/proc/<pid>/smaps should not be looking at VM_PFNMAP areas.
Certain tests in show_smap() (especially for huge pages) assume that the
mapped PFN's are backed with page structures. And this is not usually true
for VM_PFNMAP areas. This can result in panics on kernel page faults when
attempting to address those page structures.
VM_PFNMAP areas are used by
- graphics memory manager gpu/drm/drm_gem.c
- global reference unit sgi-gru/grufile.c
- sgi special memory char/mspec.c
- probably several out-of-tree modules
I'm copying everyone who has changed fs/proc/task_mmu.c recently, in case
of some reason to provide /proc/<pid>/smaps for these areas that I am not
aware of.
Signed-off-by: Cliff Wickman <[email protected]>
---
fs/proc/task_mmu.c | 3 +++
1 file changed, 3 insertions(+)
Index: linux/fs/proc/task_mmu.c
===================================================================
--- linux.orig/fs/proc/task_mmu.c
+++ linux/fs/proc/task_mmu.c
@@ -589,6 +589,9 @@ static int show_smap(struct seq_file *m,
.private = &mss,
};
+ if (vma->vm_flags & VM_PFNMAP)
+ return 0;
+
memset(&mss, 0, sizeof mss);
mss.vma = vma;
/* mmap_sem is held in m_start */
Hello Cliff,
On Mon, Apr 29, 2013 at 04:25:53PM -0500, Cliff Wickman wrote:
>
> /proc/<pid>/smaps should not be looking at VM_PFNMAP areas.
>
> Certain tests in show_smap() (especially for huge pages) assume that the
> mapped PFN's are backed with page structures. And this is not usually true
> for VM_PFNMAP areas. This can result in panics on kernel page faults when
> attempting to address those page structures.
I think it's strange that you mention to hugepages, because in my understanding
VM_PFNMAP and hugepage related vma (VM_HUGEPAGE or VM_HUGETLB) should not set
at the same time. In what testcase are these flags both set?
And I guess this race can also happen on reading pagemap or numa_maps because
walk_page_range() is called in those code paths. Are you sure the race doesn't
happen on these paths? If not, please add a few more flag checks for them.
Thanks,
Naoya Horiguchi
> VM_PFNMAP areas are used by
> - graphics memory manager gpu/drm/drm_gem.c
> - global reference unit sgi-gru/grufile.c
> - sgi special memory char/mspec.c
> - probably several out-of-tree modules
>
> I'm copying everyone who has changed fs/proc/task_mmu.c recently, in case
> of some reason to provide /proc/<pid>/smaps for these areas that I am not
> aware of.
>
> Signed-off-by: Cliff Wickman <[email protected]>
> ---
> fs/proc/task_mmu.c | 3 +++
> 1 file changed, 3 insertions(+)
>
> Index: linux/fs/proc/task_mmu.c
> ===================================================================
> --- linux.orig/fs/proc/task_mmu.c
> +++ linux/fs/proc/task_mmu.c
> @@ -589,6 +589,9 @@ static int show_smap(struct seq_file *m,
> .private = &mss,
> };
>
> + if (vma->vm_flags & VM_PFNMAP)
> + return 0;
> +
> memset(&mss, 0, sizeof mss);
> mss.vma = vma;
> /* mmap_sem is held in m_start */
> --
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to [email protected]
> More majordomo info at http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at http://www.tux.org/lkml/
> On Tue, Apr 30, 2013 at 01:11:45PM -0500, Naoya Horiguchi wrote:
> >
> > /proc/<pid>/smaps should not be looking at VM_PFNMAP areas.
> >
> > Certain tests in show_smap() (especially for huge pages) assume that the
> > mapped PFN's are backed with page structures. And this is not usually true
> > for VM_PFNMAP areas. This can result in panics on kernel page faults when
> > attempting to address those page structures.
>
> I think it's strange that you mention to hugepages, because in my understanding
> VM_PFNMAP and hugepage related vma (VM_HUGEPAGE or VM_HUGETLB) should not set
> at the same time. In what testcase are these flags both set?
I don't think VM_PFNMAP and VM_HUGE... set at the same time.
The problem is that a VM_PFNMAP'd area might have 2MB mappings in its
page table, but they may point to pfn's that are not backed by page
structures.
Then a sequence like:
show_smap
show_map_vma
walk_page_range
walk_pud_range
walk_pmd_range
split_huge_page_pmd(walk->mm, pmd)
__split_huge_page_pmd
page = pmd_page(*pmd)
can address (vmemmap + (pfn)) and panic
Or a sequence like this:
walk_pmd_range
walk->pmd_entry(pmd, addr, next, walk)
smaps_pte_range
smaps_pte_entry(*pte, addr, PAGE_SIZE, walk)
page = vm_normal_page(vma, addr, ptent)
return pfn_to_page(pfn)
>
> And I guess this race can also happen on reading pagemap or numa_maps because
> walk_page_range() is called in those code paths. Are you sure the race doesn't
> happen on these paths? If not, please add a few more flag checks for them.
Okay. I'll check and submit a version 2 of this patch.
-Cliff Wickman
> Thanks,
> Naoya Horiguchi
>
>
> > VM_PFNMAP areas are used by
> > - graphics memory manager gpu/drm/drm_gem.c
> > - global reference unit sgi-gru/grufile.c
> > - sgi special memory char/mspec.c
> > - probably several out-of-tree modules
> >
> > I'm copying everyone who has changed fs/proc/task_mmu.c recently, in case
> > of some reason to provide /proc/<pid>/smaps for these areas that I am not
> > aware of.
> >
> > Signed-off-by: Cliff Wickman <[email protected]>
> > ---
> > fs/proc/task_mmu.c | 3 +++
> > 1 file changed, 3 insertions(+)
> >
> > Index: linux/fs/proc/task_mmu.c
> > ===================================================================
> > --- linux.orig/fs/proc/task_mmu.c
> > +++ linux/fs/proc/task_mmu.c
> > @@ -589,6 +589,9 @@ static int show_smap(struct seq_file *m,
> > .private = &mss,
> > };
> >
> > + if (vma->vm_flags & VM_PFNMAP)
> > + return 0;
> > +
> > memset(&mss, 0, sizeof mss);
> > mss.vma = vma;
> > /* mmap_sem is held in m_start */
--
Cliff Wickman
SGI
[email protected]
(651) 683-3824