There were seven ways to log, each invented where the previous one was not
reachable:
OnionHEN_log libonion_platform, the only one with a file sink
shellui_log libonion_ipc, klog only — because it did not depend on
libonion_platform
game_log same body as shellui_log
onion_ipc_log same body again
logf libonion_ipc's IPC server, with a sink the host had to
register
log_msg onion_elfldr_server
crash_log util's fault handler, plus a second copy in common/
boot_log bootstrapper's fault handler
klog_printf / klog_puts / printf / puts, called directly
A message's destination therefore depended on which module happened to emit
it: ShellUI's ~300 records went to volatile klog and never reached a log file,
while the daemon's went to klog, stdout and disk.
Everything now goes through the LOG_* macros. All seven legacy entry points
are gone, and klog_printf survives only as a sink inside log.c.
The bare -nostdlib libraries deliberately did not get their own backend. Those
flags only affect the static library's own link step; each is linked into a
host with full libc and pthread, which is why they already called vsnprintf.
A second backend would have recreated the split this removes.
Levels were inferred from message text rather than applied uniformly — the old
entry points carried no severity, so a blanket mapping would either bury real
failures at debug or promote routine chatter to error. Result: 291 error, 63
warn, 140 info, 309 debug.
Consequences worth noting:
* ShellUI records now reach the log file, so a user report finally contains
what the injected process did. In release builds most of them are debug
and compile out, so this costs nothing on a retail unit.
* The IPC server's log-sink injection is gone; both servers registered the
same forwarder.
* common/backtrace.cpp and backtrace.hpp are deleted. That was a third copy
of the frame walking plus its own crash_log, and every caller of
printBacktrace/printBacktraceForCrash had already been commented out.
* Fault handlers use onion_log_emergency, so crash records go to the same
file as everything else instead of a separate crash log.
The shared entry point carries a printf format attribute, which the old
shellui_log explicitly avoided. That attribute found seven real format
defects: %p given typed pointers, %lx given a MonoObject*, %p given a
uint64_t, and %u given an unsigned long. Four of them only compile in debug
builds, so they had gone unnoticed.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
All 17 module CMakeLists repeated the same prologue: language standards, the
PS5 SDK header paths, a clang check, the cross-compile triple, the output
directory layout and an objcopy strip step. The triple appeared 17 times and
the strip block 6 times, each needing manual sync.
Add source/cmake/OnionTarget.cmake with onion_module_prologue,
onion_output_dirs and onion_strip_debug. Language standards stay opt-in
arguments because the loader-side C modules are C17 while the rest are C11,
and a shared default would silently downgrade them. Per-module compiler flags
are left alone since the modules genuinely differ.
Drop IS_SUBPROJECT and IS_CLANG, which were assigned in 11 and 6 modules
respectively and never read anywhere.
Add CONFIGURE_DEPENDS to the seven source-listing globs that lacked it, so a
newly added source file triggers a re-configure instead of being silently
omitted from the build.
Hoist -DDEBUG onto the shared C flags in the five modules that only passed it
to C++, so both languages in a target see the same macros. No .c file keys
off DEBUG today, so this changes no behaviour.
Compile and link flags were captured from build.ninja before and after: all
132 compile edges and 17 link edges are unchanged except for the intended
-DDEBUG addition.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Use a red-zone-safe, ABI-aligned stack window for pt_call, pt_call2, and
pt_syscall so elfldr_load cannot clobber SceShellUIMain mid-SwapBuffers.
Fix single-step return detection to compare against entry_rsp, and add
inject stage logs for the load/args window.
Rename product, namespaces, libraries, paths, and assets from Orion/orion
to Onion/onion across the tree. TCP stack-shutdown magic is now ONIO
(0x4F4E494F) to match the corrected branding.
Root fix stays: elevate with PTRACE_AUTHID only. Rename away from
set_ucred_to_debugger so the API cannot imply DEBUG_AUTHID. Remove
pt_attach ECHILD poll loops and multi-attempt toolbox inject — those
were workarounds, not the cause. Keep a single waitpid(WUNTRACED) after
PT_ATTACH.
Merge the three pt/elfldr clones (bootstrapper, libNineS, unpacker) into a
single liborion_elfldr with one authid policy (elevate once at inject entry,
never per ptrace). Wire consumers through shared headers; drop dead unpacker
loader sources. Route shellui if_exists through orion/platform and daemon
ShellUI/BigApp pid lookup through orion_proc. Document the layout and audit
status in docs/arch.md and docs/architecture-audit.md.
Drop duplicate allproc/ucred sources from NineS. Headers are shims to
orion/*, injector uses get_module_info, and CMake builds orion_proc before
NineS so inject and daemon share one process stack.
Keep IPC_loop as the only transport entry in msg.cpp. Move toolbox/fps
injection, LoadSettings, and remount/chmod/reply helpers into dedicated
TUs behind daemon_ops.hpp so inject changes no longer touch the IPC ring.
Drop unused hello_world ELF blob, sqlite headers/lib, and orphan
demo headers. Replace the 14k SCE error dump with a ~300-entry
install/BGFT table for DPI, cutting ~7MB from the source tree.
Match filename, libname, and sandboxed path basename. Prefer a non-allocating
get_module_info API and fix handle leaks on the successful get_module_handle path.
