I forgot to add Reviewed/Acked. Please ignore my previous post.
Sorry for the noise.
-----
Hi,
Here is 5th version of hugepage coredump fix.
I changed the place to put swap entry check in 3/3,
and explained more in comment.
Thanks,
Naoya Horiguchi
Currently we fail to include any data on hugepages into coredump,
because VM_DONTDUMP is set on hugetlbfs's vma. This behavior was recently
introduced by commit 314e51b98 "mm: kill vma flag VM_RESERVED and
mm->reserved_vm counter". This looks to me a serious regression,
so let's fix it.
ChangeLog v3:
- move 'return 0' into a separate patch
ChangeLog v2:
- add 'return 0' in hugepage memory check
Signed-off-by: Naoya Horiguchi <[email protected]>
Acked-by: Konstantin Khlebnikov <[email protected]>
Acked-by: Michal Hocko <[email protected]>
Reviewed-by: Rik van Riel <[email protected]>
Acked-by: KOSAKI Motohiro <[email protected]>
Cc: [email protected]
---
fs/hugetlbfs/inode.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git v3.9-rc3.orig/fs/hugetlbfs/inode.c v3.9-rc3/fs/hugetlbfs/inode.c
index 84e3d85..523464e 100644
--- v3.9-rc3.orig/fs/hugetlbfs/inode.c
+++ v3.9-rc3/fs/hugetlbfs/inode.c
@@ -110,7 +110,7 @@ static int hugetlbfs_file_mmap(struct file *file, struct vm_area_struct *vma)
* way when do_mmap_pgoff unwinds (may be important on powerpc
* and ia64).
*/
- vma->vm_flags |= VM_HUGETLB | VM_DONTEXPAND | VM_DONTDUMP;
+ vma->vm_flags |= VM_HUGETLB | VM_DONTEXPAND;
vma->vm_ops = &hugetlb_vm_ops;
if (vma->vm_pgoff & (~huge_page_mask(h) >> PAGE_SHIFT))
--
1.7.11.7
Documentation/filesystems/proc.txt says about coredump_filter bitmask,
Note bit 0-4 doesn't effect any hugetlb memory. hugetlb memory are only
effected by bit 5-6.
However current code can go into the subsequent flag checks of bit 0-4
for vma(VM_HUGETLB). So this patch inserts 'return' and makes it work
as written in the document.
Signed-off-by: Naoya Horiguchi <[email protected]>
Reviewed-by: Rik van Riel <[email protected]>
Acked-by: Michal Hocko <[email protected]>
Reviewed-by: HATAYAMA Daisuke <[email protected]>
Acked-by: KOSAKI Motohiro <[email protected]>
Cc: [email protected]
---
fs/binfmt_elf.c | 1 +
1 file changed, 1 insertion(+)
diff --git v3.9-rc3.orig/fs/binfmt_elf.c v3.9-rc3/fs/binfmt_elf.c
index 3939829..86af964 100644
--- v3.9-rc3.orig/fs/binfmt_elf.c
+++ v3.9-rc3/fs/binfmt_elf.c
@@ -1137,6 +1137,7 @@ static unsigned long vma_dump_size(struct vm_area_struct *vma,
goto whole;
if (!(vma->vm_flags & VM_SHARED) && FILTER(HUGETLB_PRIVATE))
goto whole;
+ return 0;
}
/* Do not dump I/O mapped devices or special mappings */
--
1.7.11.7
# I suspended Reviewed and Acked given for the previous version, because
# it has a non-minor change. If you want to restore it, please let me know.
-----
With applying the previous patch "hugetlbfs: stop setting VM_DONTDUMP in
initializing vma(VM_HUGETLB)" to reenable hugepage coredump, if a memory
error happens on a hugepage and the affected processes try to access
the error hugepage, we hit VM_BUG_ON(atomic_read(&page->_count) <= 0)
in get_page().
The reason for this bug is that coredump-related code doesn't recognise
"hugepage hwpoison entry" with which a pmd entry is replaced when a memory
error occurs on a hugepage.
In other words, physical address information is stored in different bit layout
between hugepage hwpoison entry and pmd entry, so follow_hugetlb_page()
which is called in get_dump_page() returns a wrong page from a given address.
The expected behavior is like this:
absent is_swap_pte FOLL_DUMP Expected behavior
-------------------------------------------------------------------
true false false hugetlb_fault
false true false hugetlb_fault
false false false return page
true false true skip page (to avoid allocation)
false true true hugetlb_fault
false false true return page
With this patch, we can call hugetlb_fault() and take proper actions
(we wait for migration entries, fail with VM_FAULT_HWPOISON_LARGE for
hwpoisoned entries,) and as the result we can dump all hugepages except
for hwpoisoned ones.
ChangeLog v5:
- improve comment and description.
ChangeLog v4:
- move is_swap_page() to right place.
ChangeLog v3:
- add comment about using is_swap_pte()
Signed-off-by: Naoya Horiguchi <[email protected]>
Cc: [email protected]
---
mm/hugetlb.c | 12 +++++++++++-
1 file changed, 11 insertions(+), 1 deletion(-)
diff --git v3.9-rc3.orig/mm/hugetlb.c v3.9-rc3/mm/hugetlb.c
index 0d1705b..bf26ee8 100644
--- v3.9-rc3.orig/mm/hugetlb.c
+++ v3.9-rc3/mm/hugetlb.c
@@ -2983,7 +2983,17 @@ long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
break;
}
- if (absent ||
+ /*
+ * We need call hugetlb_fault for both hugepages under migration
+ * (in which case hugetlb_fault waits for the migration,) and
+ * hwpoisoned hugepages (in which case we need to prevent the
+ * caller from accessing to them.) In order to do this, we use
+ * here is_swap_pte instead of is_hugetlb_entry_migration and
+ * is_hugetlb_entry_hwpoisoned. This is because it simply covers
+ * both cases, and because we can't follow correct pages
+ * directly from any kind of swap entries.
+ */
+ if (absent || is_swap_pte(huge_ptep_get(pte)) ||
((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) {
int ret;
--
1.7.11.7
On Wed, Apr 10, 2013 at 12:17 PM, Naoya Horiguchi
<[email protected]> wrote:
> # I suspended Reviewed and Acked given for the previous version, because
> # it has a non-minor change. If you want to restore it, please let me know.
> -----
> With applying the previous patch "hugetlbfs: stop setting VM_DONTDUMP in
> initializing vma(VM_HUGETLB)" to reenable hugepage coredump, if a memory
> error happens on a hugepage and the affected processes try to access
> the error hugepage, we hit VM_BUG_ON(atomic_read(&page->_count) <= 0)
> in get_page().
>
> The reason for this bug is that coredump-related code doesn't recognise
> "hugepage hwpoison entry" with which a pmd entry is replaced when a memory
> error occurs on a hugepage.
> In other words, physical address information is stored in different bit layout
> between hugepage hwpoison entry and pmd entry, so follow_hugetlb_page()
> which is called in get_dump_page() returns a wrong page from a given address.
>
> The expected behavior is like this:
>
> absent is_swap_pte FOLL_DUMP Expected behavior
> -------------------------------------------------------------------
> true false false hugetlb_fault
> false true false hugetlb_fault
> false false false return page
> true false true skip page (to avoid allocation)
> false true true hugetlb_fault
> false false true return page
>
> With this patch, we can call hugetlb_fault() and take proper actions
> (we wait for migration entries, fail with VM_FAULT_HWPOISON_LARGE for
> hwpoisoned entries,) and as the result we can dump all hugepages except
> for hwpoisoned ones.
>
> ChangeLog v5:
> - improve comment and description.
>
> ChangeLog v4:
> - move is_swap_page() to right place.
>
> ChangeLog v3:
> - add comment about using is_swap_pte()
>
> Signed-off-by: Naoya Horiguchi <[email protected]>
> Cc: [email protected]
Acked-by: KOSAKI Motohiro <[email protected]>
On Wed 10-04-13 12:17:49, Naoya Horiguchi wrote:
> # I suspended Reviewed and Acked given for the previous version, because
> # it has a non-minor change. If you want to restore it, please let me know.
> -----
> With applying the previous patch "hugetlbfs: stop setting VM_DONTDUMP in
> initializing vma(VM_HUGETLB)" to reenable hugepage coredump, if a memory
> error happens on a hugepage and the affected processes try to access
> the error hugepage, we hit VM_BUG_ON(atomic_read(&page->_count) <= 0)
> in get_page().
>
> The reason for this bug is that coredump-related code doesn't recognise
> "hugepage hwpoison entry" with which a pmd entry is replaced when a memory
> error occurs on a hugepage.
> In other words, physical address information is stored in different bit layout
> between hugepage hwpoison entry and pmd entry, so follow_hugetlb_page()
> which is called in get_dump_page() returns a wrong page from a given address.
>
> The expected behavior is like this:
>
> absent is_swap_pte FOLL_DUMP Expected behavior
> -------------------------------------------------------------------
> true false false hugetlb_fault
> false true false hugetlb_fault
> false false false return page
> true false true skip page (to avoid allocation)
> false true true hugetlb_fault
> false false true return page
>
> With this patch, we can call hugetlb_fault() and take proper actions
> (we wait for migration entries, fail with VM_FAULT_HWPOISON_LARGE for
> hwpoisoned entries,) and as the result we can dump all hugepages except
> for hwpoisoned ones.
>
> ChangeLog v5:
> - improve comment and description.
>
> ChangeLog v4:
> - move is_swap_page() to right place.
>
> ChangeLog v3:
> - add comment about using is_swap_pte()
>
> Signed-off-by: Naoya Horiguchi <[email protected]>
> Cc: [email protected]
Acked-by: Michal Hocko <[email protected]>
Thanks!
> ---
> mm/hugetlb.c | 12 +++++++++++-
> 1 file changed, 11 insertions(+), 1 deletion(-)
>
> diff --git v3.9-rc3.orig/mm/hugetlb.c v3.9-rc3/mm/hugetlb.c
> index 0d1705b..bf26ee8 100644
> --- v3.9-rc3.orig/mm/hugetlb.c
> +++ v3.9-rc3/mm/hugetlb.c
> @@ -2983,7 +2983,17 @@ long follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
> break;
> }
>
> - if (absent ||
> + /*
> + * We need call hugetlb_fault for both hugepages under migration
> + * (in which case hugetlb_fault waits for the migration,) and
> + * hwpoisoned hugepages (in which case we need to prevent the
> + * caller from accessing to them.) In order to do this, we use
> + * here is_swap_pte instead of is_hugetlb_entry_migration and
> + * is_hugetlb_entry_hwpoisoned. This is because it simply covers
> + * both cases, and because we can't follow correct pages
> + * directly from any kind of swap entries.
> + */
> + if (absent || is_swap_pte(huge_ptep_get(pte)) ||
> ((flags & FOLL_WRITE) && !pte_write(huge_ptep_get(pte)))) {
> int ret;
>
> --
> 1.7.11.7
>
> --
> To unsubscribe, send a message with 'unsubscribe linux-mm' in
> the body to [email protected]. For more info on Linux MM,
> see: http://www.linux-mm.org/ .
> Don't email: <a href=mailto:"[email protected]"> [email protected] </a>
--
Michal Hocko
SUSE Labs
On Wed, 10 Apr 2013, Naoya Horiguchi wrote:
> Currently we fail to include any data on hugepages into coredump,
> because VM_DONTDUMP is set on hugetlbfs's vma. This behavior was recently
> introduced by commit 314e51b98 "mm: kill vma flag VM_RESERVED and
> mm->reserved_vm counter". This looks to me a serious regression,
> so let's fix it.
>
> ChangeLog v3:
> - move 'return 0' into a separate patch
>
> ChangeLog v2:
> - add 'return 0' in hugepage memory check
>
> Signed-off-by: Naoya Horiguchi <[email protected]>
> Acked-by: Konstantin Khlebnikov <[email protected]>
> Acked-by: Michal Hocko <[email protected]>
> Reviewed-by: Rik van Riel <[email protected]>
> Acked-by: KOSAKI Motohiro <[email protected]>
> Cc: [email protected]
Acked-by: David Rientjes <[email protected]>
Stable for 3.7+.
On Wed, 10 Apr 2013, Naoya Horiguchi wrote:
> Documentation/filesystems/proc.txt says about coredump_filter bitmask,
>
> Note bit 0-4 doesn't effect any hugetlb memory. hugetlb memory are only
> effected by bit 5-6.
>
> However current code can go into the subsequent flag checks of bit 0-4
> for vma(VM_HUGETLB). So this patch inserts 'return' and makes it work
> as written in the document.
>
> Signed-off-by: Naoya Horiguchi <[email protected]>
> Reviewed-by: Rik van Riel <[email protected]>
> Acked-by: Michal Hocko <[email protected]>
> Reviewed-by: HATAYAMA Daisuke <[email protected]>
> Acked-by: KOSAKI Motohiro <[email protected]>
> Cc: [email protected]
Acked-by: David Rientjes <[email protected]>
Stable for 2.6.34+.
On Wed, 10 Apr 2013, Naoya Horiguchi wrote:
> # I suspended Reviewed and Acked given for the previous version, because
> # it has a non-minor change. If you want to restore it, please let me know.
> -----
> With applying the previous patch "hugetlbfs: stop setting VM_DONTDUMP in
> initializing vma(VM_HUGETLB)" to reenable hugepage coredump, if a memory
> error happens on a hugepage and the affected processes try to access
> the error hugepage, we hit VM_BUG_ON(atomic_read(&page->_count) <= 0)
> in get_page().
>
> The reason for this bug is that coredump-related code doesn't recognise
> "hugepage hwpoison entry" with which a pmd entry is replaced when a memory
> error occurs on a hugepage.
> In other words, physical address information is stored in different bit layout
> between hugepage hwpoison entry and pmd entry, so follow_hugetlb_page()
> which is called in get_dump_page() returns a wrong page from a given address.
>
> The expected behavior is like this:
>
> absent is_swap_pte FOLL_DUMP Expected behavior
> -------------------------------------------------------------------
> true false false hugetlb_fault
> false true false hugetlb_fault
> false false false return page
> true false true skip page (to avoid allocation)
> false true true hugetlb_fault
> false false true return page
>
> With this patch, we can call hugetlb_fault() and take proper actions
> (we wait for migration entries, fail with VM_FAULT_HWPOISON_LARGE for
> hwpoisoned entries,) and as the result we can dump all hugepages except
> for hwpoisoned ones.
>
> ChangeLog v5:
> - improve comment and description.
>
> ChangeLog v4:
> - move is_swap_page() to right place.
>
> ChangeLog v3:
> - add comment about using is_swap_pte()
>
> Signed-off-by: Naoya Horiguchi <[email protected]>
> Cc: [email protected]
Acked-by: David Rientjes <[email protected]>
Stable for 2.6.34+?
On Wed 10-04-13 14:49:07, David Rientjes wrote:
> On Wed, 10 Apr 2013, Naoya Horiguchi wrote:
>
> > Documentation/filesystems/proc.txt says about coredump_filter bitmask,
> >
> > Note bit 0-4 doesn't effect any hugetlb memory. hugetlb memory are only
> > effected by bit 5-6.
> >
> > However current code can go into the subsequent flag checks of bit 0-4
> > for vma(VM_HUGETLB). So this patch inserts 'return' and makes it work
> > as written in the document.
> >
> > Signed-off-by: Naoya Horiguchi <[email protected]>
> > Reviewed-by: Rik van Riel <[email protected]>
> > Acked-by: Michal Hocko <[email protected]>
> > Reviewed-by: HATAYAMA Daisuke <[email protected]>
> > Acked-by: KOSAKI Motohiro <[email protected]>
> > Cc: [email protected]
>
> Acked-by: David Rientjes <[email protected]>
>
> Stable for 2.6.34+.
I think it is only 3.7+ as well because VM_RESERVED stopped use before
(314e51b9).
--
Michal Hocko
SUSE Labs