mmc: sdhci-pxav3: fix unbalanced clock issues during probe

Commit 0dcaa2499b ("sdhci-pxav3: Fix runtime PM initialization") tries
to fix one hang issue caused by calling sdhci_add_host() on a suspended
device. The fix enables the clock twice, once by clk_prepare_enable() and
another by pm_runtime_get_sync(), meaning that the clock will never be
gated at runtime PM suspend. I observed the power consumption regression on
Marvell BG2Q SoCs.

In fact, the fix is not correct. There still be a very small window
during which a runtime suspend might somehow occur after pm_runtime_enable()
but before pm_runtime_get_sync().

This patch fixes all of the two problems by just incrementing the usage
counter before pm_runtime_enable(). It also adjust the order of disabling
runtime pm and storing the usage count in the error path to handle clock
gating properly.

Signed-off-by: Jisheng Zhang <jszhang@marvell.com>
Cc: <stable@vger.kernel.org> # v3.11+
Signed-off-by: Ulf Hansson <ulf.hansson@linaro.org>

drivers/mmc/host/sdhci-pxav3.c

@@ -365,10 +365,11 @@ static int sdhci_pxav3_probe(struct platform_device *pdev)
 		}
 	}
 
-	pm_runtime_enable(&pdev->dev);
+	pm_runtime_get_noresume(&pdev->dev);
-	pm_runtime_get_sync(&pdev->dev);
+	pm_runtime_set_active(&pdev->dev);
 	pm_runtime_set_autosuspend_delay(&pdev->dev, PXAV3_RPM_DELAY_MS);
 	pm_runtime_use_autosuspend(&pdev->dev);
+	pm_runtime_enable(&pdev->dev);
 	pm_suspend_ignore_children(&pdev->dev, 1);
 
 	ret = sdhci_add_host(host);
@@ -391,8 +392,8 @@ static int sdhci_pxav3_probe(struct platform_device *pdev)
 	return 0;
 
 err_add_host:
-	pm_runtime_put_sync(&pdev->dev);
 	pm_runtime_disable(&pdev->dev);
+	pm_runtime_put_noidle(&pdev->dev);
 err_of_parse:
 err_cd_req:
 err_mbus_win: