lkml.org 
[lkml]   [2018]   [Jan]   [18]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
    /
    Date
    From
    SubjectRe: [mm 4.15-rc8] Random oopses under memory pressure.
    On Thu, Jan 18, 2018 at 06:45:00AM -0800, Dave Hansen wrote:
    > On 01/18/2018 04:25 AM, Kirill A. Shutemov wrote:
    > > [ 10.084024] diff: -858690919
    > > [ 10.084258] hpage_nr_pages: 1
    > > [ 10.084386] check1: 0
    > > [ 10.084478] check2: 0
    > ...
    > > diff --git a/mm/page_vma_mapped.c b/mm/page_vma_mapped.c
    > > index d22b84310f6d..57b4397f1ea5 100644
    > > --- a/mm/page_vma_mapped.c
    > > +++ b/mm/page_vma_mapped.c
    > > @@ -70,6 +70,14 @@ static bool check_pte(struct page_vma_mapped_walk *pvmw)
    > > }
    > > if (pte_page(*pvmw->pte) < pvmw->page)
    > > return false;
    > > +
    > > + if (pte_page(*pvmw->pte) - pvmw->page) {
    > > + printk("diff: %d\n", pte_page(*pvmw->pte) - pvmw->page);
    > > + printk("hpage_nr_pages: %d\n", hpage_nr_pages(pvmw->page));
    > > + printk("check1: %d\n", pte_page(*pvmw->pte) - pvmw->page < 0);
    > > + printk("check2: %d\n", pte_page(*pvmw->pte) - pvmw->page >= hpage_nr_pages(pvmw->page));
    > > + BUG();
    > > + }
    >
    > This says that pte_page(*pvmw->pte) and pvmw->page are roughly 4GB away
    > from each other (858690919*4=0xccba559c0). That's not the compiler
    > being wonky, it just means that the virtual addresses of the memory
    > sections are that far apart.
    >
    > This won't happen when you have vmemmap or flatmem because the mem_map[]
    > is virtually contiguous and pointer arithmetic just works against all
    > 'struct page' pointers. But with classic sparsemem, it doesn't.
    >
    > You need to make sure that the PFNs are in the same section before you
    > can do the math that you want to do here.

    Isn't it simply that pvmw->page isn't a page or pte_page(*pvmw->pte)
    isn't a page?

    The distance cannot matter, MMU isn't involved, this is pure 64bit
    aritmetics, 1giga 1 terabyte, 48bits 5level pagetables are meaningless
    in this comparison.

    #include <stdio.h>

    int main()
    {
    volatile long i;
    struct x { char a[4000000000]; };
    for (i = 0; i < 4000000000*3; i += 4000000000) {
    printf("%ld\n", ((struct x *)0)-((((struct x *)i))));
    }
    printf("xxxx\n");
    for (i = 0; i < 4000000000; i += 1) {
    if (i==4)
    i = 4000000000;
    printf("%ld\n", ((struct x *)0)-((((struct x *)i))));
    }
    return 0;
    }

    You need to add two debug checks on "pte_page(*pvmw->pte) % 64" and
    same for pvmw->page to find out the one of the two that isn't a page.

    If both are real pages there's a bug that allocates page structs not
    naturally aligned.

    \
     
     \ /
      Last update: 2018-01-18 15:59    [W:4.118 / U:0.012 seconds]
    ©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site