[PATCH] vfs: reduce stack usage in write_cache_pages()

[PATCH] vfs: reduce stack usage in write_cache_pages()

[Denys Vlasenko]

[-- Attachment #1: Type: text/plain, Size: 456 bytes --]

Hi Al, David,

vfs function write_cache_pages() appears on the xfs
writeout path. It's *the* path which causes 4k stack
overflows on i386 with xfs.

This function uses ~100 bytes of stack on 32-bit i386.

This patch transforms it a little so that it uses
a bit less stack - minus 8 bytes to be precise.

This isn't much, but it helps not only xfs,
but all filesystems.

Only compile tested.

Signed-off-by: Denys Vlasenko <vda.linux@googlemail.com>
--
vda

[-- Attachment #2: stk3.diff --]
[-- Type: text/x-diff, Size: 2735 bytes --]

diff -urpN linux-2.6-xfs1/mm/page-writeback.c linux-2.6-xfs1.stk3/mm/page-writeback.c
--- linux-2.6-xfs1/mm/page-writeback.c	2008-03-30 03:27:55.000000000 +0200
+++ linux-2.6-xfs1.stk3/mm/page-writeback.c	2008-04-27 04:14:39.000000000 +0200
@@ -798,17 +798,14 @@ int write_cache_pages(struct address_spa
 		      struct writeback_control *wbc, writepage_t writepage,
 		      void *data)
 {
-	struct backing_dev_info *bdi = mapping->backing_dev_info;
 	int ret = 0;
 	int done = 0;
 	struct pagevec pvec;
-	int nr_pages;
 	pgoff_t index;
 	pgoff_t end;		/* Inclusive */
 	int scanned = 0;
-	int range_whole = 0;
 
-	if (wbc->nonblocking && bdi_write_congested(bdi)) {
+	if (wbc->nonblocking && bdi_write_congested(mapping->backing_dev_info)) {
 		wbc->encountered_congestion = 1;
 		return 0;
 	}
@@ -820,20 +817,30 @@ int write_cache_pages(struct address_spa
 	} else {
 		index = wbc->range_start >> PAGE_CACHE_SHIFT;
 		end = wbc->range_end >> PAGE_CACHE_SHIFT;
-		if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
-			range_whole = 1;
 		scanned = 1;
 	}
+
+	/* Minimizing stack use:
+	 * "nr_pages" hopefully won't require a stack slot -
+	 * we reuse "scanned" to keep its value.
+	 */
+
 retry:
-	while (!done && (index <= end) &&
-	       (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
-					      PAGECACHE_TAG_DIRTY,
-					      min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
+	while (!done && (index <= end)) {
 		unsigned i;
+		int nr_pages;
 
-		scanned = 1;
-		for (i = 0; i < nr_pages; i++) {
-			struct page *page = pvec.pages[i];
+		nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
+					      PAGECACHE_TAG_DIRTY,
+					      min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1);
+		if (!nr_pages)
+			break;
+		scanned = nr_pages;
+		/* "scanned" counts down to 1, and later acts as nonzero flag */
+
+		i = (unsigned)-1;
+		while (1) { /* for i in [0..nr_pages-1] */
+			struct page *page = pvec.pages[++i];
 
 			/*
 			 * At this point we hold neither mapping->tree_lock nor
@@ -872,10 +879,13 @@ retry:
 			}
 			if (ret || (--(wbc->nr_to_write) <= 0))
 				done = 1;
-			if (wbc->nonblocking && bdi_write_congested(bdi)) {
+			if (wbc->nonblocking && bdi_write_congested(mapping->backing_dev_info)) {
 				wbc->encountered_congestion = 1;
 				done = 1;
 			}
+			if (scanned == 1)
+				break;
+			scanned--;
 		}
 		pagevec_release(&pvec);
 		cond_resched();
@@ -889,7 +899,8 @@ retry:
 		index = 0;
 		goto retry;
 	}
-	if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
+	if (wbc->range_cyclic || (wbc->range_start == 0 && wbc->range_end == LLONG_MAX /* whole range */
+	                         && wbc->nr_to_write > 0))
 		mapping->writeback_index = index;
 	return ret;
 }