From: Luis R. Rodriguez Date: Thu, 16 Sep 2010 19:12:26 +0000 (-0400) Subject: ath9k: fix power save race conditions X-Git-Tag: v2.6.37-rc1~147^2~373^2^2~53 X-Git-Url: http://git.openpandora.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=8ab2cd09fecc8819bbaee2d0fd8f3a092d866ce3;p=pandora-kernel.git ath9k: fix power save race conditions ath9k has a race on putting the chip into network sleep and having registers read from hardware. The race occurs because although ath9k_ps_restore() locks its own callers it makes use of some variables which get altered in the driver at different code paths. The variables are the ps_enabled and ps_flags. This is easily reprodicible in large network environments when roaming with the wpa_supplicant simple bgscan. You'd get some 0xdeadbeef read out on certain registers such as: ath: timeout (100000 us) on reg 0x806c: 0xdeadbeef & 0x01f00000 != 0x00000000 ath: RX failed to go idle in 10 ms RXSM=0xdeadbeef ath: timeout (100000 us) on reg 0x7000: 0xdeadbeef & 0x00000003 != 0x00000000 ath: Chip reset failed The fix is to protect the ath9k_config(hw, IEEE80211_CONF_CHANGE_PS) calls with a spin_lock_irqsave() which will disable contendors for these variables from interrupt context, timers, re-entry from mac80211 on the same callback, and most importantly from ath9k_ps_restore() which is the only call which will put the device into network sleep. There are quite a few threads and bug reports on these a few of them are: https://bugs.launchpad.net/ubuntu/karmic/+source/linux/+bug/407040 http://code.google.com/p/chromium-os/issues/detail?id=5709 http://code.google.com/p/chromium-os/issues/detail?id=5943 Stable fixes apply to [2.6.32+] Cc: stable@kernel.org Cc: Paul Stewart Cc: Amod Bodas Signed-off-by: Luis R. Rodriguez Signed-off-by: John W. Linville --- Reading git-diff-tree failed