Re: [PATCH v2] bpf: Fix KASAN use-after-free Read in compute_effective_progs

[Tadeusz Struk <tadeusz.struk@linaro.org>]

Hi Andrii,
On 4/20/22 10:07, Andrii Nakryiko wrote:
>> diff --git a/kernel/bpf/cgroup.c b/kernel/bpf/cgroup.c
>> index 128028efda64..5a64cece09f3 100644
>> --- a/kernel/bpf/cgroup.c
>> +++ b/kernel/bpf/cgroup.c
>> @@ -723,10 +723,8 @@ static int __cgroup_bpf_detach(struct cgroup *cgrp, struct bpf_prog *prog,
>>          pl->link = NULL;
>>
>>          err = update_effective_progs(cgrp, atype);
>> -       if (err)
>> -               goto cleanup;
>>
>> -       /* now can actually delete it from this cgroup list */
>> +       /* now can delete it from this cgroup list */
>>          list_del(&pl->node);
>>          kfree(pl);
>>          if (list_empty(progs))
>> @@ -735,12 +733,55 @@ static int __cgroup_bpf_detach(struct cgroup *cgrp, struct bpf_prog *prog,
>>          if (old_prog)
>>                  bpf_prog_put(old_prog);
>>          static_branch_dec(&cgroup_bpf_enabled_key[atype]);
>> -       return 0;
>> +
>> +       if (!err)
>> +               return 0;
>>
>>   cleanup:
>> -       /* restore back prog or link */
>> -       pl->prog = old_prog;
>> -       pl->link = link;
>> +       /*
>> +        * If compute_effective_progs failed with -ENOMEM, i.e. alloc for
>> +        * cgrp->bpf.inactive table failed, we can recover by removing
>> +        * the detached prog from effective table and rearranging it.
>> +        */
>> +       if (err == -ENOMEM) {
>> +               struct bpf_prog_array_item *item;
>> +               struct bpf_prog *prog_tmp, *prog_detach, *prog_last;
>> +               struct bpf_prog_array *array;
>> +               int index = 0, index_detach = -1;
>> +
>> +               array = cgrp->bpf.effective[atype];
>> +               item = &array->items[0];
>> +
>> +               if (prog)
>> +                       prog_detach = prog;
>> +               else
>> +                       prog_detach = link->link.prog;
>> +
>> +               if (!prog_detach)
>> +                       return -EINVAL;
>> +
>> +               while ((prog_tmp = READ_ONCE(item->prog))) {
>> +                       if (prog_tmp == prog_detach)
>> +                               index_detach = index;
>> +                       item++;
>> +                       index++;
>> +                       prog_last = prog_tmp;
>> +               }
>> +
>> +               /* Check if we found what's needed for removing the prog */
>> +               if (index_detach == -1 || index_detach == index-1)
>> +                       return -EINVAL;
>> +
>> +               /* Remove the last program in the array */
>> +               if (bpf_prog_array_delete_safe_at(array, index-1))
>> +                       return -EINVAL;
>> +
>> +               /* and update the detached with the last just removed */
>> +               if (bpf_prog_array_update_at(array, index_detach, prog_last))
>> +                       return -EINVAL;
>> +
>> +               err = 0;
>> +       }

Thanks for feedback, and sorry for delay. I got pulled into something else.

> There are a bunch of problems with this implementation.
> 
> 1. We should do this fallback right after update_effective_progs()
> returns error, before we get to list_del(&pl->node) and subsequent
> code that does some additional things (like clearing flags and stuff).
> This additional code needs to run even if update_effective_progs()
> fails. So I suggest to extract the logic of removing program from
> effective prog arrays into a helper function and doing
> 
> err = update_effective_progs(...);
> if (err)
>      purge_effective_progs();
> 
> where purge_effective_progs() will be the logic you are adding. And it
> will be void function because it can't fail.

I have implemented that in v3, will send that out soon.

> 
> 2. We have to update not just cgrp->bpf.effective array, but all the
> descendants' lists as well. See what update_effective_progs() is
> doing, it has css_for_each_descendant_pre() iteration. You need to do
> it here as well. But instead of doing compute_effective_progs() which
> allocates a new copy of an array we'll need to update existing array
> in place.
> 
> 3. Not clear why you need to do both bpf_prog_array_delete_safe_at()
> and bpf_prog_array_update_at(), isn't delete_safe_at() enought?

I thought that we need to reshuffle the table and move the progs around,
but your are right, delete_safe_at() is enough.

-- 
Thanks,
Tadeusz