Skip to contentSkip to conversation

omap_cf works better

PATCH In review David Brownell 1 participant 2 messages May 1, 2006 – May 3, 2006 Pcmcia

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 omap_cf works better

PCMCIA core changed; this got rid of an oops, and restored some changes from upstream that got clobbered.

There's still something way goofy going on. The driver mostly behaves, but under much load something will access a bogus (and unmapped) kernel address, causing trouble. I suspect that's in the pcmcia core, not omap_cf.

Adding a "struct resource" prevented oopsing because of recent changes in PCMCIA core requirements.

Restore a changes that lost in some kernel.org merge: don't forget to unregister the PCMCIA socket when unbinding the driver.

drivers/pcmcia/omap_cf.c +9 -6
Index: osk/drivers/pcmcia/omap_cf.c
===================================================================
@@ -67,6 +67,7 @@ struct omap_cf_socket {
 	struct platform_device	*pdev;
 	unsigned long		phys_cf;
 	u_int			irq;
+	struct resource		iomem;
 };
 
 #define	POLL_INTERVAL		(2 * HZ)
@@ -112,10 +113,7 @@ static int omap_cf_get_status(struct pcm
 	if (!sp)
 		return -EINVAL;
 
-	/* FIXME power management should probably be board-specific:
-	 *  - 3VCARD vs XVCARD (OSK only handles 3VCARD)
-	 *  - POWERON (switched on/off by set_socket)
-	 */
+	/* NOTE CF is always 3VCARD */
 	if (omap_cf_present()) {
 		struct omap_cf_socket	*cf;
 
@@ -254,6 +252,9 @@ static int __init omap_cf_probe(struct d
 	default:
 		goto  fail1;
 	}
+	cf->iomem.start = cf->phys_cf;
+	cf->iomem.end = cf->iomem.end + SZ_8K - 1;
+	cf->iomem.flags = IORESOURCE_MEM;
 
 	/* pcmcia layer only remaps "real" memory */
 	cf->socket.io_offset = (unsigned long)
@@ -297,6 +298,7 @@ static int __init omap_cf_probe(struct d
 	cf->socket.features = SS_CAP_PCCARD | SS_CAP_STATIC_MAP
 				| SS_CAP_MEM_ALIGN;
 	cf->socket.map_size = SZ_2K;
+	cf->socket.io[0].res = &cf->iomem;
 
 	status = pcmcia_register_socket(&cf->socket);
 	if (status < 0)
@@ -321,6 +323,7 @@ static int __devexit omap_cf_remove(stru
 	struct omap_cf_socket *cf = dev_get_drvdata(dev);
 
 	cf->active = 0;
+	pcmcia_unregister_socket(&cf->socket);
 	del_timer_sync(&cf->timer);
 	iounmap((void __iomem *) cf->socket.io_offset);
 	release_mem_region(cf->phys_cf, SZ_8K);
@@ -334,8 +337,8 @@ static struct device_driver omap_cf_driv
 	.bus		= &platform_bus_type,
 	.probe		= omap_cf_probe,
 	.remove		= __devexit_p(omap_cf_remove),
-	.suspend 	= pcmcia_socket_dev_suspend,
-	.resume 	= pcmcia_socket_dev_resume,
+	.suspend	= pcmcia_socket_dev_suspend,
+	.resume		= pcmcia_socket_dev_resume,
 };
 
 static int __init omap_cf_init(void)
#1 lore raw
Quoted 5 lines · David Brownell

There's still something way goofy going on. The driver mostly behaves, but under much load something will access a bogus (and unmapped) kernel address, causing trouble. I suspect that's in the pcmcia core, not omap_cf.

More info on this problem. With PREEMPT enabled, the access came when testing "current" for the current preempt count, while switching IRQ handlers in the IDE system. (At least, during the time I managed to get a decent stack trace.)

With it disabled, the fault was a "scheduling while atomic" error ... but the "current->comm" was bogus, ditto the value of preempt_count(). dump_stack() didn't work...

In both cases it looked as if "current" had become garbage.

So I'm pretty comfortable thinking the issue isn't omap_cf.

But the real problem could be a variety of other places, and I don't have time to chase it down. Maybe someone else has some strange failures to report, so clues could add up. I'm not sure if the problem is only in the OMAP tree; or if it only affects CF card access.

- Dave