mmc: core: Use delayed work in clock gating framework

Current clock gating framework disables the MCI clock as soon as the
request is completed and enables it when a request arrives. This aggressive
clock gating framework, when enabled, cause following issues:

When there are back-to-back requests from the Queue layer, we unnecessarily
end up disabling and enabling the clocks between these requests since 8MCLK
clock cycles is a very short duration compared to the time delay between
back to back requests reaching the MMC layer. This overhead can effect the
overall performance depending on how long the clock enable and disable
calls take which is platform dependent. For example on some platforms we
can have clock control not on the local processor, but on a different
subsystem and the time taken to perform the clock enable/disable can add
significant overhead.

Also if the host controller driver decides to disable the host clock too
when mmc_set_ios function is called with ios.clock=0, it adds additional
delay and it is highly possible that the next request had already arrived
and unnecessarily blocked in enabling the clocks. This is seen frequently
when the processor is executing at high speeds and in multi-core platforms
thus reduces the overall throughput compared to if clock gating is
disabled.

Fix this by delaying turning off the clocks by posting request on
delayed workqueue. Also cancel the unscheduled pending work, if any,
when there is access to card.

sysfs entry is provided to tune the delay as needed, default
value set to 200ms.

Signed-off-by: Sujit Reddy Thumma <sthumma@codeaurora.org>
Acked-by: Linus Walleij <linus.walleij@linaro.org>
Signed-off-by: Chris Ball <cjb@laptop.org>

1 files changed, 53 insertions, 4 deletions

diff --git a/drivers/mmc/core/host.c b/drivers/mmc/core/host.c
index d31c78b72b0f..817a76039743 100644
--- a/drivers/mmc/core/host.c
+++ b/drivers/mmc/core/host.c
@@ -54,6 +54,31 @@ static DEFINE_IDR(mmc_host_idr);
 static DEFINE_SPINLOCK(mmc_host_lock);
 
 #ifdef CONFIG_MMC_CLKGATE
+static ssize_t clkgate_delay_show(struct device *dev,
+		struct device_attribute *attr, char *buf)
+{
+	struct mmc_host *host = cls_dev_to_mmc_host(dev);
+	return snprintf(buf, PAGE_SIZE, "%lu millisecs\n",
+			host->clkgate_delay);
+}
+
+static ssize_t clkgate_delay_store(struct device *dev,
+		struct device_attribute *attr, const char *buf, size_t count)
+{
+	struct mmc_host *host = cls_dev_to_mmc_host(dev);
+	unsigned long flags, value;
+
+	if (kstrtoul(buf, 0, &value))
+		return -EINVAL;
+
+	spin_lock_irqsave(&host->clk_lock, flags);
+	host->clkgate_delay = value;
+	spin_unlock_irqrestore(&host->clk_lock, flags);
+
+	pr_info("%s: clock gate delay set to %lu ms\n",
+			mmc_hostname(host), value);
+	return count;
+}
 
 /*
  * Enabling clock gating will make the core call out to the host
@@ -114,7 +139,7 @@ static void mmc_host_clk_gate_delayed(struct mmc_host *host)
 static void mmc_host_clk_gate_work(struct work_struct *work)
 {
 	struct mmc_host *host = container_of(work, struct mmc_host,
-					      clk_gate_work);
+					      clk_gate_work.work);
 
 	mmc_host_clk_gate_delayed(host);
 }
@@ -131,6 +156,8 @@ void mmc_host_clk_hold(struct mmc_host *host)
 {
 	unsigned long flags;
 
+	/* cancel any clock gating work scheduled by mmc_host_clk_release() */
+	cancel_delayed_work_sync(&host->clk_gate_work);
 	mutex_lock(&host->clk_gate_mutex);
 	spin_lock_irqsave(&host->clk_lock, flags);
 	if (host->clk_gated) {
@@ -180,7 +207,8 @@ void mmc_host_clk_release(struct mmc_host *host)
 	host->clk_requests--;
 	if (mmc_host_may_gate_card(host->card) &&
 	    !host->clk_requests)
-		queue_work(system_nrt_wq, &host->clk_gate_work);
+		queue_delayed_work(system_nrt_wq, &host->clk_gate_work,
+				msecs_to_jiffies(host->clkgate_delay));
 	spin_unlock_irqrestore(&host->clk_lock, flags);
 }
 
@@ -213,8 +241,13 @@ static inline void mmc_host_clk_init(struct mmc_host *host)
 	host->clk_requests = 0;
 	/* Hold MCI clock for 8 cycles by default */
 	host->clk_delay = 8;
+	/*
+	 * Default clock gating delay is 200ms.
+	 * This value can be tuned by writing into sysfs entry.
+	 */
+	host->clkgate_delay = 200;
 	host->clk_gated = false;
-	INIT_WORK(&host->clk_gate_work, mmc_host_clk_gate_work);
+	INIT_DELAYED_WORK(&host->clk_gate_work, mmc_host_clk_gate_work);
 	spin_lock_init(&host->clk_lock);
 	mutex_init(&host->clk_gate_mutex);
 }
@@ -229,7 +262,7 @@ static inline void mmc_host_clk_exit(struct mmc_host *host)
 	 * Wait for any outstanding gate and then make sure we're
 	 * ungated before exiting.
 	 */
-	if (cancel_work_sync(&host->clk_gate_work))
+	if (cancel_delayed_work_sync(&host->clk_gate_work))
 		mmc_host_clk_gate_delayed(host);
 	if (host->clk_gated)
 		mmc_host_clk_hold(host);
@@ -237,6 +270,17 @@ static inline void mmc_host_clk_exit(struct mmc_host *host)
 	WARN_ON(host->clk_requests > 1);
 }
 
+static inline void mmc_host_clk_sysfs_init(struct mmc_host *host)
+{
+	host->clkgate_delay_attr.show = clkgate_delay_show;
+	host->clkgate_delay_attr.store = clkgate_delay_store;
+	sysfs_attr_init(&host->clkgate_delay_attr.attr);
+	host->clkgate_delay_attr.attr.name = "clkgate_delay";
+	host->clkgate_delay_attr.attr.mode = S_IRUGO | S_IWUSR;
+	if (device_create_file(&host->class_dev, &host->clkgate_delay_attr))
+		pr_err("%s: Failed to create clkgate_delay sysfs entry\n",
+				mmc_hostname(host));
+}
 #else
 
 static inline void mmc_host_clk_init(struct mmc_host *host)
@@ -247,6 +291,10 @@ static inline void mmc_host_clk_exit(struct mmc_host *host)
 {
 }
 
+static inline void mmc_host_clk_sysfs_init(struct mmc_host *host)
+{
+}
+
 #endif
 
 /**
@@ -335,6 +383,7 @@ int mmc_add_host(struct mmc_host *host)
 #ifdef CONFIG_DEBUG_FS
 	mmc_add_host_debugfs(host);
 #endif
+	mmc_host_clk_sysfs_init(host);
 
 	mmc_start_host(host);
 	register_pm_notifier(&host->pm_notify);