[casaroli]

Summary

In a protected build the kernel cannot read the list of builtin applications in user space. The binfmt builtin loader uses its own copy of the list, which user space must set with boardctl(BOARDIOC_BUILTINS). NSH starts a builtin with task_spawn() and does not use the loader, so it never needed this. nxinit starts each service with posix_spawnp(), which goes through the loader, so in a protected build every builtin service fails with ENOENT.

esp32s3-devkit:knsh uses nxinit as init and never gets a shell:

Error Starting service 'console': 2

init_main() now gives the list to the kernel first, when CONFIG_BUILD_PROTECTED, CONFIG_BUILTIN and CONFIG_BOARDCTL are set.

Impact

nxinit in protected builds only. Flat and kernel builds do not compile the new code.

Testing

ESP32-S3-DevKitC with an ESP32-S3-WROOM-2-N32R8V (esp32s3-devkit:knsh, protected build, stock defconfig with nxinit as init, plus CONFIG_TESTING_OSTEST), nuttx master:

	apps master	this PR
boot	Error Starting service 'console': 2 in a loop, no shell	nxinit starts sh, NSH prompt
ostest	(no shell)	status 0

tools/checkpatch.sh -c -u -m -g passes.

[cederom]

Thank you @casaroli :-)

@JianyuWang0623 @linguini1 @michallenc @FelipeMdeO could you please take a look?

[JianyuWang0623]

Thank you @casaroli :-)

@JianyuWang0623 @linguini1 @michallenc @FelipeMdeO could you please take a look?

@cederom @casaroli

LGTM, good catch, thank you!

This reminds me that a set of changes from @anjiahao1 (with a follow-up fix from @yeungao) addresses the same problem on the kernel side: the protected kernel reads the builtin list directly through struct userspace_s, so no userspace call to BOARDIOC_BUILTINS is needed. I've opened a PR for it: apache/nuttx#20964

[FelipeMdeO]

Congratulations Casaroli.

[JianyuWang0623]

@casaroli Thanks again for tracking this down! The NuttX-side fix has now been merged: apache/nuttx#20964. The protected kernel now reads the builtin list directly through struct userspace_s, and BOARDIOC_BUILTINS has been removed.

I had a similar idea earlier (#3821), but with the kernel change in place this PR no longer seems necessary.