[PATCH] low-latency page reclaim

Convert the VM to not wait on other people's dirty data.

 - If we find a dirty page and its queue is not congested, do some writeback.

 - If we find a dirty page and its queue _is_ congested then just
   refile the page.

 - If we find a PageWriteback page then just refile the page.

 - There is additional throttling for write(2) callers.  Within
   generic_file_write(), record their backing queue in ->current.
   Within page reclaim, if this tasks encounters a page which is dirty
   or under writeback onthis queue, block on it.  This gives some more
   writer throttling and reduces the page refiling frequency.

It's somewhat CPU expensive - under really heavy load we only get a 50%
reclaim rate in pages coming off the tail of the LRU.  This can be
fixed by splitting the inactive list into reclaimable and
non-reclaimable lists.  But the CPU load isn't too bad, and latency is
much, much more important in these situations.

Example: with `mem=512m', running 4 instances of `dbench 100', 2.5.34
took 35 minutes to compile a kernel.  With this patch, it took three
minutes, 45 seconds.

I haven't done swapcache or MAP_SHARED pages yet.  If there's tons of
dirty swapcache or mmap data around we still stall heavily in page
reclaim.  That's less important.

This patch also has a tweak for swapless machines: don't even bother
bringing anon pages onto the inactive list if there is no swap online.

drivers/char/sysrq.c

@@ -28,7 +28,8 @@
 #include <linux/smp_lock.h>
 #include <linux/module.h>
 #include <linux/suspend.h>
-#include <linux/buffer_head.h>		/* for fsync_bdev()/wakeup_bdflush() */
+#include <linux/writeback.h>
+#include <linux/buffer_head.h>		/* for fsync_bdev() */
 
 #include <linux/spinlock.h>
 
@@ -227,7 +228,7 @@ static void sysrq_handle_sync(int key, struct pt_regs *pt_regs,
 			      struct tty_struct *tty) 
 {
 	emergency_sync_scheduled = EMERG_SYNC;
-	wakeup_bdflush();
+	wakeup_bdflush(0);
 }
 static struct sysrq_key_op sysrq_sync_op = {
 	handler:	sysrq_handle_sync,
@@ -239,7 +240,7 @@ static void sysrq_handle_mountro(int key, struct pt_regs *pt_regs,
 				 struct tty_struct *tty) 
 {
 	emergency_sync_scheduled = EMERG_REMOUNT;
-	wakeup_bdflush();
+	wakeup_bdflush(0);
 }
 static struct sysrq_key_op sysrq_mountro_op = {
 	handler:	sysrq_handle_mountro,

fs/buffer.c

@@ -458,19 +458,17 @@ void __invalidate_buffers(kdev_t dev, int destroy_dirty_buffers)
 }
 
 /*
- * FIXME: What is this function actually trying to do?  Why "zones[0]"?
- * Is it still correct/needed if/when blockdev mappings use GFP_HIGHUSER?
+ * Kick pdflush then try to free up some ZONE_NORMAL memory.
  */
 static void free_more_memory(void)
 {
 	struct zone *zone;
 
-	zone = contig_page_data.node_zonelists[GFP_NOFS & GFP_ZONEMASK].zones[0];
-
-	wakeup_bdflush();
-	try_to_free_pages(zone, GFP_NOFS, 0);
+	zone = contig_page_data.node_zonelists[GFP_NOFS&GFP_ZONEMASK].zones[0];
+	wakeup_bdflush(1024);
 	blk_run_queues();
 	yield();
+	try_to_free_pages(zone, GFP_NOFS, 0);
 }
 
 /*

include/linux/buffer_head.h

@@ -163,7 +163,6 @@ struct buffer_head * __getblk(struct block_device *, sector_t, int);
 void __brelse(struct buffer_head *);
 void __bforget(struct buffer_head *);
 struct buffer_head *__bread(struct block_device *, sector_t block, int size);
-void wakeup_bdflush(void);
 struct buffer_head *alloc_buffer_head(void);
 void free_buffer_head(struct buffer_head * bh);
 void FASTCALL(unlock_buffer(struct buffer_head *bh));

include/linux/sched.h

@@ -273,6 +273,7 @@ extern struct user_struct root_user;
 #define INIT_USER (&root_user)
 
 typedef struct prio_array prio_array_t;
+struct backing_dev_info;
 
 struct task_struct {
 	volatile long state;	/* -1 unrunnable, 0 runnable, >0 stopped */
@@ -398,6 +399,7 @@ struct task_struct {
 /* journalling filesystem info */
 	void *journal_info;
 	struct dentry *proc_dentry;
+	struct backing_dev_info *backing_dev_info;
 };
 
 extern void __put_task_struct(struct task_struct *tsk);

include/linux/writeback.h

@@ -63,6 +63,8 @@ static inline void wait_on_inode(struct inode *inode)
 /*
  * mm/page-writeback.c
  */
+int wakeup_bdflush(long nr_pages);
+
 /* These 5 are exported to sysctl. */
 extern int dirty_background_ratio;
 extern int dirty_async_ratio;

kernel/suspend.c

@@ -81,7 +81,6 @@ unsigned char software_suspend_enabled = 0;
 #define TIMEOUT	(6 * HZ)			/* Timeout for stopping processes */
 #define ADDRESS(x) ((unsigned long) phys_to_virt(((x) << PAGE_SHIFT)))
 
-extern void wakeup_bdflush(void);
 extern int C_A_D;
 
 /* References to section boundaries */

mm/filemap.c

@@ -1755,6 +1755,9 @@ generic_file_write_nolock(struct file *file, const struct iovec *iov,
 	if (unlikely(pos < 0))
 		return -EINVAL;
 
+	/* We can write back this queue in page reclaim */
+	current->backing_dev_info = mapping->backing_dev_info;
+
 	pagevec_init(&lru_pvec);
 
 	if (unlikely(file->f_error)) {
@@ -1959,6 +1962,7 @@ generic_file_write_nolock(struct file *file, const struct iovec *iov,
 	err = written ? written : status;
 out:
 	pagevec_lru_add(&lru_pvec);
+	current->backing_dev_info = 0;
 	return err;
 }
 

mm/mempool.c

@@ -12,7 +12,7 @@
 #include <linux/slab.h>
 #include <linux/module.h>
 #include <linux/mempool.h>
-#include <linux/buffer_head.h>		/* for wakeup_bdflush() */
+#include <linux/writeback.h>
 
 static void add_element(mempool_t *pool, void *element)
 {
@@ -210,7 +210,7 @@ void * mempool_alloc(mempool_t *pool, int gfp_mask)
 	/*
 	 * Kick the VM at this point.
 	 */
-	wakeup_bdflush();
+	wakeup_bdflush(0);
 
 	spin_lock_irqsave(&pool->lock, flags);
 	if (likely(pool->curr_nr)) {

mm/page-writeback.c

@@ -201,14 +201,19 @@ static void background_writeout(unsigned long _min_pages)
 }
 
 /*
- * Start heavy writeback of everything.
+ * Start writeback of `nr_pages' pages.  If `nr_pages' is zero, write back
+ * the whole world.  Returns 0 if a pdflush thread was dispatched.  Returns
+ * -1 if all pdflush threads were busy.
  */
-void wakeup_bdflush(void)
+int wakeup_bdflush(long nr_pages)
 {
+	if (nr_pages == 0) {
 		struct page_state ps;
 
 		get_page_state(&ps);
-	pdflush_operation(background_writeout, ps.nr_dirty);
+		nr_pages = ps.nr_dirty;
+	}
+	return pdflush_operation(background_writeout, nr_pages);
 }
 
 static struct timer_list wb_timer;

mm/swapfile.c

@@ -918,6 +918,26 @@ static int setup_swap_extents(struct swap_info_struct *sis)
 	return ret;
 }
 
+#if 0	/* We don't need this yet */
+#include <linux/backing-dev.h>
+int page_queue_congested(struct page *page)
+{
+	struct backing_dev_info *bdi;
+
+	BUG_ON(!PageLocked(page));	/* It pins the swap_info_struct */
+
+	bdi = page->mapping->backing_dev_info;
+	if (PageSwapCache(page)) {
+		swp_entry_t entry = { .val = page->index };
+		struct swap_info_struct *sis;
+
+		sis = get_swap_info_struct(swp_type(entry));
+		bdi = sis->bdev->bd_inode->i_mapping->backing_dev_info;
+	}
+	return bdi_write_congested(bdi);
+}
+#endif
+
 asmlinkage long sys_swapoff(const char * specialfile)
 {
 	struct swap_info_struct * p = NULL;

mm/vmscan.c

@@ -21,9 +21,11 @@
 #include <linux/file.h>
 #include <linux/writeback.h>
 #include <linux/suspend.h>
+#include <linux/blkdev.h>
 #include <linux/buffer_head.h>		/* for try_to_release_page() */
 #include <linux/mm_inline.h>
 #include <linux/pagevec.h>
+#include <linux/backing-dev.h>
 #include <linux/rmap-locking.h>
 
 #include <asm/pgalloc.h>
@@ -32,11 +34,11 @@
 
 /*
  * The "priority" of VM scanning is how much of the queues we
- * will scan in one go. A value of 6 for DEF_PRIORITY implies
- * that we'll scan 1/64th of the queues ("queue_length >> 6")
+ * will scan in one go. A value of 12 for DEF_PRIORITY implies
+ * that we'll scan 1/4096th of the queues ("queue_length >> 12")
  * during a normal aging round.
  */
-#define DEF_PRIORITY (6)
+#define DEF_PRIORITY 12
 
 #ifdef ARCH_HAS_PREFETCH
 #define prefetch_prev_lru_page(_page, _base, _field)			\
@@ -95,7 +97,7 @@ static inline int is_page_cache_freeable(struct page *page)
 
 static /* inline */ int
 shrink_list(struct list_head *page_list, int nr_pages,
-		unsigned int gfp_mask, int priority, int *max_scan)
+		unsigned int gfp_mask, int *max_scan)
 {
 	struct address_space *mapping;
 	LIST_HEAD(ret_pages);
@@ -117,10 +119,21 @@ shrink_list(struct list_head *page_list, int nr_pages,
 		BUG_ON(PageActive(page));
 		may_enter_fs = (gfp_mask & __GFP_FS) ||
 				(PageSwapCache(page) && (gfp_mask & __GFP_IO));
+
+		/*
+		 * If the page is mapped into pagetables then wait on it, to
+		 * throttle this allocator to the rate at which we can clear
+		 * MAP_SHARED data.  This will also throttle against swapcache
+		 * writes.
+		 */
 		if (PageWriteback(page)) {
-			if (may_enter_fs)
-				wait_on_page_writeback(page);  /* throttling */
-			else
+			if (may_enter_fs) {
+				if (page->pte.direct ||
+					page->mapping->backing_dev_info ==
+						current->backing_dev_info) {
+					wait_on_page_writeback(page);
+				}
+			}
 			goto keep_locked;
 		}
 
@@ -172,15 +185,43 @@ shrink_list(struct list_head *page_list, int nr_pages,
 		 * will write it.  So we're back to page-at-a-time writepage
 		 * in LRU order.
 		 */
-		if (PageDirty(page) && is_page_cache_freeable(page) &&
-					mapping && may_enter_fs) {
+		/*
+		 * If the page is dirty, only perform writeback if that write
+		 * will be non-blocking.  To prevent this allocation from being
+		 * stalled by pagecache activity.  But note that there may be
+		 * stalls if we need to run get_block().  We could test
+		 * PagePrivate for that.
+		 *
+		 * If this process is currently in generic_file_write() against
+		 * this page's queue, we can perform writeback even if that
+		 * will block.
+		 *
+		 * If the page is swapcache, write it back even if that would
+		 * block, for some throttling. This happens by accident, because
+		 * swap_backing_dev_info is bust: it doesn't reflect the
+		 * congestion state of the swapdevs.  Easy to fix, if needed.
+		 * See swapfile.c:page_queue_congested().
+		 */
+		if (PageDirty(page)) {
 			int (*writeback)(struct page *,
 					struct writeback_control *);
+			struct backing_dev_info *bdi;
 			const int cluster_size = SWAP_CLUSTER_MAX;
 			struct writeback_control wbc = {
 				.nr_to_write = cluster_size,
 			};
 
+			if (!is_page_cache_freeable(page))
+				goto keep_locked;
+			if (!mapping)
+				goto keep_locked;
+			if (!may_enter_fs)
+				goto keep_locked;
+			bdi = mapping->backing_dev_info;
+			if (bdi != current->backing_dev_info &&
+					bdi_write_congested(bdi))
+				goto keep_locked;
+
 			writeback = mapping->a_ops->vm_writeback;
 			if (writeback == NULL)
 				writeback = generic_vm_writeback;
@@ -279,7 +320,7 @@ shrink_list(struct list_head *page_list, int nr_pages,
  */
 static /* inline */ int
 shrink_cache(int nr_pages, struct zone *zone,
-		unsigned int gfp_mask, int priority, int max_scan)
+		unsigned int gfp_mask, int max_scan)
 {
 	LIST_HEAD(page_list);
 	struct pagevec pvec;
@@ -298,9 +339,11 @@ shrink_cache(int nr_pages, struct zone *zone,
 	spin_lock_irq(&zone->lru_lock);
 	while (max_scan > 0 && nr_pages > 0) {
 		struct page *page;
-		int n = 0;
+		int nr_taken = 0;
+		int nr_scan = 0;
 
-		while (n < nr_to_process && !list_empty(&zone->inactive_list)) {
+		while (nr_scan++ < nr_to_process &&
+				!list_empty(&zone->inactive_list)) {
 			page = list_entry(zone->inactive_list.prev,
 						struct page, lru);
 
@@ -318,18 +361,17 @@ shrink_cache(int nr_pages, struct zone *zone,
 			}
 			list_add(&page->lru, &page_list);
 			page_cache_get(page);
-			n++;
+			nr_taken++;
 		}
-		zone->nr_inactive -= n;
+		zone->nr_inactive -= nr_taken;
 		spin_unlock_irq(&zone->lru_lock);
 
-		if (list_empty(&page_list))
+		if (nr_taken == 0)
 			goto done;
 
-		max_scan -= n;
-		KERNEL_STAT_ADD(pgscan, n);
-		nr_pages = shrink_list(&page_list, nr_pages,
-					gfp_mask, priority, &max_scan);
+		max_scan -= nr_scan;
+		KERNEL_STAT_ADD(pgscan, nr_scan);
+		nr_pages = shrink_list(&page_list,nr_pages,gfp_mask,&max_scan);
 
 		if (nr_pages <= 0 && list_empty(&page_list))
 			goto done;
@@ -420,6 +462,15 @@ refill_inactive_zone(struct zone *zone, const int nr_pages_in)
 			}
 			pte_chain_unlock(page);
 		}
+		/*
+		 * FIXME: need to consider page_count(page) here if/when we
+		 * reap orphaned pages via the LRU (Daniel's locking stuff)
+		 */
+		if (total_swap_pages == 0 && !page->mapping &&
+						!PagePrivate(page)) {
+			list_add(&page->lru, &l_active);
+			continue;
+		}
 		list_add(&page->lru, &l_inactive);
 		pgdeactivate++;
 	}
@@ -470,11 +521,10 @@ refill_inactive_zone(struct zone *zone, const int nr_pages_in)
 }
 
 static /* inline */ int
-shrink_zone(struct zone *zone, int priority,
+shrink_zone(struct zone *zone, int max_scan,
 		unsigned int gfp_mask, int nr_pages)
 {
 	unsigned long ratio;
-	int max_scan;
 
 	/* This is bogus for ZONE_HIGHMEM? */
 	if (kmem_cache_reap(gfp_mask) >= nr_pages)
@@ -497,43 +547,50 @@ shrink_zone(struct zone *zone, int priority,
 		atomic_sub(SWAP_CLUSTER_MAX, &zone->refill_counter);
 		refill_inactive_zone(zone, SWAP_CLUSTER_MAX);
 	}
-
-	max_scan = zone->nr_inactive / priority;
-	nr_pages = shrink_cache(nr_pages, zone,
-				gfp_mask, priority, max_scan);
-
-	if (nr_pages <= 0)
-		return 0;
-
-	wakeup_bdflush();
-
-	shrink_dcache_memory(priority, gfp_mask);
-
-	/* After shrinking the dcache, get rid of unused inodes too .. */
-	shrink_icache_memory(1, gfp_mask);
-#ifdef CONFIG_QUOTA
-	shrink_dqcache_memory(DEF_PRIORITY, gfp_mask);
-#endif
-
+	nr_pages = shrink_cache(nr_pages, zone, gfp_mask, max_scan);
 	return nr_pages;
 }
 
 static int
 shrink_caches(struct zone *classzone, int priority,
-		int gfp_mask, int nr_pages)
+		int *total_scanned, int gfp_mask, int nr_pages)
 {
 	struct zone *first_classzone;
 	struct zone *zone;
 
 	first_classzone = classzone->zone_pgdat->node_zones;
-	zone = classzone;
-	while (zone >= first_classzone && nr_pages > 0) {
-		if (zone->free_pages <= zone->pages_high) {
-			nr_pages = shrink_zone(zone, priority,
-					gfp_mask, nr_pages);
-		}
-		zone--;
+	for (zone = classzone; zone >= first_classzone; zone--) {
+		int max_scan;
+		int to_reclaim;
+		int unreclaimed;
+
+		to_reclaim = zone->pages_high - zone->free_pages;
+		if (to_reclaim < 0)
+			continue;	/* zone has enough memory */
+
+		if (to_reclaim > SWAP_CLUSTER_MAX)
+			to_reclaim = SWAP_CLUSTER_MAX;
+
+		if (to_reclaim < nr_pages)
+			to_reclaim = nr_pages;
+
+		/*
+		 * If we cannot reclaim `nr_pages' pages by scanning twice
+		 * that many pages then fall back to the next zone.
+		 */
+		max_scan = zone->nr_inactive >> priority;
+		if (max_scan < to_reclaim * 2)
+			max_scan = to_reclaim * 2;
+		unreclaimed = shrink_zone(zone, max_scan, gfp_mask, to_reclaim);
+		nr_pages -= to_reclaim - unreclaimed;
+		*total_scanned += max_scan;
 	}
+
+	shrink_dcache_memory(priority, gfp_mask);
+	shrink_icache_memory(1, gfp_mask);
+#ifdef CONFIG_QUOTA
+	shrink_dqcache_memory(DEF_PRIORITY, gfp_mask);
+#endif
 	return nr_pages;
 }
 
@@ -564,12 +621,25 @@ try_to_free_pages(struct zone *classzone,
 	KERNEL_STAT_INC(pageoutrun);
 
 	for (priority = DEF_PRIORITY; priority; priority--) {
-		nr_pages = shrink_caches(classzone, priority,
+		int total_scanned = 0;
+
+		nr_pages = shrink_caches(classzone, priority, &total_scanned,
 					gfp_mask, nr_pages);
 		if (nr_pages <= 0)
 			return 1;
+		if (total_scanned == 0)
+			return 1;	/* All zones had enough free memory */
 		if (!(gfp_mask & __GFP_FS))
-			break;
+			break;		/* Let the caller handle it */
+		/*
+		 * Try to write back as many pages as we just scanned.  Not
+		 * sure if that makes sense, but it's an attempt to avoid
+		 * creating IO storms unnecessarily
+		 */
+		wakeup_bdflush(total_scanned);
+
+		/* Take a nap, wait for some writeback to complete */
+		blk_congestion_wait(WRITE, HZ/4);
 	}
 	if (gfp_mask & __GFP_FS)
 		out_of_memory();