lkml.org 
[lkml]   [2011]   [Nov]   [16]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH] hugetlb: Provide a default HPAGE_SHIFT if !CONFIG_HUGETLB_PAGE
On Wed, 16 Nov 2011, David Daney wrote:

> > > It is a well established kernel idiom to supply dummy values for symbols
> > > that
> > > are required to be defined in order to form a syntactically correct C
> > > program,
> > > but that are known by the programmer to be used only on dead code paths.
> > >
> > > This is exactly what we are doing here.
> > >
> > > To do otherwise requires that code be cluttered with #ifdefery.
> > >
> >
> > You're wrong,
>
> Ok, then please tell me why:
>
> 1) the dummy version of is_vm_hugetlb_page() exists in hugetlb_inline.h?
> 2) the dummy version of PageHuge() exists in hugetlb.h
> 3) the dummy version of reset_vma_resv_huge_pages()
> 4) the dummy version of hugetlb_total_pages() exists in hugetlb.h
> 5) the dummy version of follow_hugetlb_page() exists in hugetlb.h
> 6) the dummy version of follow_huge_addr() exists in hugetlb.h
> 7) the dummy version of copy_hugetlb_page_range() exists in hugetlb.h
> 8) the dummy version of hugetlb_prefault() exists in hugetlb.h
> 9) the dummy version of unmap_hugepage_range() exists in hugetlb.h
> 10) the dummy version of hugetlb_report_meminfo() exists in hugetlb.h
> 11) the dummy version of hugetlb_report_node_meminfo() exists in hugetlb.h
> 12) the dummy version of follow_huge_pmd() exists in hugetlb.h
> 13) the dummy version of follow_huge_pud() exists in hugetlb.h
> 14) the dummy version of prepare_hugepage_range() exists in hugetlb.h
> 15) the dummy version of pmd_huge() exists in hugetlb.h
> 16) the dummy version of pud_huge() exists in hugetlb.h
> 17) the dummy version of is_hugepage_only_range() exists in hugetlb.h
> 17) the dummy version of hugetlb_free_pgd_range() exists in hugetlb.h
> 18) the dummy version of hugetlb_fault() exists in hugetlb.h
> 19) the dummy version of huge_pte_offset() exists in hugetlb.h
> 20) the dummy version of dequeue_hwpoisoned_huge_page() exists in hugetlb.h
> 21) the dummy version of copy_huge_page() exists in hugetlb.h
> 22) the dummy version of hugetlb_change_protection() exists in hugetlb.h
> .
> .
> .
> [Other non HUGETLB_PAGE examples omitted for conciseness]
>

To avoid the #ifdef's, but you'll notice that they return either NULL, 0,
or are a no-op. We don't substitute real values in dummy functions where
they could be used in generic code as though they were valid. That's the
problem with defining HPAGE_SHIFT to be PAGE_SHIFT and, yes, defining
HPAGE_MASK and HPAGE_SIZE as well shouldn't be done and should be removed.

We expect HPAGE_* to represent hugepages, not the native page size of the
bare metal. This is why we do things like

#define HPAGE_PMD_SHIFT ({ BUG(); 0; })
#define HPAGE_PMD_MASK ({ BUG(); 0; })
#define HPAGE_PMD_SIZE ({ BUG(); 0; })

for CONFIG_TRANSPARENT_HUGEPAGE=n. We don't want to pretend that they're
valid outside the context of hugepages.

We also never use HPAGE_SHIFT, HPAGE_MASK, or HPAGE_SIZE in generic code
that isn't dependent on CONFIG_HUGETLB_PAGE. If you'd like to submit a
patch to fix this in the mips tree, then that would be good, and if you'd
like to submit a patch to remove these dummy definitions all together by
auditing other arch code, then that would be great.


\
 
 \ /
  Last update: 2011-11-17 05:59    [W:0.071 / U:0.200 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site