mbox series

[v2,bpf-next,00/10] BPF verifier retval logic fixes

Message ID 20231129003620.1049610-1-andrii@kernel.org (mailing list archive)
Headers show
Series BPF verifier retval logic fixes | expand

Message

Andrii Nakryiko Nov. 29, 2023, 12:36 a.m. UTC
This patch set fixes BPF verifier logic around validating and enforcing return
values for BPF programs that have specific range of expected return values.
Both sync and async callbacks have similar logic and are fixes as well.
A few tests are added that would fail without the fixes in this patch set.

Also, while at it, we update retval checking logic to use umin/umax range
instead of tnum, avoiding future potential issues if expected range cannot be
represented precisely by tnum (e.g., [0, 2] is not representable by tnum and
is treated as [0, 3]).

There is a little bit of refactoring to unify async callback and program exit
logic to avoid duplication of checks as much as possible.

v1->v2:
  - drop tnum from retval checks (Eduard);
  - use smin/smax instead of umin/umax (Alexei).

Andrii Nakryiko (10):
  bpf: provide correct register name for exception callback retval check
  bpf: enforce precision of R0 on callback return
  bpf: enforce exact retval range on subprog/callback exit
  selftests/bpf: add selftest validating callback result is enforced
  bpf: enforce precise retval range on program exit
  bpf: unify async callback and program retval checks
  bpf: enforce precision of R0 on program/async callback return
  selftests/bpf: validate async callback return value check correctness
  selftests/bpf: adjust global_func15 test to validate prog exit
    precision
  bpf: simplify tnum output if a fully known constant

 include/linux/bpf_verifier.h                  |   7 +-
 kernel/bpf/log.c                              |  13 ++
 kernel/bpf/tnum.c                             |   6 -
 kernel/bpf/verifier.c                         | 120 ++++++++++--------
 .../selftests/bpf/progs/exceptions_assert.c   |   2 +-
 .../selftests/bpf/progs/exceptions_fail.c     |   2 +-
 .../selftests/bpf/progs/test_global_func15.c  |  34 ++++-
 .../selftests/bpf/progs/timer_failure.c       |  36 ++++--
 .../selftests/bpf/progs/user_ringbuf_fail.c   |   2 +-
 .../bpf/progs/verifier_cgroup_inv_retcode.c   |   8 +-
 .../bpf/progs/verifier_direct_packet_access.c |   2 +-
 .../selftests/bpf/progs/verifier_int_ptr.c    |   2 +-
 .../bpf/progs/verifier_netfilter_retcode.c    |   2 +-
 .../selftests/bpf/progs/verifier_stack_ptr.c  |   4 +-
 .../bpf/progs/verifier_subprog_precision.c    |  50 ++++++++
 15 files changed, 211 insertions(+), 79 deletions(-)

Comments

Shung-Hsi Yu Nov. 29, 2023, 11:27 a.m. UTC | #1
On Tue, Nov 28, 2023 at 04:36:10PM -0800, Andrii Nakryiko wrote:
> This patch set fixes BPF verifier logic around validating and enforcing return
> values for BPF programs that have specific range of expected return values.
> Both sync and async callbacks have similar logic and are fixes as well.
> A few tests are added that would fail without the fixes in this patch set.
> 
> Also, while at it, we update retval checking logic to use umin/umax range

Looks like this should be change to smin/smax as well

> instead of tnum, avoiding future potential issues if expected range cannot be
> represented precisely by tnum (e.g., [0, 2] is not representable by tnum and
> is treated as [0, 3]).
> 
> There is a little bit of refactoring to unify async callback and program exit
> logic to avoid duplication of checks as much as possible.
> 
> v1->v2:
>   - drop tnum from retval checks (Eduard);
>   - use smin/smax instead of umin/umax (Alexei).

...
Andrii Nakryiko Nov. 29, 2023, 4:23 p.m. UTC | #2
On Wed, Nov 29, 2023 at 3:28 AM Shung-Hsi Yu <shung-hsi.yu@suse.com> wrote:
>
> On Tue, Nov 28, 2023 at 04:36:10PM -0800, Andrii Nakryiko wrote:
> > This patch set fixes BPF verifier logic around validating and enforcing return
> > values for BPF programs that have specific range of expected return values.
> > Both sync and async callbacks have similar logic and are fixes as well.
> > A few tests are added that would fail without the fixes in this patch set.
> >
> > Also, while at it, we update retval checking logic to use umin/umax range
>
> Looks like this should be change to smin/smax as well
>

yep, thanks, I fixed up few more places where I missed umin/umax ->
smin/smax updates


> > instead of tnum, avoiding future potential issues if expected range cannot be
> > represented precisely by tnum (e.g., [0, 2] is not representable by tnum and
> > is treated as [0, 3]).
> >
> > There is a little bit of refactoring to unify async callback and program exit
> > logic to avoid duplication of checks as much as possible.
> >
> > v1->v2:
> >   - drop tnum from retval checks (Eduard);
> >   - use smin/smax instead of umin/umax (Alexei).
>
> ...