[PATCH v2 3/3] writeback: use a per-sb counter to drain inode wb switches at umount

[Baokun Li <libaokun@linux.alibaba.com>]

Tracking in-flight inode wb switches with a single global counter
(isw_nr_in_flight) plus a synchronize_rcu() based wait in
cgroup_writeback_umount() forces every umount to take a global hit
whenever any other superblock on the system has wb switches in flight,
even if the superblock being unmounted has none of its own.

Replace the global synchronize_rcu()/flush_workqueue() pair with a
per-sb counter, s_isw_nr_in_flight:

  - inode_prepare_wbs_switch() increments before checking SB_ACTIVE
    and grabbing the inode; failure paths decrement before returning.
  - process_inode_switch_wbs() decrements after the matching iput().
  - cgroup_writeback_umount() waits for the per-sb counter to reach zero.

The smp_mb() pair between inode_prepare_wbs_switch() and
cgroup_writeback_umount() keeps the SB_ACTIVE / counter ordering:
either the umounter sees a non-zero counter and waits, or the switcher
sees SB_ACTIVE cleared and aborts before grabbing the inode.

The existing global isw_nr_in_flight is left in place, since it is
still used to throttle in-flight switches via WB_FRN_MAX_IN_FLIGHT.

The rcu_read_lock() extension in inode_switch_wbs() and
cleanup_offline_cgwb() introduced by the race fix is no longer
needed and is reverted.  The synchronize_rcu() in
cgroup_writeback_umount() is dropped as well.

inode_prepare_wbs_switch() also gains a lockless SB_ACTIVE check at
the top to skip the atomic_inc/smp_mb cost for inodes whose sb is
already being torn down.

The following numbers were measured with 4 background superblocks
each churning "create memcg -> write 1 MiB -> rmdir memcg" to keep
the global isw_nr_in_flight non-zero.  A separate idle target sb is
mounted and umounted in a loop (N=100); only its umount latency is
measured.

Idle target umount latency under cross-sb cgwb-switch pressure:

                              p50      p95      p99      max
  global synchronize_rcu()  64.4 ms  95.8 ms 101.4 ms 110.5 ms
  per-sb counter (this)      5.3 ms   6.9 ms   7.4 ms   7.7 ms
  no-pressure baseline       5.2 ms   5.9 ms   6.0 ms   6.1 ms

8 concurrent umounts of idle sbs under the same pressure:

                              p50      p95      max
  global synchronize_rcu()  57.9 ms  82.1 ms  90.0 ms
  per-sb counter (this)      7.5 ms   7.8 ms   8.0 ms

In-kernel cgroup_writeback_umount() time over 286 calls (bpftrace):

  global synchronize_rcu()     8717 ms total (~30 ms / call)
  per-sb counter (this)        1.16 ms total (~4 us / call)

When s_isw_nr_in_flight is zero the loop body is never entered.

Signed-off-by: Baokun Li <libaokun@linux.alibaba.com>
---
 fs/fs-writeback.c              | 74 ++++++++++++++--------------------
 include/linux/fs/super_types.h |  8 ++++
 2 files changed, 38 insertions(+), 44 deletions(-)

diff --git a/fs/fs-writeback.c b/fs/fs-writeback.c
index 9ae290547eb2..4282fcfe027b 100644
--- a/fs/fs-writeback.c
+++ b/fs/fs-writeback.c
@@ -554,8 +554,12 @@ static void process_inode_switch_wbs(struct bdi_writeback *new_wb,
 		wb_put_many(old_wb, nr_switched);
 	}
 
-	for (inodep = isw->inodes; *inodep; inodep++)
+	for (inodep = isw->inodes; *inodep; inodep++) {
+		struct super_block *sb = (*inodep)->i_sb;
+
 		iput(*inodep);
+		atomic_dec(&sb->s_isw_nr_in_flight);
+	}
 	wb_put(new_wb);
 	kfree(isw);
 	atomic_dec(&isw_nr_in_flight);
@@ -598,16 +602,19 @@ void inode_switch_wbs_work_fn(struct work_struct *work)
 static bool inode_prepare_wbs_switch(struct inode *inode,
 				     struct bdi_writeback *new_wb)
 {
+	/* Avoid the atomic_inc/smp_mb dance once SB_ACTIVE is gone. */
+	if (!(inode->i_sb->s_flags & SB_ACTIVE))
+		return false;
+
 	/*
-	 * Paired with smp_mb() in cgroup_writeback_umount().
-	 * isw_nr_in_flight must be increased before checking SB_ACTIVE and
-	 * grabbing an inode, otherwise isw_nr_in_flight can be observed as 0
-	 * in cgroup_writeback_umount() and the isw_wq will be not flushed.
+	 * Pairs with smp_mb() in cgroup_writeback_umount(): the umounter either
+	 * sees a non-zero counter and waits, or we see SB_ACTIVE clear below.
 	 */
+	atomic_inc(&inode->i_sb->s_isw_nr_in_flight);
 	smp_mb();
 
 	if (IS_DAX(inode))
-		return false;
+		goto out_dec;
 
 	/* while holding I_WB_SWITCH, no one else can update the association */
 	spin_lock(&inode->i_lock);
@@ -615,13 +622,17 @@ static bool inode_prepare_wbs_switch(struct inode *inode,
 	    inode_state_read(inode) & (I_WB_SWITCH | I_FREEING | I_WILL_FREE) ||
 	    inode_to_wb(inode) == new_wb) {
 		spin_unlock(&inode->i_lock);
-		return false;
+		goto out_dec;
 	}
 	inode_state_set(inode, I_WB_SWITCH);
 	__iget(inode);
 	spin_unlock(&inode->i_lock);
 
 	return true;
+
+out_dec:
+	atomic_dec(&inode->i_sb->s_isw_nr_in_flight);
+	return false;
 }
 
 static void wb_queue_isw(struct bdi_writeback *wb,
@@ -665,6 +676,7 @@ static void inode_switch_wbs(struct inode *inode, int new_wb_id)
 	memcg_css = css_from_id(new_wb_id, &memory_cgrp_subsys);
 	if (memcg_css && !css_tryget(memcg_css))
 		memcg_css = NULL;
+	rcu_read_unlock();
 	if (!memcg_css)
 		goto out_free;
 
@@ -679,18 +691,10 @@ static void inode_switch_wbs(struct inode *inode, int new_wb_id)
 	isw->inodes[0] = inode;
 
 	trace_inode_switch_wbs_queue(inode->i_wb, new_wb, 1);
-	/*
-	 * Paired with synchronize_rcu() in cgroup_writeback_umount().
-	 * Holding rcu_read_lock across wb_queue_isw() ensures
-	 * synchronize_rcu() cannot return until the work is queued, so
-	 * the subsequent flush_workqueue() will wait for the switch.
-	 */
 	wb_queue_isw(new_wb, isw);
-	rcu_read_unlock();
 	return;
 
 out_free:
-	rcu_read_unlock();
 	atomic_dec(&isw_nr_in_flight);
 	if (new_wb)
 		wb_put(new_wb);
@@ -748,14 +752,6 @@ bool cleanup_offline_cgwb(struct bdi_writeback *wb)
 		new_wb = &wb->bdi->wb; /* wb_get() is noop for bdi's wb */
 
 	nr = 0;
-	/*
-	 * Paired with synchronize_rcu() in cgroup_writeback_umount().
-	 * Holding rcu_read_lock across the SB_ACTIVE check, the inode grab
-	 * and wb_queue_isw() ensures synchronize_rcu() cannot return until
-	 * the work is queued, so the subsequent flush_workqueue() will wait
-	 * for the switch.
-	 */
-	rcu_read_lock();
 	spin_lock(&wb->list_lock);
 	/*
 	 * In addition to the inodes that have completed writeback, also switch
@@ -773,7 +769,6 @@ bool cleanup_offline_cgwb(struct bdi_writeback *wb)
 
 	/* no attached inodes? bail out */
 	if (nr == 0) {
-		rcu_read_unlock();
 		atomic_dec(&isw_nr_in_flight);
 		wb_put(new_wb);
 		kfree(isw);
@@ -782,7 +777,6 @@ bool cleanup_offline_cgwb(struct bdi_writeback *wb)
 
 	trace_inode_switch_wbs_queue(wb, new_wb, nr);
 	wb_queue_isw(new_wb, isw);
-	rcu_read_unlock();
 
 	return restart;
 }
@@ -1215,38 +1209,30 @@ int cgroup_writeback_by_id(u64 bdi_id, int memcg_id,
 }
 
 /**
- * cgroup_writeback_umount - flush inode wb switches for umount
+ * cgroup_writeback_umount - wait for in-flight inode wb switches on @sb
  * @sb: target super_block
  *
- * This function is called when a super_block is about to be destroyed and
- * flushes in-flight inode wb switches.  An inode wb switch goes through
- * RCU and then workqueue, so the two need to be flushed in order to ensure
- * that all previously scheduled switches are finished.  As wb switches are
- * rare occurrences and synchronize_rcu() can take a while, perform
- * flushing iff wb switches are in flight.
+ * Wait until every inode wb switch that already passed the SB_ACTIVE
+ * check on this superblock has been completed by the worker.  Since
+ * SB_ACTIVE is cleared before this is called, no new switches can start
+ * for @sb, so s_isw_nr_in_flight will monotonically drop to zero.
  */
 void cgroup_writeback_umount(struct super_block *sb)
 {
-
 	if (!(sb->s_bdi->capabilities & BDI_CAP_WRITEBACK))
 		return;
 
 	/*
-	 * SB_ACTIVE should be reliably cleared before checking
-	 * isw_nr_in_flight, see generic_shutdown_super().
+	 * Pairs with smp_mb() in inode_prepare_wbs_switch(): we either observe
+	 * a non-zero counter and wait, or the switcher sees SB_ACTIVE clear
+	 * (cleared by generic_shutdown_super()) and bails before grabbing the
+	 * inode.
 	 */
 	smp_mb();
 
-	if (atomic_read(&isw_nr_in_flight)) {
-		/*
-		 * Paired with rcu_read_lock() in inode_switch_wbs() and
-		 * cleanup_offline_cgwb().  synchronize_rcu() waits for any
-		 * in-flight switcher that already passed the SB_ACTIVE check
-		 * to finish queueing its work, so flush_workqueue() below
-		 * will then drain it.
-		 */
-		synchronize_rcu();
+	while (atomic_read(&sb->s_isw_nr_in_flight)) {
 		flush_workqueue(isw_wq);
+		cond_resched();
 	}
 }
 
diff --git a/include/linux/fs/super_types.h b/include/linux/fs/super_types.h
index 383050e7fdf5..1ab4e2265129 100644
--- a/include/linux/fs/super_types.h
+++ b/include/linux/fs/super_types.h
@@ -274,6 +274,14 @@ struct super_block {
 
 	/* number of fserrors that are being sent to fsnotify/filesystems */
 	refcount_t				s_pending_errors;
+
+#ifdef CONFIG_CGROUP_WRITEBACK
+	/*
+	 * Number of in-flight inode wb switches for this sb.  Drained by
+	 * cgroup_writeback_umount() before tear-down.
+	 */
+	atomic_t				s_isw_nr_in_flight;
+#endif
 } __randomize_layout;
 
 /*
-- 
2.43.7