Re: [PATCH v2] mm/swap: fix swap_info_struct race between swapoff and get_swap_pages()

[Rongwei Wang <rongwei.wang@linux.alibaba.com>]

Oh, sorry, I miss this email just now, that because of I'm also replying 
your previous email.

On 2023/4/6 20:12, Aaron Lu wrote:
> On Wed, Apr 05, 2023 at 12:08:47AM +0800, Rongwei Wang wrote:
>> Hello
>>
>> I have fix up some stuff base on Patch v1. And in order to help all readers
>> and reviewers to
>>
>> reproduce this bug, share a reproducer here:
> I reproduced this problem under a VM this way:
>
> $ sudo ./stress-ng --swap 1
> // on another terminal
> $ for i in `seq 8`; do ./swap & done
> Looks simpler than yours :-)
Cool, indeed become simpler.
> (Didn't realize you have posted your reproducer here since I'm not CCed
> and just found it after invented mine)
> Then the warning message normally appear within a few seconds.
>
> Here is the code for the above swap prog:
>
> #include <stdio.h>
> #include <stddef.h>
> #include <sys/mman.h>
>
> #define SIZE 0x100000
>
> int main(void)
> {
>          int i, ret;
>          void *p;
>
>          p = mmap(NULL, SIZE, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANON, -1, 0);
>          if (p == MAP_FAILED) {
>                  perror("mmap");
>                  return -1;
>          }
>
>          ret = 0;
>          while (1) {
>                  for (i = 0; i < SIZE; i += 0x1000)
>                          ((char *)p)[i] = 1;
>                  ret = madvise(p, SIZE, MADV_PAGEOUT);
>                  if (ret != 0) {
>                          perror("madvise");
>                          break;
>                  }
>          }
>
>          return ret;
> }
>
> Unfortunately, this test prog did not work on kernels before v5.4 because
> MADV_PAGEOUT is introduced in v5.4. I tested on v5.4 and the problem is
> also there.

Maybe that is this bug can not be found since now. And we found this is 
triggered by stress-ng-swap and stress-ng-madvise (PAGEOUT) firstly. It 
seems this is that reason.

It seems MADV_COLD is also introduced together with MADV_PAGEOUT. I have 
no idea and have to depend on you.:-)

>
> Haven't found a way to trigger swap with swap device come and go on
> kernels before v5.4; tried putting the test prog in a memcg with memory
> limit but then the prog is easily killed due to nowhere to swap out.
>
Personally, I do not intend to continuing searching for the method to 
reproduce before v5.4. Of course, if you have idea, I can try.


Thanks:-)