bpf: Fix 32 bit src register truncation on div/mod

bpf: Fix 32 bit src register truncation on div/mod

Commit e88b2c6e5a4d9ce30d75391e4d950da74bb2bd90 upstream.

While reviewing a different fix, John and I noticed an oddity in one of the
BPF program dumps that stood out, for example:

  # bpftool p d x i 13
   0: (b7) r0 = 808464450
   1: (b4) w4 = 808464432
   2: (bc) w0 = w0
   3: (15) if r0 == 0x0 goto pc+1
   4: (9c) w4 %= w0
  [...]

In line 2 we noticed that the mov32 would 32 bit truncate the original src
register for the div/mod operation. While for the two operations the dst
register is typically marked unknown e.g. from adjust_scalar_min_max_vals()
the src register is not, and thus verifier keeps tracking original bounds,
simplified:

  0: R1=ctx(id=0,off=0,imm=0) R10=fp0
  0: (b7) r0 = -1
  1: R0_w=invP-1 R1=ctx(id=0,off=0,imm=0) R10=fp0
  1: (b7) r1 = -1
  2: R0_w=invP-1 R1_w=invP-1 R10=fp0
  2: (3c) w0 /= w1
  3: R0_w=invP(id=0,umax_value=4294967295,var_off=(0x0; 0xffffffff)) R1_w=invP-1 R10=fp0
  3: (77) r1 >>= 32
  4: R0_w=invP(id=0,umax_value=4294967295,var_off=(0x0; 0xffffffff)) R1_w=invP4294967295 R10=fp0
  4: (bf) r0 = r1
  5: R0_w=invP4294967295 R1_w=invP4294967295 R10=fp0
  5: (95) exit
  processed 6 insns (limit 1000000) max_states_per_insn 0 total_states 0 peak_states 0 mark_read 0

Runtime result of r0 at exit is 0 instead of expected -1. Remove the
verifier mov32 src rewrite in div/mod and replace it with a jmp32 test
instead. After the fix, we result in the following code generation when
having dividend r1 and divisor r6:

  div, 64 bit:                             div, 32 bit:

   0: (b7) r6 = 8                           0: (b7) r6 = 8
   1: (b7) r1 = 8                           1: (b7) r1 = 8
   2: (55) if r6 != 0x0 goto pc+2           2: (56) if w6 != 0x0 goto pc+2
   3: (ac) w1 ^= w1                         3: (ac) w1 ^= w1
   4: (05) goto pc+1                        4: (05) goto pc+1
   5: (3f) r1 /= r6                         5: (3c) w1 /= w6
   6: (b7) r0 = 0                           6: (b7) r0 = 0
   7: (95) exit                             7: (95) exit

  mod, 64 bit:                             mod, 32 bit:

   0: (b7) r6 = 8                           0: (b7) r6 = 8
   1: (b7) r1 = 8                           1: (b7) r1 = 8
   2: (15) if r6 == 0x0 goto pc+1           2: (16) if w6 == 0x0 goto pc+1
   3: (9f) r1 %= r6                         3: (9c) w1 %= w6
   4: (b7) r0 = 0                           4: (b7) r0 = 0
   5: (95) exit                             5: (95) exit

x86 in particular can throw a 'divide error' exception for div
instruction not only for divisor being zero, but also for the case
when the quotient is too large for the designated register. For the
edx:eax and rdx:rax dividend pair it is not an issue in x86 BPF JIT
since we always zero edx (rdx). Hence really the only protection
needed is against divisor being zero.

[Salvatore Bonaccorso: This is an earlier version of the patch provided
by Daniel Borkmann 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.]

Fixes: 68fda450a7df ("bpf: fix 32-bit divide by zero")
Co-developed-by: John Fastabend <john.fastabend@gmail.com>
Signed-off-by: John Fastabend <john.fastabend@gmail.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
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>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>