fs/binfmt_aout.c: fix pointer warnings
[linux-2.6.git] / fs / fs-writeback.c
index 5887328..76fc4d5 100644 (file)
@@ -41,23 +41,24 @@ struct wb_writeback_args {
        long nr_pages;
        struct super_block *sb;
        enum writeback_sync_modes sync_mode;
-       int for_kupdate;
-       int range_cyclic;
+       int for_kupdate:1;
+       int range_cyclic:1;
+       int for_background:1;
 };
 
 /*
  * Work items for the bdi_writeback threads
  */
 struct bdi_work {
-       struct list_head list;
-       struct rcu_head rcu_head;
+       struct list_head list;          /* pending work list */
+       struct rcu_head rcu_head;       /* for RCU free/clear of work */
 
-       unsigned long seen;
-       atomic_t pending;
+       unsigned long seen;             /* threads that have seen this work */
+       atomic_t pending;               /* number of threads still to do work */
 
-       struct wb_writeback_args args;
+       struct wb_writeback_args args;  /* writeback arguments */
 
-       unsigned long state;
+       unsigned long state;            /* flag bits, see WS_* */
 };
 
 enum {
@@ -97,6 +98,11 @@ static void bdi_work_clear(struct bdi_work *work)
 {
        clear_bit(WS_USED_B, &work->state);
        smp_mb__after_clear_bit();
+       /*
+        * work can have disappeared at this point. bit waitq functions
+        * should be able to tolerate this, provided bdi_sched_wait does
+        * not dereference it's pointer argument.
+       */
        wake_up_bit(&work->state, WS_USED_B);
 }
 
@@ -113,6 +119,7 @@ static void bdi_work_free(struct rcu_head *head)
 static void wb_work_complete(struct bdi_work *work)
 {
        const enum writeback_sync_modes sync_mode = work->args.sync_mode;
+       int onstack = bdi_work_on_stack(work);
 
        /*
         * For allocated work, we can clear the done/seen bit right here.
@@ -120,9 +127,9 @@ static void wb_work_complete(struct bdi_work *work)
         * to after the RCU grace period, since the stack could be invalidated
         * as soon as bdi_work_clear() has done the wakeup.
         */
-       if (!bdi_work_on_stack(work))
+       if (!onstack)
                bdi_work_clear(work);
-       if (sync_mode == WB_SYNC_NONE || bdi_work_on_stack(work))
+       if (sync_mode == WB_SYNC_NONE || onstack)
                call_rcu(&work->rcu_head, bdi_work_free);
 }
 
@@ -151,10 +158,10 @@ static void bdi_queue_work(struct backing_dev_info *bdi, struct bdi_work *work)
        BUG_ON(!bdi->wb_cnt);
 
        /*
-        * Make sure stores are seen before it appears on the list
+        * list_add_tail_rcu() contains the necessary barriers to
+        * make sure the above stores are seen before the item is
+        * noticed on the list
         */
-       smp_mb();
-
        spin_lock(&bdi->wb_lock);
        list_add_tail_rcu(&work->list, &bdi->work_list);
        spin_unlock(&bdi->wb_lock);
@@ -168,13 +175,7 @@ static void bdi_queue_work(struct backing_dev_info *bdi, struct bdi_work *work)
        else {
                struct bdi_writeback *wb = &bdi->wb;
 
-               /*
-                * End work now if this wb has no dirty IO pending. Otherwise
-                * wakeup the handling thread
-                */
-               if (!wb_has_dirty_io(wb))
-                       wb_clear_pending(wb, work);
-               else if (wb->task)
+               if (wb->task)
                        wake_up_process(wb->task);
        }
 }
@@ -241,6 +242,7 @@ static void bdi_sync_writeback(struct backing_dev_info *bdi,
 /**
  * bdi_start_writeback - start writeback
  * @bdi: the backing device to write from
+ * @sb: write inodes from this super_block
  * @nr_pages: the number of pages to write
  *
  * Description:
@@ -249,14 +251,25 @@ static void bdi_sync_writeback(struct backing_dev_info *bdi,
  *   completion. Caller need not hold sb s_umount semaphore.
  *
  */
-void bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages)
+void bdi_start_writeback(struct backing_dev_info *bdi, struct super_block *sb,
+                        long nr_pages)
 {
        struct wb_writeback_args args = {
+               .sb             = sb,
                .sync_mode      = WB_SYNC_NONE,
                .nr_pages       = nr_pages,
                .range_cyclic   = 1,
        };
 
+       /*
+        * We treat @nr_pages=0 as the special case to do background writeback,
+        * ie. to sync pages until the background dirty threshold is reached.
+        */
+       if (!nr_pages) {
+               args.nr_pages = LONG_MAX;
+               args.for_background = 1;
+       }
+
        bdi_alloc_queue_work(bdi, &args);
 }
 
@@ -310,7 +323,7 @@ static bool inode_dirtied_after(struct inode *inode, unsigned long t)
         * For inodes being constantly redirtied, dirtied_when can get stuck.
         * It _appears_ to be in the future, but is actually in distant past.
         * This test is necessary to prevent such wrapped-around relative times
-        * from permanently stopping the whole pdflush writeback.
+        * from permanently stopping the whole bdi writeback.
         */
        ret = ret && time_before_eq(inode->dirtied_when, jiffies);
 #endif
@@ -324,13 +337,38 @@ static void move_expired_inodes(struct list_head *delaying_queue,
                               struct list_head *dispatch_queue,
                                unsigned long *older_than_this)
 {
+       LIST_HEAD(tmp);
+       struct list_head *pos, *node;
+       struct super_block *sb = NULL;
+       struct inode *inode;
+       int do_sb_sort = 0;
+
        while (!list_empty(delaying_queue)) {
-               struct inode *inode = list_entry(delaying_queue->prev,
-                                               struct inode, i_list);
+               inode = list_entry(delaying_queue->prev, struct inode, i_list);
                if (older_than_this &&
                    inode_dirtied_after(inode, *older_than_this))
                        break;
-               list_move(&inode->i_list, dispatch_queue);
+               if (sb && sb != inode->i_sb)
+                       do_sb_sort = 1;
+               sb = inode->i_sb;
+               list_move(&inode->i_list, &tmp);
+       }
+
+       /* just one sb in list, splice to dispatch_queue and we're done */
+       if (!do_sb_sort) {
+               list_splice(&tmp, dispatch_queue);
+               return;
+       }
+
+       /* Move inodes from one superblock together */
+       while (!list_empty(&tmp)) {
+               inode = list_entry(tmp.prev, struct inode, i_list);
+               sb = inode->i_sb;
+               list_for_each_prev_safe(pos, node, &tmp) {
+                       inode = list_entry(pos, struct inode, i_list);
+                       if (inode->i_sb == sb)
+                               list_move(&inode->i_list, dispatch_queue);
+               }
        }
 }
 
@@ -343,10 +381,10 @@ static void queue_io(struct bdi_writeback *wb, unsigned long *older_than_this)
        move_expired_inodes(&wb->b_dirty, &wb->b_io, older_than_this);
 }
 
-static int write_inode(struct inode *inode, int sync)
+static int write_inode(struct inode *inode, struct writeback_control *wbc)
 {
        if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode))
-               return inode->i_sb->s_op->write_inode(inode, sync);
+               return inode->i_sb->s_op->write_inode(inode, wbc);
        return 0;
 }
 
@@ -383,7 +421,6 @@ static int
 writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
 {
        struct address_space *mapping = inode->i_mapping;
-       int wait = wbc->sync_mode == WB_SYNC_ALL;
        unsigned dirty;
        int ret;
 
@@ -401,7 +438,7 @@ writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
                 * We'll have another go at writing back this inode when we
                 * completed a full scan of b_io.
                 */
-               if (!wait) {
+               if (wbc->sync_mode != WB_SYNC_ALL) {
                        requeue_io(inode);
                        return 0;
                }
@@ -423,15 +460,20 @@ writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
 
        ret = do_writepages(mapping, wbc);
 
-       /* Don't write the inode if only I_DIRTY_PAGES was set */
-       if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
-               int err = write_inode(inode, wait);
+       /*
+        * Make sure to wait on the data before writing out the metadata.
+        * This is important for filesystems that modify metadata on data
+        * I/O completion.
+        */
+       if (wbc->sync_mode == WB_SYNC_ALL) {
+               int err = filemap_fdatawait(mapping);
                if (ret == 0)
                        ret = err;
        }
 
-       if (wait) {
-               int err = filemap_fdatawait(mapping);
+       /* Don't write the inode if only I_DIRTY_PAGES was set */
+       if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
+               int err = write_inode(inode, wbc);
                if (ret == 0)
                        ret = err;
        }
@@ -439,8 +481,18 @@ writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
        spin_lock(&inode_lock);
        inode->i_state &= ~I_SYNC;
        if (!(inode->i_state & (I_FREEING | I_CLEAR))) {
-               if (!(inode->i_state & I_DIRTY) &&
-                   mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
+               if ((inode->i_state & I_DIRTY_PAGES) && wbc->for_kupdate) {
+                       /*
+                        * More pages get dirtied by a fast dirtier.
+                        */
+                       goto select_queue;
+               } else if (inode->i_state & I_DIRTY) {
+                       /*
+                        * At least XFS will redirty the inode during the
+                        * writeback (delalloc) and on io completion (isize).
+                        */
+                       redirty_tail(inode);
+               } else if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) {
                        /*
                         * We didn't write back all the pages.  nfs_writepages()
                         * sometimes bales out without doing anything. Redirty
@@ -462,6 +514,7 @@ writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
                                 * soon as the queue becomes uncongested.
                                 */
                                inode->i_state |= I_DIRTY_PAGES;
+select_queue:
                                if (wbc->nr_to_write <= 0) {
                                        /*
                                         * slice used up: queue for next turn
@@ -484,12 +537,6 @@ writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
                                inode->i_state |= I_DIRTY_PAGES;
                                redirty_tail(inode);
                        }
-               } else if (inode->i_state & I_DIRTY) {
-                       /*
-                        * Someone redirtied the inode while were writing back
-                        * the pages.
-                        */
-                       redirty_tail(inode);
                } else if (atomic_read(&inode->i_count)) {
                        /*
                         * The inode is clean, inuse
@@ -506,6 +553,17 @@ writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
        return ret;
 }
 
+static void unpin_sb_for_writeback(struct super_block **psb)
+{
+       struct super_block *sb = *psb;
+
+       if (sb) {
+               up_read(&sb->s_umount);
+               put_super(sb);
+               *psb = NULL;
+       }
+}
+
 /*
  * For WB_SYNC_NONE writeback, the caller does not have the sb pinned
  * before calling writeback. So make sure that we do pin it, so it doesn't
@@ -515,11 +573,20 @@ writeback_single_inode(struct inode *inode, struct writeback_control *wbc)
  * 1 if we failed.
  */
 static int pin_sb_for_writeback(struct writeback_control *wbc,
-                                  struct inode *inode)
+                               struct inode *inode, struct super_block **psb)
 {
        struct super_block *sb = inode->i_sb;
 
        /*
+        * If this sb is already pinned, nothing more to do. If not and
+        * *psb is non-NULL, unpin the old one first
+        */
+       if (sb == *psb)
+               return 0;
+       else if (*psb)
+               unpin_sb_for_writeback(psb);
+
+       /*
         * Caller must already hold the ref for this
         */
        if (wbc->sync_mode == WB_SYNC_ALL) {
@@ -532,7 +599,7 @@ static int pin_sb_for_writeback(struct writeback_control *wbc,
        if (down_read_trylock(&sb->s_umount)) {
                if (sb->s_root) {
                        spin_unlock(&sb_lock);
-                       return 0;
+                       goto pinned;
                }
                /*
                 * umounted, drop rwsem again and fall through to failure
@@ -543,25 +610,15 @@ static int pin_sb_for_writeback(struct writeback_control *wbc,
        sb->s_count--;
        spin_unlock(&sb_lock);
        return 1;
-}
-
-static void unpin_sb_for_writeback(struct writeback_control *wbc,
-                                  struct inode *inode)
-{
-       struct super_block *sb = inode->i_sb;
-
-       if (wbc->sync_mode == WB_SYNC_ALL)
-               return;
-
-       up_read(&sb->s_umount);
-       put_super(sb);
+pinned:
+       *psb = sb;
+       return 0;
 }
 
 static void writeback_inodes_wb(struct bdi_writeback *wb,
                                struct writeback_control *wbc)
 {
-       struct super_block *sb = wbc->sb;
-       const int is_blkdev_sb = sb_is_blkdev_sb(sb);
+       struct super_block *sb = wbc->sb, *pin_sb = NULL;
        const unsigned long start = jiffies;    /* livelock avoidance */
 
        spin_lock(&inode_lock);
@@ -582,36 +639,11 @@ static void writeback_inodes_wb(struct bdi_writeback *wb,
                        continue;
                }
 
-               if (!bdi_cap_writeback_dirty(wb->bdi)) {
-                       redirty_tail(inode);
-                       if (is_blkdev_sb) {
-                               /*
-                                * Dirty memory-backed blockdev: the ramdisk
-                                * driver does this.  Skip just this inode
-                                */
-                               continue;
-                       }
-                       /*
-                        * Dirty memory-backed inode against a filesystem other
-                        * than the kernel-internal bdev filesystem.  Skip the
-                        * entire superblock.
-                        */
-                       break;
-               }
-
                if (inode->i_state & (I_NEW | I_WILL_FREE)) {
                        requeue_io(inode);
                        continue;
                }
 
-               if (wbc->nonblocking && bdi_write_congested(wb->bdi)) {
-                       wbc->encountered_congestion = 1;
-                       if (!is_blkdev_sb)
-                               break;          /* Skip a congested fs */
-                       requeue_io(inode);
-                       continue;               /* Skip a congested blockdev */
-               }
-
                /*
                 * Was this inode dirtied after sync_sb_inodes was called?
                 * This keeps sync from extra jobs and livelock.
@@ -619,7 +651,7 @@ static void writeback_inodes_wb(struct bdi_writeback *wb,
                if (inode_dirtied_after(inode, start))
                        break;
 
-               if (pin_sb_for_writeback(wbc, inode)) {
+               if (pin_sb_for_writeback(wbc, inode, &pin_sb)) {
                        requeue_io(inode);
                        continue;
                }
@@ -628,7 +660,6 @@ static void writeback_inodes_wb(struct bdi_writeback *wb,
                __iget(inode);
                pages_skipped = wbc->pages_skipped;
                writeback_single_inode(inode, wbc);
-               unpin_sb_for_writeback(wbc, inode);
                if (wbc->pages_skipped != pages_skipped) {
                        /*
                         * writeback is not making progress due to locked
@@ -648,6 +679,8 @@ static void writeback_inodes_wb(struct bdi_writeback *wb,
                        wbc->more_io = 1;
        }
 
+       unpin_sb_for_writeback(&pin_sb);
+
        spin_unlock(&inode_lock);
        /* Leave any unwritten inodes on b_io */
 }
@@ -702,10 +735,12 @@ static long wb_writeback(struct bdi_writeback *wb,
                .sync_mode              = args->sync_mode,
                .older_than_this        = NULL,
                .for_kupdate            = args->for_kupdate,
+               .for_background         = args->for_background,
                .range_cyclic           = args->range_cyclic,
        };
        unsigned long oldest_jif;
        long wrote = 0;
+       struct inode *inode;
 
        if (wbc.for_kupdate) {
                wbc.older_than_this = &oldest_jif;
@@ -719,24 +754,19 @@ static long wb_writeback(struct bdi_writeback *wb,
 
        for (;;) {
                /*
-                * Don't flush anything for non-integrity writeback where
-                * no nr_pages was given
+                * Stop writeback when nr_pages has been consumed
                 */
-               if (!args->for_kupdate && args->nr_pages <= 0 &&
-                    args->sync_mode == WB_SYNC_NONE)
+               if (args->nr_pages <= 0)
                        break;
 
                /*
-                * If no specific pages were given and this is just a
-                * periodic background writeout and we are below the
-                * background dirty threshold, don't do anything
+                * For background writeout, stop when we are below the
+                * background dirty threshold
                 */
-               if (args->for_kupdate && args->nr_pages <= 0 &&
-                   !over_bground_thresh())
+               if (args->for_background && !over_bground_thresh())
                        break;
 
                wbc.more_io = 0;
-               wbc.encountered_congestion = 0;
                wbc.nr_to_write = MAX_WRITEBACK_PAGES;
                wbc.pages_skipped = 0;
                writeback_inodes_wb(wb, &wbc);
@@ -744,13 +774,32 @@ static long wb_writeback(struct bdi_writeback *wb,
                wrote += MAX_WRITEBACK_PAGES - wbc.nr_to_write;
 
                /*
-                * If we ran out of stuff to write, bail unless more_io got set
+                * If we consumed everything, see if we have more
+                */
+               if (wbc.nr_to_write <= 0)
+                       continue;
+               /*
+                * Didn't write everything and we don't have more IO, bail
                 */
-               if (wbc.nr_to_write > 0 || wbc.pages_skipped > 0) {
-                       if (wbc.more_io && !wbc.for_kupdate)
-                               continue;
+               if (!wbc.more_io)
                        break;
+               /*
+                * Did we write something? Try for more
+                */
+               if (wbc.nr_to_write < MAX_WRITEBACK_PAGES)
+                       continue;
+               /*
+                * Nothing written. Wait for some inode to
+                * become available for writeback. Otherwise
+                * we'll just busyloop.
+                */
+               spin_lock(&inode_lock);
+               if (!list_empty(&wb->b_more_io))  {
+                       inode = list_entry(wb->b_more_io.prev,
+                                               struct inode, i_list);
+                       inode_wait_for_writeback(inode);
                }
+               spin_unlock(&inode_lock);
        }
 
        return wrote;
@@ -758,7 +807,11 @@ static long wb_writeback(struct bdi_writeback *wb,
 
 /*
  * Return the next bdi_work struct that hasn't been processed by this
- * wb thread yet
+ * wb thread yet. ->seen is initially set for each thread that exists
+ * for this device, when a thread first notices a piece of work it
+ * clears its bit. Depending on writeback type, the thread will notify
+ * completion on either receiving the work (WB_SYNC_NONE) or after
+ * it is done (WB_SYNC_ALL).
  */
 static struct bdi_work *get_next_work_item(struct backing_dev_info *bdi,
                                           struct bdi_writeback *wb)
@@ -768,8 +821,9 @@ static struct bdi_work *get_next_work_item(struct backing_dev_info *bdi,
        rcu_read_lock();
 
        list_for_each_entry_rcu(work, &bdi->work_list, list) {
-               if (!test_and_clear_bit(wb->nr, &work->seen))
+               if (!test_bit(wb->nr, &work->seen))
                        continue;
+               clear_bit(wb->nr, &work->seen);
 
                ret = work;
                break;
@@ -880,8 +934,7 @@ int bdi_writeback_task(struct bdi_writeback *wb)
                }
 
                wait_jiffies = msecs_to_jiffies(dirty_writeback_interval * 10);
-               set_current_state(TASK_INTERRUPTIBLE);
-               schedule_timeout(wait_jiffies);
+               schedule_timeout_interruptible(wait_jiffies);
                try_to_freeze();
        }
 
@@ -1056,9 +1109,6 @@ EXPORT_SYMBOL(__mark_inode_dirty);
  * If older_than_this is non-NULL, then only write out inodes which
  * had their first dirtying at a time earlier than *older_than_this.
  *
- * If we're a pdlfush thread, then implement pdflush collision avoidance
- * against the entire list.
- *
  * If `bdi' is non-zero then we're being asked to writeback a specific queue.
  * This function assumes that the blockdev superblock's inodes are backed by
  * a variety of queues, so all inodes are searched.  For other superblocks,
@@ -1137,11 +1187,28 @@ void writeback_inodes_sb(struct super_block *sb)
        nr_to_write = nr_dirty + nr_unstable +
                        (inodes_stat.nr_inodes - inodes_stat.nr_unused);
 
-       bdi_writeback_all(sb, nr_to_write);
+       bdi_start_writeback(sb->s_bdi, sb, nr_to_write);
 }
 EXPORT_SYMBOL(writeback_inodes_sb);
 
 /**
+ * writeback_inodes_sb_if_idle -       start writeback if none underway
+ * @sb: the superblock
+ *
+ * Invoke writeback_inodes_sb if no writeback is currently underway.
+ * Returns 1 if writeback was started, 0 if not.
+ */
+int writeback_inodes_sb_if_idle(struct super_block *sb)
+{
+       if (!writeback_in_progress(sb->s_bdi)) {
+               writeback_inodes_sb(sb);
+               return 1;
+       } else
+               return 0;
+}
+EXPORT_SYMBOL(writeback_inodes_sb_if_idle);
+
+/**
  * sync_inodes_sb      -       sync sb inode pages
  * @sb: the superblock
  *