Compare commits

...

1 Commits

Author SHA1 Message Date
coah
7b5bf4854a fix(cdb): correct crypt-buffer ownership slot types 2026-10-02 21:48:46 -05:00
2 changed files with 31 additions and 7 deletions

View File

@@ -8,13 +8,11 @@ void CDBCryptBufSysInit(CDBCryptBuf* cryptBuf) {
cryptBuf->allocated = FALSE;
}
// Some weird crap is going on with these functions.
CDBErr CDBCryptBufAllocate(CDBCryptBuf* cryptBuf) {
CDBErr CDBCryptBufAllocate(CDBCryptBuf** cryptBuf) {
CDBErr err = CDB_ERROR_OK;
if (s_cryptBuf->allocated == FALSE) {
cryptBuf->buffer32[0] = (u32)&s_cryptBuf[err];
*cryptBuf = &s_cryptBuf[err];
s_cryptBuf[err].allocated = TRUE;
return CDB_ERROR_OK;
} else {
@@ -23,12 +21,12 @@ CDBErr CDBCryptBufAllocate(CDBCryptBuf* cryptBuf) {
}
}
CDBErr CDBCryptBufFree(CDBCryptBuf* cryptBuf) {
CDBErr CDBCryptBufFree(CDBCryptBuf** cryptBuf) {
CDBErr err = CDB_ERROR_OK;
if ((CDBCryptBuf*)*cryptBuf->buffer32 == s_cryptBuf) {
if (*cryptBuf == s_cryptBuf) {
s_cryptBuf[err].allocated = FALSE;
*cryptBuf->buffer32 = 0;
*cryptBuf = NULL;
return CDB_ERROR_OK;
} else {
CDBReportError("failed to free crypt buffer\n");

View File

@@ -0,0 +1,26 @@
# CDB crypt-buffer ownership contract
Base: 20ba34a8, 43U. Focused source correction after the read-only CDBRecordEncrypt audit found only register-allocation differences in that function.
## Evidence and change
CDBRecord.c already declares CDBCryptBufAllocate and CDBCryptBufFree with CDBCryptBuf** parameters. Its callers supply the address of the record's crypt-buffer pointer slot. CDBCrypt.c instead declared CDBCryptBuf* and accessed buffer32[0] as a u32-encoded pointer, producing incompatible cross-translation-unit function types and describing the caller's pointer slot as an entire crypt-buffer object.
Both definitions now take CDBCryptBuf** and use direct typed pointer loads/stores. Allocation assigns the pool pointer; free compares the owned pointer and clears the slot to NULL. The existing pool-index expression, flag update order, diagnostics and failure behavior are retained. The obsolete placeholder comment is removed. No header, payload layout, pool extent, Encrypt implementation, or linking change.
## Validation
- Configure: `python3 configure.py --version 43U --wrapper ../toolchain/wibo-build/wibo`
- Build: `WIBO_SJIS_MISSING_IMPORTS=1 ../.venv/bin/ninja`
- Entire compiled CDBCrypt.o is byte-identical to the baseline, including all code/data, symbols and relocations
- CDBCryptBufAllocate: 36/36 instructions, ctxdiff zero
- CDBCryptBufFree: 33/33 instructions, ctxdiff zero
- Unit remains 100% fuzzy, 3/3 exact functions, 296/296 code bytes and 80/80 data bytes
- Full-project before/after objdiff reports are identical; no metric gain is claimed
- Pool identical: 2/2 strings; literal-reference audit: 2 arguments across all 3 functions, no candidates/errors/skips
- Host test compiles the production source with the actual CDBCryptBuf layout and minimal platform/error stubs. AddressSanitizer and UBSan pass 50,036 ownership transitions, including every combination of six representative allocation flags and null/owned/foreign pointer slots, 10,000 randomized state cases, and 10,000 allocate/reallocate/free/refree cycles. Checks cover pointer-slot canaries, complete payload preservation, allocation flag, return value and diagnostic count
- LeakSanitizer is disabled because it cannot run under this execution environment's ptrace; the harness makes no heap allocations. Address and undefined-behavior instrumentation remain enabled
- Full 43U build and build/43U/ok pass; DOL SHA1 is 26116613f624061ba99c8d1a299aaa6efa85670d
- git diff --check passes
Private reproducible evidence is in /tmp/cdb-buffer-ownership (baseline object, full reports, host validator and build log). Original objects/disassembly from the preceding read-only audit remain private under /tmp/cdb-record-encrypt-audit.