bpf: Clarify sanitize_check_bounds()

bpf: Clarify sanitize_check_bounds()

As is, it appears as if pointer arithmetic is allowed for everything
except PTR_TO_{STACK,MAP_VALUE} if one only looks at
sanitize_check_bounds(). However, this is misleading as the function
only works together with retrieve_ptr_limit() and the two must be kept
in sync. This patch documents the interdependency and adds a check to
ensure they stay in sync.

adjust_ptr_min_max_vals(): Because the preceding switch returns -EACCES
for every opcode except for ADD/SUB, the sanitize_needed() following the
sanitize_check_bounds() call is always true if reached. This means,
unless sanitize_check_bounds() detected that the pointer goes OOB
because of the ADD/SUB and returns -EACCES, sanitize_ptr_alu() always
executes after sanitize_check_bounds().

The following shows that this also implies that retrieve_ptr_limit()
runs in all relevant cases.

Note that there are two calls to sanitize_ptr_alu(), these are simply
needed to easily calculate the correct alu_limit as explained in
commit 7fedb63a8307 ("bpf: Tighten speculative pointer arithmetic
mask"). The truncation-simulation is already performed on the first
call.

In the second sanitize_ptr_alu(commit_window = true), we always run
retrieve_ptr_limit(), unless:

* can_skip_alu_sanititation() is true, notably `BPF_SRC(insn->code) ==
  BPF_K`. BPF_K is fine because it means that there is no scalar
  register (which could be subject to speculative scalar confusion due
  to Spectre v4) that goes into the ALU operation. The pointer register
  can not be subject to v4-based value confusion due to the nospec
  added. Thus, in this case it would have been fine to also skip
  sanitize_check_bounds().

* If we are on a speculative path (`vstate->speculative`) and in the
  second "commit" phase, sanitize_ptr_alu() always just returns 0. This
  makes sense because there are no ALU sanitization limits to be learned
  from speculative paths. Furthermore, because the sanitization will
  ensure that pointer arithmetic stays in (architectural) bounds, the
  sanitize_check_bounds() on the speculative path could also be skipped.

The second case needs more attention: Assume we have some ALU operation
that is used with scalars architecturally, but with a
non-PTR_TO_{STACK,MAP_VALUE} pointer (e.g., PTR_TO_PACKET)
speculatively. It might appear as if this would allow an unsanitized
pointer ALU operations, but this can not happen because one of the
following two always holds:

* The type mismatch stems from Spectre v4, then it is prevented by a
  nospec after the possibly-bypassed store involving the pointer. There
  is no speculative path simulated for this case thus it never happens.

* The type mismatch stems from a Spectre v1 gadget like the following:

    r1 = slow(0)
    r4 = fast(0)
    r3 = SCALAR // Spectre v4 scalar confusion
    if (r1) {
      r2 = PTR_TO_PACKET
    } else {
      r2 = 42
    }
    if (r4) {
      r2 += r3
      *r2
    }

  If `r2 = PTR_TO_PACKET` is indeed dead code, it will be sanitized to
  `goto -1` (as is the case for the r4-if block). If it is not (e.g., if
  `r1 = r4 = 1` is possible), it will also be explored on an
  architectural path and retrieve_ptr_limit() will reject it.

To summarize, the exception for `vstate->speculative` is safe.

Back to retrieve_ptr_limit(): It only allows the ALU operation if the
involved pointer register (can be either source or destination for ADD)
is PTR_TO_STACK or PTR_TO_MAP_VALUE. Otherwise, it returns -EOPNOTSUPP.

Therefore, sanitize_check_bounds() returning 0 for
non-PTR_TO_{STACK,MAP_VALUE} is fine because retrieve_ptr_limit() also
runs for all relevant cases and prevents unsafe operations.

To summarize, we allow unsanitized pointer arithmetic with 64-bit
ADD/SUB for the following instructions if the requirements from
retrieve_ptr_limit() AND sanitize_check_bounds() hold:

* ptr -=/+= imm32 (i.e. `BPF_SRC(insn->code) == BPF_K`)

* PTR_TO_{STACK,MAP_VALUE} -= scalar

* PTR_TO_{STACK,MAP_VALUE} += scalar

* scalar += PTR_TO_{STACK,MAP_VALUE}

To document the interdependency between sanitize_check_bounds() and
retrieve_ptr_limit(), add a verifier_bug_if() to make sure they stay in
sync.

Signed-off-by: Luis Gerhorst <luis.gerhorst@fau.de>
Reported-by: Kumar Kartikeya Dwivedi <memxor@gmail.com>
Link: https://lore.kernel.org/bpf/CAP01T76HZ+s5h+_REqRFkRjjoKwnZZn9YswpSVinGicah1pGJw@mail.gmail.com/
Link: https://lore.kernel.org/bpf/CAP01T75oU0zfZCiymEcH3r-GQ5A6GOc6GmYzJEnMa3=53XuUQQ@mail.gmail.com/
Link: https://lore.kernel.org/r/20250603204557.332447-1-luis.gerhorst@fau.de
Signed-off-by: Alexei Starovoitov <ast@kernel.org>