arm64: bpf: fix signedness bug in loading 64-bit immediate

Consider "(u64)insn1.imm << 32 | imm" in the arm64 JIT. Since imm is signed 32-bit, it is sign-extended to 64-bit, losing the high 32 bits. The fix is to convert imm to u32 first and zero-extend it to u64.

Also extend test_bpf to catch this JIT bug; the interpreter is correct.

Before:
test_bpf: #58 load 64-bit immediate ret -1 != 1 FAIL (1 times)

After:
test_bpf: #58 load 64-bit immediate 74 PASS

Fixes 30d3d94cc3d5 ("arm64: bpf: add 'load 64-bit immediate' instruction")

Cc Zi Shen Lim <[email protected]>

Cc Alexei Starovoitov <[email protected]>

Cc Catalin Marinas <[email protected]>

Cc Will Deacon <[email protected]>

Signed-off-by Xi Wang <[email protected]>

---
 arch/arm64/net/bpf_jit_comp.c | 2 +-
 lib/test_bpf.c                | 3 ++-
 2 files changed, 3 insertions(+), 2 deletions(-)

arch/arm64/net/bpf_jit_comp.c +0 -0

index edba042b2325..14cdc099fda0 100644

arch/arm64/net/bpf_jit_comp.c +0 -0

 		emit_a64_mov_i64(dst, imm64, ctx);
 
 		return 1;

lib/test_bpf.c +0 -0

index 80d78c51f65f..9f6849891b5f 100644

arch/arm64/net/bpf_jit_comp.c emit_cond_jmp: · line 490

Quoted 2 lines

arch/arm64/net/bpf_jit_comp.c +1 -1

-             imm64 = (u64)insn1.imm << 32 | imm;
+             imm64 = ((u64)(u32)insn1.imm) << 32 | (u64)(u32)imm;

This seems a bit convoluted to me. Don't you just need to add a (u32) cast to imm and that's it? The (u64)(u32) looks redundant.

lib/test_bpf.c static struct bpf_test tests[] = { · line 1760

Quoted 4 lines

lib/test_bpf.c +2 -1

-                     BPF_ALU64_IMM(BPF_MOV, R0, 1),
+                     BPF_LD_IMM64(R0, 0x1ffffffffLL),
+                     BPF_ALU64_IMM(BPF_RSH, R0, 32), /* R0 = 1 */
                      BPF_EXIT_INSN(),

This hunk should probably be a separate patch, unless you get Alexei's ack for me to take it via the arm64 tree too.

Quoted 5 lines · Will Deacon

Review

Quoted 2 lines

(quoted diff) +1 -1

-             imm64 = (u64)insn1.imm << 32 | imm;
+             imm64 = ((u64)(u32)insn1.imm) << 32 | (u64)(u32)imm;

This seems a bit convoluted to me. Don't you just need to add a (u32) cast to imm and that's it? The (u64)(u32) looks redundant.

You're right - the second (u64) is redundant; the hope was to make the code easier to understand. It's from the interpreter code in kernel/core/bpf.c, which uses (u64)(u32) as well.

Quoted 5 lines

Quoted 4 lines

(quoted diff) +2 -1

-                     BPF_ALU64_IMM(BPF_MOV, R0, 1),
+                     BPF_LD_IMM64(R0, 0x1ffffffffLL),
+                     BPF_ALU64_IMM(BPF_RSH, R0, 32), /* R0 = 1 */
                      BPF_EXIT_INSN(),

Quoted 2 lines

This hunk should probably be a separate patch, unless you get Alexei's ack for me to take it via the arm64 tree too.

I would be happy to split this into a separate patch if that works better, or simply drop this part.

From 1e4df6b7208140f3c49f316d33a409d3a161f350 Mon Sep 17 00:00:00 2001

Ok, I plan to apply the patch below for 4.1.