On Fri, Sep 11, 2015 at 03:22:14PM +0200, Vlastimil Babka wrote:
> On 09/03/2015 05:16 PM, Kirill A. Shutemov wrote:
> >This patch brings back mlocked THP. Instead of forbidding mlocked pages
> >altogether, we just avoid mlocking PTE-mapped THPs and munlock THPs on
> >split_huge_pmd().
> >
> >This means PTE-mapped THPs will be on normal lru lists and will be
> >split under memory pressure by vmscan. After the split vmscan will
> >detect unevictable small pages and mlock them.
>
> Yeah that sounds like a compromise that should work.
>
> >This way we can void leaking mlocked pages into non-VM_LOCKED VMAs.
>
> avoid
>
> But mlocked page in non-mlocked VMA's is a normal thing for shared pages
> when only one of the sharing mm's did mlock(), right? So this description
> doesn't explain the whole issue. I admit I forgot the exact details already
> :(
Right. I'm as always bad on documentation.
Before THP refcounting rework, THP was not allowed to cross VMA boundary.
So, if we have THP and we split it, PG_mlocked can be safely transfered to
small pages.
With new THP refcounting and naive approach to mlocking we can end up with
this scenario:
1. we have a mlocked THP, which belong to one VM_LOCKED VMA.
2. the process does munlock() on the *part* of the THP:
- the VMA is split into two, one of them VM_LOCKED;
- huge PMD split into PTE table;
- THP is still mlocked;
3. split_huge_page():
- it transfers PG_mlocked to *all* small pages regrardless if it
blong to any VM_LOCKED VMA.
We probably could munlock() all small pages on split_huge_page(), but I
think we have accounting issue already on step two.
> >Not-Yet-Signed-off-by: Kirill A. Shutemov <kirill.shute...@linux.intel.com>
> >---
> >
> >I'm not yet 100% certain that this approch is correct. Review would be
> >appriciated.
> >More testing is required.
> >
> >---
> > mm/gup.c | 6 ++++--
> > mm/huge_memory.c | 33 +++++++++++++++++++++++-------
> > mm/memory.c | 3 +--
> > mm/mlock.c | 61
> > +++++++++++++++++++++++++++++++++++++-------------------
> > 4 files changed, 71 insertions(+), 32 deletions(-)
> >
> >diff --git a/mm/gup.c b/mm/gup.c
> >index 70d65e4015a4..e95b0cb6ed81 100644
> >--- a/mm/gup.c
> >+++ b/mm/gup.c
> >@@ -143,6 +143,10 @@ retry:
> > mark_page_accessed(page);
> > }
> > if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
> >+ /* Do not mlock pte-mapped THP */
> >+ if (PageTransCompound(page))
> >+ goto out;
> >+
> > /*
> > * The preliminary mapping check is mainly to avoid the
> > * pointless overhead of lock_page on the ZERO_PAGE
> >@@ -920,8 +924,6 @@ long populate_vma_page_range(struct vm_area_struct *vma,
> > gup_flags = FOLL_TOUCH | FOLL_POPULATE | FOLL_MLOCK;
> > if (vma->vm_flags & VM_LOCKONFAULT)
> > gup_flags &= ~FOLL_POPULATE;
> >- if (vma->vm_flags & VM_LOCKED)
> >- gup_flags |= FOLL_SPLIT;
> > /*
> > * We want to touch writable mappings with a write fault in order
> > * to break COW, except for shared mappings because these don't COW
> >diff --git a/mm/huge_memory.c b/mm/huge_memory.c
> >index 2cc99f9096a8..d714de02473b 100644
> >--- a/mm/huge_memory.c
> >+++ b/mm/huge_memory.c
> >@@ -846,8 +846,6 @@ int do_huge_pmd_anonymous_page(struct mm_struct *mm,
> >struct vm_area_struct *vma,
> >
> > if (haddr < vma->vm_start || haddr + HPAGE_PMD_SIZE > vma->vm_end)
> > return VM_FAULT_FALLBACK;
> >- if (vma->vm_flags & VM_LOCKED)
> >- return VM_FAULT_FALLBACK;
> > if (unlikely(anon_vma_prepare(vma)))
> > return VM_FAULT_OOM;
> > if (unlikely(khugepaged_enter(vma, vma->vm_flags)))
> >@@ -1316,7 +1314,16 @@ struct page *follow_trans_huge_pmd(struct
> >vm_area_struct *vma,
> > update_mmu_cache_pmd(vma, addr, pmd);
> > }
> > if ((flags & FOLL_MLOCK) && (vma->vm_flags & VM_LOCKED)) {
> >- if (page->mapping && trylock_page(page)) {
> >+ /*
> >+ * We don't mlock() pte-mapped THPs. This way we can avoid
> >+ * leaking mlocked pages into non-VM_LOCKED VMAs.
> >+ * In most cases the pmd is the only mapping of the page: we
> >+ * break COW for the mlock(). The only scenario when we have
>
> I don't understand what's meant by "we break COW for the mlock()"?
mm/gup.c:
880 long populate_vma_page_range(struct vm_area_struct *vma,
881 unsigned long start, unsigned long end, int *nonblocking)
.....
894 /*
895 * We want to touch writable mappings with a write fault in order
896 * to break COW, except for shared mappings because these don't COW
897 * and we would not want to dirty them for nothing.
898 */
899 if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
900 gup_flags |= FOLL_WRITE;
> >+ * the page shared here is if we mlocking read-only mapping
> >+ * shared over fork(). We skip mlocking such pages.
>
> Why do we skip them? There's no PTE mapping involved, just multiple PMD
> mappings? Why are those a problem?
We don't have a way to protect against parallel split_huge_pmd(). :(
--
Kirill A. Shutemov
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
