Re: kernel BUG at mm/huge_memory.c:2613!

[Andrea Arcangeli <aarcange@redhat.com>]

Hello,

On Mon, Jun 22, 2020 at 04:30:41PM +0100, Robin Murphy wrote:
> On 2020-06-22 13:46, Joerg Roedel wrote:
> > + Robin
> > 
> > Robin, any idea on this?
> 
> After a bit of archaeology, this dates back to the original review:
> 
> https://lore.kernel.org/linux-arm-kernel/54C285D4.3070802@arm.com/
> https://lore.kernel.org/linux-arm-kernel/54DA2666.9030003@arm.com/
> 
> In summary: originally this inherited from other arch code that did 
> simply strip __GFP_COMP; that was deemed questionable because of the 
> nonsensical comment about CONFIG_HUGETLBFS that was stuck to it; the 
> current code is like it is because in 5 and a half years nobody said 
> that it's wrong :)
> 
> If there actually *are* good reasons for stripping __GFP_COMP, then I've 
> certainly no objection to doing so.

The main question is if there's any good reasons for not forbidding
__GFP_COMP to be specified in the callers. The reason given in the
comment isn't convincing.

I don't see how a caller that gets a pointer can care about how the
page structure looks like and in turn why it's asking for __GFP_COMP.

As far as I can tell there are two orthogonal issues in play here:

1) The comment about __GFP_COMP facilitating the sound driver to do
   partial mapping doesn't make much sense. It's probably best to
   WARN_ON immediately in dma_alloc_coherent if __GFP_COMP is
   specified, not only down the call stack in the
   __iommu_dma_alloc_pages() path.

   Note: the CMA paths would already ignore __GFP_COMP if it's
   specified so that __GFP_COMP request can already be ignored. It
   sounds preferable to warn the caller it's asking something it can't
   get, than to silently ignore __GFP_COMP.

   On a side note: hugetlbfs/THP pages can only be allocated with
   __GFP_COMP because for example put_page() must work on all tail
   pages (you can't call compound_head() unless the tail page is part
   of a compound page). But for private driver pages mapped by
   remap_pfn_range, any full or partial mapping is done manually and
   nobody can call GUP on VM_PFNMAP|VM_IO anyway (there's not even the
   requirement of a page struct backing those mappings in fact).

2) __iommu_dma_alloc_pages cannot use __GFP_COMP if it intends to
   return an array of small pages, which is the only thing that the
   current sg_alloc_table_from_pages() supports in input. split_page
   will work as expected to generate small pages from non-compound
   order>0 pages, incidentally it's implement on mm/page_alloc.c, not
   in huge_memory.c.

   split_huge_page as opposed is not intended to be used on newly
   allocated compound page. Maybe we should renamed it to
   split_trans_huge_page to make it more explicit, since it won't even
   work on hugetlbfs (compound) pages.

Thanks,
Andrea