Re: a racy access flag clearing warning when calling mmap system call

[chenjiankang]

> On Fri, Dec 08, 2017 at 11:19:52AM +0800, chenjiankang wrote:
>> å 2017/12/7 21:23, Will Deacon åé:
>>> diff --git a/arch/arm64/include/asm/pgtable.h b/arch/arm64/include/asm/pgtable.h
>>> index 149d05fb9421..8fe103b1e101 100644
>>> --- a/arch/arm64/include/asm/pgtable.h
>>> +++ b/arch/arm64/include/asm/pgtable.h
>>> @@ -42,6 +42,8 @@
>>>  #include <asm/cmpxchg.h>
>>>  #include <asm/fixmap.h>
>>>  #include <linux/mmdebug.h>
>>> +#include <linux/mm_types.h>
>>> +#include <linux/sched.h>
>>>  
>>>  extern void __pte_error(const char *file, int line, unsigned long val);
>>>  extern void __pmd_error(const char *file, int line, unsigned long val);
>>> @@ -207,9 +209,6 @@ static inline void set_pte(pte_t *ptep, pte_t pte)
>>>  	}
>>>  }
>>>  
>>> -struct mm_struct;
>>> -struct vm_area_struct;
>>> -
>>>  extern void __sync_icache_dcache(pte_t pteval, unsigned long addr);
>>>  
>>>  /*
>>> @@ -238,7 +237,8 @@ static inline void set_pte_at(struct mm_struct *mm, unsigned long addr,
>>>  	 * hardware updates of the pte (ptep_set_access_flags safely changes
>>>  	 * valid ptes without going through an invalid entry).
>>>  	 */
>>> -	if (pte_valid(*ptep) && pte_valid(pte)) {
>>> +	if (IS_ENABLED(CONFIG_DEBUG_VM) && pte_valid(*ptep) && pte_valid(pte) &&
>>> +	   (mm == current->active_mm || atomic_read(&mm->mm_users) > 1)) {
>>>  		VM_WARN_ONCE(!pte_young(pte),
>>>  			     "%s: racy access flag clearing: 0x%016llx -> 0x%016llx",
>>>  			     __func__, pte_val(*ptep), pte_val(pte));
> [...]

Hi Will:
     I contruct a simple use case that can reproduce this fail;

like this: 
   #define LEN 1024*1024*100
                       
int main(void){        
        int* addr = NULL; 
        int pid = -1; 
        addr = (int*)mmap(NULL, LEN, PROT_READ | PROT_WRITE, MAP_ANONYMOUS|MAP_PRIVATE, -1, 0); 
        madvise(addr, LEN, MADV_HUGEPAGE);
        memset(addr, 1, LEN);
        pid = fork();                                                                                                                
                       
        if(pid==0){ 
                printf("wow! I am a child!\n");
        }              
        else {         
                printf("I am a father!\n");
                mmap(addr, 1024*1024*10, PROT_READ | PROT_WRITE, MAP_ANONYMOUS | MAP_PRIVATE | MAP_FIXED, -1, 0); 
        }              
                       
        return 0;   
} 

And then, I use the will's modification,which can solve this problem;
Will, this patch should send a upstream?

>>     From the print information, the only difference between pte and ptep is the PTE_SPECIAL bit.
>> And the the PTE access bit is all zero.
>>     diff below. Whether the access bit of the new ptep should be judged to eliminate the 
>> false positive?
> [...]
>> diff --git a/arch/arm64/include/asm/pgtable.h b/arch/arm64/include/asm/pgtable.h
>> index 2987d5a..3c1b0c6 100644
>> --- a/arch/arm64/include/asm/pgtable.h
>> +++ b/arch/arm64/include/asm/pgtable.h
>> @@ -206,7 +206,7 @@ static inline void set_pte_at(struct mm_struct *mm, unsigned long addr,
>>          * valid ptes without going through an invalid entry).
>>          */
>>         if (IS_ENABLED(CONFIG_ARM64_HW_AFDBM) && pte_valid(*ptep)) {
>> -               VM_WARN_ONCE(!pte_young(pte),
>> +               VM_WARN_ONCE(!pte_young(pte) && pte_young(*ptep),
>>                              "%s: racy access flag clearing: %016llx -> %016llx",
>>                              __func__, pte_val(*ptep), pte_val(pte));
> 
> It's actually the other way around: *ptep being "old" (AF = 0) could at
> any point be made "young" by the hardware (AF = 1). This is racing with
> the software update which keeps the AF bit 0.
>