bpf: Fix sync_linked_regs regarding BPF_ADD_CONST32 zext propagation

Jenny reported that in sync_linked_regs() the BPF_ADD_CONST32 flag is
checked on known_reg (the register narrowed by a conditional branch)
instead of reg (the linked target register created by an alu32 operation).

Example case with reg:

1. r6 = bpf_get_prandom_u32()
2. r7 = r6 (linked, same id)
3. w7 += 5 (alu32 -- r7 gets BPF_ADD_CONST32, zero-extended by CPU)
4. if w6 < 0xFFFFFFFC goto safe (narrows r6 to [0xFFFFFFFC, 0xFFFFFFFF])
5. sync_linked_regs() propagates to r7 but does NOT call zext_32_to_64()
6. Verifier thinks r7 is [0x100000001, 0x100000004] instead of [1, 4]

Since known_reg above does not have BPF_ADD_CONST32 set above, zext_32_to_64()
is never called on alu32-derived linked registers. This causes the verifier
to track incorrect 64-bit bounds, while the CPU correctly zero-extends the
32-bit result.

The code checking known_reg->id was correct however (see scalars_alu32_wrap
selftest case), but the real fix needs to handle both directions - zext
propagation should be done when either register has BPF_ADD_CONST32, since
the linked relationship involves a 32-bit operation regardless of which
side has the flag.

Example case with known_reg (exercised also by scalars_alu32_wrap):

1. r1 = r0; w1 += 0x100 (alu32 -- r1 gets BPF_ADD_CONST32)
2. if r1 > 0x80 - known_reg = r1 (has BPF_ADD_CONST32), reg = r0 (doesn't)

Hence, fix it by checking for (reg->id | known_reg->id) & BPF_ADD_CONST32.

Moreover, sync_linked_regs() also has a soundness issue when two linked
registers used different ALU widths: one with BPF_ADD_CONST32 and the
other with BPF_ADD_CONST64. The delta relationship between linked registers
assumes the same arithmetic width though. When one register went through
alu32 (CPU zero-extends the 32-bit result) and the other went through
alu64 (no zero-extension), the propagation produces incorrect bounds.

Example:

r6 = bpf_get_prandom_u32() // fully unknown
if r6 >= 0x100000000 goto out // constrain r6 to [0, U32_MAX]
r7 = r6
w7 += 1 // alu32: r7.id = N | BPF_ADD_CONST32
r8 = r6
r8 += 2 // alu64: r8.id = N | BPF_ADD_CONST64
if r7 < 0xFFFFFFFF goto out // narrows r7 to [0xFFFFFFFF, 0xFFFFFFFF]

At the branch on r7, sync_linked_regs() runs with known_reg=r7
(BPF_ADD_CONST32) and reg=r8 (BPF_ADD_CONST64). The delta path
computes:

r8 = r7 + (delta_r8 - delta_r7) = 0xFFFFFFFF + (2 - 1) = 0x100000000

Then, because known_reg->id has BPF_ADD_CONST32, zext_32_to_64(r8) is
called, truncating r8 to [0, 0]. But r8 used a 64-bit ALU op -- the
CPU does NOT zero-extend it. The actual CPU value of r8 is
0xFFFFFFFE + 2 = 0x100000000, not 0. The verifier now underestimates
r8's 64-bit bounds, which is a soundness violation.

Fix sync_linked_regs() by skipping propagation when the two registers
have mixed ALU widths (one BPF_ADD_CONST32, the other BPF_ADD_CONST64).

Lastly, fix regsafe() used for path pruning: the existing checks used
"& BPF_ADD_CONST" to test for offset linkage, which treated
BPF_ADD_CONST32 and BPF_ADD_CONST64 as equivalent.

Fixes: 7a433e519364 ("bpf: Support negative offsets, BPF_SUB, and alu32 for linked register tracking")
Reported-by: Jenny Guanni Qu <qguanni@gmail.com>
Co-developed-by: Puranjay Mohan <puranjay@kernel.org>
Signed-off-by: Puranjay Mohan <puranjay@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Eduard Zingerman <eddyz87@gmail.com>
Link: https://lore.kernel.org/r/20260319211507.213816-1-daniel@iogearbox.net
Signed-off-by: Alexei Starovoitov <ast@kernel.org>