mirror of
https://github.com/fluffos/fluffos
synced 2026-08-12 18:26:06 -04:00
1672e9a138
coverage along the failure axes RENAME + NORMALIZE: - lex.l -> lexer.l, lex.h -> lexer.h, lex.autogen.cc -> lexer.autogen.cc (matching lexer_rules/lexer_utils); the flex POST_BUILD copy runs the same NORMALIZE path-sanitizing step as bison's outputs, with a symmetric DENORMALIZE no-flex fallback. LEXTOKENSTREAM REMOVED (the scanner IS the interface): - It had become a forwarding wrapper; compile_file and stage_output drive the reentrant scanner directly, with the scanner guard ordered BEFORE the cleanup guard. FIVE REAL BUGS -- two caught in production on a live mud: 1. Unwind-order UAF: the wrapper was declared after the cleanup DEFER, so exception unwinds destroyed the scanner before teardown deleted the flex buffers. Fixed by guard ordering. 2. Stale active-scanner UAF (pinned by FatalAbortThenRecompile): an aborted compile skips end_new_file; the NEXT compile's current_line save read destroyed scanner guts. Owners now call lpc_lex_scanner_destroyed() before yylex_destroy. 3. Release-only spin: lpc_dump_stage_tokens skipped the shared-string table reset; a second dump walked stale hash chains into a freed A_STRINGS block. The stage env now mirrors the harness. 4. FIELD BUG -- #include with trailing comment: the include parser assumed the closing delimiter was the line's LAST character, so `#include <mudlib.h> /* ... */` swallowed the comment into the filename (classic-mudlib ftpdconf.h failed). Scan to the closing delimiter; trailing text is ignored (historical behavior); missing close is a clean error. 5. FIELD BUG -- teardown-vs-mem_block ordering UAF: an abort with a LIVE include buffer (inherit or fatal INSIDE an included file) had clean_parser free the mem_block areas BEFORE buffer teardown, whose include-accounting pops write into them (ASan WRITE on a live mud). Teardown now runs first in clean_parser; the success epilog is ordered symmetrically. Also: compile_file_fd's explicit fd flag (a -1 sentinel silently compiled the empty view), parse loop skipped on load failure, unified stage_output cleanup. COVERAGE (all portable: tmpfile(), no /tmp/mkstemp/open_memstream): 16 new unit tests -- fd entry success/bad-fd/pipe-growth, fatal aborts from the main buffer, inside an include, and inside a #if expression (suppress_expansion unstranded), unterminated text block/template, include depth limit, empty and no-final-newline includes, trailing block/line-comment includes with clean-name diagnostics, stage-output pp/tokens/load-failure. Two LPC testsuite pins run under ASan on every platform: include_trailing_comment.c (ftpdconf.h shape) and inherit_via_include.c (the abort-and-reload with the inherit inside an included file). Post-CI portability/leak follow-ups folded in: the pipe-growth test is fork-free (fits the pipe buffer; _pipe on Windows -- sys/wait.h broke MinGW), and the two test environments share one init guard (mixing the full boot with the tokenizer harness in one process double-called init_strings, orphaning a 64KB string table -- LeakSanitizer-caught in the CI sanitizer job). 297 tests under ASan/UBSan (leak detection on), driver-autotest x3. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
4 lines
174 B
C
4 lines
174 B
C
// Include target for include_trailing_comment.c -- resolved via the
|
|
// config's angle-bracket include path.
|
|
#define TRAILING_PIN_VALUE 42
|
|
#define TRAILING_PIN_NAME "pinned"
|