Re: [PATCH] perf unwind-libdw: Fix reading the stack of a 32-bit task

[Stian Halseth <stian@itx.no>]

[-- Attachment #1: Type: text/plain, Size: 1383 bytes --]

On Wed, 2026-09-23 at 19:45 +0000, sashiko-bot@kernel.org wrote:
> 
> [Severity: Low]
> Does this bitwise logic truncate the offset to a 4-byte boundary?

Yes, it does, but only for an address that isn't 4-byte aligned, and I
don't think that can happen here.

libdw only uses memory_read() to restore registers, and the save slots
are always word-aligned on a 32-bit ABI.

In theory, a malformed CFI could ask for one, though. But if that were
the case, the read would stay inside the stack dump, and the unwind
would be wrong either way. On a strict-alignment machine like sparc,
the native reads would trap on such an address as well.

I'd rather avoid adding handling for a case that shouldn't occur, butif preferred, I can make the swapped case refuse an unaligned read 
instead of returning the aligned word.
> 
> If a DWARF CFI expression requests a stack read at an unaligned
> address
> during cross-endian unwinding, the calculation discards the unaligned
> bits.
> 
> While register unwinding typically targets 4-byte aligned addresses
> on
> 32-bit architectures, could a malformed or custom ELF with non-
> standard
> CFI trigger an unaligned read? If so, would this extract the wrong
> bytes
> instead of the requested data, potentially corrupting the unwind
> results?
> 
> > +	return bswap_32(u.val32[(offset & 4) / 4]);
> > +}
> 

[-- Attachment #2: This is a digitally signed message part --]
[-- Type: application/pgp-signature, Size: 228 bytes --]