Re: [PATCH] Fix OOPS in mmap_region() when merging adjacentVM_LOCKED file segments

From: Mel Gorman
Date: Mon Feb 02 2009 - 13:58:24 EST


On Mon, Feb 02, 2009 at 11:10:42PM +0900, KOSAKI Motohiro wrote:
> (cc to mel)
>
> > On Mon, 2 Feb 2009, KOSAKI Motohiro wrote:
> > > >
> > > > - if (flags & MAP_NORESERVE)
> > > > + /*
> > > > + * Set 'VM_NORESERVE' if we should not account for the
> > > > + * memory use of this mapping. We only honor MAP_NORESERVE
> > > > + * if we're allowed to overcommit memory.
> > > > + */
> > > > + if ((flags & MAP_NORESERVE) && sysctl_overcommit_memory != OVERCOMMIT_NEVER)
> > >
> > > I afraid this line a bit.
> > > if following scenario happend, we can lost VM_NORESERVE?
> > >
> > > 1. admin set overcommit_memory to "never"
> > > 2. mmap
> > > 3. admin set overcommit_memory to "guess"
> >
> > I still haven't reviewed it fully myself (and note that what
> > Linus put in his tree is not identical to this posted patch),
> > but I do believe this is okay.
> >
> > When admin changes overcommit_memory, we don't make a pass across
> > every vma of every mm in the system, to adjust all the accounting
> > of VM_NORESERVE areas; so I think it's quite reasonable to take
> > VM_NORESERVE as reflecting the policy in force when that vma was
> > created. And nothing is displaying the VM_NORESERVE flag.
>
> hmhm, I see.
>
>
> > Ah, you're actually thinking of
> > 4. mprotect
> > with the original flags (!VM_WRITE) such that no VM_ACCOUNT was done,
> > and now VM_WRITE is added and the accounting is done despite it having
> > been mapped MAP_NORESERVE originally. Whereas before Linus's change,
> > VM_NORESERVE would have still exempted it.
> >
> > Well... I don't think I care!
>
> Yeah.
>
> FWIW, we don't need VM_NORESERVE checking now because VM_NORESERVE and VM_ACCOUNT
> are exclusive condition now :)
>
>
>
> > But I wonder what the hugetlb situation is: that
> > if (!accountable)
> > vm_flags |= VM_NORESERVE;
> > looks suspicious to me, they look as if they're exempting all
> > the hugetlb pages from its accounting, whereas !accountable was
> > only supposed to exempt them from mmap_region()'s own accounting.
>
> HAHAHA, Indeed.
>

Candidate patch for clearing that up as been posted. Thanks for cc'ing me
on this as I would have missed it.

> when hugepage shared read-only mapping -> hugepage shared writable maping,
> following code seems to cause calling vm_enough_memory() although hugepage.
>
> ========================================================
> mprotect_fixup(struct vm_area_struct *vma, struct vm_area_struct **pprev,
> unsigned long start, unsigned long end, unsigned long newflags)
> {
> if (newflags & VM_WRITE) {
> if (!(oldflags & (VM_ACCOUNT|VM_WRITE|
> VM_SHARED|VM_NORESERVE))) {
> charged = nrpages;
> if (security_vm_enough_memory(charged))
> return -ENOMEM;
> newflags |= VM_ACCOUNT;
> }
> }
> ==========================================================
>
> mel, what do you think this?
>

I think there is a problem there all right. VM_ACCOUNT will not be set
with the other patch applied but VM_NORESERVE might be. If so, we
potentially set VM_ACCOUNT on a hugetlbfs mapping and probably make a
mess out of Committed_AS later. Maybe something like the following?

================
Do not account for address space usage when making hugetlbfs mappings RW

hugetlbfs accounts for its address space usage separate from the VM
core. VM_ACCOUNT should not be set for its mappings but it is possible it gets
set if a user creates a RO hugetlbfs mapping MAP_NORESERVE and then calls
mprotect(). This patch stops VM_ACCOUNT being set for hugetlbfs mappings
during mprotect().

Credit goes to Kosaki Motohiro for spotting this.

Signed-off-by: Mel Gorman <mel@xxxxxxxxx>

diff --git a/mm/mprotect.c b/mm/mprotect.c
index abe2694..31ddc6a 100644
--- a/mm/mprotect.c
+++ b/mm/mprotect.c
@@ -153,7 +153,7 @@ mprotect_fixup(struct vm_area_struct *vma, struct vm_area_struct **pprev,
* but (without finer accounting) cannot reduce our commit if we
* make it unwritable again.
*/
- if (newflags & VM_WRITE) {
+ if (newflags & VM_WRITE && !(vma->vm_flags & VM_HUGETLB)) {
if (!(oldflags & (VM_ACCOUNT|VM_WRITE|
VM_SHARED|VM_NORESERVE))) {
charged = nrpages;

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/