[PATCH bpf-next 2/3] bpf: Keep im address consistent between dry run and real patching

[Pu Lehui]

On 2024/1/30 1:58, Song Liu wrote:
> On Tue, Jan 23, 2024 at 2:32 AM Pu Lehui <pulehui at huaweicloud.com> wrote:
>>
>> From: Pu Lehui <pulehui at huawei.com>
>>
>> In __arch_prepare_bpf_trampoline, we emit instructions to store the
>> address of im to register and then pass it to __bpf_tramp_enter and
>> __bpf_tramp_exit functions. Currently we use fake im in
>> arch_bpf_trampoline_size for the dry run, and then allocate new im for
>> the real patching. This is fine for architectures that use fixed
>> instructions to generate addresses. However, for architectures that use
>> dynamic instructions to generate addresses, this may make the front and
>> rear images inconsistent, leading to patching overflow. We can extract
>> the im allocation ahead of the dry run and pass the allocated im to
>> arch_bpf_trampoline_size, so that we can ensure that im is consistent in
>> dry run and real patching.
> 
> IIUC, this is required because emit_imm() for riscv may generate variable
> size instructions (depends on the value of im). I wonder we can fix this by
> simply set a special value for fake im in arch_bpf_trampoline_size() to
> so that emit_imm() always gives biggest value for the fake im.
> 

Hi Song,

Thanks for your review. Yes, I had the same idea as you at first, emit 
biggist count instructions when ctx->insns is NULL, but this may lead to 
memory waste. So try moving out of IM to get a fixed IM address, maybe 
other architectures require it too. If you feel it is inappropriate, I 
will withdraw it.

>>
>> Signed-off-by: Pu Lehui <pulehui at huawei.com>
>> ---
> [...]
>>
>>   static int bpf_trampoline_update(struct bpf_trampoline *tr, bool lock_direct_mutex)
>> @@ -432,23 +425,27 @@ static int bpf_trampoline_update(struct bpf_trampoline *tr, bool lock_direct_mut
>>                  tr->flags |= BPF_TRAMP_F_ORIG_STACK;
>>   #endif
>>
>> -       size = arch_bpf_trampoline_size(&tr->func.model, tr->flags,
>> +       im = kzalloc(sizeof(*im), GFP_KERNEL);
>> +       if (!im) {
>> +               err = -ENOMEM;
>> +               goto out;
>> +       }
>> +
>> +       size = arch_bpf_trampoline_size(im, &tr->func.model, tr->flags,
>>                                          tlinks, tr->func.addr);
>>          if (size < 0) {
>>                  err = size;
>> -               goto out;
>> +               goto out_free_im;
>>          }
>>
>>          if (size > PAGE_SIZE) {
>>                  err = -E2BIG;
>> -               goto out;
>> +               goto out_free_im;
>>          }
>>
>> -       im = bpf_tramp_image_alloc(tr->key, size);
>> -       if (IS_ERR(im)) {
>> -               err = PTR_ERR(im);
>> -               goto out;
>> -       }
>> +       err = bpf_tramp_image_alloc(im, tr->key, size);
>> +       if (err < 0)
>> +               goto out_free_im;
> 
> I feel this change just makes bpf_trampoline_update() even
> more confusing.
> 
>>
>>          err = arch_prepare_bpf_trampoline(im, im->image, im->image + size,
>>                                            &tr->func.model, tr->flags, tlinks,
>> @@ -496,6 +493,8 @@ static int bpf_trampoline_update(struct bpf_trampoline *tr, bool lock_direct_mut
>>
>>   out_free:
>>          bpf_tramp_image_free(im);
>> +out_free_im:
>> +       kfree_rcu(im, rcu);
> 
> If we goto out_free above, we will call kfree_rcu(im, rcu)
> twice, right? Once in bpf_tramp_image_free(), and again
> here.
> 

Oops, sorry, forgot to remove kfree_rcu in bpf_tramp_image_free in this 
version.

> Thanks,
> Song
> 
> [...]
>