Revert "writeback: do not sync data dirtied after sync start"
authorJan Kara <jack@suse.cz>
Fri, 21 Feb 2014 10:19:04 +0000 (11:19 +0100)
committerJan Kara <jack@suse.cz>
Sat, 22 Feb 2014 01:02:28 +0000 (02:02 +0100)
This reverts commit c4a391b53a72d2df4ee97f96f78c1d5971b47489. Dave
Chinner <david@fromorbit.com> has reported the commit may cause some
inodes to be left out from sync(2). This is because we can call
redirty_tail() for some inode (which sets i_dirtied_when to current time)
after sync(2) has started or similarly requeue_inode() can set
i_dirtied_when to current time if writeback had to skip some pages. The
real problem is in the functions clobbering i_dirtied_when but fixing
that isn't trivial so revert is a safer choice for now.

CC: stable@vger.kernel.org # >= 3.13
Signed-off-by: Jan Kara <jack@suse.cz>
fs/fs-writeback.c
fs/sync.c
fs/xfs/xfs_super.c
include/linux/writeback.h
include/trace/events/writeback.h

index e0259a163f98e69000c28bdb78fcf2b77c2bed2d..d754e3cf99a85af0c433e71846db6ee051751ebd 100644 (file)
 struct wb_writeback_work {
        long nr_pages;
        struct super_block *sb;
-       /*
-        * Write only inodes dirtied before this time. Don't forget to set
-        * older_than_this_is_set when you set this.
-        */
-       unsigned long older_than_this;
+       unsigned long *older_than_this;
        enum writeback_sync_modes sync_mode;
        unsigned int tagged_writepages:1;
        unsigned int for_kupdate:1;
        unsigned int range_cyclic:1;
        unsigned int for_background:1;
        unsigned int for_sync:1;        /* sync(2) WB_SYNC_ALL writeback */
-       unsigned int older_than_this_is_set:1;
        enum wb_reason reason;          /* why was writeback initiated? */
 
        struct list_head list;          /* pending work list */
@@ -252,10 +247,10 @@ static int move_expired_inodes(struct list_head *delaying_queue,
        int do_sb_sort = 0;
        int moved = 0;
 
-       WARN_ON_ONCE(!work->older_than_this_is_set);
        while (!list_empty(delaying_queue)) {
                inode = wb_inode(delaying_queue->prev);
-               if (inode_dirtied_after(inode, work->older_than_this))
+               if (work->older_than_this &&
+                   inode_dirtied_after(inode, *work->older_than_this))
                        break;
                list_move(&inode->i_wb_list, &tmp);
                moved++;
@@ -742,8 +737,6 @@ static long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages,
                .sync_mode      = WB_SYNC_NONE,
                .range_cyclic   = 1,
                .reason         = reason,
-               .older_than_this = jiffies,
-               .older_than_this_is_set = 1,
        };
 
        spin_lock(&wb->list_lock);
@@ -802,13 +795,12 @@ static long wb_writeback(struct bdi_writeback *wb,
 {
        unsigned long wb_start = jiffies;
        long nr_pages = work->nr_pages;
+       unsigned long oldest_jif;
        struct inode *inode;
        long progress;
 
-       if (!work->older_than_this_is_set) {
-               work->older_than_this = jiffies;
-               work->older_than_this_is_set = 1;
-       }
+       oldest_jif = jiffies;
+       work->older_than_this = &oldest_jif;
 
        spin_lock(&wb->list_lock);
        for (;;) {
@@ -842,10 +834,10 @@ static long wb_writeback(struct bdi_writeback *wb,
                 * safe.
                 */
                if (work->for_kupdate) {
-                       work->older_than_this = jiffies -
+                       oldest_jif = jiffies -
                                msecs_to_jiffies(dirty_expire_interval * 10);
                } else if (work->for_background)
-                       work->older_than_this = jiffies;
+                       oldest_jif = jiffies;
 
                trace_writeback_start(wb->bdi, work);
                if (list_empty(&wb->b_io))
@@ -1357,21 +1349,18 @@ EXPORT_SYMBOL(try_to_writeback_inodes_sb);
 
 /**
  * sync_inodes_sb      -       sync sb inode pages
- * @sb:                        the superblock
- * @older_than_this:   timestamp
+ * @sb: the superblock
  *
  * This function writes and waits on any dirty inode belonging to this
- * superblock that has been dirtied before given timestamp.
+ * super_block.
  */
-void sync_inodes_sb(struct super_block *sb, unsigned long older_than_this)
+void sync_inodes_sb(struct super_block *sb)
 {
        DECLARE_COMPLETION_ONSTACK(done);
        struct wb_writeback_work work = {
                .sb             = sb,
                .sync_mode      = WB_SYNC_ALL,
                .nr_pages       = LONG_MAX,
-               .older_than_this = older_than_this,
-               .older_than_this_is_set = 1,
                .range_cyclic   = 0,
                .done           = &done,
                .reason         = WB_REASON_SYNC,
index e8ba024a055b5a55d6320ba38edf1338c7d90e65..b28d1dd10e8b70194a604a1fca654237edd817be 100644 (file)
--- a/fs/sync.c
+++ b/fs/sync.c
  * wait == 1 case since in that case write_inode() functions do
  * sync_dirty_buffer() and thus effectively write one block at a time.
  */
-static int __sync_filesystem(struct super_block *sb, int wait,
-                            unsigned long start)
+static int __sync_filesystem(struct super_block *sb, int wait)
 {
        if (wait)
-               sync_inodes_sb(sb, start);
+               sync_inodes_sb(sb);
        else
                writeback_inodes_sb(sb, WB_REASON_SYNC);
 
@@ -48,7 +47,6 @@ static int __sync_filesystem(struct super_block *sb, int wait,
 int sync_filesystem(struct super_block *sb)
 {
        int ret;
-       unsigned long start = jiffies;
 
        /*
         * We need to be protected against the filesystem going from
@@ -62,17 +60,17 @@ int sync_filesystem(struct super_block *sb)
        if (sb->s_flags & MS_RDONLY)
                return 0;
 
-       ret = __sync_filesystem(sb, 0, start);
+       ret = __sync_filesystem(sb, 0);
        if (ret < 0)
                return ret;
-       return __sync_filesystem(sb, 1, start);
+       return __sync_filesystem(sb, 1);
 }
 EXPORT_SYMBOL_GPL(sync_filesystem);
 
 static void sync_inodes_one_sb(struct super_block *sb, void *arg)
 {
        if (!(sb->s_flags & MS_RDONLY))
-               sync_inodes_sb(sb, *((unsigned long *)arg));
+               sync_inodes_sb(sb);
 }
 
 static void sync_fs_one_sb(struct super_block *sb, void *arg)
@@ -104,10 +102,9 @@ static void fdatawait_one_bdev(struct block_device *bdev, void *arg)
 SYSCALL_DEFINE0(sync)
 {
        int nowait = 0, wait = 1;
-       unsigned long start = jiffies;
 
        wakeup_flusher_threads(0, WB_REASON_SYNC);
-       iterate_supers(sync_inodes_one_sb, &start);
+       iterate_supers(sync_inodes_one_sb, NULL);
        iterate_supers(sync_fs_one_sb, &nowait);
        iterate_supers(sync_fs_one_sb, &wait);
        iterate_bdevs(fdatawrite_one_bdev, NULL);
index f317488263dd975eb85eb44d9746abfb974ef46f..d971f4932b5d8bb92d5efe645f372462abfbbe81 100644 (file)
@@ -913,7 +913,7 @@ xfs_flush_inodes(
        struct super_block      *sb = mp->m_super;
 
        if (down_read_trylock(&sb->s_umount)) {
-               sync_inodes_sb(sb, jiffies);
+               sync_inodes_sb(sb);
                up_read(&sb->s_umount);
        }
 }
index fc0e4320aa6d50d0e5f966e3f8b649506d461ac7..021b8a319b9e2cf7f0f60a5f6fedcfe3367f8016 100644 (file)
@@ -97,7 +97,7 @@ void writeback_inodes_sb_nr(struct super_block *, unsigned long nr,
 int try_to_writeback_inodes_sb(struct super_block *, enum wb_reason reason);
 int try_to_writeback_inodes_sb_nr(struct super_block *, unsigned long nr,
                                  enum wb_reason reason);
-void sync_inodes_sb(struct super_block *sb, unsigned long older_than_this);
+void sync_inodes_sb(struct super_block *);
 void wakeup_flusher_threads(long nr_pages, enum wb_reason reason);
 void inode_wait_for_writeback(struct inode *inode);
 
index c7bbbe794e65cdd0a41c7235bcee6e9718970876..464ea82e10dbf1f1a519b75c52f9e129a5a715e0 100644 (file)
@@ -287,11 +287,11 @@ TRACE_EVENT(writeback_queue_io,
                __field(int,            reason)
        ),
        TP_fast_assign(
-               unsigned long older_than_this = work->older_than_this;
+               unsigned long *older_than_this = work->older_than_this;
                strncpy(__entry->name, dev_name(wb->bdi->dev), 32);
-               __entry->older  = older_than_this;
+               __entry->older  = older_than_this ?  *older_than_this : 0;
                __entry->age    = older_than_this ?
-                                 (jiffies - older_than_this) * 1000 / HZ : -1;
+                                 (jiffies - *older_than_this) * 1000 / HZ : -1;
                __entry->moved  = moved;
                __entry->reason = work->reason;
        ),