Re: [PATCH v4 7/9] kentry: Add debugging checks for proper kentry API usage

[Thomas Gleixner]

On Fri, Mar 19 2021 at 17:17, Thomas Gleixner wrote:

> On Wed, Mar 17 2021 at 11:12, Andy Lutomirski wrote:
>> +
>> +#define DEBUG_ENTRY_WARN_ONCE(condition, format...) do {} while (0)
>
> So we have a stub for !DEBUG
>
>> +static __always_inline void kentry_cpu_depth_add(unsigned int n) {}
>> +static void kentry_cpu_depth_check(unsigned int n) {}
>> +static __always_inline void kentry_cpu_depth_sub(unsigned int n) {}
>> +
>> +#endif
>> +
>>  /* See comment for enter_from_user_mode() in entry-common.h */
>>  static __always_inline void __enter_from_user_mode(struct pt_regs *regs)
>>  {
>> +	kentry_cpu_depth_add(1);
>>  	arch_check_user_regs(regs);
>>  	lockdep_hardirqs_off(CALLER_ADDR0);
>>  
>> @@ -22,6 +78,14 @@ static __always_inline void __enter_from_user_mode(struct pt_regs *regs)
>>  
>>  	instrumentation_begin();
>>  	trace_hardirqs_off_finish();
>> +
>> +#ifdef CONFIG_DEBUG_ENTRY
>
> Why do we need that #ifdeffery all over the place?
>
>> +	DEBUG_ENTRY_WARN_ONCE(
>> +		this_cpu_read(kentry_cpu_depth) != 1,
>> +		"kentry: __enter_from_user_mode() called while kentry thought the CPU was in the kernel (%u)",
>> +		this_cpu_read(kentry_cpu_depth));

Because you directly access kentry_cpu_depth which makes the compiler
unhappy.

And of course at the other place where you guard it with IS_ENABLED() it
fails to build with CONFIG_DEBUG_ENTRY=n

kernel/entry/common.c:158:10: error: ‘struct task_struct’ has no member named ‘kentry_in_syscall’
   current->kentry_in_syscall = true;

This is V4 of this series... Oh well.

     tglx