1===================== 2Split page table lock 3===================== 4 5Originally, mm->page_table_lock spinlock protected all page tables of the 6mm_struct. But this approach leads to poor page fault scalability of 7multi-threaded applications due high contention on the lock. To improve 8scalability, split page table lock was introduced. 9 10With split page table lock we have separate per-table lock to serialize 11access to the table. At the moment we use split lock for PTE and PMD 12tables. 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 the taken 18 lock; 19 - pte_unmap_unlock() 20 unlocks and unmaps PTE table; 21 - pte_alloc_map_lock() 22 allocates PTE table if needed and take the lock, returns pointer 23 to taken lock or NULL if allocation failed; 24 - pte_lockptr() 25 returns pointer to PTE table lock; 26 - pmd_lock() 27 takes PMD table lock, returns pointer to taken lock; 28 - pmd_lockptr() 29 returns pointer to PMD table lock; 30 31Split page table lock for PTE tables is enabled compile-time if 32CONFIG_SPLIT_PTLOCK_CPUS (usually 4) is less or equal to NR_CPUS. 33If split lock is disabled, all tables are guarded by mm->page_table_lock. 34 35Split page table lock for PMD tables is enabled, if it's enabled for PTE 36tables and the architecture supports it (see below). 37 38Hugetlb and split page table lock 39================================= 40 41Hugetlb can support several page sizes. We use split lock only for PMD 42level, but not for PUD. 43 44Hugetlb-specific helpers: 45 46 - huge_pte_lock() 47 takes pmd split lock for PMD_SIZE page, mm->page_table_lock 48 otherwise; 49 - huge_pte_lockptr() 50 returns pointer to table lock; 51 52Support of split page table lock by an architecture 53=================================================== 54 55There's no need in special enabling of PTE split page table lock: everything 56required is done by pgtable_pte_page_ctor() and pgtable_pte_page_dtor(), which 57must be called on PTE table allocation / freeing. 58 59Make sure the architecture doesn't use slab allocator for page table 60allocation: slab uses page->slab_cache for its pages. 61This field shares storage with page->ptl. 62 63PMD split lock only makes sense if you have more than two page table 64levels. 65 66PMD split lock enabling requires pgtable_pmd_page_ctor() call on PMD table 67allocation and pgtable_pmd_page_dtor() on freeing. 68 69Allocation usually happens in pmd_alloc_one(), freeing in pmd_free() and 70pmd_free_tlb(), but make sure you cover all PMD table allocation / freeing 71paths: i.e X86_PAE preallocate few PMDs on pgd_alloc(). 72 73With everything in place you can set CONFIG_ARCH_ENABLE_SPLIT_PMD_PTLOCK. 74 75NOTE: pgtable_pte_page_ctor() and pgtable_pmd_page_ctor() can fail -- it must 76be handled properly. 77 78page->ptl 79========= 80 81page->ptl is used to access split page table lock, where 'page' is struct 82page of page containing the table. It shares storage with page->private 83(and few other fields in union). 84 85To avoid increasing size of struct page and have best performance, we use a 86trick: 87 88 - if spinlock_t fits into long, we use page->ptr as spinlock, so we 89 can avoid indirect access and save a cache line. 90 - if size of spinlock_t is bigger then size of long, we use page->ptl as 91 pointer to spinlock_t and allocate it dynamically. This allows to use 92 split lock with enabled DEBUG_SPINLOCK or DEBUG_LOCK_ALLOC, but costs 93 one more cache line for indirect access; 94 95The spinlock_t allocated in pgtable_pte_page_ctor() for PTE table and in 96pgtable_pmd_page_ctor() for PMD table. 97 98Please, never access page->ptl directly -- use appropriate helper. 99