backing-dev: ensure wakeup_timer is deleted
[linux-2.6.git] / mm / backing-dev.c
index 9c1c199..253b071 100644 (file)
 
 static atomic_long_t bdi_seq = ATOMIC_LONG_INIT(0);
 
-void default_unplug_io_fn(struct backing_dev_info *bdi, struct page *page)
-{
-}
-EXPORT_SYMBOL(default_unplug_io_fn);
-
 struct backing_dev_info default_backing_dev_info = {
        .name           = "default",
        .ra_pages       = VM_MAX_READAHEAD * 1024 / PAGE_CACHE_SIZE,
        .state          = 0,
        .capabilities   = BDI_CAP_MAP_COPY,
-       .unplug_io_fn   = default_unplug_io_fn,
 };
 EXPORT_SYMBOL_GPL(default_backing_dev_info);
 
 struct backing_dev_info noop_backing_dev_info = {
        .name           = "noop",
+       .capabilities   = BDI_CAP_NO_ACCT_AND_WRITEBACK,
 };
 EXPORT_SYMBOL_GPL(noop_backing_dev_info);
 
@@ -50,6 +45,17 @@ static struct timer_list sync_supers_timer;
 static int bdi_sync_supers(void *);
 static void sync_supers_timer_fn(unsigned long);
 
+void bdi_lock_two(struct bdi_writeback *wb1, struct bdi_writeback *wb2)
+{
+       if (wb1 < wb2) {
+               spin_lock(&wb1->list_lock);
+               spin_lock_nested(&wb2->list_lock, 1);
+       } else {
+               spin_lock(&wb2->list_lock);
+               spin_lock_nested(&wb1->list_lock, 1);
+       }
+}
+
 #ifdef CONFIG_DEBUG_FS
 #include <linux/debugfs.h>
 #include <linux/seq_file.h>
@@ -68,37 +74,46 @@ static int bdi_debug_stats_show(struct seq_file *m, void *v)
        unsigned long background_thresh;
        unsigned long dirty_thresh;
        unsigned long bdi_thresh;
-       unsigned long nr_dirty, nr_io, nr_more_io, nr_wb;
+       unsigned long nr_dirty, nr_io, nr_more_io;
        struct inode *inode;
 
-       nr_wb = nr_dirty = nr_io = nr_more_io = 0;
-       spin_lock(&inode_lock);
-       list_for_each_entry(inode, &wb->b_dirty, i_list)
+       nr_dirty = nr_io = nr_more_io = 0;
+       spin_lock(&wb->list_lock);
+       list_for_each_entry(inode, &wb->b_dirty, i_wb_list)
                nr_dirty++;
-       list_for_each_entry(inode, &wb->b_io, i_list)
+       list_for_each_entry(inode, &wb->b_io, i_wb_list)
                nr_io++;
-       list_for_each_entry(inode, &wb->b_more_io, i_list)
+       list_for_each_entry(inode, &wb->b_more_io, i_wb_list)
                nr_more_io++;
-       spin_unlock(&inode_lock);
+       spin_unlock(&wb->list_lock);
 
-       get_dirty_limits(&background_thresh, &dirty_thresh, &bdi_thresh, bdi);
+       global_dirty_limits(&background_thresh, &dirty_thresh);
+       bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh);
 
 #define K(x) ((x) << (PAGE_SHIFT - 10))
        seq_printf(m,
-                  "BdiWriteback:     %8lu kB\n"
-                  "BdiReclaimable:   %8lu kB\n"
-                  "BdiDirtyThresh:   %8lu kB\n"
-                  "DirtyThresh:      %8lu kB\n"
-                  "BackgroundThresh: %8lu kB\n"
-                  "b_dirty:          %8lu\n"
-                  "b_io:             %8lu\n"
-                  "b_more_io:        %8lu\n"
-                  "bdi_list:         %8u\n"
-                  "state:            %8lx\n",
+                  "BdiWriteback:       %10lu kB\n"
+                  "BdiReclaimable:     %10lu kB\n"
+                  "BdiDirtyThresh:     %10lu kB\n"
+                  "DirtyThresh:        %10lu kB\n"
+                  "BackgroundThresh:   %10lu kB\n"
+                  "BdiWritten:         %10lu kB\n"
+                  "BdiWriteBandwidth:  %10lu kBps\n"
+                  "b_dirty:            %10lu\n"
+                  "b_io:               %10lu\n"
+                  "b_more_io:          %10lu\n"
+                  "bdi_list:           %10u\n"
+                  "state:              %10lx\n",
                   (unsigned long) K(bdi_stat(bdi, BDI_WRITEBACK)),
                   (unsigned long) K(bdi_stat(bdi, BDI_RECLAIMABLE)),
-                  K(bdi_thresh), K(dirty_thresh),
-                  K(background_thresh), nr_dirty, nr_io, nr_more_io,
+                  K(bdi_thresh),
+                  K(dirty_thresh),
+                  K(background_thresh),
+                  (unsigned long) K(bdi_stat(bdi, BDI_WRITTEN)),
+                  (unsigned long) K(bdi->write_bandwidth),
+                  nr_dirty,
+                  nr_io,
+                  nr_more_io,
                   !list_empty(&bdi->bdi_list), bdi->state);
 #undef K
 
@@ -236,46 +251,23 @@ static int __init default_bdi_init(void)
        sync_supers_tsk = kthread_run(bdi_sync_supers, NULL, "sync_supers");
        BUG_ON(IS_ERR(sync_supers_tsk));
 
-       init_timer(&sync_supers_timer);
        setup_timer(&sync_supers_timer, sync_supers_timer_fn, 0);
        bdi_arm_supers_timer();
 
        err = bdi_init(&default_backing_dev_info);
        if (!err)
                bdi_register(&default_backing_dev_info, NULL, "default");
+       err = bdi_init(&noop_backing_dev_info);
 
        return err;
 }
 subsys_initcall(default_bdi_init);
 
-static void bdi_wb_init(struct bdi_writeback *wb, struct backing_dev_info *bdi)
-{
-       memset(wb, 0, sizeof(*wb));
-
-       wb->bdi = bdi;
-       wb->last_old_flush = jiffies;
-       INIT_LIST_HEAD(&wb->b_dirty);
-       INIT_LIST_HEAD(&wb->b_io);
-       INIT_LIST_HEAD(&wb->b_more_io);
-}
-
 int bdi_has_dirty_io(struct backing_dev_info *bdi)
 {
        return wb_has_dirty_io(&bdi->wb);
 }
 
-static void bdi_flush_io(struct backing_dev_info *bdi)
-{
-       struct writeback_control wbc = {
-               .sync_mode              = WB_SYNC_NONE,
-               .older_than_this        = NULL,
-               .range_cyclic           = 1,
-               .nr_to_write            = 1024,
-       };
-
-       writeback_inodes_wb(&bdi->wb, &wbc);
-}
-
 /*
  * kupdated() used to do this. We cannot do it from the bdi_forker_thread()
  * or we risk deadlocking on ->s_umount. The longer term solution would be
@@ -316,6 +308,45 @@ static void sync_supers_timer_fn(unsigned long unused)
        bdi_arm_supers_timer();
 }
 
+static void wakeup_timer_fn(unsigned long data)
+{
+       struct backing_dev_info *bdi = (struct backing_dev_info *)data;
+
+       spin_lock_bh(&bdi->wb_lock);
+       if (bdi->wb.task) {
+               trace_writeback_wake_thread(bdi);
+               wake_up_process(bdi->wb.task);
+       } else {
+               /*
+                * When bdi tasks are inactive for long time, they are killed.
+                * In this case we have to wake-up the forker thread which
+                * should create and run the bdi thread.
+                */
+               trace_writeback_wake_forker_thread(bdi);
+               wake_up_process(default_backing_dev_info.wb.task);
+       }
+       spin_unlock_bh(&bdi->wb_lock);
+}
+
+/*
+ * This function is used when the first inode for this bdi is marked dirty. It
+ * wakes-up the corresponding bdi thread which should then take care of the
+ * periodic background write-out of dirty inodes. Since the write-out would
+ * starts only 'dirty_writeback_interval' centisecs from now anyway, we just
+ * set up a timer which wakes the bdi thread up later.
+ *
+ * Note, we wouldn't bother setting up the timer, but this function is on the
+ * fast-path (used by '__mark_inode_dirty()'), so we save few context switches
+ * by delaying the wake-up.
+ */
+void bdi_wakeup_thread_delayed(struct backing_dev_info *bdi)
+{
+       unsigned long timeout;
+
+       timeout = msecs_to_jiffies(dirty_writeback_interval * 10);
+       mod_timer(&bdi->wb.wakeup_timer, jiffies + timeout);
+}
+
 /*
  * Calculate the longest interval (jiffies) bdi threads are allowed to be
  * inactive.
@@ -328,11 +359,22 @@ static unsigned long bdi_longest_inactive(void)
        return max(5UL * 60 * HZ, interval);
 }
 
+/*
+ * Clear pending bit and wakeup anybody waiting for flusher thread creation or
+ * shutdown
+ */
+static void bdi_clear_pending(struct backing_dev_info *bdi)
+{
+       clear_bit(BDI_pending, &bdi->state);
+       smp_mb__after_clear_bit();
+       wake_up_bit(&bdi->state, BDI_pending);
+}
+
 static int bdi_forker_thread(void *ptr)
 {
        struct bdi_writeback *me = ptr;
 
-       current->flags |= PF_FLUSHER | PF_SWAPWRITE;
+       current->flags |= PF_SWAPWRITE;
        set_freezable();
 
        /*
@@ -353,10 +395,19 @@ static int bdi_forker_thread(void *ptr)
                 * Temporary measure, we want to make sure we don't see
                 * dirty data on the default backing_dev_info
                 */
-               if (wb_has_dirty_io(me) || !list_empty(&me->bdi->work_list))
+               if (wb_has_dirty_io(me) || !list_empty(&me->bdi->work_list)) {
+                       del_timer(&me->wakeup_timer);
                        wb_do_writeback(me, 0);
+               }
 
                spin_lock_bh(&bdi_lock);
+               /*
+                * In the following loop we are going to check whether we have
+                * some work to do without any synchronization with tasks
+                * waking us up to do work for them. So we have to set task
+                * state already here so that we don't miss wakeups coming
+                * after we verify some condition.
+                */
                set_current_state(TASK_INTERRUPTIBLE);
 
                list_for_each_entry(bdi, &bdi_list, bdi_list) {
@@ -387,6 +438,7 @@ static int bdi_forker_thread(void *ptr)
                        }
 
                        spin_lock(&bdi->wb_lock);
+
                        /*
                         * If there is no work to do and the bdi thread was
                         * inactive long enough - kill it. The wb_lock is taken
@@ -414,46 +466,51 @@ static int bdi_forker_thread(void *ptr)
                switch (action) {
                case FORK_THREAD:
                        __set_current_state(TASK_RUNNING);
-                       task = kthread_run(bdi_writeback_thread, &bdi->wb, "flush-%s",
-                                          dev_name(bdi->dev));
+                       task = kthread_create(bdi_writeback_thread, &bdi->wb,
+                                             "flush-%s", dev_name(bdi->dev));
                        if (IS_ERR(task)) {
                                /*
                                 * If thread creation fails, force writeout of
-                                * the bdi from the thread.
+                                * the bdi from the thread. Hopefully 1024 is
+                                * large enough for efficient IO.
                                 */
-                               bdi_flush_io(bdi);
+                               writeback_inodes_wb(&bdi->wb, 1024);
                        } else {
                                /*
                                 * The spinlock makes sure we do not lose
                                 * wake-ups when racing with 'bdi_queue_work()'.
+                                * And as soon as the bdi thread is visible, we
+                                * can start it.
                                 */
-                               spin_lock(&bdi->wb_lock);
+                               spin_lock_bh(&bdi->wb_lock);
                                bdi->wb.task = task;
-                               spin_unlock(&bdi->wb_lock);
+                               spin_unlock_bh(&bdi->wb_lock);
+                               wake_up_process(task);
                        }
+                       bdi_clear_pending(bdi);
                        break;
 
                case KILL_THREAD:
                        __set_current_state(TASK_RUNNING);
                        kthread_stop(task);
+                       bdi_clear_pending(bdi);
                        break;
 
                case NO_ACTION:
-                       if (dirty_writeback_interval)
-                               schedule_timeout(msecs_to_jiffies(dirty_writeback_interval * 10));
+                       if (!wb_has_dirty_io(me) || !dirty_writeback_interval)
+                               /*
+                                * There are no dirty data. The only thing we
+                                * should now care about is checking for
+                                * inactive bdi threads and killing them. Thus,
+                                * let's sleep for longer time, save energy and
+                                * be friendly for battery-driven devices.
+                                */
+                               schedule_timeout(bdi_longest_inactive());
                        else
-                               schedule();
+                               schedule_timeout(msecs_to_jiffies(dirty_writeback_interval * 10));
                        try_to_freeze();
-                       /* Back to the main loop */
-                       continue;
+                       break;
                }
-
-               /*
-                * Clear pending bit and wakeup anybody waiting to tear us down.
-                */
-               clear_bit(BDI_pending, &bdi->state);
-               smp_mb__after_clear_bit();
-               wake_up_bit(&bdi->state, BDI_pending);
        }
 
        return 0;
@@ -468,30 +525,23 @@ static void bdi_remove_from_list(struct backing_dev_info *bdi)
        list_del_rcu(&bdi->bdi_list);
        spin_unlock_bh(&bdi_lock);
 
-       synchronize_rcu();
+       synchronize_rcu_expedited();
 }
 
 int bdi_register(struct backing_dev_info *bdi, struct device *parent,
                const char *fmt, ...)
 {
        va_list args;
-       int ret = 0;
        struct device *dev;
 
        if (bdi->dev)   /* The driver needs to use separate queues per device */
-               goto exit;
+               return 0;
 
        va_start(args, fmt);
        dev = device_create_vargs(bdi_class, parent, MKDEV(0, 0), bdi, fmt, args);
        va_end(args);
-       if (IS_ERR(dev)) {
-               ret = PTR_ERR(dev);
-               goto exit;
-       }
-
-       spin_lock_bh(&bdi_lock);
-       list_add_tail_rcu(&bdi->bdi_list, &bdi_list);
-       spin_unlock_bh(&bdi_lock);
+       if (IS_ERR(dev))
+               return PTR_ERR(dev);
 
        bdi->dev = dev;
 
@@ -505,20 +555,19 @@ int bdi_register(struct backing_dev_info *bdi, struct device *parent,
 
                wb->task = kthread_run(bdi_forker_thread, wb, "bdi-%s",
                                                dev_name(dev));
-               if (IS_ERR(wb->task)) {
-                       wb->task = NULL;
-                       ret = -ENOMEM;
-
-                       bdi_remove_from_list(bdi);
-                       goto exit;
-               }
+               if (IS_ERR(wb->task))
+                       return PTR_ERR(wb->task);
        }
 
        bdi_debug_register(bdi, dev_name(dev));
        set_bit(BDI_registered, &bdi->state);
+
+       spin_lock_bh(&bdi_lock);
+       list_add_tail_rcu(&bdi->bdi_list, &bdi_list);
+       spin_unlock_bh(&bdi_lock);
+
        trace_writeback_bdi_register(bdi);
-exit:
-       return ret;
+       return 0;
 }
 EXPORT_SYMBOL(bdi_register);
 
@@ -569,7 +618,7 @@ static void bdi_prune_sb(struct backing_dev_info *bdi)
        spin_lock(&sb_lock);
        list_for_each_entry(sb, &super_blocks, s_list) {
                if (sb->s_bdi == bdi)
-                       sb->s_bdi = NULL;
+                       sb->s_bdi = &default_backing_dev_info;
        }
        spin_unlock(&sb_lock);
 }
@@ -577,8 +626,10 @@ static void bdi_prune_sb(struct backing_dev_info *bdi)
 void bdi_unregister(struct backing_dev_info *bdi)
 {
        if (bdi->dev) {
+               bdi_set_min_ratio(bdi, 0);
                trace_writeback_bdi_unregister(bdi);
                bdi_prune_sb(bdi);
+               del_timer_sync(&bdi->wb.wakeup_timer);
 
                if (!bdi_cap_flush_forker(bdi))
                        bdi_wb_shutdown(bdi);
@@ -589,6 +640,24 @@ void bdi_unregister(struct backing_dev_info *bdi)
 }
 EXPORT_SYMBOL(bdi_unregister);
 
+static void bdi_wb_init(struct bdi_writeback *wb, struct backing_dev_info *bdi)
+{
+       memset(wb, 0, sizeof(*wb));
+
+       wb->bdi = bdi;
+       wb->last_old_flush = jiffies;
+       INIT_LIST_HEAD(&wb->b_dirty);
+       INIT_LIST_HEAD(&wb->b_io);
+       INIT_LIST_HEAD(&wb->b_more_io);
+       spin_lock_init(&wb->list_lock);
+       setup_timer(&wb->wakeup_timer, wakeup_timer_fn, (unsigned long)bdi);
+}
+
+/*
+ * Initial write bandwidth: 100 MB/s
+ */
+#define INIT_BW                (100 << (20 - PAGE_SHIFT))
+
 int bdi_init(struct backing_dev_info *bdi)
 {
        int i, err;
@@ -611,6 +680,13 @@ int bdi_init(struct backing_dev_info *bdi)
        }
 
        bdi->dirty_exceeded = 0;
+
+       bdi->bw_time_stamp = jiffies;
+       bdi->written_stamp = 0;
+
+       bdi->write_bandwidth = INIT_BW;
+       bdi->avg_write_bandwidth = INIT_BW;
+
        err = prop_local_init_percpu(&bdi->completions);
 
        if (err) {
@@ -634,15 +710,24 @@ void bdi_destroy(struct backing_dev_info *bdi)
        if (bdi_has_dirty_io(bdi)) {
                struct bdi_writeback *dst = &default_backing_dev_info.wb;
 
-               spin_lock(&inode_lock);
+               bdi_lock_two(&bdi->wb, dst);
                list_splice(&bdi->wb.b_dirty, &dst->b_dirty);
                list_splice(&bdi->wb.b_io, &dst->b_io);
                list_splice(&bdi->wb.b_more_io, &dst->b_more_io);
-               spin_unlock(&inode_lock);
+               spin_unlock(&bdi->wb.list_lock);
+               spin_unlock(&dst->list_lock);
        }
 
        bdi_unregister(bdi);
 
+       /*
+        * If bdi_unregister() had already been called earlier, the
+        * wakeup_timer could still be armed because bdi_prune_sb()
+        * can race with the bdi_wakeup_thread_delayed() calls from
+        * __mark_inode_dirty().
+        */
+       del_timer_sync(&bdi->wb.wakeup_timer);
+
        for (i = 0; i < NR_BDI_STAT_ITEMS; i++)
                percpu_counter_destroy(&bdi->bdi_stat[i]);
 
@@ -681,6 +766,7 @@ static wait_queue_head_t congestion_wqh[2] = {
                __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[0]),
                __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[1])
        };
+static atomic_t nr_bdi_congested[2];
 
 void clear_bdi_congested(struct backing_dev_info *bdi, int sync)
 {
@@ -688,7 +774,8 @@ void clear_bdi_congested(struct backing_dev_info *bdi, int sync)
        wait_queue_head_t *wqh = &congestion_wqh[sync];
 
        bit = sync ? BDI_sync_congested : BDI_async_congested;
-       clear_bit(bit, &bdi->state);
+       if (test_and_clear_bit(bit, &bdi->state))
+               atomic_dec(&nr_bdi_congested[sync]);
        smp_mb__after_clear_bit();
        if (waitqueue_active(wqh))
                wake_up(wqh);
@@ -700,7 +787,8 @@ void set_bdi_congested(struct backing_dev_info *bdi, int sync)
        enum bdi_state bit;
 
        bit = sync ? BDI_sync_congested : BDI_async_congested;
-       set_bit(bit, &bdi->state);
+       if (!test_and_set_bit(bit, &bdi->state))
+               atomic_inc(&nr_bdi_congested[sync]);
 }
 EXPORT_SYMBOL(set_bdi_congested);
 
@@ -716,13 +804,72 @@ EXPORT_SYMBOL(set_bdi_congested);
 long congestion_wait(int sync, long timeout)
 {
        long ret;
+       unsigned long start = jiffies;
        DEFINE_WAIT(wait);
        wait_queue_head_t *wqh = &congestion_wqh[sync];
 
        prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
        ret = io_schedule_timeout(timeout);
        finish_wait(wqh, &wait);
+
+       trace_writeback_congestion_wait(jiffies_to_usecs(timeout),
+                                       jiffies_to_usecs(jiffies - start));
+
        return ret;
 }
 EXPORT_SYMBOL(congestion_wait);
 
+/**
+ * wait_iff_congested - Conditionally wait for a backing_dev to become uncongested or a zone to complete writes
+ * @zone: A zone to check if it is heavily congested
+ * @sync: SYNC or ASYNC IO
+ * @timeout: timeout in jiffies
+ *
+ * In the event of a congested backing_dev (any backing_dev) and the given
+ * @zone has experienced recent congestion, this waits for up to @timeout
+ * jiffies for either a BDI to exit congestion of the given @sync queue
+ * or a write to complete.
+ *
+ * In the absence of zone congestion, cond_resched() is called to yield
+ * the processor if necessary but otherwise does not sleep.
+ *
+ * The return value is 0 if the sleep is for the full timeout. Otherwise,
+ * it is the number of jiffies that were still remaining when the function
+ * returned. return_value == timeout implies the function did not sleep.
+ */
+long wait_iff_congested(struct zone *zone, int sync, long timeout)
+{
+       long ret;
+       unsigned long start = jiffies;
+       DEFINE_WAIT(wait);
+       wait_queue_head_t *wqh = &congestion_wqh[sync];
+
+       /*
+        * If there is no congestion, or heavy congestion is not being
+        * encountered in the current zone, yield if necessary instead
+        * of sleeping on the congestion queue
+        */
+       if (atomic_read(&nr_bdi_congested[sync]) == 0 ||
+                       !zone_is_reclaim_congested(zone)) {
+               cond_resched();
+
+               /* In case we scheduled, work out time remaining */
+               ret = timeout - (jiffies - start);
+               if (ret < 0)
+                       ret = 0;
+
+               goto out;
+       }
+
+       /* Sleep until uncongested or a write happens */
+       prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
+       ret = io_schedule_timeout(timeout);
+       finish_wait(wqh, &wait);
+
+out:
+       trace_writeback_wait_iff_congested(jiffies_to_usecs(timeout),
+                                       jiffies_to_usecs(jiffies - start));
+
+       return ret;
+}
+EXPORT_SYMBOL(wait_iff_congested);