writeback: don't drain bdi_writeback_congested on bdi destruction
authorTejun Heo <tj@kernel.org>
Thu, 2 Jul 2015 00:53:37 +0000 (20:53 -0400)
committerJens Axboe <axboe@fb.com>
Thu, 2 Jul 2015 14:46:00 +0000 (08:46 -0600)
52ebea749aae ("writeback: make backing_dev_info host cgroup-specific
bdi_writebacks") made bdi (backing_dev_info) host per-cgroup wb's
(bdi_writeback's).  As the congested state needs to be per-wb and
referenced from blkcg side and multiple wbs, the patch made all
non-root cong's (bdi_writeback_congested's) reference counted and
indexed on bdi.

When a bdi is destroyed, cgwb_bdi_destroy() tries to drain all
non-root cong's; however, this can hang indefinitely because wb's can
also be referenced from blkcg_gq's which are destroyed after bdi
destruction is complete.

This patch fixes the bug by updating bdi destruction to not wait for
cong's to drain.  A cong is unlinked from bdi->cgwb_congested_tree on
bdi destuction regardless of its reference count as the bdi may go
away any point after destruction.  wb_congested_put() checks whether
the cong is already unlinked on release.

Signed-off-by: Tejun Heo <tj@kernel.org>
Reported-by: Jon Christopherson <jon@jons.org>
Link: https://bugzilla.kernel.org/show_bug.cgi?id=100681
Fixes: 52ebea749aae ("writeback: make backing_dev_info host cgroup-specific bdi_writebacks")
Tested-by: Jon Christopherson <jon@jons.org>
Signed-off-by: Jens Axboe <axboe@fb.com>
mm/backing-dev.c

index 51cc461e7256cb4c5e61b40cbbbdf40561f722e0..dac5bf59309d04d119b699ba47d849993edbec09 100644 (file)
@@ -425,7 +425,6 @@ retry:
                new_congested = NULL;
                rb_link_node(&congested->rb_node, parent, node);
                rb_insert_color(&congested->rb_node, &bdi->cgwb_congested_tree);
-               atomic_inc(&bdi->usage_cnt);
                goto found;
        }
 
@@ -456,7 +455,6 @@ found:
  */
 void wb_congested_put(struct bdi_writeback_congested *congested)
 {
-       struct backing_dev_info *bdi = congested->bdi;
        unsigned long flags;
 
        local_irq_save(flags);
@@ -465,12 +463,15 @@ void wb_congested_put(struct bdi_writeback_congested *congested)
                return;
        }
 
-       rb_erase(&congested->rb_node, &congested->bdi->cgwb_congested_tree);
+       /* bdi might already have been destroyed leaving @congested unlinked */
+       if (congested->bdi) {
+               rb_erase(&congested->rb_node,
+                        &congested->bdi->cgwb_congested_tree);
+               congested->bdi = NULL;
+       }
+
        spin_unlock_irqrestore(&cgwb_lock, flags);
        kfree(congested);
-
-       if (atomic_dec_and_test(&bdi->usage_cnt))
-               wake_up_all(&cgwb_release_wait);
 }
 
 static void cgwb_release_workfn(struct work_struct *work)
@@ -675,13 +676,22 @@ static int cgwb_bdi_init(struct backing_dev_info *bdi)
 static void cgwb_bdi_destroy(struct backing_dev_info *bdi)
 {
        struct radix_tree_iter iter;
+       struct bdi_writeback_congested *congested, *congested_n;
        void **slot;
 
        WARN_ON(test_bit(WB_registered, &bdi->wb.state));
 
        spin_lock_irq(&cgwb_lock);
+
        radix_tree_for_each_slot(slot, &bdi->cgwb_tree, &iter, 0)
                cgwb_kill(*slot);
+
+       rbtree_postorder_for_each_entry_safe(congested, congested_n,
+                                       &bdi->cgwb_congested_tree, rb_node) {
+               rb_erase(&congested->rb_node, &bdi->cgwb_congested_tree);
+               congested->bdi = NULL;  /* mark @congested unlinked */
+       }
+
        spin_unlock_irq(&cgwb_lock);
 
        /*