Re: [PATCH bpf-next v4 1/2] bpf: Fix attaching fentry/fexit/fmod_ret/lsm to modules

[Yonghong Song <yhs@meta.com>]

On 12/13/22 2:59 AM, Viktor Malik wrote:
> On 12/12/22 18:08, Yonghong Song wrote:
>>
>>
>> On 12/12/22 4:59 AM, Viktor Malik wrote:
>>> When attaching fentry/fexit/fmod_ret/lsm to a function located in a
>>> module without specifying the target program, the verifier tries to find
>>> the address to attach to in kallsyms. This is always done by searching
>>> the entire kallsyms, not respecting the module in which the function is
>>> located.
>>>
>>> This approach causes an incorrect attachment address to be computed if
>>> the function to attach to is shadowed by a function of the same name
>>> located earlier in kallsyms.
>>>
>>> Since the attachment must contain the BTF of the program to attach to,
>>> we may extract the module from it and search for the function address in
>>> the module.
>>>
>>> Signed-off-by: Viktor Malik <vmalik@redhat.com>
>>> ---
>>>   kernel/bpf/verifier.c | 16 +++++++++++++++-
>>>   1 file changed, 15 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
>>> index a5255a0dcbb6..d646c5263bc5 100644
>>> --- a/kernel/bpf/verifier.c
>>> +++ b/kernel/bpf/verifier.c
>>> @@ -24,6 +24,7 @@
>>>   #include <linux/bpf_lsm.h>
>>>   #include <linux/btf_ids.h>
>>>   #include <linux/poison.h>
>>> +#include "../module/internal.h"
>>>   #include "disasm.h"
>>> @@ -16478,6 +16479,7 @@ int bpf_check_attach_target(struct 
>>> bpf_verifier_log *log,
>>>       const char *tname;
>>>       struct btf *btf;
>>>       long addr = 0;
>>> +    struct module *mod;
>>>       if (!btf_id) {
>>>           bpf_log(log, "Tracing programs must provide btf_id\n");
>>> @@ -16645,7 +16647,19 @@ int bpf_check_attach_target(struct 
>>> bpf_verifier_log *log,
>>>               else
>>>                   addr = (long) tgt_prog->aux->func[subprog]->bpf_func;
>>>           } else {
>>> -            addr = kallsyms_lookup_name(tname);
>>> +            if (btf_is_module(btf)) {
>>> +                preempt_disable();
>>> +                mod = btf_try_get_module(btf);
>>> +                if (mod) {
>>> +                    addr = find_kallsyms_symbol_value(mod, tname);
>>> +                    module_put(mod);
>>> +                } else {
>>> +                    addr = 0;
>>> +                }
>>> +                preempt_enable();
>>
>> What if module is unloaded right after preempt_enabled so 'addr' 
>> becomes invalid? Is this a corner case we should consider?
> 
> IIUC, if 'addr' becomes invalid, the attachment will eventually fail.
> 
> So I'd say that there's no need to consider that case here, it's not
> considered for kallsyms_lookup_name below (which may call
> module_kallsyms_lookup_name) either.

The below kallsyms_lookup_name(tname) works for vmlinux and vmlinux
function won't go away.

The following is what I understand for module address:

    bpf_tracing_prog_attach:
        ...
        bpf_check_attach_target (get addr etc. into tgt_info.
        ...
        bpf_trampoline_link_prog
            __bpf_trampoline_link_prog
                bpf_trampoline_update
                    register_fentry
                        bpf_trampoline_module_get


static int bpf_trampoline_module_get(struct bpf_trampoline *tr)
{
         struct module *mod;
         int err = 0;

         preempt_disable();
         mod = __module_text_address((unsigned long) tr->func.addr);
         if (mod && !try_module_get(mod))
                 err = -ENOENT;
         preempt_enable();
         tr->mod = mod;
         return err;
}

We try to grab a module reference here based on the func addr.
But there is a risk that module (m1) might be unloaded and a different
module (m2) might be loaded which occupies the address space of
m1. In such cases, we might have issues.

To resolve this issue, we might want to grab the reference
earlier for find_kallsyms_symbol_value and won't release it
until the program is detached. try_module_get() then will
be unnecessary in this particular case. But care must be
taken for other code paths.
                       >
>>
>>> +            } else {
>>> +                addr = kallsyms_lookup_name(tname);
>>> +            }
>>>               if (!addr) {
>>>                   bpf_log(log,
>>>                       "The address of function %s cannot be found\n",
>>
>