Naoya Horiguchi wrote: > On Mon, Sep 16, 2013 at 01:42:05PM +0300, Kirill A. Shutemov wrote: > > Naoya Horiguchi wrote: > > > Hi, > > > > > > Kirill posted split_ptl patchset for thp today, so in this version > > > I post only hugetlbfs part. I added Kconfig variables in following > > > Kirill's patches (although without CONFIG_SPLIT_*_PTLOCK_CPUS.) > > > > > > This patch changes many lines, but all are in hugetlbfs specific code, > > > so I think we can apply this independent of thp patches. > > > ----- > > > From: Naoya Horiguchi <n-horiguchi@xxxxxxxxxxxxx> > > > Date: Fri, 13 Sep 2013 18:12:30 -0400 > > > Subject: [PATCH v4] hugetlbfs: support split page table lock > > > > > > Currently all of page table handling by hugetlbfs code are done under > > > mm->page_table_lock. So when a process have many threads and they heavily > > > access to the memory, lock contention happens and impacts the performance. > > > > > > This patch makes hugepage support split page table lock so that we use > > > page->ptl of the leaf node of page table tree which is pte for normal pages > > > but can be pmd and/or pud for hugepages of some architectures. > > > > > > ChangeLog v4: > > > - introduce arch dependent macro ARCH_ENABLE_SPLIT_HUGETLB_PTLOCK > > > (only defined for x86 for now) > > > - rename USE_SPLIT_PTLOCKS_HUGETLB to USE_SPLIT_HUGETLB_PTLOCKS > > > > > > ChangeLog v3: > > > - disable split ptl for ppc with USE_SPLIT_PTLOCKS_HUGETLB. > > > - remove replacement in some architecture dependent code. This is justified > > > because an allocation of pgd/pud/pmd/pte entry can race with other > > > allocation, not with read/write access, so we can use different locks. > > > http://thread.gmane.org/gmane.linux.kernel.mm/106292/focus=106458 > > > > > > ChangeLog v2: > > > - add split ptl on other archs missed in v1 > > > - drop changes on arch/{powerpc,tile}/mm/hugetlbpage.c > > > > > > Signed-off-by: Naoya Horiguchi <n-horiguchi@xxxxxxxxxxxxx> > > > --- > > > arch/x86/Kconfig | 4 +++ > > > include/linux/hugetlb.h | 20 +++++++++++ > > > include/linux/mm_types.h | 2 ++ > > > mm/Kconfig | 3 ++ > > > mm/hugetlb.c | 92 +++++++++++++++++++++++++++++------------------- > > > mm/mempolicy.c | 5 +-- > > > mm/migrate.c | 4 +-- > > > mm/rmap.c | 2 +- > > > 8 files changed, 91 insertions(+), 41 deletions(-) > > > > > > diff --git a/arch/x86/Kconfig b/arch/x86/Kconfig > > > index 6a5cf6a..5b83d14 100644 > > > --- a/arch/x86/Kconfig > > > +++ b/arch/x86/Kconfig > > > @@ -1884,6 +1884,10 @@ config ARCH_ENABLE_SPLIT_PMD_PTLOCK > > > def_bool y > > > depends on X86_64 || X86_PAE > > > > > > +config ARCH_ENABLE_SPLIT_HUGETLB_PTLOCK > > > + def_bool y > > > + depends on X86_64 || X86_PAE > > > + > > > menu "Power management and ACPI options" > > > > > > config ARCH_HIBERNATION_HEADER > > > diff --git a/include/linux/hugetlb.h b/include/linux/hugetlb.h > > > index 0393270..2cdac68 100644 > > > --- a/include/linux/hugetlb.h > > > +++ b/include/linux/hugetlb.h > > > @@ -80,6 +80,24 @@ extern const unsigned long hugetlb_zero, hugetlb_infinity; > > > extern int sysctl_hugetlb_shm_group; > > > extern struct list_head huge_boot_pages; > > > > > > +#if USE_SPLIT_HUGETLB_PTLOCKS > > > +#define huge_pte_lockptr(mm, ptep) ({__pte_lockptr(virt_to_page(ptep)); }) > > > +#else /* !USE_SPLIT_HUGETLB_PTLOCKS */ > > > +#define huge_pte_lockptr(mm, ptep) ({&(mm)->page_table_lock; }) > > > +#endif /* USE_SPLIT_HUGETLB_PTLOCKS */ > > > + > > > +#define huge_pte_offset_lock(mm, address, ptlp) \ > > > +({ \ > > > + pte_t *__pte = huge_pte_offset(mm, address); \ > > > + spinlock_t *__ptl = NULL; \ > > > + if (__pte) { \ > > > + __ptl = huge_pte_lockptr(mm, __pte); \ > > > + *(ptlp) = __ptl; \ > > > + spin_lock(__ptl); \ > > > + } \ > > > + __pte; \ > > > +}) > > > + > > > > [ Disclaimer: I don't know much about hugetlb. ] > > > > I don't think it's correct. Few points: > > > > - Hugetlb supports multiple page sizes: on x86_64 2M (PMD) and 1G (PUD). > > My patchset only implements it for PMD. We don't even initialize > > spinlock in struct page for PUD. > > In hugetlbfs code, we use huge_pte_offset() to get leaf level entries > which can be pud or pmd in x86. huge_pte_lockptr() uses this function, > so we can always get the correct ptl regardless of hugepage sizes. > As for spinlock initialization, you're right. I'll add it on huge_pte_alloc(). Please, don't. If USE_SPLIT_PMD_PTLOCKS is true, pmd_alloc_one() will do it for you already for PMD table. For pud it should be done in pud_alloc_one(), not in hugetlb code. We already have too many special cases for hugetlb. Don't contribute to the mess. > > - If we enable split PMD lock we should use it *globally*. With you patch > > we can end up with different locks used by hugetlb and rest of kernel > > to protect the same PMD table if USE_SPLIT_HUGETLB_PTLOCKS != > > USE_SPLIT_PMD_PTLOCKS. It's just broken. > > I don't think so. Thp specific operations (like thp allocation, split, > and collapse) are never called on the virtual address range covered by > vma(VM_HUGETLB) by checking VM_HUGETLB. So no one tries to lock/unlock > a ptl concurrently from thp context and hugetlbfs context. Two vma's can be next to each other and share the same PMD table (not entries) and in this case I don't see what will serialize pmd_alloc() if USE_SPLIT_HUGETLB_PTLOCKS != USE_SPLIT_PMD_PTLOCKS. -- Kirill A. Shutemov -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>