The routine b43_is_hw_radio_enabled() has long been a problem.
For PPC architecture with PHY Revision < 3, a read of the register
B43_MMIO_HWENABLED_LO will cause a CPU fault unless b43_status()
returns a value of 2 (B43_STAT_STARTED) (BUG 14181). Fixing that
results in Bug 14538 in which the driver is unable to reassociate
after resuming from hibernation because b43_status() returns 0.
The correct fix would be to determine why the status is 0; however,
I have not yet found why that happens. The correct value is found for
my device, which has PHY revision >= 3.
Returning TRUE when the PHY revision < 3 and b43_status() returns 0 fixes
the regression for 2.6.32.
This patch fixes the problem in Red Hat Bugzilla #538523.
Signed-off-by: Larry Finger <[email protected]>
Tested-by: Christian Casteyde <[email protected]>
---
John,
Your suggested change was a good one and has been implemented. This is
2.6.32 material and should be backported to stable; however, it may not
apply cleanly. Should I make it Cc: stable, or submit it to GregKH in the
proper form once it is committed to mainline?
Larry
---
Index: wireless-testing/drivers/net/wireless/b43/rfkill.c
===================================================================
--- wireless-testing.orig/drivers/net/wireless/b43/rfkill.c
+++ wireless-testing/drivers/net/wireless/b43/rfkill.c
@@ -33,8 +33,14 @@ bool b43_is_hw_radio_enabled(struct b43_
& B43_MMIO_RADIO_HWENABLED_HI_MASK))
return 1;
} else {
- if (b43_status(dev) >= B43_STAT_STARTED &&
- b43_read16(dev, B43_MMIO_RADIO_HWENABLED_LO)
+ /* To prevent CPU fault on PPC, do not read a register
+ * unless the interface is started; however, on resume
+ * for hibernation, this routine is entered early. When
+ * that happens, unconditionally return TRUE.
+ */
+ if (b43_status(dev) < B43_STAT_STARTED)
+ return 1;
+ if (b43_read16(dev, B43_MMIO_RADIO_HWENABLED_LO)
& B43_MMIO_RADIO_HWENABLED_LO_MASK)
return 1;
}