On Mon, Oct 15, 2012 at 04:58:55PM +0200, Jean-Christophe PLAGNIOL-VILLARD wrote:

On 11:44 Mon 15 Oct     , Russell King - ARM Linux wrote:

quoted

On Mon, Oct 15, 2012 at 11:37:25AM +0100, Russell King - ARM Linux wrote:

quoted

On Mon, Oct 15, 2012 at 10:43:48AM +0100, Russell King - ARM Linux wrote:

quoted

1. Unregister the device _BEFORE_ taking away any resources it may
   be using.
2. Don't check clks against NULL.

Signed-off-by: Russell King <redacted>

Looking at this driver some more, who the hell came up with the sdhci
registration interface?  It violates one of the most fundamental
principles of kernel driver programming.  You do _NOT_ publish your
driver interfaces _UNTIL_ you have finished setting your device up.
Otherwise, in a preemptible or SMP kernel, your driver can be used
before the initialization has completed.

As this driver calls sdhci_pltfm_register() before it has obtained the
clock for the interface, and this function does:
	sdhci_pltfm_init
	sdhci_add_host
		mmc_add_host
			mmc_start_host
				mmc_power_up
					mmc_set_ios
						sdhci_set_ios

See, we're trying to power up and clock the card _before_ the dove
sdhci driver has even claimed the clock let alone enabled it.  This
is total bollocks.  The sdhci platform interface is total crap for
creating this broken design in the first place.  This is why MMC has
the init + add interfaces, they're there to allow drivers to do stuff
the Right(tm) way and avoid shit like the above.

This should have been picked up at review time before the driver went
into mainline.  In any case, it needs to be fixed.

Here's an updated patch which just about fixes the sdhci-dove driver.
I would not be surprised given the idiotic sdhci-pltfm API if many
other drivers suffered the same bug.

8<====
From: Russell King <redacted>
Subject: [PATCH] MMC: fix sdhci-dove probe/removal

1. Never ever publish a device in the system before it has been setup
   to a usable state.
2. Unregister the device _BEFORE_ taking away any resources it may be
   using.
3. Don't check clks against NULL.

Signed-off-by: Russell King <redacted>
---
 drivers/mmc/host/sdhci-dove.c |   36 ++++++++++++++++++------------------
 1 files changed, 18 insertions(+), 18 deletions(-)

diff --git a/drivers/mmc/host/sdhci-dove.c b/drivers/mmc/host/sdhci-dove.c
index a6e53a1..7d3a4e4 100644
--- a/drivers/mmc/host/sdhci-dove.c
+++ b/drivers/mmc/host/sdhci-dove.c

@@ -83,30 +83,31 @@ static int __devinit sdhci_dove_probe(struct platform_device *pdev)
 	struct sdhci_dove_priv *priv;
 	int ret;
 
-	ret = sdhci_pltfm_register(pdev, &sdhci_dove_pdata);
-	if (ret)
-		goto sdhci_dove_register_fail;
-
 	priv = devm_kzalloc(&pdev->dev, sizeof(struct sdhci_dove_priv),
 			    GFP_KERNEL);
 	if (!priv) {
 		dev_err(&pdev->dev, "unable to allocate private data");
-		ret = -ENOMEM;
-		goto sdhci_dove_allocate_fail;
+		return -ENOMEM;
 	}
 
+	priv->clk = clk_get(&pdev->dev, NULL);

you have devm_clk_get too

maybe you could use it here too

This isn't a cleanup patch, this is a patch just to fix some stupid bugs
in this code.  If someone wants to convert that afterwards, it should be
an entirely separate patch.