Daniel Borkmann <daniel@xxxxxxxxxxxxx> writes:
> On 3/21/24 1:46 PM, Puranjay Mohan wrote:
>> With BPF_PROBE_MEM, BPF allows de-referencing an untrusted pointer. To
>> thwart invalid memory accesses, the JITs add an exception table entry
>> for all such accesses. But in case the src_reg + offset overflows and
>> turns into a userspace address, the BPF program might read that memory if
>> the user has mapped it.
>>
>> There are architectural features that prevent the kernel from accessing
>> userspace memory, like Privileged Access Never (PAN) on ARM64,
>> Supervisor Mode Access Prevention (SMAP) on x86-64, Supervisor User
>> Memory access (SUM) on RISC-V, etc. But BPF should not rely on the
>> existence of these features.
>>
>> Make the verifier add guard instructions around such memory accesses and
>> skip the load if the address falls into the userspace region.
>>
>> The JITs need to implement bpf_arch_uaddress_limit() to define where
>> the userspace addresses end for that architecture or TASK_SIZE is taken
>> as default.
>>
>> The implementation is as follows:
>>
>> REG_AX = SRC_REG
>> if(offset)
>> REG_AX += offset;
>> REG_AX >>= 32;
>> if (REG_AX <= (uaddress_limit >> 32))
>> DST_REG = 0;
>> else
>> DST_REG = *(size *)(SRC_REG + offset);
>>
>> Comparing just the upper 32 bits of the load address with the upper
>> 32 bits of uaddress_limit implies that the values are being aligned down
>> to a 4GB boundary before comparison.
>>
>> The above means that all loads with address <= uaddress_limit + 4GB are
>> skipped. This is acceptable because there is a large hole (much larger
>> than 4GB) between userspace and kernel space memory, therefore a
>> correctly functioning BPF program should not access this 4GB memory
>> above the userspace.
>>
>> Let's analyze what this patch does to the following fentry program
>> dereferencing an untrusted pointer:
>>
>> SEC("fentry/tcp_v4_connect")
>> int BPF_PROG(fentry_tcp_v4_connect, struct sock *sk)
>> {
>> *(volatile long *)sk;
>> return 0;
>> }
>>
>> BPF Program before | BPF Program after
>> ------------------ | -----------------
>>
>> 0: (79) r1 = *(u64 *)(r1 +0) 0: (79) r1 = *(u64 *)(r1 +0)
>> -----------------------------------------------------------------------
>> 1: (79) r1 = *(u64 *)(r1 +0) --\ 1: (bf) r11 = r1
>> ----------------------------\ \ 2: (77) r11 >>= 32
>> 2: (b7) r0 = 0 \ \ 3: (b5) if r11 <= 0x8000 goto pc+2
>> 3: (95) exit \ \-> 4: (79) r1 = *(u64 *)(r1 +0)
>> \ 5: (05) goto pc+1
>> \ 6: (b7) r1 = 0
>> \--------------------------------------
>> 7: (b7) r0 = 0
>> 8: (95) exit
>>
>> As you can see from above, in the best case (off=0), 5 extra instructions
>> are emitted.
>>
>> Now, we analyse the same program after it has gone through the JITs of
>> X86-64, ARM64, and RISC-V architectures. We follow the single load
>> instruction that has the untrusted pointer and see what instrumentation
>> has been added around it.
>>
>> x86-64 JIT
>> ==========
>> JIT's Instrumentation Verifier's Instrumentation
>> (upstream) (This patch)
>> --------------------- --------------------------
>>
>> 0: nopl 0x0(%rax,%rax,1) 0: nopl 0x0(%rax,%rax,1)
>> 5: xchg %ax,%ax 5: xchg %ax,%ax
>> 7: push %rbp 7: push %rbp
>> 8: mov %rsp,%rbp 8: mov %rsp,%rbp
>> b: mov 0x0(%rdi),%rdi b: mov 0x0(%rdi),%rdi
>> ------------------------------------------------------------------------
>> f: movabs $0x800000000000,%r11 f: mov %rdi,%r10
>> 19: cmp %r11,%rdi 12: shr $0x20,%r10
>> 1c: jb 0x000000000000002a 16: cmp $0x8000,%r10
>> 1e: mov %rdi,%r11 1d: jbe 0x0000000000000025
>> 21: add $0x0,%r11 /--> 1f: mov 0x0(%rdi),%rdi
>> 28: jae 0x000000000000002e / 23: jmp 0x0000000000000027
>> 2a: xor %edi,%edi / 25: xor %edi,%edi
>> 2c: jmp 0x0000000000000032 / /------------------------------------
>> 2e: mov 0x0(%rdi),%rdi ---/ / 27: xor %eax,%eax
>> ---------------------------------/ 29: leave
>> 32: xor %eax,%eax 2a: ret
>> 34: leave
>> 35: ret
>>
>> The x86-64 JIT already emits some instructions to protect against user
>> memory access. The implementation in this patch leads to a smaller
>> number of instructions being emitted. In the worst case the JIT will
>> emit 9 extra instructions and this patch decreases it to 7.
>>
>> ARM64 JIT
>> =========
>>
>> No Intrumentation Verifier's Instrumentation
>> (upstream) (This patch)
>> ----------------- --------------------------
>>
>> 0: add x9, x30, #0x0 0: add x9, x30, #0x0
>> 4: nop 4: nop
>> 8: paciasp 8: paciasp
>> c: stp x29, x30, [sp, #-16]! c: stp x29, x30, [sp, #-16]!
>> 10: mov x29, sp 10: mov x29, sp
>> 14: stp x19, x20, [sp, #-16]! 14: stp x19, x20, [sp, #-16]!
>> 18: stp x21, x22, [sp, #-16]! 18: stp x21, x22, [sp, #-16]!
>> 1c: stp x25, x26, [sp, #-16]! 1c: stp x25, x26, [sp, #-16]!
>> 20: stp x27, x28, [sp, #-16]! 20: stp x27, x28, [sp, #-16]!
>> 24: mov x25, sp 24: mov x25, sp
>> 28: mov x26, #0x0 28: mov x26, #0x0
>> 2c: sub x27, x25, #0x0 2c: sub x27, x25, #0x0
>> 30: sub sp, sp, #0x0 30: sub sp, sp, #0x0
>> 34: ldr x0, [x0] 34: ldr x0, [x0]
>> --------------------------------------------------------------------------------
>> 38: ldr x0, [x0] ----------\ 38: add x9, x0, #0x0
>> -----------------------------------\\ 3c: lsr x9, x9, #32
>> 3c: mov x7, #0x0 \\ 40: cmp x9, #0x10, lsl #12
>> 40: mov sp, sp \\ 44: b.ls 0x0000000000000050
>> 44: ldp x27, x28, [sp], #16 \\--> 48: ldr x0, [x0]
>> 48: ldp x25, x26, [sp], #16 \ 4c: b 0x0000000000000054
>> 4c: ldp x21, x22, [sp], #16 \ 50: mov x0, #0x0
>> 50: ldp x19, x20, [sp], #16 \---------------------------------------
>> 54: ldp x29, x30, [sp], #16 54: mov x7, #0x0
>> 58: add x0, x7, #0x0 58: mov sp, sp
>> 5c: autiasp 5c: ldp x27, x28, [sp], #16
>> 60: ret 60: ldp x25, x26, [sp], #16
>> 64: nop 64: ldp x21, x22, [sp], #16
>> 68: ldr x10, 0x0000000000000070 68: ldp x19, x20, [sp], #16
>> 6c: br x10 6c: ldp x29, x30, [sp], #16
>> 70: add x0, x7, #0x0
>> 74: autiasp
>> 78: ret
>> 7c: nop
>> 80: ldr x10, 0x0000000000000088
>> 84: br x10
>>
>> There are 6 extra instructions added in ARM64 in the best case. This will
>> become 7 in the worst case (off != 0).
>>
>> RISC-V JIT (RISCV_ISA_C Disabled)
>> ==========
>>
>> No Intrumentation Verifier's Instrumentation
>> (upstream) (This patch)
>> ----------------- --------------------------
>>
>> 0: nop 0: nop
>> 4: nop 4: nop
>> 8: li a6, 33 8: li a6, 33
>> c: addi sp, sp, -16 c: addi sp, sp, -16
>> 10: sd s0, 8(sp) 10: sd s0, 8(sp)
>> 14: addi s0, sp, 16 14: addi s0, sp, 16
>> 18: ld a0, 0(a0) 18: ld a0, 0(a0)
>> ---------------------------------------------------------------
>> 1c: ld a0, 0(a0) --\ 1c: mv t0, a0
>> --------------------------\ \ 20: srli t0, t0, 32
>> 20: li a5, 0 \ \ 24: lui t1, 4096
>> 24: ld s0, 8(sp) \ \ 28: sext.w t1, t1
>> 28: addi sp, sp, 16 \ \ 2c: bgeu t1, t0, 12
>> 2c: sext.w a0, a5 \ \--> 30: ld a0, 0(a0)
>> 30: ret \ 34: j 8
>> \ 38: li a0, 0
>> \------------------------------
>> 3c: li a5, 0
>> 40: ld s0, 8(sp)
>> 44: addi sp, sp, 16
>> 48: sext.w a0, a5
>> 4c: ret
>>
>> There are 7 extra instructions added in RISC-V.
>>
>> Fixes: 800834285361 ("bpf, arm64: Add BPF exception tables")
>> Reported-by: Breno Leitao <leitao@xxxxxxxxxx>
>> Suggested-by: Alexei Starovoitov <ast@xxxxxxxxxx>
>> Signed-off-by: Puranjay Mohan <puranjay12@xxxxxxxxx>
>> ---
>> V3: https://lore.kernel.org/bpf/20240321120842.78983-1-puranjay12@xxxxxxxxx/
>> Changes in V4:
>> - Disable this feature on architectures that don't define
>> CONFIG_ARCH_HAS_NON_OVERLAPPING_ADDRESS_SPACE.
>> - By doing the above, we don't need anything explicitly for s390x.
>>
>> V2: https://lore.kernel.org/bpf/20240321101058.68530-1-puranjay12@xxxxxxxxx/
>> Changes in V3:
>> - Return 0 from bpf_arch_uaddress_limit() in disabled case because it
>> returns u64.
>> - Modify the check in verifier to no do instrumentation when uaddress_limit
>> is 0.
>>
>> V1: https://lore.kernel.org/bpf/20240320105436.4781-1-puranjay12@xxxxxxxxx/
>> Changes in V2:
>> - Disable this feature on s390x.
>> ---
>> arch/x86/net/bpf_jit_comp.c | 72 +++++--------------------------------
>> include/linux/filter.h | 1 +
>> kernel/bpf/core.c | 9 +++++
>> kernel/bpf/verifier.c | 30 ++++++++++++++++
>> 4 files changed, 48 insertions(+), 64 deletions(-)
>>
>> diff --git a/arch/x86/net/bpf_jit_comp.c b/arch/x86/net/bpf_jit_comp.c
>> index 4900b1ee019f..9b3136187938 100644
>> --- a/arch/x86/net/bpf_jit_comp.c
>> +++ b/arch/x86/net/bpf_jit_comp.c
>> @@ -1327,7 +1327,6 @@ static int do_jit(struct bpf_prog *bpf_prog, int *addrs, u8 *image, u8 *rw_image
>> u8 b2 = 0, b3 = 0;
>> u8 *start_of_ldx;
>> s64 jmp_offset;
>> - s16 insn_off;
>> u8 jmp_cond;
>> u8 *func;
>> int nops;
>> @@ -1802,78 +1801,18 @@ st: if (is_imm8(insn->off))
>> case BPF_LDX | BPF_PROBE_MEMSX | BPF_B:
>> case BPF_LDX | BPF_PROBE_MEMSX | BPF_H:
>> case BPF_LDX | BPF_PROBE_MEMSX | BPF_W:
>> - insn_off = insn->off;
>> -
>> - if (BPF_MODE(insn->code) == BPF_PROBE_MEM ||
>> - BPF_MODE(insn->code) == BPF_PROBE_MEMSX) {
>> - /* Conservatively check that src_reg + insn->off is a kernel address:
>> - * src_reg + insn->off >= TASK_SIZE_MAX + PAGE_SIZE
>> - * src_reg is used as scratch for src_reg += insn->off and restored
>> - * after emit_ldx if necessary
>> - */
>> -
>> - u64 limit = TASK_SIZE_MAX + PAGE_SIZE;
>> - u8 *end_of_jmp;
>> -
>> - /* At end of these emitted checks, insn->off will have been added
>> - * to src_reg, so no need to do relative load with insn->off offset
>> - */
>> - insn_off = 0;
>> -
>> - /* movabsq r11, limit */
>> - EMIT2(add_1mod(0x48, AUX_REG), add_1reg(0xB8, AUX_REG));
>> - EMIT((u32)limit, 4);
>> - EMIT(limit >> 32, 4);
>> -
>> - if (insn->off) {
>> - /* add src_reg, insn->off */
>> - maybe_emit_1mod(&prog, src_reg, true);
>> - EMIT2_off32(0x81, add_1reg(0xC0, src_reg), insn->off);
>> - }
>> -
>> - /* cmp src_reg, r11 */
>> - maybe_emit_mod(&prog, src_reg, AUX_REG, true);
>> - EMIT2(0x39, add_2reg(0xC0, src_reg, AUX_REG));
>> -
>> - /* if unsigned '>=', goto load */
>> - EMIT2(X86_JAE, 0);
>> - end_of_jmp = prog;
>> -
>> - /* xor dst_reg, dst_reg */
>> - emit_mov_imm32(&prog, false, dst_reg, 0);
>> - /* jmp byte_after_ldx */
>> - EMIT2(0xEB, 0);
>> -
>> - /* populate jmp_offset for JAE above to jump to start_of_ldx */
>> - start_of_ldx = prog;
>> - end_of_jmp[-1] = start_of_ldx - end_of_jmp;
>> - }
>> + start_of_ldx = prog;
>> if (BPF_MODE(insn->code) == BPF_PROBE_MEMSX ||
>> BPF_MODE(insn->code) == BPF_MEMSX)
>> - emit_ldsx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn_off);
>> + emit_ldsx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn->off);
>> else
>> - emit_ldx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn_off);
>> + emit_ldx(&prog, BPF_SIZE(insn->code), dst_reg, src_reg, insn->off);
>> if (BPF_MODE(insn->code) == BPF_PROBE_MEM ||
>> BPF_MODE(insn->code) == BPF_PROBE_MEMSX) {
>> struct exception_table_entry *ex;
>> u8 *_insn = image + proglen + (start_of_ldx - temp);
>> s64 delta;
>>
>> - /* populate jmp_offset for JMP above */
>> - start_of_ldx[-1] = prog - start_of_ldx;
>> -
>> - if (insn->off && src_reg != dst_reg) {
>> - /* sub src_reg, insn->off
>> - * Restore src_reg after "add src_reg, insn->off" in prev
>> - * if statement. But if src_reg == dst_reg, emit_ldx
>> - * above already clobbered src_reg, so no need to restore.
>> - * If add src_reg, insn->off was unnecessary, no need to
>> - * restore either.
>> - */
>> - maybe_emit_1mod(&prog, src_reg, true);
>> - EMIT2_off32(0x81, add_1reg(0xE8, src_reg), insn->off);
>> - }
>> -
>> if (!bpf_prog->aux->extable)
>> break;
>>
>> @@ -3473,3 +3412,8 @@ bool bpf_jit_supports_ptr_xchg(void)
>> {
>> return true;
>> }
>> +
>> +u64 bpf_arch_uaddress_limit(void)
>> +{
>> + return TASK_SIZE_MAX + PAGE_SIZE;
>> +}
>> diff --git a/include/linux/filter.h b/include/linux/filter.h
>> index c0d51bff8f96..cf12bfa2a78c 100644
>> --- a/include/linux/filter.h
>> +++ b/include/linux/filter.h
>> @@ -965,6 +965,7 @@ bool bpf_jit_supports_far_kfunc_call(void);
>> bool bpf_jit_supports_exceptions(void);
>> bool bpf_jit_supports_ptr_xchg(void);
>> bool bpf_jit_supports_arena(void);
>> +u64 bpf_arch_uaddress_limit(void);
>> void arch_bpf_stack_walk(bool (*consume_fn)(void *cookie, u64 ip, u64 sp, u64 bp), void *cookie);
>> bool bpf_helper_changes_pkt_data(void *func);
>>
>> diff --git a/kernel/bpf/core.c b/kernel/bpf/core.c
>> index 5aacb1d3c4cc..a04695ca82b9 100644
>> --- a/kernel/bpf/core.c
>> +++ b/kernel/bpf/core.c
>> @@ -2958,6 +2958,15 @@ bool __weak bpf_jit_supports_arena(void)
>> return false;
>> }
>>
>> +u64 __weak bpf_arch_uaddress_limit(void)
>> +{
>> +#if defined(CONFIG_64BIT) && defined(CONFIG_ARCH_HAS_NON_OVERLAPPING_ADDRESS_SPACE)
>> + return TASK_SIZE;
>> +#else
>> + return 0;
>> +#endif
>> +}
>> +
>> /* Return TRUE if the JIT backend satisfies the following two conditions:
>> * 1) JIT backend supports atomic_xchg() on pointer-sized words.
>> * 2) Under the specific arch, the implementation of xchg() is the same
>> diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
>> index de7813947981..7ce56da6cfa4 100644
>> --- a/kernel/bpf/verifier.c
>> +++ b/kernel/bpf/verifier.c
>> @@ -19657,6 +19657,36 @@ static int do_misc_fixups(struct bpf_verifier_env *env)
>> goto next_insn;
>> }
>>
>> + /* Make it impossible to de-reference a userspace address */
>> + if (BPF_CLASS(insn->code) == BPF_LDX &&
>> + (BPF_MODE(insn->code) == BPF_PROBE_MEM ||
>> + BPF_MODE(insn->code) == BPF_PROBE_MEMSX)) {
>> + struct bpf_insn *patch = &insn_buf[0];
>> + u64 uaddress_limit = bpf_arch_uaddress_limit();
>> +
>> + if (!uaddress_limit)
>> + goto next_insn;
>> +
>> + *patch++ = BPF_MOV64_REG(BPF_REG_AX, insn->src_reg);
>> + if (insn->off)
>> + *patch++ = BPF_ALU64_IMM(BPF_ADD, BPF_REG_AX, insn->off);
>> + *patch++ = BPF_ALU64_IMM(BPF_RSH, BPF_REG_AX, 32);
>> + *patch++ = BPF_JMP_IMM(BPF_JLE, BPF_REG_AX, uaddress_limit >> 32, 2);
>> + *patch++ = *insn;
>> + *patch++ = BPF_JMP_IMM(BPF_JA, 0, 0, 1);
>> + *patch++ = BPF_MOV64_IMM(insn->dst_reg, 0);
>
> But how does this address other cases where we could fault e.g. non-canonical,
> vsyscall page, etc? Technically, we would have to call to copy_from_kernel_nofault_allowed()
> to really address all the cases aside from the overflow (good catch btw!) where kernel
> turns into user address.
So, we are trying to ~simulate a call to
copy_from_kernel_nofault_allowed() here. If the address under
consideration is below TASK_SIZE (TASK_SIZE + 4GB to be precise) then we
skip that load because that address could be mapped by the user.
If the address is above TASK_SIZE + 4GB, we allow the load and it could
cause a fault if the address is invalid, non-canonical etc. Taking the
fault is fine because JIT will add an exception table entry for
for that load with BPF_PBOBE_MEM.
The vsyscall page is special, this approach skips all loads from this
page. I am not sure if that is acceptable.
Thanks,
Puranjay
