Web lists-archives.com

Re: [Question] There is a thp count left when the process exits


I can not find the possibility when I check the code. because the mmap_sem and spin_lock will
protect the concurrence.

I will be appreciated if anyone has some clue.

zhong jiang

On 2018/11/8 23:01, zhong jiang wrote:
> Hi, 
> Recently,  I hit  the following issue in linux 3.10 stable.  and hard to recur.
> bad pmd ffff880c13ecea80(80000017b16000e7)
> Call Trace:
>   [<ffffffff8164195f>] dump_stack+0x19/0x1b
>   [<ffffffff8107b230>] warn_slowpath_common+0x70/0xb0
>   [<ffffffff8107b37a>] warn_slowpath_null+0x1a/0x20
>   [<ffffffff811a2b86>] exit_mmap+0x196/0x1a0
>   [<ffffffff810782e7>] mmput+0x67/0xf0
>   [<ffffffff81081b2c>] do_exit+0x28c/0xa60
>   [<ffffffff810a9dc0>] ? hrtimer_get_res+0x50/0x50
>   [<ffffffff8108237f>] do_group_exit+0x3f/0xa0
>   [<ffffffff81093240>] get_signal_to_deliver+0x1d0/0x6d0
>   [<ffffffff81014427>] do_signal+0x57/0x6b0
>   [<ffffffff810e4f92>] ? futex_wait_queue_me+0xa2/0x120
>   [<ffffffff8164d323>] ? __do_page_fault+0x183/0x470
>   [<ffffffff81014adf>] do_notify_resume+0x5f/0xb0
>   [<ffffffff816520bd>] int_signal+0x12/0x17
> BUG: Bad rss-counter state mm:ffff8820136b5dc0 idx:1 val:512
> The pmd entry show that it is still a thp. but It fails to check and clear the pmd.
> hence,   page fault will produce a new page for pmd when accessing the page,
> which thp count will reduplicative increase.
> Thanks,
> zhong jiang