Re: [PATCH] x86: fix PAE pmd_bad bootup warning

[Nishanth Aravamudan <nacc@us.ibm.com>]

On 08.05.2008 [18:51:11 +0200], Hans Rosenfeld wrote:
> On Thu, May 08, 2008 at 09:33:52AM -0700, Nishanth Aravamudan wrote:
> > So this seems to lend credence to Dave's hypothesis. Without, as you
> > were trying before, teaching pagemap all about hugepages, what are our
> > options?
> > 
> > Can we just skip over the current iteration of the PMD loop (would we
> > need something similar for the PTE loop for power?) if pmd_huge(pmd)?
> 
> Allowing huge pages in the page walker would affect both walk_pmd_range
> and walk_pud_range. Then either the users of the page walker need to
> know how to handle huge pages themselves (in the pmd_entry and pud_entry
> callback functions), or the page walker treats huge pages as any other
> pages (calling the pte_entry callback function).

Right, I agree *if* we allow huge pages in the walker. But AIUI, things
are broken now with hugepages in the process' address space. This is a
bug upstream and leads to hugepages leaking out of the kernel when
/proc/pid/pagemap is read. Why not, instead (as a short-term fix), skip
hugepage mappings altogether in the page-walker code?

Hrm, upon further investigation, this seems to be a pretty clear
limitation of walk_page_range(). One that is avoided in the other two
callers, i.e.

static int show_smap(struct seq_file *m, void *v)
{
...
	if (vma->vm_mm && !is_vm_hugetlb_page(vma))
		walk_page_range(vma->vm_mm, vma->vm_start, vma->vm_end,
				&smaps_walk, &mss);
...
}

static ssize_t clear_refs_write(struct file *file, const char __user *buf,
				size_t count, loff_t *ppos)
{
...
		for (vma = mm->mmap; vma; vma = vma->vm_next)
			if (!is_vm_hugetlb_page(vma))
				walk_page_range(mm, vma->vm_start, vma->vm_end,
						&clear_refs_walk, vma);
...
}

No such protection exists for

static ssize_t pagemap_read(struct file *file, char __user *buf,
			    size_t count, loff_t *ppos);

So, is there any way to either add a is_vm_hugetlb_page(vma) check into
pagemap_read()? Or can we modify walk_page_range to take the a vma and
skip the walking if is_vm_hugetlb_page(vma) is set [to avoid
complications down the road until hugepage walking is fixed]. I guess
the latter isn't possible for pagemap_read(), since we are just looking
at arbitrary addresses in the process space?

Dunno, seems quite clear that the bug is in pagemap_read(), not any
hugepage code, and that the simplest fix is to make pagemap_read() do
what the other walker-callers do, and skip hugepage regions.

Thanks,
Nish

-- 
Nishanth Aravamudan <nacc@us.ibm.com>
IBM Linux Technology Center