writeback: implement foreign cgroup inode detection

As concurrent write sharing of an inode is expected to be very rare
and memcg only tracks page ownership on first-use basis severely
confining the usefulness of such sharing, cgroup writeback tracks
ownership per-inode.  While the support for concurrent write sharing
of an inode is deemed unnecessary, an inode being written to by
different cgroups at different points in time is a lot more common,
and, more importantly, charging only by first-use can too readily lead
to grossly incorrect behaviors (single foreign page can lead to
gigabytes of writeback to be incorrectly attributed).

To resolve this issue, cgroup writeback detects the majority dirtier
of an inode and will transfer the ownership to it.  To avoid
unnnecessary oscillation, the detection mechanism keeps track of
history and gives out the switch verdict only if the foreign usage
pattern is stable over a certain amount of time and/or writeback
attempts.

The detection mechanism has fairly low space and computation overhead.
It adds 8 bytes to struct inode (one int and two u16's) and minimal
amount of calculation per IO.  The detection mechanism converges to
the correct answer usually in several seconds of IO time when there's
a clear majority dirtier.  Even when there isn't, it can reach an
acceptable answer fairly quickly under most circumstances.

Please see wb_detach_inode() for more details.

This patch only implements detection.  Following patches will
implement actual switching.

v2: wbc_account_io() now checks whether the wbc is associated with a
    wb before dereferencing it.  This can happen when pageout() is
    writing pages directly without going through the usual writeback
    path.  As pageout() path is single-threaded, we don't want it to
    be blocked behind a slow cgroup and ultimately want it to delegate
    actual writing to the usual writeback path.

Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Jan Kara <jack@suse.cz>
Cc: Wu Fengguang <fengguang.wu@intel.com>
Cc: Greg Thelen <gthelen@google.com>
Signed-off-by: Jens Axboe <axboe@fb.com>

1 files changed, 175 insertions, 2 deletions

diff --git a/fs/fs-writeback.c b/fs/fs-writeback.c
index 755e8ef..f98d403 100644
--- a/fs/fs-writeback.c
+++ b/fs/fs-writeback.c
@@ -214,6 +214,20 @@ static void wb_wait_for_completion(struct backing_dev_info *bdi,
 
 #ifdef CONFIG_CGROUP_WRITEBACK
 
+/* parameters for foreign inode detection, see wb_detach_inode() */
+#define WB_FRN_TIME_SHIFT	13	/* 1s = 2^13, upto 8 secs w/ 16bit */
+#define WB_FRN_TIME_AVG_SHIFT	3	/* avg = avg * 7/8 + new * 1/8 */
+#define WB_FRN_TIME_CUT_DIV	2	/* ignore rounds < avg / 2 */
+#define WB_FRN_TIME_PERIOD	(2 * (1 << WB_FRN_TIME_SHIFT))	/* 2s */
+
+#define WB_FRN_HIST_SLOTS	16	/* inode->i_wb_frn_history is 16bit */
+#define WB_FRN_HIST_UNIT	(WB_FRN_TIME_PERIOD / WB_FRN_HIST_SLOTS)
+					/* each slot's duration is 2s / 16 */
+#define WB_FRN_HIST_THR_SLOTS	(WB_FRN_HIST_SLOTS / 2)
+					/* if foreign slots >= 8, switch */
+#define WB_FRN_HIST_MAX_SLOTS	(WB_FRN_HIST_THR_SLOTS / 2 + 1)
+					/* one round can affect upto 5 slots */
+
 void __inode_attach_wb(struct inode *inode, struct page *page)
 {
 	struct backing_dev_info *bdi = inode_to_bdi(inode);
@@ -258,24 +272,183 @@ void wbc_attach_and_unlock_inode(struct writeback_control *wbc,
 				 struct inode *inode)
 {
 	wbc->wb = inode_to_wb(inode);
+	wbc->inode = inode;
+
+	wbc->wb_id = wbc->wb->memcg_css->id;
+	wbc->wb_lcand_id = inode->i_wb_frn_winner;
+	wbc->wb_tcand_id = 0;
+	wbc->wb_bytes = 0;
+	wbc->wb_lcand_bytes = 0;
+	wbc->wb_tcand_bytes = 0;
+
 	wb_get(wbc->wb);
 	spin_unlock(&inode->i_lock);
 }
 
 /**
- * wbc_detach_inode - disassociate wbc from its target inode
- * @wbc: writeback_control of interest
+ * wbc_detach_inode - disassociate wbc from inode and perform foreign detection
+ * @wbc: writeback_control of the just finished writeback
  *
  * To be called after a writeback attempt of an inode finishes and undoes
  * wbc_attach_and_unlock_inode().  Can be called under any context.
+ *
+ * As concurrent write sharing of an inode is expected to be very rare and
+ * memcg only tracks page ownership on first-use basis severely confining
+ * the usefulness of such sharing, cgroup writeback tracks ownership
+ * per-inode.  While the support for concurrent write sharing of an inode
+ * is deemed unnecessary, an inode being written to by different cgroups at
+ * different points in time is a lot more common, and, more importantly,
+ * charging only by first-use can too readily lead to grossly incorrect
+ * behaviors (single foreign page can lead to gigabytes of writeback to be
+ * incorrectly attributed).
+ *
+ * To resolve this issue, cgroup writeback detects the majority dirtier of
+ * an inode and transfers the ownership to it.  To avoid unnnecessary
+ * oscillation, the detection mechanism keeps track of history and gives
+ * out the switch verdict only if the foreign usage pattern is stable over
+ * a certain amount of time and/or writeback attempts.
+ *
+ * On each writeback attempt, @wbc tries to detect the majority writer
+ * using Boyer-Moore majority vote algorithm.  In addition to the byte
+ * count from the majority voting, it also counts the bytes written for the
+ * current wb and the last round's winner wb (max of last round's current
+ * wb, the winner from two rounds ago, and the last round's majority
+ * candidate).  Keeping track of the historical winner helps the algorithm
+ * to semi-reliably detect the most active writer even when it's not the
+ * absolute majority.
+ *
+ * Once the winner of the round is determined, whether the winner is
+ * foreign or not and how much IO time the round consumed is recorded in
+ * inode->i_wb_frn_history.  If the amount of recorded foreign IO time is
+ * over a certain threshold, the switch verdict is given.
  */
 void wbc_detach_inode(struct writeback_control *wbc)
 {
+	struct bdi_writeback *wb = wbc->wb;
+	struct inode *inode = wbc->inode;
+	u16 history = inode->i_wb_frn_history;
+	unsigned long avg_time = inode->i_wb_frn_avg_time;
+	unsigned long max_bytes, max_time;
+	int max_id;
+
+	/* pick the winner of this round */
+	if (wbc->wb_bytes >= wbc->wb_lcand_bytes &&
+	    wbc->wb_bytes >= wbc->wb_tcand_bytes) {
+		max_id = wbc->wb_id;
+		max_bytes = wbc->wb_bytes;
+	} else if (wbc->wb_lcand_bytes >= wbc->wb_tcand_bytes) {
+		max_id = wbc->wb_lcand_id;
+		max_bytes = wbc->wb_lcand_bytes;
+	} else {
+		max_id = wbc->wb_tcand_id;
+		max_bytes = wbc->wb_tcand_bytes;
+	}
+
+	/*
+	 * Calculate the amount of IO time the winner consumed and fold it
+	 * into the running average kept per inode.  If the consumed IO
+	 * time is lower than avag / WB_FRN_TIME_CUT_DIV, ignore it for
+	 * deciding whether to switch or not.  This is to prevent one-off
+	 * small dirtiers from skewing the verdict.
+	 */
+	max_time = DIV_ROUND_UP((max_bytes >> PAGE_SHIFT) << WB_FRN_TIME_SHIFT,
+				wb->avg_write_bandwidth);
+	if (avg_time)
+		avg_time += (max_time >> WB_FRN_TIME_AVG_SHIFT) -
+			    (avg_time >> WB_FRN_TIME_AVG_SHIFT);
+	else
+		avg_time = max_time;	/* immediate catch up on first run */
+
+	if (max_time >= avg_time / WB_FRN_TIME_CUT_DIV) {
+		int slots;
+
+		/*
+		 * The switch verdict is reached if foreign wb's consume
+		 * more than a certain proportion of IO time in a
+		 * WB_FRN_TIME_PERIOD.  This is loosely tracked by 16 slot
+		 * history mask where each bit represents one sixteenth of
+		 * the period.  Determine the number of slots to shift into
+		 * history from @max_time.
+		 */
+		slots = min(DIV_ROUND_UP(max_time, WB_FRN_HIST_UNIT),
+			    (unsigned long)WB_FRN_HIST_MAX_SLOTS);
+		history <<= slots;
+		if (wbc->wb_id != max_id)
+			history |= (1U << slots) - 1;
+
+		/*
+		 * Switch if the current wb isn't the consistent winner.
+		 * If there are multiple closely competing dirtiers, the
+		 * inode may switch across them repeatedly over time, which
+		 * is okay.  The main goal is avoiding keeping an inode on
+		 * the wrong wb for an extended period of time.
+		 */
+		if (hweight32(history) > WB_FRN_HIST_THR_SLOTS) {
+			/* switch */
+			max_id = 0;
+			avg_time = 0;
+			history = 0;
+		}
+	}
+
+	/*
+	 * Multiple instances of this function may race to update the
+	 * following fields but we don't mind occassional inaccuracies.
+	 */
+	inode->i_wb_frn_winner = max_id;
+	inode->i_wb_frn_avg_time = min(avg_time, (unsigned long)U16_MAX);
+	inode->i_wb_frn_history = history;
+
 	wb_put(wbc->wb);
 	wbc->wb = NULL;
 }
 
 /**
+ * wbc_account_io - account IO issued during writeback
+ * @wbc: writeback_control of the writeback in progress
+ * @page: page being written out
+ * @bytes: number of bytes being written out
+ *
+ * @bytes from @page are about to written out during the writeback
+ * controlled by @wbc.  Keep the book for foreign inode detection.  See
+ * wbc_detach_inode().
+ */
+void wbc_account_io(struct writeback_control *wbc, struct page *page,
+		    size_t bytes)
+{
+	int id;
+
+	/*
+	 * pageout() path doesn't attach @wbc to the inode being written
+	 * out.  This is intentional as we don't want the function to block
+	 * behind a slow cgroup.  Ultimately, we want pageout() to kick off
+	 * regular writeback instead of writing things out itself.
+	 */
+	if (!wbc->wb)
+		return;
+
+	rcu_read_lock();
+	id = mem_cgroup_css_from_page(page)->id;
+	rcu_read_unlock();
+
+	if (id == wbc->wb_id) {
+		wbc->wb_bytes += bytes;
+		return;
+	}
+
+	if (id == wbc->wb_lcand_id)
+		wbc->wb_lcand_bytes += bytes;
+
+	/* Boyer-Moore majority vote algorithm */
+	if (!wbc->wb_tcand_bytes)
+		wbc->wb_tcand_id = id;
+	if (id == wbc->wb_tcand_id)
+		wbc->wb_tcand_bytes += bytes;
+	else
+		wbc->wb_tcand_bytes -= min(bytes, wbc->wb_tcand_bytes);
+}
+
+/**
  * inode_congested - test whether an inode is congested
  * @inode: inode to test for congestion
  * @cong_bits: mask of WB_[a]sync_congested bits to test