在 2021/7/14 7:17, Alexei Starovoitov 写道:
On Sun, Jul 11, 2021 at 7:17 PM He Fengqing <hefengqing@xxxxxxxxxx> wrote:
在 2021/7/9 23:12, Alexei Starovoitov 写道:
On Fri, Jul 9, 2021 at 4:11 AM He Fengqing <hefengqing@xxxxxxxxxx> wrote:
在 2021/7/8 11:09, Alexei Starovoitov 写道:
On Wed, Jul 7, 2021 at 8:00 PM He Fengqing <hefengqing@xxxxxxxxxx> wrote:
Ok, I will change this in next version.
before you spam the list with the next version
please explain why any of these changes are needed?
I don't see an explanation in the patches and I don't see a bug in the code.
Did you check what is the prog clone ?
When is it constructed? Why verifier has anything to do with it?
.
I'm sorry, I didn't describe these errors clearly.
bpf_check(bpf_verifier_env)
|
|->do_misc_fixups(env)
| |
| |->bpf_patch_insn_data(env)
| | |
| | |->bpf_patch_insn_single(env->prog)
| | | |
| | | |->bpf_prog_realloc(env->prog)
| | | | |
| | | | |->construct new_prog
| | | | | free old_prog(env->prog)
| | | | |
| | | | |->return new_prog;
| | | |
| | | |->return new_prog;
| | |
| | |->adjust_insn_aux_data
| | | |
| | | |->return ENOMEM;
| | |
| | |->return NULL;
| |
| |->return ENOMEM;
bpf_verifier_env->prog had been freed in bpf_prog_realloc function.
There are two errors here, the first is memleak in the
bpf_patch_insn_data function, and the second is use after free in the
bpf_check function.
memleak in bpf_patch_insn_data:
Look at the call chain above, if adjust_insn_aux_data function return
ENOMEM, bpf_patch_insn_data will return NULL, but we do not free the
new_prog.
So in the patch 2, before bpf_patch_insn_data return NULL, we free the
new_prog.
use after free in bpf_check:
If bpf_patch_insn_data function return NULL, we will not assign new_prog
to the bpf_verifier_env->prog, but bpf_verifier_env->prog has been freed
in the bpf_prog_realloc function. Then in bpf_check function, we will
use bpf_verifier_env->prog after do_misc_fixups function.
In the patch 3, I added a free_old parameter to bpf_prog_realloc, in
this scenario we don't free old_prog. Instead, we free it in the
do_misc_fixups function when bpf_patch_insn_data return a valid new_prog.
Thanks for explaining.
Why not to make adjust_insn_aux_data() in bpf_patch_insn_data() first then?
Just changing the order will resolve both issues, no?
.
adjust_insn_aux_data() need the new constructed new_prog as an input
parameter, so we must call bpf_patch_insn_single() before
adjust_insn_aux_data().
Right. I forgot about insn_has_def32() logic and
commit b325fbca4b13 ("bpf: verifier: mark patched-insn with
sub-register zext flag")
that added that extra parameter.
But we can make adjust_insn_aux_data() never return ENOMEM. In
bpf_patch_insn_data(), first we pre-malloc memory for new aux_data, then
call bpf_patch_insn_single() to constructed the new_prog, at last call
adjust_insn_aux_data() functin. In this way, adjust_insn_aux_data()
never fails.
bpf_patch_insn_data(env) {
struct bpf_insn_aux_data *new_data = vzalloc();
struct bpf_prog *new_prog;
if (new_data == NULL)
return NULL;
new_prog = bpf_patch_insn_single(env->prog);
if (new_prog == NULL) {
vfree(new_data);
return NULL;
}
adjust_insn_aux_data(new_prog, new_data);
return new_prog;
}
What do you think about it?
That's a good idea. Let's do that. The new size for vzalloc is easy to compute.
What should be the commit in the Fixes tag?
commit 8041902dae52 ("bpf: adjust insn_aux_data when patching insns")
right?
Ok, I will add this in the commit message.
4 year old bug then.
I wonder why syzbot with malloc error injection didn't catch it sooner.
.
