block: relax jiffies rounding for timeouts

In doing high IOPS testing, blk-mq is generally pretty well optimized.
There are a few things that stuck out as using more CPU than what is
really warranted, and one thing is the round_jiffies_up() that we do
twice for each request. That accounts for about 0.8% of the CPU in
my testing.

We can make this cheaper by avoiding an integer division, by just adding
a rough HZ mask that we can AND with instead. The timeouts are only on a
second granularity already, we don't have to be that accurate here and
this patch barely changes that. All we care about is nice grouping.

Signed-off-by: Jens Axboe <[email protected]>

block/blk-timeout.c

@@ -88,11 +88,29 @@ void blk_abort_request(struct request *req)
 }
 EXPORT_SYMBOL_GPL(blk_abort_request);
 
+static unsigned long blk_timeout_mask __read_mostly;
+
+int __init blk_timeout_init(void)
+{
+	blk_timeout_mask = roundup_pow_of_two(HZ) - 1;
+	return 0;
+}
+
+late_initcall(blk_timeout_init);
+
+/*
+ * Just a rough estimate, we don't care about specific values for timeouts.
+ */
+static inline unsigned long blk_round_jiffies(unsigned long j)
+{
+	return (j + blk_timeout_mask) + 1;
+}
+
 unsigned long blk_rq_timeout(unsigned long timeout)
 {
 	unsigned long maxt;
 
-	maxt = round_jiffies_up(jiffies + BLK_MAX_TIMEOUT);
+	maxt = blk_round_jiffies(jiffies + BLK_MAX_TIMEOUT);
 	if (time_after(timeout, maxt))
 		timeout = maxt;
 
@@ -129,7 +147,7 @@ void blk_add_timer(struct request *req)
 	 * than an existing one, modify the timer. Round up to next nearest
 	 * second.
 	 */
-	expiry = blk_rq_timeout(round_jiffies_up(expiry));
+	expiry = blk_rq_timeout(blk_round_jiffies(expiry));
 
 	if (!timer_pending(&q->timeout) ||
 	    time_before(expiry, q->timeout.expires)) {