Return-path: Received: from mail2.candelatech.com ([208.74.158.173]:57952 "EHLO mail2.candelatech.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S932367AbeB1TMX (ORCPT ); Wed, 28 Feb 2018 14:12:23 -0500 From: greearb@candelatech.com To: linux-wireless@vger.kernel.org Cc: ath10k@lists.infradead.org, Ben Greear Subject: [PATCH] ath10k: Attempt to work around napi_synchronize hang. Date: Wed, 28 Feb 2018 11:12:17 -0800 Message-Id: <1519845137-15365-1-git-send-email-greearb@candelatech.com> (sfid-20180228_201232_293570_57EF3A2C) Sender: linux-wireless-owner@vger.kernel.org List-ID: From: Ben Greear Calling napi_disable twice in a row (w/out starting it and/or without having NAPI active leads to deadlock because napi_disable sets NAPI_STATE_SCHED and NAPI_STATE_NPSVC when it returns, as far as I can tell. So, guard this call to napi_disable. I believe the failure case is something like this: rmmod ath10k_pci ath10k_core Firmware crashes before hif_stop is called by the rmmod path The crash handling logic calls hif_stop Then rmmod gets around to calling hif_stop, but spins endlessly in napi_synchronize. I think one way this could happen is that ath10k_stop checks for state != ATH10K_STATE_OFF, but STATE_RESTARTING is also a possibility. That might be how we can have hif_stop called twice without a hif_start in between. --Ben Signed-off-by: Ben Greear --- * since RFC: Added similar code to ahb This seems needed back to at least 4.9 kernels. drivers/net/wireless/ath/ath10k/ahb.c | 9 +++++++-- drivers/net/wireless/ath/ath10k/core.h | 1 + drivers/net/wireless/ath/ath10k/pci.c | 25 +++++++++++++++++++++++-- 3 files changed, 31 insertions(+), 4 deletions(-) diff --git a/drivers/net/wireless/ath/ath10k/ahb.c b/drivers/net/wireless/ath/ath10k/ahb.c index da770af..f826c59 100644 --- a/drivers/net/wireless/ath/ath10k/ahb.c +++ b/drivers/net/wireless/ath/ath10k/ahb.c @@ -641,6 +641,8 @@ static int ath10k_ahb_hif_start(struct ath10k *ar) ath10k_dbg(ar, ATH10K_DBG_BOOT, "boot ahb hif start\n"); napi_enable(&ar->napi); + ar->napi_enabled = true; + ath10k_ce_enable_interrupts(ar); ath10k_pci_enable_legacy_irq(ar); @@ -660,8 +662,11 @@ static void ath10k_ahb_hif_stop(struct ath10k *ar) ath10k_pci_flush(ar); - napi_synchronize(&ar->napi); - napi_disable(&ar->napi); + if (ar->napi_enabled) { + napi_synchronize(&ar->napi); + napi_disable(&ar->napi); + ar->napi_enabled = false; + } } static int ath10k_ahb_hif_power_up(struct ath10k *ar) diff --git a/drivers/net/wireless/ath/ath10k/core.h b/drivers/net/wireless/ath/ath10k/core.h index 72b4495..c7ba49f 100644 --- a/drivers/net/wireless/ath/ath10k/core.h +++ b/drivers/net/wireless/ath/ath10k/core.h @@ -1205,6 +1205,7 @@ struct ath10k { /* NAPI */ struct net_device napi_dev; struct napi_struct napi; + bool napi_enabled; struct work_struct stop_scan_work; diff --git a/drivers/net/wireless/ath/ath10k/pci.c b/drivers/net/wireless/ath/ath10k/pci.c index 398e413..9131e15 100644 --- a/drivers/net/wireless/ath/ath10k/pci.c +++ b/drivers/net/wireless/ath/ath10k/pci.c @@ -1956,6 +1956,7 @@ static int ath10k_pci_hif_start(struct ath10k *ar) ath10k_dbg(ar, ATH10K_DBG_BOOT, "boot hif start\n"); napi_enable(&ar->napi); + ar->napi_enabled = true; ath10k_pci_irq_enable(ar); ath10k_pci_rx_post(ar); @@ -2086,8 +2087,28 @@ static void ath10k_pci_hif_stop(struct ath10k *ar) ath10k_pci_irq_disable(ar); ath10k_pci_irq_sync(ar); ath10k_pci_flush(ar); - napi_synchronize(&ar->napi); - napi_disable(&ar->napi); + + /* Calling napi_disable twice in a row (w/out starting it and/or without + * having NAPI active leads to deadlock because napi_disable sets + * NAPI_STATE_SCHED and NAPI_STATE_NPSVC when it returns, as far as I + * can tell. So, guard this call to napi_disable. I believe the + * failure case is something like this: + * rmmod ath10k_pci ath10k_core + * Firmware crashes before hif_stop is called by the rmmod path + * The crash handling logic calls hif_stop + * Then rmmod gets around to calling hif_stop, but spins endlessly + * in napi_synchronize. + * + * I think one way this could happen is that ath10k_stop checks + * for state != ATH10K_STATE_OFF, but STATE_RESTARTING is also + * a possibility. That might be how we can have hif_stop called twice + * without a hif_start in between. --Ben + */ + if (ar->napi_enabled) { + napi_synchronize(&ar->napi); + napi_disable(&ar->napi); + ar->napi_enabled = false; + } spin_lock_irqsave(&ar_pci->ps_lock, flags); WARN_ON(ar_pci->ps_wake_refcount > 0); -- 2.4.11