On 11/22/2024 08:41 AM, John Fastabend wrote:
Tiezhu Yang wrote:
(1) Description of Problem:
When testing BPF JIT with the latest compiler toolchains on LoongArch,
there exist some strange failed test cases, dmesg shows something like
this:
# dmesg -t | grep FAIL | head -1
... ret -3 != -3 (0xfffffffd != 0xfffffffd)FAIL ...
...
(5) Final Solution:
Keep a5 zero-extended, but explicitly sign-extend a0 (which is used
outside BPF land). Because libbpf currently defines the return value
of an ebpf program as a 32-bit unsigned integer, just use addi.w to
extend bit 31 into bits 63 through 32 of a5 to a0. This is similar
with commit 2f1b0d3d7331 ("riscv, bpf: Sign-extend return values").
Fixes: 5dc615520c4d ("LoongArch: Add BPF JIT support")
Signed-off-by: Tiezhu Yang <yangtiezhu@xxxxxxxxxxx>
---
arch/loongarch/net/bpf_jit.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/arch/loongarch/net/bpf_jit.c b/arch/loongarch/net/bpf_jit.c
index 7dbefd4ba210..dd350cba1252 100644
--- a/arch/loongarch/net/bpf_jit.c
+++ b/arch/loongarch/net/bpf_jit.c
@@ -179,7 +179,7 @@ static void __build_epilogue(struct jit_ctx *ctx, bool is_tail_call)
if (!is_tail_call) {
/* Set return value */
- move_reg(ctx, LOONGARCH_GPR_A0, regmap[BPF_REG_0]);
+ emit_insn(ctx, addiw, LOONGARCH_GPR_A0, regmap[BPF_REG_0], 0);
Not overly familiar with this JIT but just to check this wont be used
for BPF 2 BPF calls correct?
I am not sure I understand your comment correctly, but with and without
this patch, the LoongArch JIT uses a5 as a dedicated register for BPF
return values, a5 is kept as zero-extended for bpf2bpf, just make a0
(which is used outside BPF land) as sign-extend, all of the test cases
in test_bpf.ko passed with "echo 1 > /proc/sys/net/core/bpf_jit_enable".
Thanks,
Tiezhu
