JFS: clear PAGECACHE_TAG_DIRTY for no-write pages
authorDave Kleikamp <shaggy@linux.vnet.ibm.com>
Thu, 3 Jan 2008 19:09:33 +0000 (13:09 -0600)
committerDave Kleikamp <shaggy@linux.vnet.ibm.com>
Thu, 3 Jan 2008 19:09:33 +0000 (13:09 -0600)
When JFS decides to drop a dirty metapage, it simply clears the META_dirty
bit and leave alone the PG_dirty and PAGECACHE_TAG_DIRTY bits.

When such no-write page goes to metapage_writepage(), the `relic'
PAGECACHE_TAG_DIRTY tag should be cleared, to prevent pdflush from
repeatedly trying to sync them.  This is done through
set_page_writeback(), so call it should be called in all cases.  If
no I/O is initiated, end_page_writeback() should be called immediately.

This is how __block_write_full_page() does things.

Signed-off-by: Dave Kleikamp <shaggy@linux.vnet.ibm.com>
CC: Fengguang Wu <wfg@mail.ustc.edu.cn>
fs/jfs/jfs_metapage.c

index f5cd8d3..b27fa11 100644 (file)
@@ -360,6 +360,7 @@ static int metapage_writepage(struct page *page, struct writeback_control *wbc)
        struct metapage *mp;
        int redirty = 0;
        sector_t lblock;
+       int nr_underway = 0;
        sector_t pblock;
        sector_t next_block = 0;
        sector_t page_start;
@@ -371,6 +372,7 @@ static int metapage_writepage(struct page *page, struct writeback_control *wbc)
                     (PAGE_CACHE_SHIFT - inode->i_blkbits);
        BUG_ON(!PageLocked(page));
        BUG_ON(PageWriteback(page));
+       set_page_writeback(page);
 
        for (offset = 0; offset < PAGE_CACHE_SIZE; offset += PSIZE) {
                mp = page_to_mp(page, offset);
@@ -413,11 +415,10 @@ static int metapage_writepage(struct page *page, struct writeback_control *wbc)
                        if (!bio->bi_size)
                                goto dump_bio;
                        submit_bio(WRITE, bio);
+                       nr_underway++;
                        bio = NULL;
-               } else {
-                       set_page_writeback(page);
+               } else
                        inc_io(page);
-               }
                xlen = (PAGE_CACHE_SIZE - offset) >> inode->i_blkbits;
                pblock = metapage_get_blocks(inode, lblock, &xlen);
                if (!pblock) {
@@ -449,12 +450,16 @@ static int metapage_writepage(struct page *page, struct writeback_control *wbc)
                        goto dump_bio;
 
                submit_bio(WRITE, bio);
+               nr_underway++;
        }
        if (redirty)
                redirty_page_for_writepage(wbc, page);
 
        unlock_page(page);
 
+       if (nr_underway == 0)
+               end_page_writeback(page);
+
        return 0;
 add_failed:
        /* We should never reach here, since we're only adding one vec */