ata_piix: parallel scanning on PATA needs an extra locking

Hi,

This one is quite late but it fixes rather nasty 2.6.31 regression..

From: Bartlomiej Zolnierkiewicz <[email protected]>
Subject: [PATCH] ata_piix: parallel scanning on PATA needs an extra locking

Commit log for commit 517d3cc15b36392e518abab6bacbb72089658313 ("[libata] ata_piix: Enable parallel scan") says:

This patch turns on parallel scanning for the ata_piix driver. This driver is used on most netbooks (no AHCI for cheap storage it seems). The scan is the dominating time factor in the kernel boot for these devices; with this flag it gets cut in half for the device I used for testing (eeepc).
Alan took a look at the driver source and concluded that it ought to be safe to do for this driver. Alan has also checked with the hardware team.

and it is all true but once we put all things together additional constraints for PATA controllers show up (some hardware registers have per-host not per-port atomicity) and we risk misprogramming the controller.

I used the following test to check whether the issue is real:

@@ -736,8 +736,20 @@ static void piix_set_piomode(struct ata_
(timings[pio][1] << 8);
}
pci_write_config_word(dev, master_port, master_data);
- if (is_slave)
+ if (is_slave) {
+ if (ap->port_no == 0) {
+ u8 tmp = slave_data;
+
+ while (slave_data == tmp) {
+ pci_read_config_byte(dev, slave_port, &tmp);
+ msleep(50);
+ }
+
+ dev_printk(KERN_ERR, &dev->dev, "PATA parallel scan "
+ "race detected\n");
+ }
pci_write_config_byte(dev, slave_port, slave_data);
+ }

/* Ensure the UDMA bit is off - it will be turned back on if
UDMA is selected */

and it indeed triggered the error message.

Lets fix all such races by adding an extra locking to ->set_piomode and ->set_dmamode methods for PATA controllers.

---
 drivers/ata/ata_piix.c |   14 +++++++++++++-
 1 file changed, 13 insertions(+), 1 deletion(-)

Index: b/drivers/ata/ata_piix.c
===================================================================
@@ -664,6 +664,8 @@ static int piix_pata_prereset(struct ata
 	return ata_sff_prereset(link, deadline);
 }
 
+static DEFINE_SPINLOCK(piix_lock);
+
 /**
  *	piix_set_piomode - Initialize host controller PATA PIO timings
  *	@ap: Port whose timings we are configuring

@@ -677,8 +679,9 @@ static int piix_pata_prereset(struct ata
 
 static void piix_set_piomode(struct ata_port *ap, struct ata_device *adev)
 {
-	unsigned int pio	= adev->pio_mode - XFER_PIO_0;
 	struct pci_dev *dev	= to_pci_dev(ap->host->dev);
+	unsigned long flags;
+	unsigned int pio	= adev->pio_mode - XFER_PIO_0;
 	unsigned int is_slave	= (adev->devno != 0);
 	unsigned int master_port= ap->port_no ? 0x42 : 0x40;
 	unsigned int slave_port	= 0x44;

@@ -708,6 +711,8 @@ static void piix_set_piomode(struct ata_
 	if (adev->class == ATA_DEV_ATA)
 		control |= 4;	/* PPE enable */
 
+	spin_lock_irqsave(&piix_lock, flags);
+
 	/* PIO configuration clears DTE unconditionally.  It will be
 	 * programmed in set_dmamode which is guaranteed to be called
 	 * after set_piomode if any DMA mode is available.

@@ -747,6 +752,8 @@ static void piix_set_piomode(struct ata_
 		udma_enable &= ~(1 << (2 * ap->port_no + adev->devno));
 		pci_write_config_byte(dev, 0x48, udma_enable);
 	}
+
+	spin_unlock_irqrestore(&piix_lock, flags);
 }
 
 /**

@@ -764,6 +771,7 @@ static void piix_set_piomode(struct ata_
 static void do_pata_set_dmamode(struct ata_port *ap, struct ata_device *adev, int isich)
 {
 	struct pci_dev *dev	= to_pci_dev(ap->host->dev);
+	unsigned long flags;
 	u8 master_port		= ap->port_no ? 0x42 : 0x40;
 	u16 master_data;
 	u8 speed		= adev->dma_mode;

@@ -777,6 +785,8 @@ static void do_pata_set_dmamode(struct a
 			    { 2, 1 },
 			    { 2, 3 }, };
 
+	spin_lock_irqsave(&piix_lock, flags);
+
 	pci_read_config_word(dev, master_port, &master_data);
 	if (ap->udma_mask)
 		pci_read_config_byte(dev, 0x48, &udma_enable);

@@ -867,6 +877,8 @@ static void do_pata_set_dmamode(struct a
 	/* Don't scribble on 0x48 if the controller does not support UDMA */
 	if (ap->udma_mask)
 		pci_write_config_byte(dev, 0x48, udma_enable);
+
+	spin_unlock_irqrestore(&piix_lock, flags);
 }
 
 /**

Would it not be better to take the host lock in libata-core for these cases so that we fix all the adapters in one swoop. Even if we are doing this lock taking in the controller specific code it seems the right lock to use ?

Looks fine as a temproary quickfix tho

Yes, taking the host lock would be a preferred solution in the long-term and would allow us to enable parallel scanning for many other controllers later.

The downside is that it requires somebody to audit all host drivers before making such change and much more testing time..

That was the goal of the patch (I verified the issue only yesterday).