[PATCH v3] bpf: fix x86 call stack alignment for external calls

[Thomas Monjalon]

21/01/2026 11:16, Marat Khalili:
> Correctly align stack pointer on x86 JIT if external calls are present.
> 
> According to x86-64 ABI (https://gitlab.com/x86-psABIs/x86-64-ABI,
> section 3.2.2 The Stack Frame) stack needs to be 16 (or more) bytes
> aligned immediately before the call instruction is executed. Once
> control has been transferred to the function entry point it is always
> off by 8 bytes. It means that JIT-compiled BPF function will always have
> its stack misaligned for any nested call unless it performs operations
> with the stack; even if it does use stack there is still 50% chance of
> stack being misaligned since it uses it in multiples of 8.
> 
> To solve the issue mark RBP as used whenever we have external function
> calls, and align RSP using AND instruction at the end of the prolog.
> Marking RBP as used triggers stack pointer saving in prolog and
> restoration in epilog.
> 
> Add tests for external calls from BPF program demonstrating the problem:
> * direct verification of a local variable alignment;
> * operations with 128-bit integers;
> * aligned and unaligned SSE2 instructions;
> * memcpy and rte_memcpy (may use vector instructions in their code).
> 
> (Such variety is needed because not all of these tests are available or
> reproduce the problem on all targets even when the problem exists.)
> 
> Fixes: cc752e43e079 ("bpf: add JIT compilation for x86_64 ISA")
> Cc: stable at dpdk.org
> 
> Signed-off-by: Marat Khalili <marat.khalili at huawei.com>
> Acked-by: Konstantin Ananyev <konstantin.ananyev at huawei.com>
> Tested-by: Konstantin Ananyev <konstantin.ananyev at huawei.com>

Applied, thanks.