Return-path: Received: from mail-wi0-f174.google.com ([209.85.212.174]:38629 "EHLO mail-wi0-f174.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751886AbbFWH25 (ORCPT ); Tue, 23 Jun 2015 03:28:57 -0400 Received: by wibdq8 with SMTP id dq8so7532923wib.1 for ; Tue, 23 Jun 2015 00:28:56 -0700 (PDT) MIME-Version: 1.0 In-Reply-To: References: <1434973385-15865-1-git-send-email-janusz.dziedzic@tieto.com> <1434973385-15865-2-git-send-email-janusz.dziedzic@tieto.com> Date: Tue, 23 Jun 2015 09:28:55 +0200 Message-ID: (sfid-20150623_092901_592208_82F6FE01) Subject: Re: [PATCH 1/3] ath9k: handle RoC abort correctly From: Janusz Dziedzic To: Krishna Chaitanya Cc: Michal Kazior , linux-wireless , ath9k-devel , Felix Fietkau , Sujith Manoharan Content-Type: text/plain; charset=UTF-8 Sender: linux-wireless-owner@vger.kernel.org List-ID: On 22 June 2015 at 16:01, Krishna Chaitanya wrote: > On Mon, Jun 22, 2015 at 6:32 PM, Michal Kazior wrote: >> On 22 June 2015 at 13:56, Krishna Chaitanya wrote: >>> On Mon, Jun 22, 2015 at 5:13 PM, Janusz Dziedzic >>> wrote: >>>> In case we will get ROC abort we should not call >>>> ieee80211_remain_on_channel_expired(). >>>> >>>> In other case I hit such warning on MIPS and >>>> p2p negotiation failed (tested with use_chanctx=1). >>>> >>>> ath: phy0: Starting RoC period >>>> ath: phy0: Channel definition created: 2412 MHz >>>> ath: phy0: Assigned next_chan to 2412 MHz >>>> ath: phy0: Offchannel duration for chan 2412 MHz : 506632 >>>> ath: phy0: ath_chanctx_set_next: current: 2412 MHz, next: 2412 MHz >>>> ath: phy0: Stopping current chanctx: 2412 >>>> ath: phy0: Flush timeout: 200 >>>> ath: phy0: ath_chanctx_set_next: Set channel 2412 MHz >>>> ath: phy0: Set channel: 2412 MHz width: 0 >>>> ath: phy0: Reset to 2412 MHz, HT40: 0 fastcc: 0 >>>> ath: phy0: cur_chan: 2412 MHz, event: ATH_CHANCTX_EVENT_TSF_TIMER, state: ATH_CHANCTX_STATE_IDLE >>>> ath: phy0: ath_offchannel_channel_change: offchannel state: ATH_OFFCHANNEL_ROC_START >>>> ath: phy0: cur_chan: 2412 MHz, event: ATH_CHANCTX_EVENT_SWITCH, state: ATH_CHANCTX_STATE_IDLE >>>> ath: phy0: Cancel RoC >>>> ath: phy0: RoC aborted >>>> ath: phy0: RoC request on vif: 00:03:7f:4e:a0:cd, type: 1 duration: 500 >>>> ath: phy0: Starting RoC period >>>> ath: phy0: Channel definition created: 2412 MHz >>>> ath: phy0: Assigned next_chan to 2412 MHz >>>> ath: phy0: Offchannel duration for chan 2412 MHz : 506705 >>>> ath: phy0: ath_chanctx_set_next: current: 2412 MHz, next: 2412 MHz >>>> ath: phy0: ath_offchannel_channel_change: offchannel state: ATH_OFFCHANNEL_ROC_START >>>> ath: phy0: cur_chan: 2412 MHz, event: ATH_CHANCTX_EVENT_SWITCH, state: ATH_CHANCTX_STATE_IDLE >>>> ------------[ cut here ]------------ >>>> WARNING: CPU: 0 PID: 3312 at drivers/net/wireless/ath/ath9k/main.c:2319 >>>> Modules linked in: ath9k ath9k_common ath9k_hw ath mac80211 cfg80211 >>>> >>>> Signed-off-by: Janusz Dziedzic >>>> --- >>>> drivers/net/wireless/ath/ath9k/channel.c | 3 ++- >>>> 1 file changed, 2 insertions(+), 1 deletion(-) >>>> >>>> diff --git a/drivers/net/wireless/ath/ath9k/channel.c b/drivers/net/wireless/ath/ath9k/channel.c >>>> index 2066650..e211325 100644 >>>> --- a/drivers/net/wireless/ath/ath9k/channel.c >>>> +++ b/drivers/net/wireless/ath/ath9k/channel.c >>>> @@ -926,7 +926,8 @@ void ath_roc_complete(struct ath_softc *sc, bool abort) >>>> >>>> sc->offchannel.roc_vif = NULL; >>>> sc->offchannel.roc_chan = NULL; >>>> - ieee80211_remain_on_channel_expired(sc->hw); >>>> + if (!abort) >>>> + ieee80211_remain_on_channel_expired(sc->hw); >>>> ath_offchannel_next(sc); >>>> ath9k_ps_restore(sc); >>>> } >>> If HW aborts RoC in middle, should not we inform mac80211 >>> that RoC is expired? >> >> Good point. The ath_roc_complete() can be called with abort=true from >> ath9k_cancel_pending_offchannel() as well. I guess ath_roc_complete() >> needs a "reason" argument (instead of "abort") with: expired, aborted, >> cancelled values. ieee80211_remain_on_channel_expired() should be >> called whenever reason != cancelled. > Agree, make sense. >> By the way - is ath_roc_complete() lock protected properly? It looks >> like it isn't from a quick glance. Neither sdata lock nor local->mtx >> can be implied in all contexts and sc->mutex isn't always held while >> it's called, hmm.. or am I missing something? >> >>> Also the we are clearing roc_vif independent of abort, so the warning >>> indicates that roc_complete has not come from FW, may be we should >>> understand that first? >> >> There's no FW in ath9k. >> >> The problem is the following sequence: >> 1. mac80211 requests roc A >> 2. mac80211 cancels roc A >> a. ath9k calls expired() and hw_roc_done work is scheduled >> 3. mac80211 requests roc B >> 4. mac80211 starts to process the scheduled hw_roc_done >> 5. mac80211 thinks roc B has expired >> 6. mac80211 requests roc C >> 7. ath9k WARN_ON is hit >> >> There's a race between (3) and (4). Depending on circumstances (3) and >> (4) may be reordered so the current code doesn't fail all the time. > Ok i understand, but if we get roc_complete for B before 6, then it works > fine at least at ath9k level, C will be unblocked. > > Anyways, handling the cancel case should resolve it along with proper locking. Thanks for comments, will send v2 BR Janusz