Skip to contentSkip to conversation

Add PM support to sis900 network driver

PATCH In review Daniele Venzano 2 participants 6 messages Nov 2, 2003 – Nov 3, 2003 Network drivers

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 Add PM support to sis900 network driver

The attached patch adds support for suspend/resume to the sis900 driver. With this patch on resume the NIC is fully configured and operational, before a module reload was needed because of the complete lack of suspend/resume callbacks.

I added two functions, sis900_suspend and sis900_resume, with their pointers in struct pci_driver. A vector of 16 u32 was then needed to the to keep PCI data during suspend. I added it in struct sis900_private. I updated the revision number to reflect my changes.
Looking at the code I also killed three typos.

The patch doesn't touch any other code.

Since I don't know anything on ethernet drivers the rule 'works for me' is fully valid.

Please consider this patch for inclusion when the freeze is over, the sis900 driver is not updated since 2002 (looking at the changelog) and the email address listed in the MAINTAINERS file bounces.

I'll try to keep the patch updated on new kernel releases at:
http://teg.homeunix.org/kernel_patches.html

Bye.

Signature · 3 lines
------------------------------
Daniele Venzano
Web: http://teg.homeunix.org
drivers/net/sis900.c +75 -6
diff -Naur -X /home/venza/dontdiff linux-2.6.0-test9/drivers/net/sis900.c linux-2.6.0-test9-dv/drivers/net/sis900.c
@@ -18,10 +18,11 @@
    preliminary Rev. 1.0 Jan. 18, 1998
    http://www.sis.com.tw/support/databook.htm
 
+   Rev 1.08.07 Nov.  2 2003 Daniele Venzano <[email protected]> add suspend/resume support
    Rev 1.08.06 Sep. 24 2002 Mufasa Yang bug fix for Tx timeout & add SiS963 support
-   Rev 1.08.05 Jun. 6 2002 Mufasa Yang bug fix for read_eeprom & Tx descriptor over-boundary 
+   Rev 1.08.05 Jun.  6 2002 Mufasa Yang bug fix for read_eeprom & Tx descriptor over-boundary 
    Rev 1.08.04 Apr. 25 2002 Mufasa Yang <[email protected]> added SiS962 support
-   Rev 1.08.03 Feb. 1 2002 Matt Domsch <[email protected]> update to use library crc32 function
+   Rev 1.08.03 Feb.  1 2002 Matt Domsch <[email protected]> update to use library crc32 function
    Rev 1.08.02 Nov. 30 2001 Hui-Fen Hsu workaround for EDB & bug fix for dhcp problem
    Rev 1.08.01 Aug. 25 2001 Hui-Fen Hsu update for 630ET & workaround for ICS1893 PHY
    Rev 1.08.00 Jun. 11 2001 Hui-Fen Hsu workaround for RTL8201 PHY and some bug fix
@@ -72,7 +73,7 @@
 #include "sis900.h"
 
 #define SIS900_MODULE_NAME "sis900"
-#define SIS900_DRV_VERSION "v1.08.06 9/24/2002"
+#define SIS900_DRV_VERSION "v1.08.07 11/02/2003"
 
 static char version[] __devinitdata =
 KERN_INFO "sis900.c: " SIS900_DRV_VERSION "\n";
@@ -169,6 +170,7 @@
 
 	unsigned int tx_full;			/* The Tx queue is full.    */
 	u8 host_bridge_rev;
+	u32 pci_state[16];
 };
 
 MODULE_AUTHOR("Jim Huang <[email protected]>, Ollie Lho <[email protected]>");
@@ -305,7 +307,7 @@
 		*( ((u16 *)net_dev->dev_addr) + i) = inw(ioaddr + rfdr);
 	}
 
-	/* enable packet filitering */
+	/* enable packet filtering */
 	outl(rfcrSave | RFEN, rfcr + ioaddr);
 
 	return 1;
@@ -994,7 +996,7 @@
 		}
 	}
 
-	/* enable packet filitering */
+	/* enable packet filtering */
 	outl(rfcrSave | RFEN, rfcr + ioaddr);
 }
 
@@ -1466,7 +1468,7 @@
  *	@net_dev: the net device to transmit with
  *
  *	Set the transmit buffer descriptor, 
- *	and write TxENA to enable transimt state machine.
+ *	and write TxENA to enable transmit state machine.
  *	tell upper layer if the buffer is full
  */
 
@@ -2184,11 +2186,78 @@
 	pci_set_drvdata(pci_dev, NULL);
 }
 
+#ifdef CONFIG_PM
+
+static int sis900_suspend(struct pci_dev *pci_dev, u32 state)
+{
+	struct net_device *net_dev = pci_get_drvdata(pci_dev);
+	struct sis900_private *sis_priv = net_dev->priv;
+	long ioaddr = net_dev->base_addr;
+	unsigned long flags;
+
+	if(!netif_running(net_dev))
+		return 0;
+	netif_stop_queue(net_dev);
+	
+	netif_device_detach(net_dev);
+	spin_lock_irqsave(&sis_priv->lock, flags);
+
+	/* Stop the chip's Tx and Rx Status Machine */
+	outl(RxDIS | TxDIS | inl(ioaddr + cr), ioaddr + cr);
+	
+	pci_set_power_state(pci_dev, 3);
+	pci_save_state(pci_dev, sis_priv->pci_state);
+
+	spin_unlock_irqrestore(&sis_priv->lock, flags);
+	return 0;
+}
+
+static int sis900_resume(struct pci_dev *pci_dev)
+{
+	struct net_device *net_dev = pci_get_drvdata(pci_dev);
+	struct sis900_private *sis_priv = net_dev->priv;
+	long ioaddr = net_dev->base_addr;
+
+	if(!netif_running(net_dev))
+		return 0;
+	pci_restore_state(pci_dev, sis_priv->pci_state);
+	pci_set_power_state(pci_dev, 0);
+
+//	/* Soft reset the chip */
+//	sis900_reset(net_dev);
+	sis900_init_rxfilter(net_dev);
+
+	sis900_init_tx_ring(net_dev);
+	sis900_init_rx_ring(net_dev);
+
+	set_rx_mode(net_dev);
+
+	netif_device_attach(net_dev);
+	netif_start_queue(net_dev);
+
+	/* Workaround for EDB */
+	sis900_set_mode(ioaddr, HW_SPEED_10_MBPS, FDX_CAPABLE_HALF_SELECTED);
+
+	/* Enable all known interrupts by setting the interrupt mask. */
+	outl((RxSOVR|RxORN|RxERR|RxOK|TxURN|TxERR|TxIDLE), ioaddr + imr);
+	outl(RxENA | inl(ioaddr + cr), ioaddr + cr);
+	outl(IE, ioaddr + ier);
+
+	sis900_check_mode(net_dev, sis_priv->mii);
+
+	return 0;
+}
+#endif /* CONFIG_PM */
+
 static struct pci_driver sis900_pci_driver = {
 	.name		= SIS900_MODULE_NAME,
 	.id_table	= sis900_pci_tbl,
 	.probe		= sis900_probe,
 	.remove		= __devexit_p(sis900_remove),
+#ifdef CONFIG_PM
+	.suspend	= sis900_suspend,
+	.resume		= sis900_resume,
+#endif /* CONFIG_PM */
 };
 
 static int __init sis900_init_module(void)
#1 lore raw
Quoted 2 lines · Daniele Venzano

The attached patch adds support for suspend/resume to the sis900 driver.

...

Quoted 10 lines

+static int sis900_suspend(struct pci_dev *pci_dev, u32 state)
+{
+ struct net_device *net_dev = pci_get_drvdata(pci_dev);
+ struct sis900_private *sis_priv = net_dev->priv;
+ long ioaddr = net_dev->base_addr;
+ unsigned long flags;
+
+ if(!netif_running(net_dev))
+ return 0;
+ netif_stop_queue(net_dev);

Quoted 8 lines

+
+ netif_device_detach(net_dev);
+ spin_lock_irqsave(&sis_priv->lock, flags);
+
+ /* Stop the chip's Tx and Rx Status Machine */
+ outl(RxDIS | TxDIS | inl(ioaddr + cr), ioaddr + cr);
+
+ pci_set_power_state(pci_dev, 3);

pci_set_power_state() can sleep, so we shouldn't be calling it under spin_lock_irqsave(). Is it necessary to hold the lock here?

#2 lore raw
Quoted 10 lines · Andrew Morton
Quoted 6 lines

+ spin_lock_irqsave(&sis_priv->lock, flags);
+
+ /* Stop the chip's Tx and Rx Status Machine */
+ outl(RxDIS | TxDIS | inl(ioaddr + cr), ioaddr + cr);
+
+ pci_set_power_state(pci_dev, 3);

pci_set_power_state() can sleep, so we shouldn't be calling it under spin_lock_irqsave(). Is it necessary to hold the lock here?

I don't know enough of the driver to express an opinion, but probably the lock is needed only for the outl call, if needed at all. If nothing new comes out, tomorrow I'll rediff and send a new patch.

Signature · 3 lines
------------------------------
Daniele Venzano
Web: http://teg.homeunix.org
#3 lore raw
In reply to Andrew Morton
Quoted 3 lines · Andrew Morton

pci_set_power_state() can sleep, so we shouldn't be calling it under spin_lock_irqsave(). Is it necessary to hold the lock here?

New patch with locking completely removed, since in a similar function none was used.

I think also the 8139too driver has the same locking problem in rtl8139_suspend, do you want a patch ?

Signature · 3 lines
------------------------------
Daniele Venzano
Web: http://teg.homeunix.org
drivers/net/sis900.c +70 -6
diff -Naur -X /home/venza/dontdiff linux-2.6.0-test9/drivers/net/sis900.c linux-2.6.0-test9-dv/drivers/net/sis900.c
@@ -18,10 +18,11 @@
    preliminary Rev. 1.0 Jan. 18, 1998
    http://www.sis.com.tw/support/databook.htm
 
+   Rev 1.08.07 Nov.  2 2003 Daniele Venzano <[email protected]> add suspend/resume support
    Rev 1.08.06 Sep. 24 2002 Mufasa Yang bug fix for Tx timeout & add SiS963 support
-   Rev 1.08.05 Jun. 6 2002 Mufasa Yang bug fix for read_eeprom & Tx descriptor over-boundary 
+   Rev 1.08.05 Jun.  6 2002 Mufasa Yang bug fix for read_eeprom & Tx descriptor over-boundary 
    Rev 1.08.04 Apr. 25 2002 Mufasa Yang <[email protected]> added SiS962 support
-   Rev 1.08.03 Feb. 1 2002 Matt Domsch <[email protected]> update to use library crc32 function
+   Rev 1.08.03 Feb.  1 2002 Matt Domsch <[email protected]> update to use library crc32 function
    Rev 1.08.02 Nov. 30 2001 Hui-Fen Hsu workaround for EDB & bug fix for dhcp problem
    Rev 1.08.01 Aug. 25 2001 Hui-Fen Hsu update for 630ET & workaround for ICS1893 PHY
    Rev 1.08.00 Jun. 11 2001 Hui-Fen Hsu workaround for RTL8201 PHY and some bug fix
@@ -72,7 +73,7 @@
 #include "sis900.h"
 
 #define SIS900_MODULE_NAME "sis900"
-#define SIS900_DRV_VERSION "v1.08.06 9/24/2002"
+#define SIS900_DRV_VERSION "v1.08.07 11/02/2003"
 
 static char version[] __devinitdata =
 KERN_INFO "sis900.c: " SIS900_DRV_VERSION "\n";
@@ -169,6 +170,7 @@
 
 	unsigned int tx_full;			/* The Tx queue is full.    */
 	u8 host_bridge_rev;
+	u32 pci_state[16];
 };
 
 MODULE_AUTHOR("Jim Huang <[email protected]>, Ollie Lho <[email protected]>");
@@ -305,7 +307,7 @@
 		*( ((u16 *)net_dev->dev_addr) + i) = inw(ioaddr + rfdr);
 	}
 
-	/* enable packet filitering */
+	/* enable packet filtering */
 	outl(rfcrSave | RFEN, rfcr + ioaddr);
 
 	return 1;
@@ -994,7 +996,7 @@
 		}
 	}
 
-	/* enable packet filitering */
+	/* enable packet filtering */
 	outl(rfcrSave | RFEN, rfcr + ioaddr);
 }
 
@@ -1466,7 +1468,7 @@
  *	@net_dev: the net device to transmit with
  *
  *	Set the transmit buffer descriptor, 
- *	and write TxENA to enable transimt state machine.
+ *	and write TxENA to enable transmit state machine.
  *	tell upper layer if the buffer is full
  */
 
@@ -2184,11 +2186,73 @@
 	pci_set_drvdata(pci_dev, NULL);
 }
 
+#ifdef CONFIG_PM
+
+static int sis900_suspend(struct pci_dev *pci_dev, u32 state)
+{
+	struct net_device *net_dev = pci_get_drvdata(pci_dev);
+	struct sis900_private *sis_priv = net_dev->priv;
+	long ioaddr = net_dev->base_addr;
+
+	if(!netif_running(net_dev))
+		return 0;
+	netif_stop_queue(net_dev);
+	
+	netif_device_detach(net_dev);
+
+	/* Stop the chip's Tx and Rx Status Machine */
+	outl(RxDIS | TxDIS | inl(ioaddr + cr), ioaddr + cr);
+	
+	pci_set_power_state(pci_dev, 3);
+	pci_save_state(pci_dev, sis_priv->pci_state);
+
+	return 0;
+}
+
+static int sis900_resume(struct pci_dev *pci_dev)
+{
+	struct net_device *net_dev = pci_get_drvdata(pci_dev);
+	struct sis900_private *sis_priv = net_dev->priv;
+	long ioaddr = net_dev->base_addr;
+
+	if(!netif_running(net_dev))
+		return 0;
+	pci_restore_state(pci_dev, sis_priv->pci_state);
+	pci_set_power_state(pci_dev, 0);
+
+	sis900_init_rxfilter(net_dev);
+
+	sis900_init_tx_ring(net_dev);
+	sis900_init_rx_ring(net_dev);
+
+	set_rx_mode(net_dev);
+
+	netif_device_attach(net_dev);
+	netif_start_queue(net_dev);
+
+	/* Workaround for EDB */
+	sis900_set_mode(ioaddr, HW_SPEED_10_MBPS, FDX_CAPABLE_HALF_SELECTED);
+
+	/* Enable all known interrupts by setting the interrupt mask. */
+	outl((RxSOVR|RxORN|RxERR|RxOK|TxURN|TxERR|TxIDLE), ioaddr + imr);
+	outl(RxENA | inl(ioaddr + cr), ioaddr + cr);
+	outl(IE, ioaddr + ier);
+
+	sis900_check_mode(net_dev, sis_priv->mii);
+
+	return 0;
+}
+#endif /* CONFIG_PM */
+
 static struct pci_driver sis900_pci_driver = {
 	.name		= SIS900_MODULE_NAME,
 	.id_table	= sis900_pci_tbl,
 	.probe		= sis900_probe,
 	.remove		= __devexit_p(sis900_remove),
+#ifdef CONFIG_PM
+	.suspend	= sis900_suspend,
+	.resume		= sis900_resume,
+#endif /* CONFIG_PM */
 };
 
 static int __init sis900_init_module(void)
#4 lore raw
Quoted 8 lines · Daniele Venzano
Quoted 3 lines · Andrew Morton

pci_set_power_state() can sleep, so we shouldn't be calling it under spin_lock_irqsave(). Is it necessary to hold the lock here?

New patch with locking completely removed, since in a similar function none was used.

OK. I think. Net driver suspend handlers in general seem a bit racy wrt interrupt activity as well as SMP. Maybe I'm missing something.

Quoted 2 lines

I think also the 8139too driver has the same locking problem in rtl8139_suspend, do you want a patch ?

Wouldn't hurt, thanks. It's one way to wake Jeff up ;)

8139too just does netif_device_detach(), whereas your sis900 patch does netif_stop_queue() and then netif_device_detach().

I don't know which is right, really. 8139too will end up with a non-stopped queue if __LINK_STATE_PRESENT is clear. The sis900 approach is certainly safe enough, but it'd be nice to know what netif_device_detach() is trying to do there.

#5 lore raw
Quoted 2 lines · Andrew Morton

OK. I think. Net driver suspend handlers in general seem a bit racy wrt interrupt activity as well as SMP. Maybe I'm missing something.

Looking at the logs I just discovered the following line buried in the middle of all that Bad: scheduling while atomic during resume:
eth0: Abnormal interrupt, status 0x03008200

This seems to happen with all the patches I sent you, but then the network card is working correctly, it seems a leftover handled correctly on the first interrupt..

Quoted 4 lines
Quoted 2 lines

I think also the 8139too driver has the same locking problem in rtl8139_suspend, do you want a patch ?

Wouldn't hurt, thanks. It's one way to wake Jeff up ;)

Attached you find a patch that just moves up the lock release before the call to pci_set_power_state. (and Jeff is CCed)

Quoted 7 lines

8139too just does netif_device_detach(), whereas your sis900 patch does netif_stop_queue() and then netif_device_detach().

I don't know which is right, really. 8139too will end up with a non-stopped queue if __LINK_STATE_PRESENT is clear. The sis900 approach is certainly safe enough, but it'd be nice to know what netif_device_detach() is trying to do there.

I tried to comment out the stop_queue call, and suspending with the network cable unplugged. The card keeps working on resume, as I was expecting.

Perhaps we could keep only the stop_queue call and still remain on the safe side, certainly doing two times the same call is dumb. With this setup (only stop_queue) I did not get any Abnormal interrupt message, but I don't know if they are related.

So in the end you get two patches, the 8139too and the sis900 with no locking and no netif_device_detach.

Both are available at http://teg.homeunix.org/kernel_patches.html

Signature · 3 lines
------------------------------
Daniele Venzano
Web: http://teg.homeunix.org
drivers/net/8139too.c +2 -1
@@ -2474,10 +2474,11 @@
 	tp->stats.rx_missed_errors += RTL_R32 (RxMissed);
 	RTL_W32 (RxMissed, 0);
 
+	spin_unlock_irqrestore (&tp->lock, flags);
+
 	pci_set_power_state (pdev, 3);
 	pci_save_state (pdev, tp->pci_state);
 
-	spin_unlock_irqrestore (&tp->lock, flags);
 	return 0;
 }
 
drivers/net/sis900.c +69 -6

124 lines not on this page · raw message