From: Daniel Borkmann daniel@iogearbox.net

Commit f6b1b3bf0d5f681631a293cfe1ca934b81716f1e upstream.

One of the ugly leftovers from the early eBPF days is that div/mod operations based on registers have a hard-coded src_reg == 0 test in the interpreter as well as in JIT code generators that would return from the BPF program with exit code 0. This was basically adopted from cBPF interpreter for historical reasons.

There are multiple reasons why this is very suboptimal and prone to bugs. To name one: the return code mapping for such abnormal program exit of 0 does not always match with a suitable program type's exit code mapping. For example, '0' in tc means action 'ok' where the packet gets passed further up the stack, which is just undesirable for such cases (e.g. when implementing policy) and also does not match with other program types.

While trying to work out an exception handling scheme, I also noticed that programs crafted like the following will currently pass the verifier:

0: (bf) r6 = r1 1: (85) call pc+8 caller: R6=ctx(id=0,off=0,imm=0) R10=fp0,call_-1 callee: frame1: R1=ctx(id=0,off=0,imm=0) R10=fp0,call_1 10: (b4) (u32) r2 = (u32) 0 11: (b4) (u32) r3 = (u32) 1 12: (3c) (u32) r3 /= (u32) r2 13: (61) r0 = *(u32 *)(r1 +76) 14: (95) exit returning from callee: frame1: R0_w=pkt(id=0,off=0,r=0,imm=0) R1=ctx(id=0,off=0,imm=0) R2_w=inv0 R3_w=inv(id=0,umax_value=4294967295,var_off=(0x0; 0xffffffff)) R10=fp0,call_1 to caller at 2: R0_w=pkt(id=0,off=0,r=0,imm=0) R6=ctx(id=0,off=0,imm=0) R10=fp0,call_-1

from 14 to 2: R0=pkt(id=0,off=0,r=0,imm=0) R6=ctx(id=0,off=0,imm=0) R10=fp0,call_-1 2: (bf) r1 = r6 3: (61) r1 = *(u32 *)(r1 +80) 4: (bf) r2 = r0 5: (07) r2 += 8 6: (2d) if r2 > r1 goto pc+1 R0=pkt(id=0,off=0,r=8,imm=0) R1=pkt_end(id=0,off=0,imm=0) R2=pkt(id=0,off=8,r=8,imm=0) R6=ctx(id=0,off=0,imm=0) R10=fp0,call_-1 7: (71) r0 = *(u8 *)(r0 +0) 8: (b7) r0 = 1 9: (95) exit

from 6 to 8: safe processed 16 insns (limit 131072), stack depth 0+0

Basically what happens is that in the subprog we make use of a div/mod by 0 exception and in the 'normal' subprog's exit path we just return skb->data back to the main prog. This has the implication that the verifier thinks we always get a pkt pointer in R0 while we still have the implicit 'return 0' from the div as an alternative unconditional return path earlier. Thus, R0 then contains 0, meaning back in the parent prog we get the address range of [0x0, skb->data_end] as read and writeable. Similar can be crafted with other pointer register types.

Since i) BPF_ABS/IND is not allowed in programs that contain BPF to BPF calls (and generally it's also disadvised to use in native eBPF context), ii) unknown opcodes don't return zero anymore, iii) we don't return an exception code in dead branches, the only last missing case affected and to fix is the div/mod handling.

What we would really need is some infrastructure to propagate exceptions all the way to the original prog unwinding the current stack and returning that code to the caller of the BPF program. In user space such exception handling for similar runtimes is typically implemented with setjmp(3) and longjmp(3) as one possibility which is not available in the kernel, though (kgdb used to implement it in kernel long time ago). I implemented a PoC exception handling mechanism into the BPF interpreter with porting setjmp()/longjmp() into x86_64 and adding a new internal BPF_ABRT opcode that can use a program specific exception code for all exception cases we have (e.g. div/mod by 0, unknown opcodes, etc). While this seems to work in the constrained BPF environment (meaning, here, we don't need to deal with state e.g. from memory allocations that we would need to undo before going into exception state), it still has various drawbacks: i) we would need to implement the setjmp()/longjmp() for every arch supported in the kernel and for x86_64, arm64, sparc64 JITs currently supporting calls, ii) it has unconditional additional cost on main program entry to store CPU register state in initial setjmp() call, and we would need some way to pass the jmp_buf down into ___bpf_prog_run() for main prog and all subprogs, but also storing on stack is not really nice (other option would be per-cpu storage for this, but it also has the drawback that we need to disable preemption for every BPF program types). All in all this approach would add a lot of complexity.

Another poor-man's solution would be to have some sort of additional shared register or scratch buffer to hold state for exceptions, and test that after every call return to chain returns and pass R0 all the way down to BPF prog caller. This is also problematic in various ways: i) an additional register doesn't map well into JITs, and some other scratch space could only be on per-cpu storage, which, again has the side-effect that this only works when we disable preemption, or somewhere in the input context which is not available everywhere either, and ii) this adds significant runtime overhead by putting conditionals after each and every call, as well as implementation complexity.

Yet another option is to teach verifier that div/mod can return an integer, which however is also complex to implement as verifier would need to walk such fake 'mov r0,<code>; exit;' sequeuence and there would still be no guarantee for having propagation of this further down to the BPF caller as proper exception code. For parent prog, it is also is not distinguishable from a normal return of a constant scalar value.

The approach taken here is a completely different one with little complexity and no additional overhead involved in that we make use of the fact that a div/mod by 0 is undefined behavior. Instead of bailing out, we adapt the same behavior as on some major archs like ARMv8 [0] into eBPF as well: X div 0 results in 0, and X mod 0 results in X. aarch64 and aarch32 ISA do not generate any traps or otherwise aborts of program execution for unsigned divides. I verified this also with a test program compiled by gcc and clang, and the behavior matches with the spec. Going forward we adapt the eBPF verifier to emit such rewrites once div/mod by register was seen. cBPF is not touched and will keep existing 'return 0' semantics. Given the options, it seems the most suitable from all of them, also since major archs have similar schemes in place. Given this is all in the realm of undefined behavior, we still have the option to adapt if deemed necessary and this way we would also have the option of more flexibility from LLVM code generation side (which is then fully visible to verifier). Thus, this patch i) fixes the panic seen in above program and ii) doesn't bypass the verifier observations.

[0] ARM Architecture Reference Manual, ARMv8 [ARM DDI 0487B.b] http://infocenter.arm.com/help/topic/com.arm.doc.ddi0487b.b/DDI0487B_b_armv8... 1) aarch64 instruction set: section C3.4.7 and C6.2.279 (UDIV) "A division by zero results in a zero being written to the destination register, without any indication that the division by zero occurred." 2) aarch32 instruction set: section F1.4.8 and F5.1.263 (UDIV) "For the SDIV and UDIV instructions, division by zero always returns a zero result."

Fixes: f4d7e40a5b71 ("bpf: introduce function calls (verification)") Signed-off-by: Daniel Borkmann daniel@iogearbox.net Acked-by: Alexei Starovoitov ast@kernel.org Signed-off-by: Alexei Starovoitov ast@kernel.org Signed-off-by: Thadeu Lima de Souza Cascardo cascardo@canonical.com Signed-off-by: Edward Liaw edliaw@google.com --- kernel/bpf/core.c | 8 -------- kernel/bpf/verifier.c | 38 ++++++++++++++++++++++++++++++-------- net/core/filter.c | 9 ++++++++- 3 files changed, 38 insertions(+), 17 deletions(-)

diff --git a/kernel/bpf/core.c b/kernel/bpf/core.c index 4ddb846693bb..2ca36bb440de 100644 --- a/kernel/bpf/core.c +++ b/kernel/bpf/core.c @@ -1055,14 +1055,10 @@ static unsigned int ___bpf_prog_run(u64 *regs, const struct bpf_insn *insn, (*(s64 *) &DST) >>= IMM; CONT; ALU64_MOD_X: - if (unlikely(SRC == 0)) - return 0; div64_u64_rem(DST, SRC, &tmp); DST = tmp; CONT; ALU_MOD_X: - if (unlikely((u32)SRC == 0)) - return 0; tmp = (u32) DST; DST = do_div(tmp, (u32) SRC); CONT; @@ -1075,13 +1071,9 @@ static unsigned int ___bpf_prog_run(u64 *regs, const struct bpf_insn *insn, DST = do_div(tmp, (u32) IMM); CONT; ALU64_DIV_X: - if (unlikely(SRC == 0)) - return 0; DST = div64_u64(DST, SRC); CONT; ALU_DIV_X: - if (unlikely((u32)SRC == 0)) - return 0; tmp = (u32) DST; do_div(tmp, (u32) SRC); DST = (u32) tmp; diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c index e8d9ddd5cb18..836791403129 100644 --- a/kernel/bpf/verifier.c +++ b/kernel/bpf/verifier.c @@ -4839,15 +4839,37 @@ static int fixup_bpf_calls(struct bpf_verifier_env *env) struct bpf_insn_aux_data *aux;

for (i = 0; i < insn_cnt; i++, insn++) { - if (insn->code == (BPF_ALU | BPF_MOD | BPF_X) || + if (insn->code == (BPF_ALU64 | BPF_MOD | BPF_X) || + insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) || + insn->code == (BPF_ALU | BPF_MOD | BPF_X) || insn->code == (BPF_ALU | BPF_DIV | BPF_X)) { - /* due to JIT bugs clear upper 32-bits of src register - * before div/mod operation - */ - insn_buf[0] = BPF_MOV32_REG(insn->src_reg, insn->src_reg); - insn_buf[1] = *insn; - cnt = 2; - new_prog = bpf_patch_insn_data(env, i + delta, insn_buf, cnt); + bool is64 = BPF_CLASS(insn->code) == BPF_ALU64; + struct bpf_insn mask_and_div[] = { + BPF_MOV32_REG(insn->src_reg, insn->src_reg), + /* Rx div 0 -> 0 */ + BPF_JMP_IMM(BPF_JNE, insn->src_reg, 0, 2), + BPF_ALU32_REG(BPF_XOR, insn->dst_reg, insn->dst_reg), + BPF_JMP_IMM(BPF_JA, 0, 0, 1), + *insn, + }; + struct bpf_insn mask_and_mod[] = { + BPF_MOV32_REG(insn->src_reg, insn->src_reg), + /* Rx mod 0 -> Rx */ + BPF_JMP_IMM(BPF_JEQ, insn->src_reg, 0, 1), + *insn, + }; + struct bpf_insn *patchlet; + + if (insn->code == (BPF_ALU64 | BPF_DIV | BPF_X) || + insn->code == (BPF_ALU | BPF_DIV | BPF_X)) { + patchlet = mask_and_div + (is64 ? 1 : 0); + cnt = ARRAY_SIZE(mask_and_div) - (is64 ? 1 : 0); + } else { + patchlet = mask_and_mod + (is64 ? 1 : 0); + cnt = ARRAY_SIZE(mask_and_mod) - (is64 ? 1 : 0); + } + + new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt); if (!new_prog) return -ENOMEM;

diff --git a/net/core/filter.c b/net/core/filter.c index 29d85a20f4fc..fc665885d57c 100644 --- a/net/core/filter.c +++ b/net/core/filter.c @@ -458,8 +458,15 @@ static int bpf_convert_filter(struct sock_filter *prog, int len, break;

if (fp->code == (BPF_ALU | BPF_DIV | BPF_X) || - fp->code == (BPF_ALU | BPF_MOD | BPF_X)) + fp->code == (BPF_ALU | BPF_MOD | BPF_X)) { *insn++ = BPF_MOV32_REG(BPF_REG_X, BPF_REG_X); + /* Error with exception code on div/mod by 0. + * For cBPF programs, this was always return 0. + */ + *insn++ = BPF_JMP_IMM(BPF_JNE, BPF_REG_X, 0, 2); + *insn++ = BPF_ALU32_REG(BPF_XOR, BPF_REG_A, BPF_REG_A); + *insn++ = BPF_EXIT_INSN(); + }

*insn = BPF_RAW_INSN(fp->code, BPF_REG_A, BPF_REG_X, 0, fp->k); break;

From: Daniel Borkmann daniel@iogearbox.net

Commit 9b00f1b78809309163dda2d044d9e94a3c0248a3 upstream.

Recently noticed that when mod32 with a known src reg of 0 is performed, then the dst register is 32-bit truncated in verifier:

0: R1=ctx(id=0,off=0,imm=0) R10=fp0 0: (b7) r0 = 0 1: R0_w=inv0 R1=ctx(id=0,off=0,imm=0) R10=fp0 1: (b7) r1 = -1 2: R0_w=inv0 R1_w=inv-1 R10=fp0 2: (b4) w2 = -1 3: R0_w=inv0 R1_w=inv-1 R2_w=inv4294967295 R10=fp0 3: (9c) w1 %= w0 4: R0_w=inv0 R1_w=inv(id=0,umax_value=4294967295,var_off=(0x0; 0xffffffff)) R2_w=inv4294967295 R10=fp0 4: (b7) r0 = 1 5: R0_w=inv1 R1_w=inv(id=0,umax_value=4294967295,var_off=(0x0; 0xffffffff)) R2_w=inv4294967295 R10=fp0 5: (1d) if r1 == r2 goto pc+1 R0_w=inv1 R1_w=inv(id=0,umax_value=4294967295,var_off=(0x0; 0xffffffff)) R2_w=inv4294967295 R10=fp0 6: R0_w=inv1 R1_w=inv(id=0,umax_value=4294967295,var_off=(0x0; 0xffffffff)) R2_w=inv4294967295 R10=fp0 6: (b7) r0 = 2 7: R0_w=inv2 R1_w=inv(id=0,umax_value=4294967295,var_off=(0x0; 0xffffffff)) R2_w=inv4294967295 R10=fp0 7: (95) exit 7: R0=inv1 R1=inv(id=0,umin_value=4294967295,umax_value=4294967295,var_off=(0x0; 0xffffffff)) R2=inv4294967295 R10=fp0 7: (95) exit

However, as a runtime result, we get 2 instead of 1, meaning the dst register does not contain (u32)-1 in this case. The reason is fairly straight forward given the 0 test leaves the dst register as-is:

# ./bpftool p d x i 23 0: (b7) r0 = 0 1: (b7) r1 = -1 2: (b4) w2 = -1 3: (16) if w0 == 0x0 goto pc+1 4: (9c) w1 %= w0 5: (b7) r0 = 1 6: (1d) if r1 == r2 goto pc+1 7: (b7) r0 = 2 8: (95) exit

This was originally not an issue given the dst register was marked as completely unknown (aka 64 bit unknown). However, after 468f6eafa6c4 ("bpf: fix 32-bit ALU op verification") the verifier casts the register output to 32 bit, and hence it becomes 32 bit unknown. Note that for the case where the src register is unknown, the dst register is marked 64 bit unknown. After the fix, the register is truncated by the runtime and the test passes:

# ./bpftool p d x i 23 0: (b7) r0 = 0 1: (b7) r1 = -1 2: (b4) w2 = -1 3: (16) if w0 == 0x0 goto pc+2 4: (9c) w1 %= w0 5: (05) goto pc+1 6: (bc) w1 = w1 7: (b7) r0 = 1 8: (1d) if r1 == r2 goto pc+1 9: (b7) r0 = 2 10: (95) exit

Semantics also match with {R,W}x mod{64,32} 0 -> {R,W}x. Invalid div has always been {R,W}x div{64,32} 0 -> 0. Rewrites are as follows:

mod32: mod64:

(16) if w0 == 0x0 goto pc+2 (15) if r0 == 0x0 goto pc+1 (9c) w1 %= w0 (9f) r1 %= r0 (05) goto pc+1 (bc) w1 = w1

[Salvatore Bonaccorso: This is an earlier version based on work by Daniel and John which does not rely on availability of the BPF_JMP32 instruction class. This means it is not even strictly a backport of the upstream commit mentioned but based on Daniel's and John's work to address the issue and was finalized by Thadeu Lima de Souza Cascardo.]

Fixes: 468f6eafa6c4 ("bpf: fix 32-bit ALU op verification") Signed-off-by: Daniel Borkmann daniel@iogearbox.net Reviewed-by: John Fastabend john.fastabend@gmail.com Tested-by: Salvatore Bonaccorso carnil@debian.org Signed-off-by: Thadeu Lima de Souza Cascardo cascardo@canonical.com Signed-off-by: Edward Liaw edliaw@google.com --- kernel/bpf/verifier.c | 9 +++++---- 1 file changed, 5 insertions(+), 4 deletions(-)

diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c index 9f04d413df92..a55e264cdb54 100644 --- a/kernel/bpf/verifier.c +++ b/kernel/bpf/verifier.c @@ -4846,7 +4846,7 @@ static int fixup_bpf_calls(struct bpf_verifier_env *env) bool is64 = BPF_CLASS(insn->code) == BPF_ALU64; struct bpf_insn mask_and_div[] = { BPF_MOV_REG(BPF_CLASS(insn->code), BPF_REG_AX, insn->src_reg), - /* Rx div 0 -> 0 */ + /* [R,W]x div 0 -> 0 */ BPF_JMP_IMM(BPF_JEQ, BPF_REG_AX, 0, 2), BPF_RAW_REG(*insn, insn->dst_reg, BPF_REG_AX), BPF_JMP_IMM(BPF_JA, 0, 0, 1), @@ -4854,9 +4854,10 @@ static int fixup_bpf_calls(struct bpf_verifier_env *env) }; struct bpf_insn mask_and_mod[] = { BPF_MOV_REG(BPF_CLASS(insn->code), BPF_REG_AX, insn->src_reg), - /* Rx mod 0 -> Rx */ - BPF_JMP_IMM(BPF_JEQ, BPF_REG_AX, 0, 1), + BPF_JMP_IMM(BPF_JEQ, BPF_REG_AX, 0, 1 + (is64 ? 0 : 1)), BPF_RAW_REG(*insn, insn->dst_reg, BPF_REG_AX), + BPF_JMP_IMM(BPF_JA, 0, 0, 1), + BPF_MOV32_REG(insn->dst_reg, insn->dst_reg), }; struct bpf_insn *patchlet;

@@ -4866,7 +4867,7 @@ static int fixup_bpf_calls(struct bpf_verifier_env *env) cnt = ARRAY_SIZE(mask_and_div); } else { patchlet = mask_and_mod; - cnt = ARRAY_SIZE(mask_and_mod); + cnt = ARRAY_SIZE(mask_and_mod) - (is64 ? 2 : 0); }

new_prog = bpf_patch_insn_data(env, i + delta, patchlet, cnt);