Undefined behavior in tidb_query_datatype's RowSlice

[brson]

Bug Report

RowSlice does an unsafe pointer cast that is often unaligned. This is undefined behavior, which could potentially lead to miscompilation.

Here is the source:

tikv/components/tidb_query_datatype/src/codec/row/v2/row_slice.rs

Line 149 in 7565efc

let slice = unsafe { std::slice::from_raw_parts(buf.as_ptr() as *const T, len) };

The fix isn't simple, so I'm just filing a bug cc @breeswish.

Potential fixes are to accumulate the casted items into a new buffer instead of slicing the original buffer, but that costs an allocation. Another fix would be to change the serialization such that the cast always ends up aligned, while adding an assertion here that the alignment is correct.

Tested against commit adf3a94

Found with miri (cc @oli-obk 👍 ).

[breezewish]

Maybe we can drop the usage of slice at all. The slice is constructed to reuse the binary search method from std. We can write our own binary search using explicitly unaligned memory access.

[breezewish]

@zhongzc PTAL

[zhongzc]

Maybe we can drop the usage of slice at all. The slice is constructed to reuse the binary search method from std. We can write our own binary search using explicitly unaligned memory access.

Writing our own binary search has to deal with the endian at every comparison. Or any other good idea?