[1/2] MIPS: eBPF: Always return sign extended 32b values

Commit Message

Paul Burton Feb. 15, 2019, 8:14 p.m. UTC

The function prototype used to call JITed eBPF code (ie. the type of the
struct bpf_prog bpf_func field) returns an unsigned int. The MIPS n64
ABI that MIPS64 kernels target defines that 32 bit integers should
always be sign extended when passed in registers as either arguments or
return values.

This means that when returning any value which may not already be sign
extended (ie. of type REG_64BIT or REG_32BIT_ZERO_EX) we need to perform
that sign extension in order to comply with the n64 ABI. Without this we
see strange looking test failures from test_bpf.ko, such as:

  test_bpf: #65 ALU64_MOV_X:
    dst = 4294967295 jited:1 ret -1 != -1 FAIL (1 times)

Although the return value printed matches the expected value, this is
only because printf is only examining the least significant 32 bits of
the 64 bit register value we returned. The register holding the expected
value is sign extended whilst the v0 register was set to a zero extended
value by our JITed code, so when compared by a conditional branch
instruction the values are not equal.

We already handle this when the return value register is of type
REG_32BIT_ZERO_EX, so simply extend this to also cover REG_64BIT.

Signed-off-by: Paul Burton <[e-mail]>
Fixes: b6bd53f9c4e8 ("MIPS: Add missing file for eBPF JIT.")
Cc: [e-mail] # v4.13+
---

 arch/mips/net/ebpf_jit.c | 9 ++++++---
 1 file changed, 6 insertions(+), 3 deletions(-)

Comments

Paul Burton Feb. 19, 2019, 8:48 p.m. UTC | #1

Hello,

Paul Burton wrote:
> The function prototype used to call JITed eBPF code (ie. the type of the
> struct bpf_prog bpf_func field) returns an unsigned int. The MIPS n64
> ABI that MIPS64 kernels target defines that 32 bit integers should
> always be sign extended when passed in registers as either arguments or
> return values.
> 
> This means that when returning any value which may not already be sign
> extended (ie. of type REG_64BIT or REG_32BIT_ZERO_EX) we need to perform
> that sign extension in order to comply with the n64 ABI. Without this we
> see strange looking test failures from test_bpf.ko, such as:
> 
> test_bpf: #65 ALU64_MOV_X:
> dst = 4294967295 jited:1 ret -1 != -1 FAIL (1 times)
> 
> Although the return value printed matches the expected value, this is
> only because printf is only examining the least significant 32 bits of
> the 64 bit register value we returned. The register holding the expected
> value is sign extended whilst the v0 register was set to a zero extended
> value by our JITed code, so when compared by a conditional branch
> instruction the values are not equal.
> 
> We already handle this when the return value register is of type
> REG_32BIT_ZERO_EX, so simply extend this to also cover REG_64BIT.
> 
> Signed-off-by: Paul Burton <[e-mail]>
> Fixes: b6bd53f9c4e8 ("MIPS: Add missing file for eBPF JIT.")
> Cc: [e-mail] # v4.13+

Series applied to mips-next.

Thanks,
    Paul

[ This message was auto-generated; if you believe anything is incorrect
  then please email [e-mail] to report it. ]

Patch

diff --git a/arch/mips/net/ebpf_jit.c b/arch/mips/net/ebpf_jit.c
index b16710a8a9e7..715415fa2345 100644
--- a/arch/mips/net/ebpf_jit.c
+++ b/arch/mips/net/ebpf_jit.c
@@ -343,12 +343,15 @@  static int build_int_epilogue(struct jit_ctx *ctx, int dest_reg)
 	const struct bpf_prog *prog = ctx->skf;
 	int stack_adjust = ctx->stack_size;
 	int store_offset = stack_adjust - 8;
+	enum reg_val_type td;
 	int r0 = MIPS_R_V0;
 
-	if (dest_reg == MIPS_R_RA &&
-	    get_reg_val_type(ctx, prog->len, BPF_REG_0) == REG_32BIT_ZERO_EX)
+	if (dest_reg == MIPS_R_RA) {
 		/* Don't let zero extended value escape. */
-		emit_instr(ctx, sll, r0, r0, 0);
+		td = get_reg_val_type(ctx, prog->len, BPF_REG_0);
+		if (td == REG_64BIT || td == REG_32BIT_ZERO_EX)
+			emit_instr(ctx, sll, r0, r0, 0);
+	}
 
 	if (ctx->flags & EBPF_SAVE_RA) {
 		emit_instr(ctx, ld, MIPS_R_RA, store_offset, MIPS_R_SP);