bpf: fix mixed signed/unsigned derived min/max value bounds

bpf: fix mixed signed/unsigned derived min/max value bounds

Edward reported that there's an issue in min/max value bounds
tracking when signed and unsigned compares both provide hints
on limits when having unknown variables. E.g. a program such
as the following should have been rejected:

   0: (7a) *(u64 *)(r10 -8) = 0
   1: (bf) r2 = r10
   2: (07) r2 += -8
   3: (18) r1 = 0xffff8a94cda93400
   5: (85) call bpf_map_lookup_elem#1
   6: (15) if r0 == 0x0 goto pc+7
  R0=map_value(ks=8,vs=8,id=0),min_value=0,max_value=0 R10=fp
   7: (7a) *(u64 *)(r10 -16) = -8
   8: (79) r1 = *(u64 *)(r10 -16)
   9: (b7) r2 = -1
  10: (2d) if r1 > r2 goto pc+3
  R0=map_value(ks=8,vs=8,id=0),min_value=0,max_value=0 R1=inv,min_value=0
  R2=imm-1,max_value=18446744073709551615,min_align=1 R10=fp
  11: (65) if r1 s> 0x1 goto pc+2
  R0=map_value(ks=8,vs=8,id=0),min_value=0,max_value=0 R1=inv,min_value=0,max_value=1
  R2=imm-1,max_value=18446744073709551615,min_align=1 R10=fp
  12: (0f) r0 += r1
  13: (72) *(u8 *)(r0 +0) = 0
  R0=map_value_adj(ks=8,vs=8,id=0),min_value=0,max_value=1 R1=inv,min_value=0,max_value=1
  R2=imm-1,max_value=18446744073709551615,min_align=1 R10=fp
  14: (b7) r0 = 0
  15: (95) exit

What happens is that in the first part ...

   8: (79) r1 = *(u64 *)(r10 -16)
   9: (b7) r2 = -1
  10: (2d) if r1 > r2 goto pc+3

... r1 carries an unsigned value, and is compared as unsigned
against a register carrying an immediate. Verifier deduces in
reg_set_min_max() that since the compare is unsigned and operation
is greater than (>), that in the fall-through/false case, r1's
minimum bound must be 0 and maximum bound must be r2. Latter is
larger than the bound and thus max value is reset back to being
'invalid' aka BPF_REGISTER_MAX_RANGE. Thus, r1 state is now
'R1=inv,min_value=0'. The subsequent test ...

  11: (65) if r1 s> 0x1 goto pc+2

... is a signed compare of r1 with immediate value 1. Here,
verifier deduces in reg_set_min_max() that since the compare
is signed this time and operation is greater than (>), that
in the fall-through/false case, we can deduce that r1's maximum
bound must be 1, meaning with prior test, we result in r1 having
the following state: R1=inv,min_value=0,max_value=1. Given that
the actual value this holds is -8, the bounds are wrongly deduced.
When this is being added to r0 which holds the map_value(_adj)
type, then subsequent store access in above case will go through
check_mem_access() which invokes check_map_access_adj(), that
will then probe whether the map memory is in bounds based
on the min_value and max_value as well as access size since
the actual unknown value is min_value <= x <= max_value; commit
fce366a9dd0d ("bpf, verifier: fix alu ops against map_value{,
_adj} register types") provides some more explanation on the
semantics.

It's worth to note in this context that in the current code,
min_value and max_value tracking are used for two things, i)
dynamic map value access via check_map_access_adj() and since
commit 06c1c049721a ("bpf: allow helpers access to variable memory")
ii) also enforced at check_helper_mem_access() when passing a
memory address (pointer to packet, map value, stack) and length
pair to a helper and the length in this case is an unknown value
defining an access range through min_value/max_value in that
case. The min_value/max_value tracking is /not/ used in the
direct packet access case to track ranges. However, the issue
also affects case ii), for example, the following crafted program
based on the same principle must be rejected as well:

   0: (b7) r2 = 0
   1: (bf) r3 = r10
   2: (07) r3 += -512
   3: (7a) *(u64 *)(r10 -16) = -8
   4: (79) r4 = *(u64 *)(r10 -16)
   5: (b7) r6 = -1
   6: (2d) if r4 > r6 goto pc+5
  R1=ctx R2=imm0,min_value=0,max_value=0,min_align=2147483648 R3=fp-512
  R4=inv,min_value=0 R6=imm-1,max_value=18446744073709551615,min_align=1 R10=fp
   7: (65) if r4 s> 0x1 goto pc+4
  R1=ctx R2=imm0,min_value=0,max_value=0,min_align=2147483648 R3=fp-512
  R4=inv,min_value=0,max_value=1 R6=imm-1,max_value=18446744073709551615,min_align=1
  R10=fp
   8: (07) r4 += 1
   9: (b7) r5 = 0
  10: (6a) *(u16 *)(r10 -512) = 0
  11: (85) call bpf_skb_load_bytes#26
  12: (b7) r0 = 0
  13: (95) exit

Meaning, while we initialize the max_value stack slot that the
verifier thinks we access in the [1,2] range, in reality we
pass -7 as length which is interpreted as u32 in the helper.
Thus, this issue is relevant also for the case of helper ranges.
Resetting both bounds in check_reg_overflow() in case only one
of them exceeds limits is also not enough as similar test can be
created that uses values which are within range, thus also here
learned min value in r1 is incorrect when mixed with later signed
test to create a range:

   0: (7a) *(u64 *)(r10 -8) = 0
   1: (bf) r2 = r10
   2: (07) r2 += -8
   3: (18) r1 = 0xffff880ad081fa00
   5: (85) call bpf_map_lookup_elem#1
   6: (15) if r0 == 0x0 goto pc+7
  R0=map_value(ks=8,vs=8,id=0),min_value=0,max_value=0 R10=fp
   7: (7a) *(u64 *)(r10 -16) = -8
   8: (79) r1 = *(u64 *)(r10 -16)
   9: (b7) r2 = 2
  10: (3d) if r2 >= r1 goto pc+3
  R0=map_value(ks=8,vs=8,id=0),min_value=0,max_value=0 R1=inv,min_value=3
  R2=imm2,min_value=2,max_value=2,min_align=2 R10=fp
  11: (65) if r1 s> 0x4 goto pc+2
  R0=map_value(ks=8,vs=8,id=0),min_value=0,max_value=0
  R1=inv,min_value=3,max_value=4 R2=imm2,min_value=2,max_value=2,min_align=2 R10=fp
  12: (0f) r0 += r1
  13: (72) *(u8 *)(r0 +0) = 0
  R0=map_value_adj(ks=8,vs=8,id=0),min_value=3,max_value=4
  R1=inv,min_value=3,max_value=4 R2=imm2,min_value=2,max_value=2,min_align=2 R10=fp
  14: (b7) r0 = 0
  15: (95) exit

This leaves us with two options for fixing this: i) to invalidate
all prior learned information once we switch signed context, ii)
to track min/max signed and unsigned boundaries separately as
done in [0]. (Given latter introduces major changes throughout
the whole verifier, it's rather net-next material, thus this
patch follows option i), meaning we can derive bounds either
from only signed tests or only unsigned tests.) There is still the
case of adjust_reg_min_max_vals(), where we adjust bounds on ALU
operations, meaning programs like the following where boundaries
on the reg get mixed in context later on when bounds are merged
on the dst reg must get rejected, too:

   0: (7a) *(u64 *)(r10 -8) = 0
   1: (bf) r2 = r10
   2: (07) r2 += -8
   3: (18) r1 = 0xffff89b2bf87ce00
   5: (85) call bpf_map_lookup_elem#1
   6: (15) if r0 == 0x0 goto pc+6
  R0=map_value(ks=8,vs=8,id=0),min_value=0,max_value=0 R10=fp
   7: (7a) *(u64 *)(r10 -16) = -8
   8: (79) r1 = *(u64 *)(r10 -16)
   9: (b7) r2 = 2
  10: (3d) if r2 >= r1 goto pc+2
  R0=map_value(ks=8,vs=8,id=0),min_value=0,max_value=0 R1=inv,min_value=3
  R2=imm2,min_value=2,max_value=2,min_align=2 R10=fp
  11: (b7) r7 = 1
  12: (65) if r7 s> 0x0 goto pc+2
  R0=map_value(ks=8,vs=8,id=0),min_value=0,max_value=0 R1=inv,min_value=3
  R2=imm2,min_value=2,max_value=2,min_align=2 R7=imm1,max_value=0 R10=fp
  13: (b7) r0 = 0
  14: (95) exit

  from 12 to 15: R0=map_value(ks=8,vs=8,id=0),min_value=0,max_value=0
  R1=inv,min_value=3 R2=imm2,min_value=2,max_value=2,min_align=2 R7=imm1,min_value=1 R10=fp
  15: (0f) r7 += r1
  16: (65) if r7 s> 0x4 goto pc+2
  R0=map_value(ks=8,vs=8,id=0),min_value=0,max_value=0 R1=inv,min_value=3
  R2=imm2,min_value=2,max_value=2,min_align=2 R7=inv,min_value=4,max_value=4 R10=fp
  17: (0f) r0 += r7
  18: (72) *(u8 *)(r0 +0) = 0
  R0=map_value_adj(ks=8,vs=8,id=0),min_value=4,max_value=4 R1=inv,min_value=3
  R2=imm2,min_value=2,max_value=2,min_align=2 R7=inv,min_value=4,max_value=4 R10=fp
  19: (b7) r0 = 0
  20: (95) exit

Meaning, in adjust_reg_min_max_vals() we must also reset range
values on the dst when src/dst registers have mixed signed/
unsigned derived min/max value bounds with one unbounded value
as otherwise they can be added together deducing false boundaries.
Once both boundaries are established from either ALU ops or
compare operations w/o mixing signed/unsigned insns, then they
can safely be added to other regs also having both boundaries
established. Adding regs with one unbounded side to a map value
where the bounded side has been learned w/o mixing ops is
possible, but the resulting map value won't recover from that,
meaning such op is considered invalid on the time of actual
access. Invalid bounds are set on the dst reg in case i) src reg,
or ii) in case dst reg already had them. The only way to recover
would be to perform i) ALU ops but only 'add' is allowed on map
value types or ii) comparisons, but these are disallowed on
pointers in case they span a range. This is fine as only BPF_JEQ
and BPF_JNE may be performed on PTR_TO_MAP_VALUE_OR_NULL registers
which potentially turn them into PTR_TO_MAP_VALUE type depending
on the branch, so only here min/max value cannot be invalidated
for them.

In terms of state pruning, value_from_signed is considered
as well in states_equal() when dealing with adjusted map values.
With regards to breaking existing programs, there is a small
risk, but use-cases are rather quite narrow where this could
occur and mixing compares probably unlikely.

Joint work with Josef and Edward.

  [0] https://lists.iovisor.org/pipermail/iovisor-dev/2017-June/000822.html

Fixes: 484611357c19 ("bpf: allow access into map value arrays")
Reported-by: Edward Cree <ecree@solarflare.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Edward Cree <ecree@solarflare.com>
Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: David S. Miller <davem@davemloft.net>