load-buffer performs an invalid freeFile: src/cmd/cmd.zig:524-537
Severity: CRITICAL
Status: ✅ FIXED
The new regression test immediately exposed a second half of this bug:
pushOwned(data.items) handed the buffer stack a sub-span of a larger
capacity allocation; BufferEntry.deinit later frees by .len, which fails
the DebugAllocator canary check (“Invalid free”) even before the static-pointer
deferred deinit fired.
cmdLoadBuffer now uses data.toOwnedSlice() so the stack receives an
exactly-sized allocation, with errdefers covering name/data on every failure
path after ownership starts transferring.
Test: load-buffer transfers ownership without invalid free — bug #357
(real temp file, full parse → exec → server.deinit cycle under
testing.allocator’s ownership checks).
loadb <file> succeeds functionally but its cleanup path frees memory that was
never heap-allocated with a length that was never allocated — DebugAllocator
asserts / crashes; libc malloc corrupts the heap.
var data: std.ArrayList(u8) = .empty;
defer data.deinit(server.allocator); // ① deferred free
...
server.buffers.pushOwned(name, data.items) catch return .err;
data.items = &.{}; // ② ptr → comptime static, capacity kept
return .ok;
In Zig 0.16 unmanaged deinit executes gpa.free(self.items.ptr[0..self.capacity]).
After line ② items.ptr points at static storage while capacity retains the
≥4096 value from the appends, so ① frees a non-heap pointer with a bogus span.
No test covers this path (only table lookup("load-buffer")).
Replace data.items = &.{}; with data = .empty; (resets ptr and
capacity), or drop the defer once ownership has transferred and document it.
Regression test: run cmdLoadBuffer against a temp file under
testing.allocator and verify leak-check passes.