writeback: fix time ordering of the per superblock inode lists 8
[firefly-linux-kernel-4.4.55.git] / fs / fs-writeback.c
index a4b142a6a2c7f4c870a322bf9d3c9d00be1e8a5f..1f22fb5217c0188b90ab6af20b4a7305e7748ca1 100644 (file)
@@ -14,6 +14,7 @@
  */
 
 #include <linux/kernel.h>
+#include <linux/module.h>
 #include <linux/spinlock.h>
 #include <linux/sched.h>
 #include <linux/fs.h>
@@ -118,7 +119,7 @@ void __mark_inode_dirty(struct inode *inode, int flags)
                        goto out;
 
                /*
-                * If the inode was already on s_dirty or s_io, don't
+                * If the inode was already on s_dirty/s_io/s_more_io, don't
                 * reposition it (that would break s_dirty time-ordering).
                 */
                if (!was_dirty) {
@@ -139,6 +140,65 @@ static int write_inode(struct inode *inode, int sync)
        return 0;
 }
 
+/*
+ * Redirty an inode: set its when-it-was dirtied timestamp and move it to the
+ * furthest end of its superblock's dirty-inode list.
+ *
+ * Before stamping the inode's ->dirtied_when, we check to see whether it is
+ * already the most-recently-dirtied inode on the s_dirty list.  If that is
+ * the case then the inode must have been redirtied while it was being written
+ * out and we don't reset its dirtied_when.
+ */
+static void redirty_tail(struct inode *inode)
+{
+       struct super_block *sb = inode->i_sb;
+
+       if (!list_empty(&sb->s_dirty)) {
+               struct inode *tail_inode;
+
+               tail_inode = list_entry(sb->s_dirty.next, struct inode, i_list);
+               if (!time_after_eq(inode->dirtied_when,
+                               tail_inode->dirtied_when))
+                       inode->dirtied_when = jiffies;
+       }
+       list_move(&inode->i_list, &sb->s_dirty);
+}
+
+/*
+ * requeue inode for re-scanning after sb->s_io list is exhausted.
+ */
+static void requeue_io(struct inode *inode)
+{
+       list_move(&inode->i_list, &inode->i_sb->s_more_io);
+}
+
+/*
+ * Move expired dirty inodes from @delaying_queue to @dispatch_queue.
+ */
+static void move_expired_inodes(struct list_head *delaying_queue,
+                              struct list_head *dispatch_queue,
+                               unsigned long *older_than_this)
+{
+       while (!list_empty(delaying_queue)) {
+               struct inode *inode = list_entry(delaying_queue->prev,
+                                               struct inode, i_list);
+               if (older_than_this &&
+                       time_after(inode->dirtied_when, *older_than_this))
+                       break;
+               list_move(&inode->i_list, dispatch_queue);
+       }
+}
+
+/*
+ * Queue all expired dirty inodes for io, eldest first.
+ */
+static void queue_io(struct super_block *sb,
+                               unsigned long *older_than_this)
+{
+       list_splice_init(&sb->s_more_io, sb->s_io.prev);
+       move_expired_inodes(&sb->s_dirty, &sb->s_io, older_than_this);
+}
+
 /*
  * Write a single inode's dirty pages and inode data out to disk.
  * If `wait' is set, wait on the writeout.
@@ -154,7 +214,6 @@ __sync_single_inode(struct inode *inode, struct writeback_control *wbc)
 {
        unsigned dirty;
        struct address_space *mapping = inode->i_mapping;
-       struct super_block *sb = inode->i_sb;
        int wait = wbc->sync_mode == WB_SYNC_ALL;
        int ret;
 
@@ -190,17 +249,25 @@ __sync_single_inode(struct inode *inode, struct writeback_control *wbc)
                        /*
                         * We didn't write back all the pages.  nfs_writepages()
                         * sometimes bales out without doing anything. Redirty
-                        * the inode.  It is still on sb->s_io.
+                        * the inode; Move it from s_io onto s_more_io/s_dirty.
+                        */
+                       /*
+                        * akpm: if the caller was the kupdate function we put
+                        * this inode at the head of s_dirty so it gets first
+                        * consideration.  Otherwise, move it to the tail, for
+                        * the reasons described there.  I'm not really sure
+                        * how much sense this makes.  Presumably I had a good
+                        * reasons for doing it this way, and I'd rather not
+                        * muck with it at present.
                         */
                        if (wbc->for_kupdate) {
                                /*
-                                * For the kupdate function we leave the inode
-                                * at the head of sb_dirty so it will get more
-                                * writeout as soon as the queue becomes
-                                * uncongested.
+                                * For the kupdate function we move the inode
+                                * to s_more_io so it will get more writeout as
+                                * soon as the queue becomes uncongested.
                                 */
                                inode->i_state |= I_DIRTY_PAGES;
-                               list_move_tail(&inode->i_list, &sb->s_dirty);
+                               requeue_io(inode);
                        } else {
                                /*
                                 * Otherwise fully redirty the inode so that
@@ -210,15 +277,14 @@ __sync_single_inode(struct inode *inode, struct writeback_control *wbc)
                                 * all the other files.
                                 */
                                inode->i_state |= I_DIRTY_PAGES;
-                               inode->dirtied_when = jiffies;
-                               list_move(&inode->i_list, &sb->s_dirty);
+                               redirty_tail(inode);
                        }
                } else if (inode->i_state & I_DIRTY) {
                        /*
                         * Someone redirtied the inode while were writing back
                         * the pages.
                         */
-                       list_move(&inode->i_list, &sb->s_dirty);
+                       redirty_tail(inode);
                } else if (atomic_read(&inode->i_count)) {
                        /*
                         * The inode is clean, inuse
@@ -254,7 +320,14 @@ __writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
                struct address_space *mapping = inode->i_mapping;
                int ret;
 
-               list_move(&inode->i_list, &inode->i_sb->s_dirty);
+               /*
+                * We're skipping this inode because it's locked, and we're not
+                * doing writeback-for-data-integrity.  Move it to s_more_io so
+                * that writeback can proceed with the other inodes on s_io.
+                * We'll have another go at writing back this inode when we
+                * completed a full scan of s_io.
+                */
+               requeue_io(inode);
 
                /*
                 * Even if we don't actually write the inode itself here,
@@ -319,7 +392,7 @@ sync_sb_inodes(struct super_block *sb, struct writeback_control *wbc)
        const unsigned long start = jiffies;    /* livelock avoidance */
 
        if (!wbc->for_kupdate || list_empty(&sb->s_io))
-               list_splice_init(&sb->s_dirty, &sb->s_io);
+               queue_io(sb, wbc->older_than_this);
 
        while (!list_empty(&sb->s_io)) {
                struct inode *inode = list_entry(sb->s_io.prev,
@@ -329,7 +402,7 @@ sync_sb_inodes(struct super_block *sb, struct writeback_control *wbc)
                long pages_skipped;
 
                if (!bdi_cap_writeback_dirty(bdi)) {
-                       list_move(&inode->i_list, &sb->s_dirty);
+                       redirty_tail(inode);
                        if (sb_is_blkdev_sb(sb)) {
                                /*
                                 * Dirty memory-backed blockdev: the ramdisk
@@ -349,14 +422,14 @@ sync_sb_inodes(struct super_block *sb, struct writeback_control *wbc)
                        wbc->encountered_congestion = 1;
                        if (!sb_is_blkdev_sb(sb))
                                break;          /* Skip a congested fs */
-                       list_move(&inode->i_list, &sb->s_dirty);
+                       requeue_io(inode);
                        continue;               /* Skip a congested blockdev */
                }
 
                if (wbc->bdi && bdi != wbc->bdi) {
                        if (!sb_is_blkdev_sb(sb))
                                break;          /* fs has the wrong queue */
-                       list_move(&inode->i_list, &sb->s_dirty);
+                       requeue_io(inode);
                        continue;               /* blockdev has wrong queue */
                }
 
@@ -364,11 +437,6 @@ sync_sb_inodes(struct super_block *sb, struct writeback_control *wbc)
                if (time_after(inode->dirtied_when, start))
                        break;
 
-               /* Was this inode dirtied too recently? */
-               if (wbc->older_than_this && time_after(inode->dirtied_when,
-                                               *wbc->older_than_this))
-                       break;
-
                /* Is another pdflush already flushing this queue? */
                if (current_is_pdflush() && !writeback_acquire(bdi))
                        break;
@@ -388,7 +456,7 @@ sync_sb_inodes(struct super_block *sb, struct writeback_control *wbc)
                         * writeback is not making progress due to locked
                         * buffers.  Skip this inode for now.
                         */
-                       list_move(&inode->i_list, &sb->s_dirty);
+                       redirty_tail(inode);
                }
                spin_unlock(&inode_lock);
                iput(inode);
@@ -406,7 +474,7 @@ sync_sb_inodes(struct super_block *sb, struct writeback_control *wbc)
  * Note:
  * We don't need to grab a reference to superblock here. If it has non-empty
  * ->s_dirty it's hadn't been killed yet and kill_super() won't proceed
- * past sync_inodes_sb() until both the ->s_dirty and ->s_io lists are
+ * past sync_inodes_sb() until the ->s_dirty/s_io/s_more_io lists are all
  * empty. Since __sync_single_inode() regains inode_lock before it finally moves
  * inode from superblock lists we are OK.
  *