During the bounds refinement, we improve the precision of various ranges
by looking at other ranges. Among others, we improve the following in
this order (other things happen between 1 and 2):
1. Improve u32 from s32 in __reg32_deduce_bounds.
2. Improve s/u64 from u32 in __reg_deduce_mixed_bounds.
3. Improve s/u64 from s32 in __reg_deduce_mixed_bounds.
In particular, if the s32 range forms a valid u32 range, we will use it
to improve the u32 range in __reg32_deduce_bounds. In
__reg_deduce_mixed_bounds, under the same condition, we will use the s32
range to improve the s/u64 ranges.
If at (1) we were able to learn from s32 to improve u32, we'll then be
able to use that in (2) to improve s/u64. Hence, as (3) happens under
the same precondition as (1), it won't improve s/u64 ranges further than
(1)+(2) did. Thus, we can get rid of (3).
In addition to the extensive suite of selftests for bounds refinement,
this patch was also tested with the Agni formal verification tool [1].
Additionally, Eduard mentioned:
The argument appears to be as follows:
Under precondition `(u32)reg->s32_min <= (u32)reg->s32_max`
__reg32_deduce_bounds produces:
reg->u32_min = max_t(u32, reg->s32_min, reg->u32_min);
reg->u32_max = min_t(u32, reg->s32_max, reg->u32_max);
And then first part of __reg_deduce_mixed_bounds assigns:
a. reg->umin umax= (reg->umin & ~0xffffffffULL) | max_t(u32, reg->s32_min, reg->u32_min);
b. reg->umax umin= (reg->umax & ~0xffffffffULL) | min_t(u32, reg->s32_max, reg->u32_max);
And then second part of __reg_deduce_mixed_bounds assigns:
c. reg->umin umax= (reg->umin & ~0xffffffffULL) | (u32)reg->s32_min;
d. reg->umax umin= (reg->umax & ~0xffffffffULL) | (u32)reg->s32_max;
But assignment (c) is a noop because:
max_t(u32, reg->s32_min, reg->u32_min) >= (u32)reg->s32_min
Hence RHS(a) >= RHS(c) and umin= does nothing.
Also assignment (d) is a noop because:
min_t(u32, reg->s32_max, reg->u32_max) <= (u32)reg->s32_max
Hence RHS(b) <= RHS(d) and umin= does nothing.
Plus the same reasoning for the part dealing with reg->s{min,max}_value:
e. reg->smin_value smax= (reg->smin_value & ~0xffffffffULL) | max_t(u32, reg->s32_min_value, reg->u32_min_value);
f. reg->smax_value smin= (reg->smax_value & ~0xffffffffULL) | min_t(u32, reg->s32_max_value, reg->u32_max_value);
vs
g. reg->smin_value smax= (reg->smin_value & ~0xffffffffULL) | (u32)reg->s32_min_value;
h. reg->smax_value smin= (reg->smax_value & ~0xffffffffULL) | (u32)reg->s32_max_value;
RHS(e) >= RHS(g) and RHS(f) <= RHS(h), hence smax=,smin= do nothing.
This appears to be correct.
Also, Shung-Hsi:
Beside going through the reasoning, I also played with CBMC a bit to
double check that as far as a single run of __reg_deduce_bounds() is
concerned (and that the register state matches certain handwavy
expectations), the change indeed still preserve the original behavior.
Signed-off-by: Paul Chaignon <paul.chaignon@gmail.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Reviewed-by: Shung-Hsi Yu <shung-hsi.yu@suse.com>
Acked-by: Eduard Zingerman <eddyz87@gmail.com>
Link: https://github.com/bpfverif/agni [1]
Link: https://lore.kernel.org/bpf/aIJwnFnFyUjNsCNa@mail.gmail.com
5345e64760