From 234baefc912309e1ee2eb05354617c52775bda87 Mon Sep 17 00:00:00 2001 From: Luis R. Rodriguez Date: Tue, 14 Sep 2010 19:03:57 -0700 Subject: [PATCH 1/2] 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 --- drivers/net/wireless/ath/ath9k/main.c | 5 ++++- drivers/net/wireless/ath/ath9k/recv.c | 3 +++ 2 files changed, 7 insertions(+), 1 deletions(-) diff --git a/drivers/net/wireless/ath/ath9k/main.c b/drivers/net/wireless/ath/ath9k/main.c index e55f718..b096c2a 100644 --- a/drivers/net/wireless/ath/ath9k/main.c +++ b/drivers/net/wireless/ath/ath9k/main.c @@ -2745,6 +2745,8 @@ static int ath9k_config(struct ieee80211_hw *hw, u32 changed) * IEEE80211_CONF_CHANGE_PS is only passed by mac80211 for STA mode. */ if (changed & IEEE80211_CONF_CHANGE_PS) { + unsigned long flags; + spin_lock_irqsave(&sc->sc_pm_lock, flags); if (conf->flags & IEEE80211_CONF_PS) { sc->sc_flags |= SC_OP_PS_ENABLED; /* @@ -2759,7 +2761,7 @@ static int ath9k_config(struct ieee80211_hw *hw, u32 changed) sc->ps_enabled = false; sc->sc_flags &= ~(SC_OP_PS_ENABLED | SC_OP_NULLFUNC_COMPLETED); - ath9k_setpower(sc, ATH9K_PM_AWAKE); + ath9k_hw_setpower(sc->sc_ah, ATH9K_PM_AWAKE); if (!(ah->caps.hw_caps & ATH9K_HW_CAP_AUTOSLEEP)) { ath9k_hw_setrxabort(sc->sc_ah, 0); @@ -2774,6 +2776,7 @@ static int ath9k_config(struct ieee80211_hw *hw, u32 changed) } } } + spin_unlock_irqrestore(&sc->sc_pm_lock, flags); } if (changed & IEEE80211_CONF_CHANGE_CHANNEL) { diff --git a/drivers/net/wireless/ath/ath9k/recv.c b/drivers/net/wireless/ath/ath9k/recv.c index 477365e..b536197 100644 --- a/drivers/net/wireless/ath/ath9k/recv.c +++ b/drivers/net/wireless/ath/ath9k/recv.c @@ -491,6 +491,7 @@ int ath_rx_tasklet(struct ath_softc *sc, int flush) struct ieee80211_hdr *hdr; int retval; bool decrypt_error = false; + unsigned long flags; spin_lock_bh(&sc->rx.rxbuflock); @@ -631,10 +632,12 @@ int ath_rx_tasklet(struct ath_softc *sc, int flush) sc->rx.rxotherant = 0; } + spin_lock_irqsave(&sc->sc_pm_lock, flags); if (unlikely(sc->sc_flags & (SC_OP_WAIT_FOR_BEACON | SC_OP_WAIT_FOR_CAB | SC_OP_WAIT_FOR_PSPOLL_DATA))) ath_rx_ps(sc, skb); + spin_unlock_irqrestore(&sc->sc_pm_lock, flags); ath_rx_send_to_mac80211(hw, sc, skb, rxs); -- 1.7.0.4 From a9fcd9d79a0b8f730a136d79a99548b821fb8d57 Mon Sep 17 00:00:00 2001 From: Senthil Balasubramanian Date: Tue, 14 Sep 2010 23:44:22 -0700 Subject: [PATCH 2/2] ath9k: fix regression which prevents chip sleep after CAB data The patch: commit 293dc5dfdbcc16cde06e40a688394cc8ab083e48 Author: Gabor Juhos Date: Fri Jun 19 12:17:48 2009 +0200 ath9k: remove ath_rx_ps_back_to_sleep helper This helper only clears the SC_OP_WAIT_FOR_{BEACON,CAB} flags. Remove it and clear these flags directly in the approptiate places instead. Changes-licensed-under: ISC Signed-off-by: Gabor Juhos Signed-off-by: John W. Linville introduced a regression which forgot to lift the beacon flag after we received all broadcast and multicast data. This meant we never went to sleep consuming about ~650mW on idle. This pretty much broke power save completely. This patch has fixes for stable kernels [2.6.32+]. Cc: stable@kernel.org Cc: Paul Stewart Cc: Sameer Nanda Cc: Gabor Juhos Cc: Amod Bodas Signed-off-by: Senthil Balasubramanian Signed-off-by: Luis R. Rodriguez --- drivers/net/wireless/ath/ath9k/recv.c | 2 +- 1 files changed, 1 insertions(+), 1 deletions(-) diff --git a/drivers/net/wireless/ath/ath9k/recv.c b/drivers/net/wireless/ath/ath9k/recv.c index b536197..d41759d 100644 --- a/drivers/net/wireless/ath/ath9k/recv.c +++ b/drivers/net/wireless/ath/ath9k/recv.c @@ -420,7 +420,7 @@ static void ath_rx_ps(struct ath_softc *sc, struct sk_buff *skb) * No more broadcast/multicast frames to be received at this * point. */ - sc->sc_flags &= ~SC_OP_WAIT_FOR_CAB; + sc->sc_flags &= ~(SC_OP_WAIT_FOR_CAB | SC_OP_WAIT_FOR_BEACON); ath_print(common, ATH_DBG_PS, "All PS CAB frames received, back to sleep\n"); } else if ((sc->sc_flags & SC_OP_WAIT_FOR_PSPOLL_DATA) && -- 1.7.0.4