Re: [PATCH v1 1/1] mm: buddy page accessed before initialized

[Pavel Tatashin]

> > Yes, but as I said, unfortunately memset(1) with CONFIG_VM_DEBUG does not
> > catch this case. So, when CONFIG_VM_DEBUG is enabled kexec reboots without
> > issues.
>
> Can we make the init pattern to catch this?

Unfortunately, that is not easy: memset() gives us only one byte to play 
with, and if we use something else that will make CONFIG_VM_DEBUG 
unacceptably slow.

One byte is not enough to trigger the pattern that satisfy 
page_is_buddy() logic. I have tried it. With kexec, however it is more 
predictable: we use the same memory during boot to allocate vmemmap, and 
therefore the struct pages are more like "valid" struct pages from the 
previous boot.

> > > > > > This is why we must initialize the computed buddy page beforehand.
> > > > >
> > > > > Ble, this is really ugly. I will think about it more.
> > > >
> > > > Another approach that I considered is to split loop inside
> > > > deferred_init_range() into two loops: one where we initialize pages by
> > > > calling __init_single_page(), another where we free them to buddy allocator
> > > > by calling deferred_free_range().
> > >
> > > Yes, that would make much more sense to me.
> >
> > Ok, so should I submit a new patch with two loops? (The logic within loops
> > is going to be the same:
>
> Could you post it please?
>   
> > if (!pfn_valid_within(pfn)) {
> > } else if (!(pfn & nr_pgmask) && !pfn_valid(pfn)) {
> > } else if (!meminit_pfn_in_nid(pfn, nid, &nid_init_state)) {
> > } else if (page && (pfn & nr_pgmask)) {
> >
> > This fix was already added into mm-tree as
> > mm-deferred_init_memmap-improvements-fix-2.patch
>
> I think Andrew can drop it and replace by a different patch.

The new patch is coming, I will test it on two machines where I observed 
the problem.

Thank you,
Pasha