Re: [PATCH v2 6/9] mm/memory: convert print_bad_pte() to print_bad_page_map()

[David Hildenbrand <david@xxxxxxxxxx>]

> > The report will now look something like (dumping pgd to pmd values):
> >
> > [   77.943408] BUG: Bad page map in process XXX  entry:80000001233f5867
> > [   77.944077] addr:00007fd84bb1c000 vm_flags:08100071 anon_vma: ...
> > [   77.945186] pgd:10a89f067 p4d:10a89f067 pud:10e5a2067 pmd:105327067
> >
> > Not using pgdp_get(), because that does not work properly on some arm
> > configs where pgd_t is an array. Note that we are dumping all levels
> > even when levels are folded for simplicity.
> Oh god. I reviewed this below. BUT OH GOD. What. Why???
Exactly.

I wish this patch wouldn't exist, but Hugh convinced me that apparently 
it can be useful in the real world.
> > Signed-off-by: David Hildenbrand <david@xxxxxxxxxx>
> > ---
> >   mm/memory.c | 120 ++++++++++++++++++++++++++++++++++++++++------------
> >   1 file changed, 94 insertions(+), 26 deletions(-)
> >
> > diff --git a/mm/memory.c b/mm/memory.c
> > index 173eb6267e0ac..08d16ed7b4cc7 100644
> > --- a/mm/memory.c
> > +++ b/mm/memory.c
> > @@ -473,22 +473,8 @@ static inline void add_mm_rss_vec(struct mm_struct *mm, 
> > int *rss)
> >                         add_mm_counter(mm, i, rss[i]);
> >   }
> >
> > -/*
> > - * This function is called to print an error when a bad pte
> > - * is found. For example, we might have a PFN-mapped pte in
> > - * a region that doesn't allow it.
> > - *
> > - * The calling function must still handle the error.
> > - */
> > -static void print_bad_pte(struct vm_area_struct *vma, unsigned long addr,
> > -                         pte_t pte, struct page *page)
> > +static bool is_bad_page_map_ratelimited(void)
> >   {
> > -       pgd_t *pgd = pgd_offset(vma->vm_mm, addr);
> > -       p4d_t *p4d = p4d_offset(pgd, addr);
> > -       pud_t *pud = pud_offset(p4d, addr);
> > -       pmd_t *pmd = pmd_offset(pud, addr);
> > -       struct address_space *mapping;
> > -       pgoff_t index;
> >         static unsigned long resume;
> >         static unsigned long nr_shown;
> >         static unsigned long nr_unshown;
> > @@ -500,7 +486,7 @@ static void print_bad_pte(struct vm_area_struct *vma, 
> > unsigned long addr,
> >         if (nr_shown == 60) {
> >                 if (time_before(jiffies, resume)) {
> >                         nr_unshown++;
> > -                       return;
> > +                       return true;
> >                 }
> >                 if (nr_unshown) {
> >                         pr_alert("BUG: Bad page map: %lu messages suppressed\n",
> > @@ -511,15 +497,87 @@ static void print_bad_pte(struct vm_area_struct *vma, 
> > unsigned long addr,
> >         }
> >         if (nr_shown++ == 0)
> >                 resume = jiffies + 60 * HZ;
> > +       return false;
> > +}
> > +
> > +static void __dump_bad_page_map_pgtable(struct mm_struct *mm, unsigned long 
> > addr)
> > +{
> > +       unsigned long long pgdv, p4dv, pudv, pmdv;
> > +       p4d_t p4d, *p4dp;
> > +       pud_t pud, *pudp;
> > +       pmd_t pmd, *pmdp;
> > +       pgd_t *pgdp;
> > +
> > +       /*
> > +        * This looks like a fully lockless walk, however, the caller is
> > +        * expected to hold the leaf page table lock in addition to other
> > +        * rmap/mm/vma locks. So this is just a re-walk to dump page table
> > +        * content while any concurrent modifications should be completely
> > +        * prevented.
> > +        */
> Hmmm :)
>
> Why aren't we trying to lock at leaf level?
Ehm, did you read the:

"the caller is expected to hold the leaf page table lock"

:)


> We need to:
>
> - Keep VMA stable which prevents unmap page table teardown and khugepaged
>    collapse.
> - (not relevant as we don't traverse PTE table but) RCU lock for PTE
>    entries to avoid MADV_DONTNEED page table withdrawal.
>
> Buuut if we're not locking at leaf level, we leave ourselves open to racing
> faults, zaps, etc. etc.
Yes. I can clarify in the comment of print_bad_page_map(), that it is 
expected to be called with the PTL (unwritten rule, not changing that).
> So perhaps this why you require such strict conditions...
>
> But can you truly be sure of these existing? And we should then assert them
> here no? For rmap though we'd need the folio/vma.
I hope you realize that this nastiness of a code is called in case our 
system is already running into something extremely unexpected and will 
probably be dead soon.
So I am not to interested in adding anything more here. If you run into 
this code you're in big trouble already.
> > +       pgdp = pgd_offset(mm, addr);
> > +       pgdv = pgd_val(*pgdp);
> Before I went and looked again at the commit msg I said:
>
>         "Shoudln't we strictly speaking use pgdp_get()? I see you use this
>          helper for other levels."
>
> But obviously yeah. You explained the insane reason why not.
Had to find out the hard way ... :)

[...]

> > +/*
> > + * This function is called to print an error when a bad page table entry (e.g.,
> > + * corrupted page table entry) is found. For example, we might have a
> > + * PFN-mapped pte in a region that doesn't allow it.
> > + *
> > + * The calling function must still handle the error.
> > + */
> We have extremely strict locking conditions for the page table traversal... but
> no mention of them here?
Yeah, I can add that.

> > +static void print_bad_page_map(struct vm_area_struct *vma,
> > +               unsigned long addr, unsigned long long entry, struct page *page)
> > +{
> > +       struct address_space *mapping;
> > +       pgoff_t index;
> > +
> > +       if (is_bad_page_map_ratelimited())
> > +               return;
> >
> >         mapping = vma->vm_file ? vma->vm_file->f_mapping : NULL;
> >         index = linear_page_index(vma, addr);
> >
> > -       pr_alert("BUG: Bad page map in process %s  pte:%08llx pmd:%08llx\n",
> > -                current->comm,
> > -                (long long)pte_val(pte), (long long)pmd_val(*pmd));
> > +       pr_alert("BUG: Bad page map in process %s  entry:%08llx", 
> > current->comm, entry);
> Sort of wonder if this is even useful if you don't know what the 'entry'
> is? But I guess the dump below will tell you.
You probably missed in the patch description:

"Whether it is a PTE or something else will usually become obvious from 
the page table dump or from the dumped stack. If ever required in the 
future, we could pass the entry level type similar to "enum rmap_level". 
For now, let's keep it simple."
> Though maybe actually useful to see flags etc. in case some horrid
> corruption happened and maybe dump isn't valid? But then the dump assumes
> strict conditions to work so... can that happen?
Not sure what you mean, can you elaborate?

If you system is falling apart completely, I'm afraid this report here 
will not safe you.
You'll probably get a flood of 60 ... if your system doesn't collapse 
before that.
> > +       __dump_bad_page_map_pgtable(vma->vm_mm, addr);
> >         if (page)
> > -               dump_page(page, "bad pte");
> > +               dump_page(page, "bad page map");
> >         pr_alert("addr:%px vm_flags:%08lx anon_vma:%px mapping:%px index:%lx\n",
> >                  (void *)addr, vma->vm_flags, vma->anon_vma, mapping, index);
> >         pr_alert("file:%pD fault:%ps mmap:%ps mmap_prepare: %ps 
> > read_folio:%ps\n",
> > @@ -597,7 +655,7 @@ struct page *vm_normal_page(struct vm_area_struct *vma, 
> > unsigned long addr,
> >                 if (is_zero_pfn(pfn))
> >                         return NULL;
> >
> > -               print_bad_pte(vma, addr, pte, NULL);
> > +               print_bad_page_map(vma, addr, pte_val(pte), NULL);
> >                 return NULL;
> >         }
> >
> > @@ -625,7 +683,7 @@ struct page *vm_normal_page(struct vm_area_struct *vma, 
> > unsigned long addr,
> >
> >   check_pfn:
> >         if (unlikely(pfn > highest_memmap_pfn)) {
> > -               print_bad_pte(vma, addr, pte, NULL);
> > +               print_bad_page_map(vma, addr, pte_val(pte), NULL);
> This is unrelated to your series, but I guess this is for cases where
> you're e.g. iomapping or such? So it's not something in the memmap but it's
> a PFN that might reference io memory or such?
No, it's just an easy check for catching corruptions. In a later patch I 
add a comment.
> >                 return NULL;
> >         }
> >
> > @@ -654,8 +712,15 @@ struct page *vm_normal_page_pmd(struct vm_area_struct 
> > *vma, unsigned long addr,
> >   {
> >         unsigned long pfn = pmd_pfn(pmd);
> >
> > -       if (unlikely(pmd_special(pmd)))
> > +       if (unlikely(pmd_special(pmd))) {
> > +               if (vma->vm_flags & (VM_PFNMAP | VM_MIXEDMAP))
> > +                       return NULL;
> I guess we'll bring this altogether in a later patch with vm_normal_page()
> as getting a little duplicative :P
Not that part, but the other part :)

> Makes me think that VM_SPECIAL is kind of badly named (other than fact
> 'special' is nebulous and overloaded in general) in that it contains stuff
> that is -VMA-special but only VM_PFNMAP | VM_MIXEDMAP really indicates
> specialness wrt to underlying folio.
It is.

> Then we have VM_IO, which strictly must not have an associated page right?
VM_IO just means read/write side-effects, I think you could have ones 
with an memmap easily ... e.g., memory section (128MiB) spanning both 
memory and MMIO regions.
Thanks!

--
Cheers,

David / dhildenb