[RFC,RESEND,bpf-next,1/4] bpf: Rollback to text_poke when arch not supported ftrace direct call

Commit Message

Pu Lehui Dec. 20, 2022, 2:13 a.m. UTC

From: Pu Lehui <[e-mail]>

The current bpf trampoline attach to kernel functions via ftrace direct
call API, while text_poke is applied for bpf2bpf attach and tail call
optimization. For architectures that do not support ftrace direct call,
text_poke is still able to attach bpf trampoline to kernel functions.
Let's relax it by rollback to text_poke when architecture not supported.

Signed-off-by: Pu Lehui <[e-mail]>
---
 kernel/bpf/trampoline.c | 8 ++------
 1 file changed, 2 insertions(+), 6 deletions(-)

Comments

Xu Kuohai Dec. 20, 2022, 2:32 a.m. UTC | #1

On 12/20/2022 10:13 AM, Pu Lehui wrote:
> From: Pu Lehui <[e-mail]>
> 
> The current bpf trampoline attach to kernel functions via ftrace direct
> call API, while text_poke is applied for bpf2bpf attach and tail call
> optimization. For architectures that do not support ftrace direct call,
> text_poke is still able to attach bpf trampoline to kernel functions.
> Let's relax it by rollback to text_poke when architecture not supported.
> 
> Signed-off-by: Pu Lehui <[e-mail]>
> ---
>   kernel/bpf/trampoline.c | 8 ++------
>   1 file changed, 2 insertions(+), 6 deletions(-)
> 
> diff --git a/kernel/bpf/trampoline.c b/kernel/bpf/trampoline.c
> index d6395215b849..386197a7952c 100644
> --- a/kernel/bpf/trampoline.c
> +++ b/kernel/bpf/trampoline.c
> @@ -228,15 +228,11 @@ static int modify_fentry(struct bpf_trampoline *tr, void *old_addr, void *new_ad
>   static int register_fentry(struct bpf_trampoline *tr, void *new_addr)
>   {
>   	void *ip = tr->func.addr;
> -	unsigned long faddr;
>   	int ret;
>   
> -	faddr = ftrace_location((unsigned long)ip);
> -	if (faddr) {
> -		if (!tr->fops)
> -			return -ENOTSUPP;
> +	if (IS_ENABLED(CONFIG_DYNAMIC_FTRACE_WITH_DIRECT_CALLS) &&
> +	    !!ftrace_location((unsigned long)ip))
>   		tr->func.ftrace_managed = true;
> -	}
>

After this patch, a kernel function with true trace_location will be patched
by bpf when CONFIG_DYNAMIC_FTRACE_WITH_DIRECT_CALLS is disabled, which means
that a kernel function may be patched by both bpf and ftrace in a mutually
unaware way. This will cause ftrace and bpf_arch_text_poke to fail in a
somewhat random way if the function to be patched was already patched
by the other.

>   	if (bpf_trampoline_module_get(tr))
>   		return -ENOENT;

Pu Lehui Dec. 21, 2022, 2:36 a.m. UTC | #2

On 2022/12/20 10:32, Xu Kuohai wrote:
> On 12/20/2022 10:13 AM, Pu Lehui wrote:
>> From: Pu Lehui <[e-mail]>
>>
>> The current bpf trampoline attach to kernel functions via ftrace direct
>> call API, while text_poke is applied for bpf2bpf attach and tail call
>> optimization. For architectures that do not support ftrace direct call,
>> text_poke is still able to attach bpf trampoline to kernel functions.
>> Let's relax it by rollback to text_poke when architecture not supported.
>>
>> Signed-off-by: Pu Lehui <[e-mail]>
>> ---
>>   kernel/bpf/trampoline.c | 8 ++------
>>   1 file changed, 2 insertions(+), 6 deletions(-)
>>
>> diff --git a/kernel/bpf/trampoline.c b/kernel/bpf/trampoline.c
>> index d6395215b849..386197a7952c 100644
>> --- a/kernel/bpf/trampoline.c
>> +++ b/kernel/bpf/trampoline.c
>> @@ -228,15 +228,11 @@ static int modify_fentry(struct bpf_trampoline 
>> *tr, void *old_addr, void *new_ad
>>   static int register_fentry(struct bpf_trampoline *tr, void *new_addr)
>>   {
>>       void *ip = tr->func.addr;
>> -    unsigned long faddr;
>>       int ret;
>> -    faddr = ftrace_location((unsigned long)ip);
>> -    if (faddr) {
>> -        if (!tr->fops)
>> -            return -ENOTSUPP;
>> +    if (IS_ENABLED(CONFIG_DYNAMIC_FTRACE_WITH_DIRECT_CALLS) &&
>> +        !!ftrace_location((unsigned long)ip))
>>           tr->func.ftrace_managed = true;
>> -    }
>>
> 
> After this patch, a kernel function with true trace_location will be 
> patched
> by bpf when CONFIG_DYNAMIC_FTRACE_WITH_DIRECT_CALLS is disabled, which 
> means
> that a kernel function may be patched by both bpf and ftrace in a mutually
> unaware way. This will cause ftrace and bpf_arch_text_poke to fail in a
> somewhat random way if the function to be patched was already patched
> by the other.

Thanks for your review. And yes, this is a backward compatible solution 
for architectures that not support DYNAMIC_FTRACE_WITH_DIRECT_CALLS.

> 
>>       if (bpf_trampoline_module_get(tr))
>>           return -ENOENT;

Björn Töpel Jan. 3, 2023, 12:05 p.m. UTC | #3

Pu Lehui <[e-mail]> writes:

> On 2022/12/20 10:32, Xu Kuohai wrote:
>> On 12/20/2022 10:13 AM, Pu Lehui wrote:
>>> From: Pu Lehui <[e-mail]>
>>>
>>> The current bpf trampoline attach to kernel functions via ftrace direct
>>> call API, while text_poke is applied for bpf2bpf attach and tail call
>>> optimization. For architectures that do not support ftrace direct call,
>>> text_poke is still able to attach bpf trampoline to kernel functions.
>>> Let's relax it by rollback to text_poke when architecture not supported.
>>>
>>> Signed-off-by: Pu Lehui <[e-mail]>
>>> ---
>>>   kernel/bpf/trampoline.c | 8 ++------
>>>   1 file changed, 2 insertions(+), 6 deletions(-)
>>>
>>> diff --git a/kernel/bpf/trampoline.c b/kernel/bpf/trampoline.c
>>> index d6395215b849..386197a7952c 100644
>>> --- a/kernel/bpf/trampoline.c
>>> +++ b/kernel/bpf/trampoline.c
>>> @@ -228,15 +228,11 @@ static int modify_fentry(struct bpf_trampoline 
>>> *tr, void *old_addr, void *new_ad
>>>   static int register_fentry(struct bpf_trampoline *tr, void *new_addr)
>>>   {
>>>       void *ip = tr->func.addr;
>>> -    unsigned long faddr;
>>>       int ret;
>>> -    faddr = ftrace_location((unsigned long)ip);
>>> -    if (faddr) {
>>> -        if (!tr->fops)
>>> -            return -ENOTSUPP;
>>> +    if (IS_ENABLED(CONFIG_DYNAMIC_FTRACE_WITH_DIRECT_CALLS) &&
>>> +        !!ftrace_location((unsigned long)ip))
>>>           tr->func.ftrace_managed = true;
>>> -    }
>>>
>> 
>> After this patch, a kernel function with true trace_location will be 
>> patched
>> by bpf when CONFIG_DYNAMIC_FTRACE_WITH_DIRECT_CALLS is disabled, which 
>> means
>> that a kernel function may be patched by both bpf and ftrace in a mutually
>> unaware way. This will cause ftrace and bpf_arch_text_poke to fail in a
>> somewhat random way if the function to be patched was already patched
>> by the other.
>
> Thanks for your review. And yes, this is a backward compatible solution 
> for architectures that not support DYNAMIC_FTRACE_WITH_DIRECT_CALLS.

It's not "backward compatible". Reiterating what Kuohai said; The BPF
trampoline must be aware of ftrace's state -- with this patch, the
trampoline can't blindly poke the text managed my ftrace.

I'd recommend to remove this patch from the series.


Björn

Pu Lehui Jan. 6, 2023, 2:35 a.m. UTC | #4

On 2023/1/3 20:05, Björn Töpel wrote:
> Pu Lehui <[e-mail]> writes:
> 
>> On 2022/12/20 10:32, Xu Kuohai wrote:
>>> On 12/20/2022 10:13 AM, Pu Lehui wrote:
>>>> From: Pu Lehui <[e-mail]>
>>>>
>>>> The current bpf trampoline attach to kernel functions via ftrace direct
>>>> call API, while text_poke is applied for bpf2bpf attach and tail call
>>>> optimization. For architectures that do not support ftrace direct call,
>>>> text_poke is still able to attach bpf trampoline to kernel functions.
>>>> Let's relax it by rollback to text_poke when architecture not supported.
>>>>
>>>> Signed-off-by: Pu Lehui <[e-mail]>
>>>> ---
>>>>    kernel/bpf/trampoline.c | 8 ++------
>>>>    1 file changed, 2 insertions(+), 6 deletions(-)
>>>>
>>>> diff --git a/kernel/bpf/trampoline.c b/kernel/bpf/trampoline.c
>>>> index d6395215b849..386197a7952c 100644
>>>> --- a/kernel/bpf/trampoline.c
>>>> +++ b/kernel/bpf/trampoline.c
>>>> @@ -228,15 +228,11 @@ static int modify_fentry(struct bpf_trampoline
>>>> *tr, void *old_addr, void *new_ad
>>>>    static int register_fentry(struct bpf_trampoline *tr, void *new_addr)
>>>>    {
>>>>        void *ip = tr->func.addr;
>>>> -    unsigned long faddr;
>>>>        int ret;
>>>> -    faddr = ftrace_location((unsigned long)ip);
>>>> -    if (faddr) {
>>>> -        if (!tr->fops)
>>>> -            return -ENOTSUPP;
>>>> +    if (IS_ENABLED(CONFIG_DYNAMIC_FTRACE_WITH_DIRECT_CALLS) &&
>>>> +        !!ftrace_location((unsigned long)ip))
>>>>            tr->func.ftrace_managed = true;
>>>> -    }
>>>>
>>>
>>> After this patch, a kernel function with true trace_location will be
>>> patched
>>> by bpf when CONFIG_DYNAMIC_FTRACE_WITH_DIRECT_CALLS is disabled, which
>>> means
>>> that a kernel function may be patched by both bpf and ftrace in a mutually
>>> unaware way. This will cause ftrace and bpf_arch_text_poke to fail in a
>>> somewhat random way if the function to be patched was already patched
>>> by the other.
>>
>> Thanks for your review. And yes, this is a backward compatible solution
>> for architectures that not support DYNAMIC_FTRACE_WITH_DIRECT_CALLS.
> 
> It's not "backward compatible". Reiterating what Kuohai said; The BPF
> trampoline must be aware of ftrace's state -- with this patch, the
> trampoline can't blindly poke the text managed my ftrace.
> 
> I'd recommend to remove this patch from the series.
> 

After deep consideration, Kuohai's catching is much more reasonable. 
Will remove it in the next.

In the meantime, I found that song and guoren have worked on supporting 
riscv ftrace with direct call [0], so we can concentrate on making 
bpf_arch_text_poke specifically for the bpf context.

However, riscv ftrace base framework will change because [0] uses t0 as 
the link register of traced function. We should consider the generality 
of riscv bpf trampoline for kernel function and bpf context. It's not 
clear if [0] will be upstreamed, so maybe we should wait for it?

[0] 
https://lore.kernel.org/linux-riscv/[e-mail]

Anyway, thanks both of you for the review.

> 
> Björn

Patch

diff --git a/kernel/bpf/trampoline.c b/kernel/bpf/trampoline.c
index d6395215b849..386197a7952c 100644
--- a/kernel/bpf/trampoline.c
+++ b/kernel/bpf/trampoline.c
@@ -228,15 +228,11 @@  static int modify_fentry(struct bpf_trampoline *tr, void *old_addr, void *new_ad
 static int register_fentry(struct bpf_trampoline *tr, void *new_addr)
 {
 	void *ip = tr->func.addr;
-	unsigned long faddr;
 	int ret;
 
-	faddr = ftrace_location((unsigned long)ip);
-	if (faddr) {
-		if (!tr->fops)
-			return -ENOTSUPP;
+	if (IS_ENABLED(CONFIG_DYNAMIC_FTRACE_WITH_DIRECT_CALLS) &&
+	    !!ftrace_location((unsigned long)ip))
 		tr->func.ftrace_managed = true;
-	}
 
 	if (bpf_trampoline_module_get(tr))
 		return -ENOENT;