREERTOS_CHECK_STACKOVERFLOW=2` or (the four
standalone esp32 CI boards) never set the choice at all. kconfig silently
discards an assignment to an unknown symbol name, so the framework
default (_CANARY) was in effect on every board despite comments (and
components/bb_health/Kconfig's documented recommendation) stating PTRVAL
was wanted. Confirmed via the merged sdkconfig.h before/after this
change, per board: before, only CONFIG_FREERTOS_CHECK_STACKOVERFLOW_CANARY
was defined; after, CONFIG_FREERTOS_CHECK_STACKOVERFLOW_PTRVAL is, on all
nine boards.

This was proven material on hardware, not cosmetic: a latent httpd stack
overflow that CANARY (context-switch-sampled) caught only ~2/3 of runs,
surfacing as an unattributed LoadProhibited wild-pointer fault, was
caught 5/5 with PTRVAL + CONFIG_FREERTOS_WATCHPOINT_END_OF_STACK=y.
Unblocks diagnosis of the class of fault behind B1-1284.

The end-of-stack watchpoint is now applied selectively, not blanket. It
costs up to 60 B of usable stack per task plus one of the CPU's limited
hardware watchpoints, and Espressif recommends HW_STACK_GUARD over it
where available. Verified empirically per chip against the pinned
ESP-IDF 5.5.4 tree's soc_caps.h (not just docs, which reflect a newer
IDF): esp32c3 and esp32p4 (elecrow-p4-hmi7) define
SOC_ASSIST_DEBUG_SUPPORTED, so CONFIG_ESP_SYSTEM_HW_STACK_GUARD is
already on by default there -- confirmed present in each board's merged
sdkconfig.h -- so the watchpoint is dropped on those two; PTRVAL stays.
esp32s3 (t... (continued)

7973 of 8576 branches covered (92.97%)

Branch coverage included in aggregate %.

13540 of 13695 relevant lines covered (98.87%)

2612.01 hits per line