Re: [RFC][PATCH] vmscan: report vm_flags in page_referenced()

[Wu Fengguang]

On Fri, May 08, 2009 at 10:01:19PM +0800, Minchan Kim wrote:
> On Fri, May 8, 2009 at 9:15 PM, Wu Fengguang <fengguang.wu@xxxxxxxxx> wrote:
> > On Fri, May 08, 2009 at 08:09:24PM +0800, Minchan Kim wrote:
> >> On Fri, May 8, 2009 at 1:17 PM, Wu Fengguang <fengguang.wu@xxxxxxxxx> wrote:
> >> > On Thu, May 07, 2009 at 11:17:46PM +0800, Peter Zijlstra wrote:
> >> >> On Thu, 2009-05-07 at 17:10 +0200, Johannes Weiner wrote:
> >> >>
> >> >> > > @@ -1269,8 +1270,15 @@ static void shrink_active_list(unsigned
> >> >> > >
> >> >> > > Â Â Â Â Â /* page_referenced clears PageReferenced */
> >> >> > > Â Â Â Â Â if (page_mapping_inuse(page) &&
> >> >> > > - Â Â Â Â Â Â page_referenced(page, 0, sc->mem_cgroup))
> >> >> > > + Â Â Â Â Â Â page_referenced(page, 0, sc->mem_cgroup)) {
> >> >> > > + Â Â Â Â Â Â Â Â struct address_space *mapping = page_mapping(page);
> >> >> > > +
> >> >> > > Â Â Â Â Â Â Â Â Â pgmoved++;
> >> >> > > + Â Â Â Â Â Â Â Â if (mapping && test_bit(AS_EXEC, &mapping->flags)) {
> >> >> > > + Â Â Â Â Â Â Â Â Â Â Â Â list_add(&page->lru, &l_active);
> >> >> > > + Â Â Â Â Â Â Â Â Â Â Â Â continue;
> >> >> > > + Â Â Â Â Â Â Â Â }
> >> >> > > + Â Â Â Â }
> >> >> >
> >> >> > Since we walk the VMAs in page_referenced anyway, wouldn't it be
> >> >> > better to check if one of them is executable? ÂThis would even work
> >> >> > for executable anon pages. ÂAfter all, there are applications that cow
> >> >> > executable mappings (sbcl and other language environments that use an
> >> >> > executable, run-time modified core image come to mind).
> >> >>
> >> >> Hmm, like provide a vm_flags mask along to page_referenced() to only
> >> >> account matching vmas... seems like a sensible idea.
> >> >
> >> > Here is a quick patch for your opinions. Compile tested.
> >> >
> >> > With the added vm_flags reporting, the mlock=>unevictable logic can
> >> > possibly be made more straightforward.
> >> >
> >> > Thanks,
> >> > Fengguang
> >> > ---
> >> > vmscan: report vm_flags in page_referenced()
> >> >
> >> > This enables more informed reclaim heuristics, eg. to protect executable
> >> > file pages more aggressively.
> >> >
> >> > Signed-off-by: Wu Fengguang <fengguang.wu@xxxxxxxxx>
> >> > ---
> >> > Âinclude/linux/rmap.h | Â Â5 +++--
> >> > Âmm/rmap.c      Â|  30 +++++++++++++++++++++---------
> >> > Âmm/vmscan.c     Â|  Â7 +++++--
> >> > Â3 files changed, 29 insertions(+), 13 deletions(-)
> >> >
> >> > --- linux.orig/include/linux/rmap.h
> >> > +++ linux/include/linux/rmap.h
> >> > @@ -83,7 +83,8 @@ static inline void page_dup_rmap(struct
> >> > Â/*
> >> > Â* Called from mm/vmscan.c to handle paging out
> >> > Â*/
> >> > -int page_referenced(struct page *, int is_locked, struct mem_cgroup *cnt);
> >> > +int page_referenced(struct page *, int is_locked,
> >> > + Â Â Â Â Â Â Â Â Â Â Â struct mem_cgroup *cnt, unsigned long *vm_flags);
> >> > Âint try_to_unmap(struct page *, int ignore_refs);
> >> >
> >> > Â/*
> >> > @@ -128,7 +129,7 @@ int page_wrprotect(struct page *page, in
> >> > Â#define anon_vma_prepare(vma) Â(0)
> >> > Â#define anon_vma_link(vma) Â Â do {} while (0)
> >> >
> >> > -#define page_referenced(page,l,cnt) TestClearPageReferenced(page)
> >> > +#define page_referenced(page, locked, cnt, flags) TestClearPageReferenced(page)
> >> > Â#define try_to_unmap(page, refs) SWAP_FAIL
> >> >
> >> > Âstatic inline int page_mkclean(struct page *page)
> >> > --- linux.orig/mm/rmap.c
> >> > +++ linux/mm/rmap.c
> >> > @@ -333,7 +333,8 @@ static int page_mapped_in_vma(struct pag
> >> > Â* repeatedly from either page_referenced_anon or page_referenced_file.
> >> > Â*/
> >> > Âstatic int page_referenced_one(struct page *page,
> >> > - Â Â Â struct vm_area_struct *vma, unsigned int *mapcount)
> >> > + Â Â Â Â Â Â Â Â Â Â Â Â Â Â Âstruct vm_area_struct *vma,
> >> > + Â Â Â Â Â Â Â Â Â Â Â Â Â Â Âunsigned int *mapcount)
> >> > Â{
> >> > Â Â Â Âstruct mm_struct *mm = vma->vm_mm;
> >> > Â Â Â Âunsigned long address;
> >> > @@ -385,7 +386,8 @@ out:
> >> > Â}
> >> >
> >> > Âstatic int page_referenced_anon(struct page *page,
> >> > - Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â struct mem_cgroup *mem_cont)
> >> > + Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â struct mem_cgroup *mem_cont,
> >> > + Â Â Â Â Â Â Â Â Â Â Â Â Â Â Â unsigned long *vm_flags)
> >> > Â{
> >> > Â Â Â Âunsigned int mapcount;
> >> > Â Â Â Âstruct anon_vma *anon_vma;
> >> > @@ -406,6 +408,7 @@ static int page_referenced_anon(struct p
> >> > Â Â Â Â Â Â Â Âif (mem_cont && !mm_match_cgroup(vma->vm_mm, mem_cont))
> >> > Â Â Â Â Â Â Â Â Â Â Â Âcontinue;
> >> > Â Â Â Â Â Â Â Âreferenced += page_referenced_one(page, vma, &mapcount);
> >> > + Â Â Â Â Â Â Â *vm_flags |= vma->vm_flags;
> >>
> >> Sometime this vma don't contain the anon page.
> >> That's why we need page_check_address.
> >> For such a case, wrong *vm_flag cause be harmful to reclaim.
> >> It can be happen in your first class citizen patch, I think.
> >
> > Yes I'm aware of that - the VMA area covers that page, but have no pte
> > actually installed for that page. That should be OK - the presentation
> > of such VMA is a good indication of it being some executable text.
> >
> 
> Sorry but I can't understand your point.
> 
> This is general interface but not only executable text.
> Sometime, The information of vma which don't really have the page can
> be passed to caller.

Right. But if the caller don't care, why bother passing the vm_flags
parameter down to page_referenced_one()? We can do that when there
comes a need, otherwise it sounds more like unnecessary overheads.

> ex) It can be happen by COW, mremap, non-linear mapping and so on.
> but I am not sure.

Hmm, this reminded me of the mlocked page protection logic in
page_referenced_one(). Why shall the "if (vma->vm_flags & VM_LOCKED)"
check be placed *after* the page_check_address() check? Is there a
case that an *existing* page frame is not mapped to the VM_LOCKED vma?
And why not to protect the page in such a case?

> I doubt  vm_flag information is useful.

Me too :) I don't expect many of the other flags to be useful.
Just that passing them out blindly could be more convenient than doing

        if (vma->vm_flags & PROT_EXEC)
                *vm_flags = PROT_EXEC;

But I do suspect passing out VM_LOCKED could help somehow.

Thanks,
Fengguang
--
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/