From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1750909AbdAXLHP (ORCPT ); Tue, 24 Jan 2017 06:07:15 -0500 Received: from foss.arm.com ([217.140.101.70]:39186 "EHLO foss.arm.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1750804AbdAXLHO (ORCPT ); Tue, 24 Jan 2017 06:07:14 -0500 Subject: Re: [PATCH 1/4] PCI: Xilinx NWL: Fix, do not check for legacy status in while loop To: Bharat Kumar Gogada , "bhelgaas@google.com" , "paul.gortmaker@windriver.com" , "robh@kernel.org" , "colin.king@canonical.com" , "linux-pci@vger.kernel.org" References: <1484997072-19276-1-git-send-email-bharatku@xilinx.com> <8520D5D51A55D047800579B094147198263CCF96@XAP-PVEXMBX02.xlnx.xilinx.com> Cc: "michal.simek@xilinx.com" , "linux-arm-kernel@lists.infradead.org" , "linux-kernel@vger.kernel.org" , Ravikiran Gummaluri , "arnd@arndb.de" From: Marc Zyngier X-Enigmail-Draft-Status: N1110 Organization: ARM Ltd Message-ID: <0def9067-5c35-f7fe-8e15-bab70d479d1e@arm.com> Date: Tue, 24 Jan 2017 11:07:09 +0000 User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:45.0) Gecko/20100101 Icedove/45.5.1 MIME-Version: 1.0 In-Reply-To: <8520D5D51A55D047800579B094147198263CCF96@XAP-PVEXMBX02.xlnx.xilinx.com> Content-Type: text/plain; charset=windows-1252 Content-Transfer-Encoding: 7bit Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On 24/01/17 10:15, Bharat Kumar Gogada wrote: >> On 21/01/17 11:11, Bharat Kumar Gogada wrote: >>> - The legacy status register value for particular INTx becomes low >>> only after DEASSERT_INTx is received. >>> - Few End Points take time for sending DEASSERT_INTx, checking legacy >>> status register in while loop causes invoking of EP handler >>> continuosly until DEASSERT_INTx is received. >>> >>> Signed-off-by: Bharat Kumar Gogada >>> --- >>> drivers/pci/host/pcie-xilinx-nwl.c | 5 +++-- >>> 1 files changed, 3 insertions(+), 2 deletions(-) >>> >>> diff --git a/drivers/pci/host/pcie-xilinx-nwl.c >>> b/drivers/pci/host/pcie-xilinx-nwl.c >>> index 43eaa4a..c8b5a33 100644 >>> --- a/drivers/pci/host/pcie-xilinx-nwl.c >>> +++ b/drivers/pci/host/pcie-xilinx-nwl.c >>> @@ -342,9 +342,10 @@ static void nwl_pcie_leg_handler(struct irq_desc >>> *desc) >>> >>> chained_irq_enter(chip, desc); >>> pcie = irq_desc_get_handler_data(desc); >>> + status = nwl_bridge_readl(pcie, MSGF_LEG_STATUS) & >>> + MSGF_LEG_SR_MASKALL; >>> >>> - while ((status = nwl_bridge_readl(pcie, MSGF_LEG_STATUS) & >>> - MSGF_LEG_SR_MASKALL) != 0) { >>> + if (status != 0) { >>> for_each_set_bit(bit, &status, INTX_NUM) { >>> virq = irq_find_mapping(pcie->legacy_irq_domain, >>> bit + 1); >>> >> >> But even if you only handle the interrupt once, it is still asserted, right? You exit >> the low-level exception handler, only to take the interrupt immediately again. So >> what is the gain here? >> > Whenever masking of INTx happens (like as in my next patch[PATCH 2/4]), the irq line goes low, > but status bit will be set until DEASSERT_INTx comes. > In this scenario if it is always in while loop, even though line went low it will not be seen > until DEASSERT_INTx, unnecessarily invoking EP driver. so instead of while loop > using if condition so that this change in line is noticed. But what guarantees that you will observe this DEASSERT if you return to the interrupted context early? From what I understand, your interrupt flow is the following: read status mask handler unmask If the EP takes time to send a deassert after the handler has poked it and the interrupt is still active, then the only thing that this patch buys you is that by returning to the interrupted context, you waste a lot of cycles, giving the device a chance to send its deassert. But that's just luck. Plug this on a fast CPU, and the same issue will resurface. It could also just hide a driver bug where the write acknowledging the interrupt has been posted, but has not taken effect yet. Thanks, M. -- Jazz is not dead. It just smells funny...