[PATCH] mm: split page table lock

Christoph Lameter demonstrated very poor scalability on the SGI 512-way, with
a many-threaded application which concurrently initializes different parts of
a large anonymous area.

This patch corrects that, by using a separate spinlock per page table page, to
guard the page table entries in that page, instead of using the mm's single
page_table_lock.  (But even then, page_table_lock is still used to guard page
table allocation, and anon_vma allocation.)

In this implementation, the spinlock is tucked inside the struct page of the
page table page: with a BUILD_BUG_ON in case it overflows - which it would in
the case of 32-bit PA-RISC with spinlock debugging enabled.

Splitting the lock is not quite for free: another cacheline access.  Ideally,
I suppose we would use split ptlock only for multi-threaded processes on
multi-cpu machines; but deciding that dynamically would have its own costs.
So for now enable it by config, at some number of cpus - since the Kconfig
language doesn't support inequalities, let preprocessor compare that with
NR_CPUS.  But I don't think it's worth being user-configurable: for good
testing of both split and unsplit configs, split now at 4 cpus, and perhaps
change that to 8 later.

There is a benefit even for singly threaded processes: kswapd can be attacking
one part of the mm while another part is busy faulting.

Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>

12 files changed, 74 insertions, 48 deletions

diff --git a/mm/Kconfig b/mm/Kconfig
index 391ffc5..f35a550 100644
--- a/mm/Kconfig
+++ b/mm/Kconfig
@@ -111,3 +111,16 @@ config SPARSEMEM_STATIC
 config SPARSEMEM_EXTREME
 	def_bool y
 	depends on SPARSEMEM && !SPARSEMEM_STATIC
+
+# Heavily threaded applications may benefit from splitting the mm-wide
+# page_table_lock, so that faults on different parts of the user address
+# space can be handled with less contention: split it at this NR_CPUS.
+# Default to 4 for wider testing, though 8 might be more appropriate.
+# ARM's adjust_pte (unused if VIPT) depends on mm-wide page_table_lock.
+# PA-RISC's debug spinlock_t is too large for the 32-bit struct page.
+#
+config SPLIT_PTLOCK_CPUS
+	int
+	default "4096" if ARM && !CPU_CACHE_VIPT
+	default "4096" if PARISC && DEBUG_SPINLOCK && !64BIT
+	default "4"
diff --git a/mm/filemap.c b/mm/filemap.c
index 8aa344e..f560b41 100644
--- a/mm/filemap.c
+++ b/mm/filemap.c
@@ -152,7 +152,7 @@ static int sync_page(void *word)
 	 * in the ->sync_page() methods make essential use of the
 	 * page_mapping(), merely passing the page down to the backing
 	 * device's unplug functions when it's non-NULL, which in turn
-	 * ignore it for all cases but swap, where only page->private is
+	 * ignore it for all cases but swap, where only page_private(page) is
 	 * of interest. When page_mapping() does go NULL, the entire
 	 * call stack gracefully ignores the page and returns.
 	 * -- wli
diff --git a/mm/memory.c b/mm/memory.c
index 8461e2d..e9ef599 100644
--- a/mm/memory.c
+++ b/mm/memory.c
@@ -114,6 +114,7 @@ static void free_pte_range(struct mmu_gather *tlb, pmd_t *pmd)
 {
 	struct page *page = pmd_page(*pmd);
 	pmd_clear(pmd);
+	pte_lock_deinit(page);
 	pte_free_tlb(tlb, page);
 	dec_page_state(nr_page_table_pages);
 	tlb->mm->nr_ptes--;
@@ -294,10 +295,12 @@ int __pte_alloc(struct mm_struct *mm, pmd_t *pmd, unsigned long address)
 	if (!new)
 		return -ENOMEM;
 
+	pte_lock_init(new);
 	spin_lock(&mm->page_table_lock);
-	if (pmd_present(*pmd))		/* Another has populated it */
+	if (pmd_present(*pmd)) {	/* Another has populated it */
+		pte_lock_deinit(new);
 		pte_free(new);
-	else {
+	} else {
 		mm->nr_ptes++;
 		inc_page_state(nr_page_table_pages);
 		pmd_populate(mm, pmd, new);
@@ -432,7 +435,7 @@ again:
 	if (!dst_pte)
 		return -ENOMEM;
 	src_pte = pte_offset_map_nested(src_pmd, addr);
-	src_ptl = &src_mm->page_table_lock;
+	src_ptl = pte_lockptr(src_mm, src_pmd);
 	spin_lock(src_ptl);
 
 	do {
@@ -1194,15 +1197,16 @@ EXPORT_SYMBOL(remap_pfn_range);
  * (but do_wp_page is only called after already making such a check;
  * and do_anonymous_page and do_no_page can safely check later on).
  */
-static inline int pte_unmap_same(struct mm_struct *mm,
+static inline int pte_unmap_same(struct mm_struct *mm, pmd_t *pmd,
 				pte_t *page_table, pte_t orig_pte)
 {
 	int same = 1;
 #if defined(CONFIG_SMP) || defined(CONFIG_PREEMPT)
 	if (sizeof(pte_t) > sizeof(unsigned long)) {
-		spin_lock(&mm->page_table_lock);
+		spinlock_t *ptl = pte_lockptr(mm, pmd);
+		spin_lock(ptl);
 		same = pte_same(*page_table, orig_pte);
-		spin_unlock(&mm->page_table_lock);
+		spin_unlock(ptl);
 	}
 #endif
 	pte_unmap(page_table);
@@ -1655,7 +1659,7 @@ static int do_swap_page(struct mm_struct *mm, struct vm_area_struct *vma,
 	pte_t pte;
 	int ret = VM_FAULT_MINOR;
 
-	if (!pte_unmap_same(mm, page_table, orig_pte))
+	if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
 		goto out;
 
 	entry = pte_to_swp_entry(orig_pte);
@@ -1773,7 +1777,7 @@ static int do_anonymous_page(struct mm_struct *mm, struct vm_area_struct *vma,
 		page_cache_get(page);
 		entry = mk_pte(page, vma->vm_page_prot);
 
-		ptl = &mm->page_table_lock;
+		ptl = pte_lockptr(mm, pmd);
 		spin_lock(ptl);
 		if (!pte_none(*page_table))
 			goto release;
@@ -1934,7 +1938,7 @@ static int do_file_page(struct mm_struct *mm, struct vm_area_struct *vma,
 	pgoff_t pgoff;
 	int err;
 
-	if (!pte_unmap_same(mm, page_table, orig_pte))
+	if (!pte_unmap_same(mm, pmd, page_table, orig_pte))
 		return VM_FAULT_MINOR;
 
 	if (unlikely(!(vma->vm_flags & VM_NONLINEAR))) {
@@ -1992,7 +1996,7 @@ static inline int handle_pte_fault(struct mm_struct *mm,
 					pte, pmd, write_access, entry);
 	}
 
-	ptl = &mm->page_table_lock;
+	ptl = pte_lockptr(mm, pmd);
 	spin_lock(ptl);
 	if (unlikely(!pte_same(*pte, entry)))
 		goto unlock;
diff --git a/mm/mremap.c b/mm/mremap.c
index 8de77b6..b535438 100644
--- a/mm/mremap.c
+++ b/mm/mremap.c
@@ -72,7 +72,7 @@ static void move_ptes(struct vm_area_struct *vma, pmd_t *old_pmd,
 	struct address_space *mapping = NULL;
 	struct mm_struct *mm = vma->vm_mm;
 	pte_t *old_pte, *new_pte, pte;
-	spinlock_t *old_ptl;
+	spinlock_t *old_ptl, *new_ptl;
 
 	if (vma->vm_file) {
 		/*
@@ -88,8 +88,15 @@ static void move_ptes(struct vm_area_struct *vma, pmd_t *old_pmd,
 			new_vma->vm_truncate_count = 0;
 	}
 
+	/*
+	 * We don't have to worry about the ordering of src and dst
+	 * pte locks because exclusive mmap_sem prevents deadlock.
+	 */
 	old_pte = pte_offset_map_lock(mm, old_pmd, old_addr, &old_ptl);
  	new_pte = pte_offset_map_nested(new_pmd, new_addr);
+	new_ptl = pte_lockptr(mm, new_pmd);
+	if (new_ptl != old_ptl)
+		spin_lock(new_ptl);
 
 	for (; old_addr < old_end; old_pte++, old_addr += PAGE_SIZE,
 				   new_pte++, new_addr += PAGE_SIZE) {
@@ -101,6 +108,8 @@ static void move_ptes(struct vm_area_struct *vma, pmd_t *old_pmd,
 		set_pte_at(mm, new_addr, new_pte, pte);
 	}
 
+	if (new_ptl != old_ptl)
+		spin_unlock(new_ptl);
 	pte_unmap_nested(new_pte - 1);
 	pte_unmap_unlock(old_pte - 1, old_ptl);
 	if (mapping)
diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index 0541288..a2995a5 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -154,7 +154,7 @@ static void prep_compound_page(struct page *page, unsigned long order)
 		struct page *p = page + i;
 
 		SetPageCompound(p);
-		p->private = (unsigned long)page;
+		set_page_private(p, (unsigned long)page);
 	}
 }
 
@@ -174,7 +174,7 @@ static void destroy_compound_page(struct page *page, unsigned long order)
 
 		if (!PageCompound(p))
 			bad_page(__FUNCTION__, page);
-		if (p->private != (unsigned long)page)
+		if (page_private(p) != (unsigned long)page)
 			bad_page(__FUNCTION__, page);
 		ClearPageCompound(p);
 	}
@@ -187,18 +187,18 @@ static void destroy_compound_page(struct page *page, unsigned long order)
  * So, we don't need atomic page->flags operations here.
  */
 static inline unsigned long page_order(struct page *page) {
-	return page->private;
+	return page_private(page);
 }
 
 static inline void set_page_order(struct page *page, int order) {
-	page->private = order;
+	set_page_private(page, order);
 	__SetPagePrivate(page);
 }
 
 static inline void rmv_page_order(struct page *page)
 {
 	__ClearPagePrivate(page);
-	page->private = 0;
+	set_page_private(page, 0);
 }
 
 /*
@@ -238,7 +238,7 @@ __find_combined_index(unsigned long page_idx, unsigned int order)
  * (a) the buddy is free &&
  * (b) the buddy is on the buddy system &&
  * (c) a page and its buddy have the same order.
- * for recording page's order, we use page->private and PG_private.
+ * for recording page's order, we use page_private(page) and PG_private.
  *
  */
 static inline int page_is_buddy(struct page *page, int order)
@@ -264,7 +264,7 @@ static inline int page_is_buddy(struct page *page, int order)
  * parts of the VM system.
  * At each level, we keep a list of pages, which are heads of continuous
  * free pages of length of (1 << order) and marked with PG_Private.Page's
- * order is recorded in page->private field.
+ * order is recorded in page_private(page) field.
  * So when we are allocating or freeing one, we can derive the state of the
  * other.  That is, if we allocate a small block, and both were   
  * free, the remainder of the region must be split into blocks.   
@@ -463,7 +463,7 @@ static void prep_new_page(struct page *page, int order)
 	page->flags &= ~(1 << PG_uptodate | 1 << PG_error |
 			1 << PG_referenced | 1 << PG_arch_1 |
 			1 << PG_checked | 1 << PG_mappedtodisk);
-	page->private = 0;
+	set_page_private(page, 0);
 	set_page_refs(page, order);
 	kernel_map_pages(page, 1 << order, 1);
 }
diff --git a/mm/page_io.c b/mm/page_io.c
index 330e00d..bb2b0d5 100644
--- a/mm/page_io.c
+++ b/mm/page_io.c
@@ -91,7 +91,8 @@ int swap_writepage(struct page *page, struct writeback_control *wbc)
 		unlock_page(page);
 		goto out;
 	}
-	bio = get_swap_bio(GFP_NOIO, page->private, page, end_swap_bio_write);
+	bio = get_swap_bio(GFP_NOIO, page_private(page), page,
+				end_swap_bio_write);
 	if (bio == NULL) {
 		set_page_dirty(page);
 		unlock_page(page);
@@ -115,7 +116,8 @@ int swap_readpage(struct file *file, struct page *page)
 
 	BUG_ON(!PageLocked(page));
 	ClearPageUptodate(page);
-	bio = get_swap_bio(GFP_KERNEL, page->private, page, end_swap_bio_read);
+	bio = get_swap_bio(GFP_KERNEL, page_private(page), page,
+				end_swap_bio_read);
 	if (bio == NULL) {
 		unlock_page(page);
 		ret = -ENOMEM;
diff --git a/mm/rmap.c b/mm/rmap.c
index a84bdfe..a33e779 100644
--- a/mm/rmap.c
+++ b/mm/rmap.c
@@ -274,7 +274,7 @@ pte_t *page_check_address(struct page *page, struct mm_struct *mm,
 		return NULL;
 	}
 
-	ptl = &mm->page_table_lock;
+	ptl = pte_lockptr(mm, pmd);
 	spin_lock(ptl);
 	if (pte_present(*pte) && page_to_pfn(page) == pte_pfn(*pte)) {
 		*ptlp = ptl;
@@ -550,7 +550,7 @@ static int try_to_unmap_one(struct page *page, struct vm_area_struct *vma)
 	update_hiwater_rss(mm);
 
 	if (PageAnon(page)) {
-		swp_entry_t entry = { .val = page->private };
+		swp_entry_t entry = { .val = page_private(page) };
 		/*
 		 * Store the swap location in the pte.
 		 * See handle_pte_fault() ...
diff --git a/mm/shmem.c b/mm/shmem.c
index 37777f4..dc25565 100644
--- a/mm/shmem.c
+++ b/mm/shmem.c
@@ -71,9 +71,6 @@
 /* Pretend that each entry is of this size in directory's i_size */
 #define BOGO_DIRENT_SIZE 20
 
-/* Keep swapped page count in private field of indirect struct page */
-#define nr_swapped		private
-
 /* Flag allocation requirements to shmem_getpage and shmem_swp_alloc */
 enum sgp_type {
 	SGP_QUICK,	/* don't try more than file page cache lookup */
@@ -324,8 +321,10 @@ static void shmem_swp_set(struct shmem_inode_info *info, swp_entry_t *entry, uns
 
 	entry->val = value;
 	info->swapped += incdec;
-	if ((unsigned long)(entry - info->i_direct) >= SHMEM_NR_DIRECT)
-		kmap_atomic_to_page(entry)->nr_swapped += incdec;
+	if ((unsigned long)(entry - info->i_direct) >= SHMEM_NR_DIRECT) {
+		struct page *page = kmap_atomic_to_page(entry);
+		set_page_private(page, page_private(page) + incdec);
+	}
 }
 
 /*
@@ -368,9 +367,8 @@ static swp_entry_t *shmem_swp_alloc(struct shmem_inode_info *info, unsigned long
 
 		spin_unlock(&info->lock);
 		page = shmem_dir_alloc(mapping_gfp_mask(inode->i_mapping) | __GFP_ZERO);
-		if (page) {
-			page->nr_swapped = 0;
-		}
+		if (page)
+			set_page_private(page, 0);
 		spin_lock(&info->lock);
 
 		if (!page) {
@@ -561,7 +559,7 @@ static void shmem_truncate(struct inode *inode)
 			diroff = 0;
 		}
 		subdir = dir[diroff];
-		if (subdir && subdir->nr_swapped) {
+		if (subdir && page_private(subdir)) {
 			size = limit - idx;
 			if (size > ENTRIES_PER_PAGE)
 				size = ENTRIES_PER_PAGE;
@@ -572,10 +570,10 @@ static void shmem_truncate(struct inode *inode)
 			nr_swaps_freed += freed;
 			if (offset)
 				spin_lock(&info->lock);
-			subdir->nr_swapped -= freed;
+			set_page_private(subdir, page_private(subdir) - freed);
 			if (offset)
 				spin_unlock(&info->lock);
-			BUG_ON(subdir->nr_swapped > offset);
+			BUG_ON(page_private(subdir) > offset);
 		}
 		if (offset)
 			offset = 0;
@@ -743,7 +741,7 @@ static int shmem_unuse_inode(struct shmem_inode_info *info, swp_entry_t entry, s
 			dir = shmem_dir_map(subdir);
 		}
 		subdir = *dir;
-		if (subdir && subdir->nr_swapped) {
+		if (subdir && page_private(subdir)) {
 			ptr = shmem_swp_map(subdir);
 			size = limit - idx;
 			if (size > ENTRIES_PER_PAGE)
diff --git a/mm/swap.c b/mm/swap.c
index 21d15f9..b895128 100644
--- a/mm/swap.c
+++ b/mm/swap.c
@@ -39,7 +39,7 @@ int page_cluster;
 void put_page(struct page *page)
 {
 	if (unlikely(PageCompound(page))) {
-		page = (struct page *)page->private;
+		page = (struct page *)page_private(page);
 		if (put_page_testzero(page)) {
 			void (*dtor)(struct page *page);
 
diff --git a/mm/swap_state.c b/mm/swap_state.c
index 132164f..cafc1ed 100644
--- a/mm/swap_state.c
+++ b/mm/swap_state.c
@@ -83,7 +83,7 @@ static int __add_to_swap_cache(struct page *page, swp_entry_t entry,
 			page_cache_get(page);
 			SetPageLocked(page);
 			SetPageSwapCache(page);
-			page->private = entry.val;
+			set_page_private(page, entry.val);
 			total_swapcache_pages++;
 			pagecache_acct(1);
 		}
@@ -126,8 +126,8 @@ void __delete_from_swap_cache(struct page *page)
 	BUG_ON(PageWriteback(page));
 	BUG_ON(PagePrivate(page));
 
-	radix_tree_delete(&swapper_space.page_tree, page->private);
-	page->private = 0;
+	radix_tree_delete(&swapper_space.page_tree, page_private(page));
+	set_page_private(page, 0);
 	ClearPageSwapCache(page);
 	total_swapcache_pages--;
 	pagecache_acct(-1);
@@ -197,7 +197,7 @@ void delete_from_swap_cache(struct page *page)
 {
 	swp_entry_t entry;
 
-	entry.val = page->private;
+	entry.val = page_private(page);
 
 	write_lock_irq(&swapper_space.tree_lock);
 	__delete_from_swap_cache(page);
diff --git a/mm/swapfile.c b/mm/swapfile.c
index 510f003..8970c0b 100644
--- a/mm/swapfile.c
+++ b/mm/swapfile.c
@@ -61,7 +61,7 @@ void swap_unplug_io_fn(struct backing_dev_info *unused_bdi, struct page *page)
 	swp_entry_t entry;
 
 	down_read(&swap_unplug_sem);
-	entry.val = page->private;
+	entry.val = page_private(page);
 	if (PageSwapCache(page)) {
 		struct block_device *bdev = swap_info[swp_type(entry)].bdev;
 		struct backing_dev_info *bdi;
@@ -69,8 +69,8 @@ void swap_unplug_io_fn(struct backing_dev_info *unused_bdi, struct page *page)
 		/*
 		 * If the page is removed from swapcache from under us (with a
 		 * racy try_to_unuse/swapoff) we need an additional reference
-		 * count to avoid reading garbage from page->private above. If
-		 * the WARN_ON triggers during a swapoff it maybe the race
+		 * count to avoid reading garbage from page_private(page) above.
+		 * If the WARN_ON triggers during a swapoff it maybe the race
 		 * condition and it's harmless. However if it triggers without
 		 * swapoff it signals a problem.
 		 */
@@ -294,7 +294,7 @@ static inline int page_swapcount(struct page *page)
 	struct swap_info_struct *p;
 	swp_entry_t entry;
 
-	entry.val = page->private;
+	entry.val = page_private(page);
 	p = swap_info_get(entry);
 	if (p) {
 		/* Subtract the 1 for the swap cache itself */
@@ -339,7 +339,7 @@ int remove_exclusive_swap_page(struct page *page)
 	if (page_count(page) != 2) /* 2: us + cache */
 		return 0;
 
-	entry.val = page->private;
+	entry.val = page_private(page);
 	p = swap_info_get(entry);
 	if (!p)
 		return 0;
@@ -1042,7 +1042,7 @@ int page_queue_congested(struct page *page)
 	BUG_ON(!PageLocked(page));	/* It pins the swap_info_struct */
 
 	if (PageSwapCache(page)) {
-		swp_entry_t entry = { .val = page->private };
+		swp_entry_t entry = { .val = page_private(page) };
 		struct swap_info_struct *sis;
 
 		sis = get_swap_info_struct(swp_type(entry));
diff --git a/mm/vmscan.c b/mm/vmscan.c
index 41d1064..135bf8c 100644
--- a/mm/vmscan.c
+++ b/mm/vmscan.c
@@ -521,7 +521,7 @@ static int shrink_list(struct list_head *page_list, struct scan_control *sc)
 
 #ifdef CONFIG_SWAP
 		if (PageSwapCache(page)) {
-			swp_entry_t swap = { .val = page->private };
+			swp_entry_t swap = { .val = page_private(page) };
 			__delete_from_swap_cache(page);
 			write_unlock_irq(&mapping->tree_lock);
 			swap_free(swap);