writeback: make bdi_has_dirty_io() take multiple bdi_writeback's into account
authorTejun Heo <tj@kernel.org>
Fri, 22 May 2015 21:13:47 +0000 (17:13 -0400)
committerJens Axboe <axboe@fb.com>
Tue, 2 Jun 2015 14:33:36 +0000 (08:33 -0600)
bdi_has_dirty_io() used to only reflect whether the root wb
(bdi_writeback) has dirty inodes.  For cgroup writeback support, it
needs to take all active wb's into account.  If any wb on the bdi has
dirty inodes, bdi_has_dirty_io() should return true.

To achieve that, as inode_wb_list_{move|del}_locked() now keep track
of the dirty state transition of each wb, the number of dirty wbs can
be counted in the bdi; however, bdi is already aggregating
wb->avg_write_bandwidth which can easily be guaranteed to be > 0 when
there are any dirty inodes by ensuring wb->avg_write_bandwidth can't
dip below 1.  bdi_has_dirty_io() can simply test whether
bdi->tot_write_bandwidth is zero or not.

While this bumps the value of wb->avg_write_bandwidth to one when it
used to be zero, this shouldn't cause any meaningful behavior
difference.

bdi_has_dirty_io() is made an inline function which tests whether
->tot_write_bandwidth is non-zero.  Also, WARN_ON_ONCE()'s on its
value are added to inode_wb_list_{move|del}_locked().

Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Jan Kara <jack@suse.cz>
Signed-off-by: Jens Axboe <axboe@fb.com>
fs/fs-writeback.c
include/linux/backing-dev-defs.h
include/linux/backing-dev.h
mm/backing-dev.c
mm/page-writeback.c

index bbccf68b38dbd57ceb3ff601453f84fcc48130c7..c98d3923d9a989d1fedd340847e060f602acbce8 100644 (file)
@@ -99,6 +99,7 @@ static bool wb_io_lists_populated(struct bdi_writeback *wb)
                return false;
        } else {
                set_bit(WB_has_dirty_io, &wb->state);
+               WARN_ON_ONCE(!wb->avg_write_bandwidth);
                atomic_long_add(wb->avg_write_bandwidth,
                                &wb->bdi->tot_write_bandwidth);
                return true;
@@ -110,8 +111,8 @@ static void wb_io_lists_depopulated(struct bdi_writeback *wb)
        if (wb_has_dirty_io(wb) && list_empty(&wb->b_dirty) &&
            list_empty(&wb->b_io) && list_empty(&wb->b_more_io)) {
                clear_bit(WB_has_dirty_io, &wb->state);
-               atomic_long_sub(wb->avg_write_bandwidth,
-                               &wb->bdi->tot_write_bandwidth);
+               WARN_ON_ONCE(atomic_long_sub_return(wb->avg_write_bandwidth,
+                                       &wb->bdi->tot_write_bandwidth) < 0);
        }
 }
 
index d631a61f40237bfde6f810447ad13d760aea63c1..8c857d723023260e6b7aa117e53b143af41a8a26 100644 (file)
@@ -98,7 +98,7 @@ struct bdi_writeback {
        unsigned long dirtied_stamp;
        unsigned long written_stamp;    /* pages written at bw_time_stamp */
        unsigned long write_bandwidth;  /* the estimated write bandwidth */
-       unsigned long avg_write_bandwidth; /* further smoothed write bw */
+       unsigned long avg_write_bandwidth; /* further smoothed write bw, > 0 */
 
        /*
         * The base dirty throttle rate, re-calculated on every 200ms.
@@ -142,7 +142,11 @@ struct backing_dev_info {
        unsigned int min_ratio;
        unsigned int max_ratio, max_prop_frac;
 
-       atomic_long_t tot_write_bandwidth; /* sum of active avg_write_bw */
+       /*
+        * Sum of avg_write_bw of wbs with dirty inodes.  > 0 if there are
+        * any dirty wbs, which is depended upon by bdi_has_dirty().
+        */
+       atomic_long_t tot_write_bandwidth;
 
        struct bdi_writeback wb;  /* the root writeback info for this bdi */
        struct bdi_writeback_congested wb_congested; /* its congested state */
index 3c8403c012ce1b78f7354f40093e5465bf1d060f..0839e44105bddd01e66f8e0dcf76769a6ed88de2 100644 (file)
@@ -29,7 +29,6 @@ void bdi_start_writeback(struct backing_dev_info *bdi, long nr_pages,
                        enum wb_reason reason);
 void bdi_start_background_writeback(struct backing_dev_info *bdi);
 void wb_workfn(struct work_struct *work);
-bool bdi_has_dirty_io(struct backing_dev_info *bdi);
 void wb_wakeup_delayed(struct bdi_writeback *wb);
 
 extern spinlock_t bdi_lock;
@@ -42,6 +41,15 @@ static inline bool wb_has_dirty_io(struct bdi_writeback *wb)
        return test_bit(WB_has_dirty_io, &wb->state);
 }
 
+static inline bool bdi_has_dirty_io(struct backing_dev_info *bdi)
+{
+       /*
+        * @bdi->tot_write_bandwidth is guaranteed to be > 0 if there are
+        * any dirty wbs.  See wb_update_write_bandwidth().
+        */
+       return atomic_long_read(&bdi->tot_write_bandwidth);
+}
+
 static inline void __add_wb_stat(struct bdi_writeback *wb,
                                 enum wb_stat_item item, s64 amount)
 {
index 161ddf12b8628179630d391830bb1237e8aaed84..d2f16fc9069aeeca97bf6bba2348c1ea94ab0d34 100644 (file)
@@ -256,11 +256,6 @@ static int __init default_bdi_init(void)
 }
 subsys_initcall(default_bdi_init);
 
-bool bdi_has_dirty_io(struct backing_dev_info *bdi)
-{
-       return wb_has_dirty_io(&bdi->wb);
-}
-
 /*
  * This function is used when the first inode for this wb is marked dirty. It
  * wakes-up the corresponding bdi thread which should then take care of the
index c95eb24dd6525f94dd0b77bfdb2115a4a85d0e53..99b88465096e6cf8399e292db5cb3c4feb7f8052 100644 (file)
@@ -881,9 +881,13 @@ static void wb_update_write_bandwidth(struct bdi_writeback *wb,
                avg += (old - avg) >> 3;
 
 out:
-       if (wb_has_dirty_io(wb))
-               atomic_long_add(avg - wb->avg_write_bandwidth,
-                               &wb->bdi->tot_write_bandwidth);
+       /* keep avg > 0 to guarantee that tot > 0 if there are dirty wbs */
+       avg = max(avg, 1LU);
+       if (wb_has_dirty_io(wb)) {
+               long delta = avg - wb->avg_write_bandwidth;
+               WARN_ON_ONCE(atomic_long_add_return(delta,
+                                       &wb->bdi->tot_write_bandwidth) <= 0);
+       }
        wb->write_bandwidth = bw;
        wb->avg_write_bandwidth = avg;
 }