104 lines · plain
1=====================2Split page table lock3=====================4 5Originally, mm->page_table_lock spinlock protected all page tables of the6mm_struct. But this approach leads to poor page fault scalability of7multi-threaded applications due high contention on the lock. To improve8scalability, split page table lock was introduced.9 10With split page table lock we have separate per-table lock to serialize11access to the table. At the moment we use split lock for PTE and PMD12tables. Access to higher level tables protected by mm->page_table_lock.13 14There are helpers to lock/unlock a table and other accessor functions:15 16 - pte_offset_map_lock()17 maps PTE and takes PTE table lock, returns pointer to PTE with18 pointer to its PTE table lock, or returns NULL if no PTE table;19 - pte_offset_map_nolock()20 maps PTE, returns pointer to PTE with pointer to its PTE table21 lock (not taken), or returns NULL if no PTE table;22 - pte_offset_map()23 maps PTE, returns pointer to PTE, or returns NULL if no PTE table;24 - pte_unmap()25 unmaps PTE table;26 - pte_unmap_unlock()27 unlocks and unmaps PTE table;28 - pte_alloc_map_lock()29 allocates PTE table if needed and takes its lock, returns pointer to30 PTE with pointer to its lock, or returns NULL if allocation failed;31 - pmd_lock()32 takes PMD table lock, returns pointer to taken lock;33 - pmd_lockptr()34 returns pointer to PMD table lock;35 36Split page table lock for PTE tables is enabled compile-time if37CONFIG_SPLIT_PTLOCK_CPUS (usually 4) is less or equal to NR_CPUS.38If split lock is disabled, all tables are guarded by mm->page_table_lock.39 40Split page table lock for PMD tables is enabled, if it's enabled for PTE41tables and the architecture supports it (see below).42 43Hugetlb and split page table lock44=================================45 46Hugetlb can support several page sizes. We use split lock only for PMD47level, but not for PUD.48 49Hugetlb-specific helpers:50 51 - huge_pte_lock()52 takes pmd split lock for PMD_SIZE page, mm->page_table_lock53 otherwise;54 - huge_pte_lockptr()55 returns pointer to table lock;56 57Support of split page table lock by an architecture58===================================================59 60There's no need in special enabling of PTE split page table lock: everything61required is done by pagetable_pte_ctor() and pagetable_pte_dtor(), which62must be called on PTE table allocation / freeing.63 64Make sure the architecture doesn't use slab allocator for page table65allocation: slab uses page->slab_cache for its pages.66This field shares storage with page->ptl.67 68PMD split lock only makes sense if you have more than two page table69levels.70 71PMD split lock enabling requires pagetable_pmd_ctor() call on PMD table72allocation and pagetable_pmd_dtor() on freeing.73 74Allocation usually happens in pmd_alloc_one(), freeing in pmd_free() and75pmd_free_tlb(), but make sure you cover all PMD table allocation / freeing76paths: i.e X86_PAE preallocate few PMDs on pgd_alloc().77 78With everything in place you can set CONFIG_ARCH_ENABLE_SPLIT_PMD_PTLOCK.79 80NOTE: pagetable_pte_ctor() and pagetable_pmd_ctor() can fail -- it must81be handled properly.82 83page->ptl84=========85 86page->ptl is used to access split page table lock, where 'page' is struct87page of page containing the table. It shares storage with page->private88(and few other fields in union).89 90To avoid increasing size of struct page and have best performance, we use a91trick:92 93 - if spinlock_t fits into long, we use page->ptr as spinlock, so we94 can avoid indirect access and save a cache line.95 - if size of spinlock_t is bigger then size of long, we use page->ptl as96 pointer to spinlock_t and allocate it dynamically. This allows to use97 split lock with enabled DEBUG_SPINLOCK or DEBUG_LOCK_ALLOC, but costs98 one more cache line for indirect access;99 100The spinlock_t allocated in pagetable_pte_ctor() for PTE table and in101pagetable_pmd_ctor() for PMD table.102 103Please, never access page->ptl directly -- use appropriate helper.104