bpf: Fix tail_call_reachable rejection for interpreter when jit failed
authorDaniel Borkmann <daniel@iogearbox.net>
Mon, 12 Jul 2021 20:57:35 +0000 (22:57 +0200)
committerAlexei Starovoitov <ast@kernel.org>
Tue, 13 Jul 2021 15:19:13 +0000 (08:19 -0700)
During testing of f263a81451c1 ("bpf: Track subprog poke descriptors correctly
and fix use-after-free") under various failure conditions, for example, when
jit_subprogs() fails and tries to clean up the program to be run under the
interpreter, we ran into the following freeze:

  [...]
  #127/8 tailcall_bpf2bpf_3:FAIL
  [...]
  [   92.041251] BUG: KASAN: slab-out-of-bounds in ___bpf_prog_run+0x1b9d/0x2e20
  [   92.042408] Read of size 8 at addr ffff88800da67f68 by task test_progs/682
  [   92.043707]
  [   92.044030] CPU: 1 PID: 682 Comm: test_progs Tainted: G   O   5.13.0-53301-ge6c08cb33a30-dirty #87
  [   92.045542] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.13.0-1ubuntu1 04/01/2014
  [   92.046785] Call Trace:
  [   92.047171]  ? __bpf_prog_run_args64+0xc0/0xc0
  [   92.047773]  ? __bpf_prog_run_args32+0x8b/0xb0
  [   92.048389]  ? __bpf_prog_run_args64+0xc0/0xc0
  [   92.049019]  ? ktime_get+0x117/0x130
  [...] // few hundred [similar] lines more
  [   92.659025]  ? ktime_get+0x117/0x130
  [   92.659845]  ? __bpf_prog_run_args64+0xc0/0xc0
  [   92.660738]  ? __bpf_prog_run_args32+0x8b/0xb0
  [   92.661528]  ? __bpf_prog_run_args64+0xc0/0xc0
  [   92.662378]  ? print_usage_bug+0x50/0x50
  [   92.663221]  ? print_usage_bug+0x50/0x50
  [   92.664077]  ? bpf_ksym_find+0x9c/0xe0
  [   92.664887]  ? ktime_get+0x117/0x130
  [   92.665624]  ? kernel_text_address+0xf5/0x100
  [   92.666529]  ? __kernel_text_address+0xe/0x30
  [   92.667725]  ? unwind_get_return_address+0x2f/0x50
  [   92.668854]  ? ___bpf_prog_run+0x15d4/0x2e20
  [   92.670185]  ? ktime_get+0x117/0x130
  [   92.671130]  ? __bpf_prog_run_args64+0xc0/0xc0
  [   92.672020]  ? __bpf_prog_run_args32+0x8b/0xb0
  [   92.672860]  ? __bpf_prog_run_args64+0xc0/0xc0
  [   92.675159]  ? ktime_get+0x117/0x130
  [   92.677074]  ? lock_is_held_type+0xd5/0x130
  [   92.678662]  ? ___bpf_prog_run+0x15d4/0x2e20
  [   92.680046]  ? ktime_get+0x117/0x130
  [   92.681285]  ? __bpf_prog_run32+0x6b/0x90
  [   92.682601]  ? __bpf_prog_run64+0x90/0x90
  [   92.683636]  ? lock_downgrade+0x370/0x370
  [   92.684647]  ? mark_held_locks+0x44/0x90
  [   92.685652]  ? ktime_get+0x117/0x130
  [   92.686752]  ? lockdep_hardirqs_on+0x79/0x100
  [   92.688004]  ? ktime_get+0x117/0x130
  [   92.688573]  ? __cant_migrate+0x2b/0x80
  [   92.689192]  ? bpf_test_run+0x2f4/0x510
  [   92.689869]  ? bpf_test_timer_continue+0x1c0/0x1c0
  [   92.690856]  ? rcu_read_lock_bh_held+0x90/0x90
  [   92.691506]  ? __kasan_slab_alloc+0x61/0x80
  [   92.692128]  ? eth_type_trans+0x128/0x240
  [   92.692737]  ? __build_skb+0x46/0x50
  [   92.693252]  ? bpf_prog_test_run_skb+0x65e/0xc50
  [   92.693954]  ? bpf_prog_test_run_raw_tp+0x2d0/0x2d0
  [   92.694639]  ? __fget_light+0xa1/0x100
  [   92.695162]  ? bpf_prog_inc+0x23/0x30
  [   92.695685]  ? __sys_bpf+0xb40/0x2c80
  [   92.696324]  ? bpf_link_get_from_fd+0x90/0x90
  [   92.697150]  ? mark_held_locks+0x24/0x90
  [   92.698007]  ? lockdep_hardirqs_on_prepare+0x124/0x220
  [   92.699045]  ? finish_task_switch+0xe6/0x370
  [   92.700072]  ? lockdep_hardirqs_on+0x79/0x100
  [   92.701233]  ? finish_task_switch+0x11d/0x370
  [   92.702264]  ? __switch_to+0x2c0/0x740
  [   92.703148]  ? mark_held_locks+0x24/0x90
  [   92.704155]  ? __x64_sys_bpf+0x45/0x50
  [   92.705146]  ? do_syscall_64+0x35/0x80
  [   92.706953]  ? entry_SYSCALL_64_after_hwframe+0x44/0xae
  [...]

Turns out that the program rejection from e411901c0b77 ("bpf: allow for tailcalls
in BPF subprograms for x64 JIT") is buggy since env->prog->aux->tail_call_reachable
is never true. Commit ebf7d1f508a7 ("bpf, x64: rework pro/epilogue and tailcall
handling in JIT") added a tracker into check_max_stack_depth() which propagates
the tail_call_reachable condition throughout the subprograms. This info is then
assigned to the subprogram's func[i]->aux->tail_call_reachable. However, in the
case of the rejection check upon JIT failure, env->prog->aux->tail_call_reachable
is used. func[0]->aux->tail_call_reachable which represents the main program's
information did not propagate this to the outer env->prog->aux, though. Add this
propagation into check_max_stack_depth() where it needs to belong so that the
check can be done reliably.

Fixes: ebf7d1f508a7 ("bpf, x64: rework pro/epilogue and tailcall handling in JIT")
Fixes: e411901c0b77 ("bpf: allow for tailcalls in BPF subprograms for x64 JIT")
Co-developed-by: John Fastabend <john.fastabend@gmail.com>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: John Fastabend <john.fastabend@gmail.com>
Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Acked-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Link: https://lore.kernel.org/bpf/618c34e3163ad1a36b1e82377576a6081e182f25.1626123173.git.daniel@iogearbox.net
kernel/bpf/verifier.c

index 42a4063de7cd24d5f946f0c95815501fef3019e2..9de3c9c3267cb10ddcaf1070defea4930dfbb596 100644 (file)
@@ -3677,6 +3677,8 @@ continue_func:
        if (tail_call_reachable)
                for (j = 0; j < frame; j++)
                        subprog[ret_prog[j]].tail_call_reachable = true;
+       if (subprog[0].tail_call_reachable)
+               env->prog->aux->tail_call_reachable = true;
 
        /* end of for() loop means the last insn of the 'subprog'
         * was reached. Doesn't matter whether it was JA or EXIT