From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: X-Spam-Checker-Version: SpamAssassin 3.4.0 (2014-02-07) on aws-us-west-2-korg-lkml-1.web.codeaurora.org X-Spam-Level: X-Spam-Status: No, score=-0.8 required=3.0 tests=DKIM_SIGNED,DKIM_VALID, DKIM_VALID_AU,HEADER_FROM_DIFFERENT_DOMAINS,MAILING_LIST_MULTI,SPF_PASS, URIBL_BLOCKED autolearn=ham autolearn_force=no version=3.4.0 Received: from mail.kernel.org (mail.kernel.org [198.145.29.99]) by smtp.lore.kernel.org (Postfix) with ESMTP id 2316BC6778C for ; Thu, 5 Jul 2018 15:45:39 +0000 (UTC) Received: from vger.kernel.org (vger.kernel.org [209.132.180.67]) by mail.kernel.org (Postfix) with ESMTP id BF3C723FF6 for ; Thu, 5 Jul 2018 15:45:38 +0000 (UTC) Authentication-Results: mail.kernel.org; dkim=pass (1024-bit key) header.d=linaro.org header.i=@linaro.org header.b="A8jksmzc" DMARC-Filter: OpenDMARC Filter v1.3.2 mail.kernel.org BF3C723FF6 Authentication-Results: mail.kernel.org; dmarc=fail (p=none dis=none) header.from=linaro.org Authentication-Results: mail.kernel.org; spf=none smtp.mailfrom=linux-kernel-owner@vger.kernel.org Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1753428AbeGEPpd (ORCPT ); Thu, 5 Jul 2018 11:45:33 -0400 Received: from mail-io0-f194.google.com ([209.85.223.194]:39508 "EHLO mail-io0-f194.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1753264AbeGEPpa (ORCPT ); Thu, 5 Jul 2018 11:45:30 -0400 Received: by mail-io0-f194.google.com with SMTP id e13-v6so8151098iof.6 for ; Thu, 05 Jul 2018 08:45:30 -0700 (PDT) DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=linaro.org; s=google; h=mime-version:in-reply-to:references:from:date:message-id:subject:to :cc; bh=c1fMqwCAJMWV4TatKeJzTL1q10tE/lGYvbJPsMXPjPI=; b=A8jksmzcaaUXmlKwh0Iv+B22WM+JTezclNlUqr5/ddQRbAXlUoVhrE149ULRHROnt7 gJLXlBbE83N68yl94MEtWvaYgIUoeHPm99Igi0htJj0vUPzJiqiaf7qjT4SYdsGFakmw TWrrvJeKSyeOhVjzZay4UXIgDaLiiPPty8S8k= X-Google-DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=1e100.net; s=20161025; h=x-gm-message-state:mime-version:in-reply-to:references:from:date :message-id:subject:to:cc; bh=c1fMqwCAJMWV4TatKeJzTL1q10tE/lGYvbJPsMXPjPI=; b=Locjw6SBju/IhdFrk78pAcfstQ6E99ViRvwcqOS1OzC0a7Ng/ijGoqVJ2Im+TbPM90 IKcCYBDmtjX/WZoQLftC9KCpMlUIOZ9tVS5xG0zrfHHvCSK6jA52MjsfuKw9bNcsR2cc iw08NGjS5abwobZQmU4o83etQhiUtFXTaXE8w7mx/8OW/3FPorRovT6x6ZMesoIvPAvV mPWqv1JITEc6ooSU9mVJRLjCuATtFRlWFKkUO4sHFYQ8a9FLvIqy3pFbERQEkGEPq2Sh RIAp2h6iFtWF/5YvX7nGCGRkoBIIaaqh43ZCls2/eNh4e4Kzkk7rzQ1dIa1IfnQBL8MX 56LQ== X-Gm-Message-State: APt69E1dNjxtXUgc2Gh+LBLGZTtzfV8o9mmVbeWLhU06NXWTgXsLbVrF 5W9OPRC6eG79fBbES2kYVwEuhKs4PdHqa302J6vssQ== X-Google-Smtp-Source: AAOMgpc597jFeyaWvOQk7ox/b2lBwvgwRxlXaFex607bSJyQX0O891xd/KcyimduBdbmCsFP9U26v49Vg6lKb9yO+7M= X-Received: by 2002:a6b:e403:: with SMTP id u3-v6mr5250590iog.131.1530805529876; Thu, 05 Jul 2018 08:45:29 -0700 (PDT) MIME-Version: 1.0 Received: by 2002:a02:818f:0:0:0:0:0 with HTTP; Thu, 5 Jul 2018 08:45:29 -0700 (PDT) In-Reply-To: <90c9c35b-f900-4d75-b813-1b6ba3ee2fc9@arm.com> References: <1530685741-20604-1-git-send-email-stefan@olimex.com> <9b8f30fd-12aa-46ba-ced9-aed38ada0059@arm.com> <20180704212951.43c6a62a@why.wild-wind.fr.eu.org> <79f3c720-cc1d-f480-8e6b-8dbc3232837a@arm.com> <90c9c35b-f900-4d75-b813-1b6ba3ee2fc9@arm.com> From: Ulf Hansson Date: Thu, 5 Jul 2018 17:45:29 +0200 Message-ID: Subject: Re: [PATCH v2 1/1] mmc: sunxi: Disable irq during pm_suspend To: Marc Zyngier Cc: Stefan Mavrodiev , Maxime Ripard , Chen-Yu Tsai , "open list:MULTIMEDIA CARD (MMC), SECURE DIGITAL (SD) AND..." , "moderated list:ARM/Allwinner sunXi SoC support" , open list Content-Type: text/plain; charset="UTF-8" Sender: linux-kernel-owner@vger.kernel.org Precedence: bulk List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On 5 July 2018 at 15:56, Marc Zyngier wrote: > On 05/07/18 13:07, Ulf Hansson wrote: >> On 5 July 2018 at 13:40, Marc Zyngier wrote: >>> On 05/07/18 12:12, Ulf Hansson wrote: >>>> On 4 July 2018 at 22:29, Marc Zyngier wrote: >>>>> On Wed, 4 Jul 2018 15:34:36 +0200 >>>>> Ulf Hansson wrote: >>>>> >>>>>> On 4 July 2018 at 13:34, Marc Zyngier wrote: >>>>>>> On 04/07/18 11:50, Ulf Hansson wrote: >>>>>>>> + Marc >>>>>>>> >>>>>>>> On 4 July 2018 at 08:28, Stefan Mavrodiev wrote: >>>>>>>>> When mmc host controller enters suspend state, the clocks are >>>>>>>>> disabled, but irqs are not. For some reason the irqchip emits >>>>>>>>> false interrupts, which causes system lock loop. >>>>>>>>> >>>>>>>>> Debug log is: >>>>>>>>> ... >>>>>>>>> sunxi-mmc 1c11000.mmc: setting clk to 52000000, rounded 51200000 >>>>>>>>> sunxi-mmc 1c11000.mmc: enabling the clock >>>>>>>>> sunxi-mmc 1c11000.mmc: cmd 13(8000014d) arg 10000 ie 0x0000bbc6 len 0 >>>>>>>>> sunxi-mmc 1c11000.mmc: irq: rq (ptrval) mi 00000004 idi 00000000 >>>>>>>>> sunxi-mmc 1c11000.mmc: cmd 6(80000146) arg 3210101 ie 0x0000bbc6 len 0 >>>>>>>>> sunxi-mmc 1c11000.mmc: irq: rq (ptrval) mi 00000004 idi 00000000 >>>>>>>>> sunxi-mmc 1c11000.mmc: cmd 13(8000014d) arg 10000 ie 0x0000bbc6 len 0 >>>>>>>>> sunxi-mmc 1c11000.mmc: irq: rq (ptrval) mi 00000004 idi 00000000 >>>>>>>>> mmc1: new DDR MMC card at address 0001 >>>>>>>>> mmcblk1: mmc1:0001 AGND3R 14.6 GiB >>>>>>>>> mmcblk1boot0: mmc1:0001 AGND3R partition 1 4.00 MiB >>>>>>>>> mmcblk1boot1: mmc1:0001 AGND3R partition 2 4.00 MiB >>>>>>>>> sunxi-mmc 1c11000.mmc: cmd 18(80003352) arg 0 ie 0x0000fbc2 len 409 >>>>>>>>> sunxi-mmc 1c11000.mmc: irq: rq (ptrval) mi 00004000 idi 00000002 >>>>>>>>> mmcblk1: p1 >>>>>>>>> sunxi-mmc 1c11000.mmc: irq: rq (null) mi 00000000 idi 00000000 >>>>>>>>> sunxi-mmc 1c11000.mmc: irq: rq (null) mi 00000000 idi 00000000 >>>>>>>>> sunxi-mmc 1c11000.mmc: irq: rq (null) mi 00000000 idi 00000000 >>>>>>>>> sunxi-mmc 1c11000.mmc: irq: rq (null) mi 00000000 idi 00000000 >>>>>>>>> and so on... >>>>>>>>> >>>>>>>>> This issue apears on eMMC cards, routed on MMC2 slot. The patch is >>>>>>>>> tested with A20-OLinuXino-MICRO/LIME/LIME2 boards. >>>>>>>>> >>>>>>>>> Fixes: 9a8e1e8cc2c0 ("mmc: sunxi: Add runtime_pm support") >>>>>>>>> Signed-off-by: Stefan Mavrodiev >>>>>>>>> --- >>>>>>>>> Changes in v2: >>>>>>>>> - Add comment why disable_irq() is necessary >>>>>>>>> >>>>>>>>> --- >>>>>>>>> drivers/mmc/host/sunxi-mmc.c | 7 +++++++ >>>>>>>>> 1 file changed, 7 insertions(+) >>>>>>>>> >>>>>>>>> diff --git a/drivers/mmc/host/sunxi-mmc.c b/drivers/mmc/host/sunxi-mmc.c >>>>>>>>> index e747259..8e7f3e3 100644 >>>>>>>>> --- a/drivers/mmc/host/sunxi-mmc.c >>>>>>>>> +++ b/drivers/mmc/host/sunxi-mmc.c >>>>>>>>> @@ -1446,6 +1446,7 @@ static int sunxi_mmc_runtime_resume(struct device *dev) >>>>>>>>> sunxi_mmc_init_host(host); >>>>>>>>> sunxi_mmc_set_bus_width(host, mmc->ios.bus_width); >>>>>>>>> sunxi_mmc_set_clk(host, &mmc->ios); >>>>>>>>> + enable_irq(host->irq); >>>>>>>>> >>>>>>>>> return 0; >>>>>>>>> } >>>>>>>>> @@ -1455,6 +1456,12 @@ static int sunxi_mmc_runtime_suspend(struct device *dev) >>>>>>>>> struct mmc_host *mmc = dev_get_drvdata(dev); >>>>>>>>> struct sunxi_mmc_host *host = mmc_priv(mmc); >>>>>>>>> >>>>>>>>> + /* >>>>>>>>> + * When clocks are off, it's possible receiving >>>>>>>>> + * fake interrupts, which will stall the system. >>>>>>>>> + * Disabling the irq will prevent this. >>>>>>>>> + */ >>>>>>>>> + disable_irq(host->irq); >>>>>>>> >>>>>>>> No, this doesn't work for shared IRQs. >>>>>>> >>>>>>> Well, in this case, it does work, because that interrupt line cannot be >>>>>>> shared with anything else, if I understand how the SoC is wired: each >>>>>>> MMC controller has a dedicated interrupt line to the GIC, and it isn't >>>>>>> shared with anything (that's on the A20 though, and I don't know about >>>>>>> other SoCs integrating the same IP). >>>>>> >>>>>> That's the problem. This may work on some SoCs but not on others. >>>>>> >>>>>>> >>>>>>>> >>>>>>>>> sunxi_mmc_reset_host(host); >>>>>>>>> sunxi_mmc_disable(host); >>>>>>>>> >>>>>>>>> -- >>>>>>>>> 2.7.4 >>>>>>>>> >>>>>>>> >>>>>>>> The only option today is to use free_irq() in runtime suspend and then >>>>>>>> re-request the irq to re-install the handler at runtime resume. >>>>>>>> >>>>>>>> That's not an optimal solution, which is pointed out in the below >>>>>>>> discussion as well. Moreover, it has also turned out using free_irq() >>>>>>>> is also problematic in cases threaded handlers are used. >>>>>>>> >>>>>>>> Here's the link to the discussion, it's not the only one I know of, so >>>>>>>> this is common problem. >>>>>>>> https://lkml.org/lkml/2016/1/28/213 >>>>>>>> >>>>>>>> Care to have a hack on the "common" solution, which in principle means >>>>>>>> adding APIs to genirq that can disable/enable handlers from being >>>>>>>> called, rather than the entire IRQ line. >>>>>>> >>>>>>> That doesn't work. You still end-up with a screaming interrupt, and you >>>>>>> will still spend 100% of your time in interrupt context for nothing. >>>>>>> >>>>>>> Eventually, the kernel will have enough (the /other/ shared handlers >>>>>>> returning IRQ_NONE all the time), and will forcefully kill that >>>>>>> particular interrupt interrupt line, meaning you end-up in the same >>>>>>> situation of having the line disabled for all the users of that >>>>>>> interrupt line. Except that now, it is disabled forever. >>>>>> >>>>>> Ahh, correct! >>>>>> >>>>>> Sounds like free_irq() is what we need. Only that it's bit heavy >>>>>> weight as we need to re-install handlers. >>>>> >>>>> BTW, free_irq() doesn't help you either in the case of a shared >>>>> handler. You'll end-up in the exact same scenario as above. >>>> >>>> In regards to the spurious interrupt storm issue, yes, I fully agree. >>>> >>>> On the other hand, in case of a shared IRQ, don't we want the genirq >>>> core to deal with disabling the IRQ, rather than the driver? >>> >>> How do you propose we do that? You have an OR gate between two device, >>> and the result of that gate is directly plugged in the interrupt controller. >>> >>> The only thing the genirq subsystem can do is take the interrupt. If >>> nobody cares, the whole interrupt *line* will eventually get disabled. >> >> Yep, something like that. That would work, right? >> >>> >>>> Also, don't forget the other related issue, which is when the IRQ >>>> handler gets invoked (not as a storm, but once is enough), either >>>> because of a spurious IRQ or because of a shared IRQ - while the >>>> device is in a low power state (runtime suspended with clock gated for >>>> example). If that happens and the handler accesses a register the >>>> handler may hang. >>> >>> Doing a free_irq() in that case is fine, as long as the rate of spurious >>> interrupts is low. >> >> Yep. >> >>> >>>>> The real solution to this is to prevent the device itself from >>>>> generating interrupts (or to forbid interrupt sharing if it isn't >>>>> possible). >>>> >>>> I fully agree that the device should be configured to not deliver >>>> interrupt, this is the first and most important step a driver should >>>> take. For example it should mask its device's IRQ register bits. >>>> >>>> However, this isn't sufficient, because of shared IRQs and buggy HWs >>>> delivering spurious IRQs. >>> >>> It *is* sufficient for shared IRQs. Actually, it is the only way to >>> sanely implement shared IRQs (you must gate the interrupt upstream of >>> the summing interrupt controller). Buggy HW is another story (and that's >>> probably the case here). >>> >>> Now: can we please get this patch merged? ;-) >> >> Right, I have applied it for fixes! > > Thanks a lot for that. > >> Thanks for the discussion! However it would be nice to reach a >> conclusion for the problem generically. > > The only thing I can come up with is to have a requester-specific > callback that would get called when doing a requester-specific > disable_irq(). This callback would have to disable the interrupt at the > source level, instead of doing it at the irqchip level (and would only > make sense for shared interrupts). > > You'd need a per-action refcount so that enable/disable can nest, and > some new APIs to request, enable and disable specific actions. > > I could look into it if there would be more than one user... Great! If you start hacking on something, please keep me in the loop. For mmc, we have all the mmc sdhci variant drivers/devices which likely suffers from similar problems, as there have been a couple of different reports. I think the link I gave was for sdhci, reported several years ago. I have also received reports for the dw_mmc host driver [1], pointing to similar problems that the irq handler can be called when the device is in a low power state and thus hanging on register accesses. At that point the reporter used CONFIG_DEBUG_SHIRQ to trigger the problems. [1] https://patchwork.kernel.org/patch/9898377/ Kind regards Uffe From mboxrd@z Thu Jan 1 00:00:00 1970 From: Ulf Hansson Subject: Re: [PATCH v2 1/1] mmc: sunxi: Disable irq during pm_suspend Date: Thu, 5 Jul 2018 17:45:29 +0200 Message-ID: References: <1530685741-20604-1-git-send-email-stefan@olimex.com> <9b8f30fd-12aa-46ba-ced9-aed38ada0059@arm.com> <20180704212951.43c6a62a@why.wild-wind.fr.eu.org> <79f3c720-cc1d-f480-8e6b-8dbc3232837a@arm.com> <90c9c35b-f900-4d75-b813-1b6ba3ee2fc9@arm.com> Mime-Version: 1.0 Content-Type: text/plain; charset="UTF-8" Return-path: In-Reply-To: <90c9c35b-f900-4d75-b813-1b6ba3ee2fc9@arm.com> Sender: linux-kernel-owner@vger.kernel.org To: Marc Zyngier Cc: Stefan Mavrodiev , Maxime Ripard , Chen-Yu Tsai , "open list:MULTIMEDIA CARD (MMC), SECURE DIGITAL (SD) AND..." , "moderated list:ARM/Allwinner sunXi SoC support" , open list List-Id: linux-mmc@vger.kernel.org On 5 July 2018 at 15:56, Marc Zyngier wrote: > On 05/07/18 13:07, Ulf Hansson wrote: >> On 5 July 2018 at 13:40, Marc Zyngier wrote: >>> On 05/07/18 12:12, Ulf Hansson wrote: >>>> On 4 July 2018 at 22:29, Marc Zyngier wrote: >>>>> On Wed, 4 Jul 2018 15:34:36 +0200 >>>>> Ulf Hansson wrote: >>>>> >>>>>> On 4 July 2018 at 13:34, Marc Zyngier wrote: >>>>>>> On 04/07/18 11:50, Ulf Hansson wrote: >>>>>>>> + Marc >>>>>>>> >>>>>>>> On 4 July 2018 at 08:28, Stefan Mavrodiev wrote: >>>>>>>>> When mmc host controller enters suspend state, the clocks are >>>>>>>>> disabled, but irqs are not. For some reason the irqchip emits >>>>>>>>> false interrupts, which causes system lock loop. >>>>>>>>> >>>>>>>>> Debug log is: >>>>>>>>> ... >>>>>>>>> sunxi-mmc 1c11000.mmc: setting clk to 52000000, rounded 51200000 >>>>>>>>> sunxi-mmc 1c11000.mmc: enabling the clock >>>>>>>>> sunxi-mmc 1c11000.mmc: cmd 13(8000014d) arg 10000 ie 0x0000bbc6 len 0 >>>>>>>>> sunxi-mmc 1c11000.mmc: irq: rq (ptrval) mi 00000004 idi 00000000 >>>>>>>>> sunxi-mmc 1c11000.mmc: cmd 6(80000146) arg 3210101 ie 0x0000bbc6 len 0 >>>>>>>>> sunxi-mmc 1c11000.mmc: irq: rq (ptrval) mi 00000004 idi 00000000 >>>>>>>>> sunxi-mmc 1c11000.mmc: cmd 13(8000014d) arg 10000 ie 0x0000bbc6 len 0 >>>>>>>>> sunxi-mmc 1c11000.mmc: irq: rq (ptrval) mi 00000004 idi 00000000 >>>>>>>>> mmc1: new DDR MMC card at address 0001 >>>>>>>>> mmcblk1: mmc1:0001 AGND3R 14.6 GiB >>>>>>>>> mmcblk1boot0: mmc1:0001 AGND3R partition 1 4.00 MiB >>>>>>>>> mmcblk1boot1: mmc1:0001 AGND3R partition 2 4.00 MiB >>>>>>>>> sunxi-mmc 1c11000.mmc: cmd 18(80003352) arg 0 ie 0x0000fbc2 len 409 >>>>>>>>> sunxi-mmc 1c11000.mmc: irq: rq (ptrval) mi 00004000 idi 00000002 >>>>>>>>> mmcblk1: p1 >>>>>>>>> sunxi-mmc 1c11000.mmc: irq: rq (null) mi 00000000 idi 00000000 >>>>>>>>> sunxi-mmc 1c11000.mmc: irq: rq (null) mi 00000000 idi 00000000 >>>>>>>>> sunxi-mmc 1c11000.mmc: irq: rq (null) mi 00000000 idi 00000000 >>>>>>>>> sunxi-mmc 1c11000.mmc: irq: rq (null) mi 00000000 idi 00000000 >>>>>>>>> and so on... >>>>>>>>> >>>>>>>>> This issue apears on eMMC cards, routed on MMC2 slot. The patch is >>>>>>>>> tested with A20-OLinuXino-MICRO/LIME/LIME2 boards. >>>>>>>>> >>>>>>>>> Fixes: 9a8e1e8cc2c0 ("mmc: sunxi: Add runtime_pm support") >>>>>>>>> Signed-off-by: Stefan Mavrodiev >>>>>>>>> --- >>>>>>>>> Changes in v2: >>>>>>>>> - Add comment why disable_irq() is necessary >>>>>>>>> >>>>>>>>> --- >>>>>>>>> drivers/mmc/host/sunxi-mmc.c | 7 +++++++ >>>>>>>>> 1 file changed, 7 insertions(+) >>>>>>>>> >>>>>>>>> diff --git a/drivers/mmc/host/sunxi-mmc.c b/drivers/mmc/host/sunxi-mmc.c >>>>>>>>> index e747259..8e7f3e3 100644 >>>>>>>>> --- a/drivers/mmc/host/sunxi-mmc.c >>>>>>>>> +++ b/drivers/mmc/host/sunxi-mmc.c >>>>>>>>> @@ -1446,6 +1446,7 @@ static int sunxi_mmc_runtime_resume(struct device *dev) >>>>>>>>> sunxi_mmc_init_host(host); >>>>>>>>> sunxi_mmc_set_bus_width(host, mmc->ios.bus_width); >>>>>>>>> sunxi_mmc_set_clk(host, &mmc->ios); >>>>>>>>> + enable_irq(host->irq); >>>>>>>>> >>>>>>>>> return 0; >>>>>>>>> } >>>>>>>>> @@ -1455,6 +1456,12 @@ static int sunxi_mmc_runtime_suspend(struct device *dev) >>>>>>>>> struct mmc_host *mmc = dev_get_drvdata(dev); >>>>>>>>> struct sunxi_mmc_host *host = mmc_priv(mmc); >>>>>>>>> >>>>>>>>> + /* >>>>>>>>> + * When clocks are off, it's possible receiving >>>>>>>>> + * fake interrupts, which will stall the system. >>>>>>>>> + * Disabling the irq will prevent this. >>>>>>>>> + */ >>>>>>>>> + disable_irq(host->irq); >>>>>>>> >>>>>>>> No, this doesn't work for shared IRQs. >>>>>>> >>>>>>> Well, in this case, it does work, because that interrupt line cannot be >>>>>>> shared with anything else, if I understand how the SoC is wired: each >>>>>>> MMC controller has a dedicated interrupt line to the GIC, and it isn't >>>>>>> shared with anything (that's on the A20 though, and I don't know about >>>>>>> other SoCs integrating the same IP). >>>>>> >>>>>> That's the problem. This may work on some SoCs but not on others. >>>>>> >>>>>>> >>>>>>>> >>>>>>>>> sunxi_mmc_reset_host(host); >>>>>>>>> sunxi_mmc_disable(host); >>>>>>>>> >>>>>>>>> -- >>>>>>>>> 2.7.4 >>>>>>>>> >>>>>>>> >>>>>>>> The only option today is to use free_irq() in runtime suspend and then >>>>>>>> re-request the irq to re-install the handler at runtime resume. >>>>>>>> >>>>>>>> That's not an optimal solution, which is pointed out in the below >>>>>>>> discussion as well. Moreover, it has also turned out using free_irq() >>>>>>>> is also problematic in cases threaded handlers are used. >>>>>>>> >>>>>>>> Here's the link to the discussion, it's not the only one I know of, so >>>>>>>> this is common problem. >>>>>>>> https://lkml.org/lkml/2016/1/28/213 >>>>>>>> >>>>>>>> Care to have a hack on the "common" solution, which in principle means >>>>>>>> adding APIs to genirq that can disable/enable handlers from being >>>>>>>> called, rather than the entire IRQ line. >>>>>>> >>>>>>> That doesn't work. You still end-up with a screaming interrupt, and you >>>>>>> will still spend 100% of your time in interrupt context for nothing. >>>>>>> >>>>>>> Eventually, the kernel will have enough (the /other/ shared handlers >>>>>>> returning IRQ_NONE all the time), and will forcefully kill that >>>>>>> particular interrupt interrupt line, meaning you end-up in the same >>>>>>> situation of having the line disabled for all the users of that >>>>>>> interrupt line. Except that now, it is disabled forever. >>>>>> >>>>>> Ahh, correct! >>>>>> >>>>>> Sounds like free_irq() is what we need. Only that it's bit heavy >>>>>> weight as we need to re-install handlers. >>>>> >>>>> BTW, free_irq() doesn't help you either in the case of a shared >>>>> handler. You'll end-up in the exact same scenario as above. >>>> >>>> In regards to the spurious interrupt storm issue, yes, I fully agree. >>>> >>>> On the other hand, in case of a shared IRQ, don't we want the genirq >>>> core to deal with disabling the IRQ, rather than the driver? >>> >>> How do you propose we do that? You have an OR gate between two device, >>> and the result of that gate is directly plugged in the interrupt controller. >>> >>> The only thing the genirq subsystem can do is take the interrupt. If >>> nobody cares, the whole interrupt *line* will eventually get disabled. >> >> Yep, something like that. That would work, right? >> >>> >>>> Also, don't forget the other related issue, which is when the IRQ >>>> handler gets invoked (not as a storm, but once is enough), either >>>> because of a spurious IRQ or because of a shared IRQ - while the >>>> device is in a low power state (runtime suspended with clock gated for >>>> example). If that happens and the handler accesses a register the >>>> handler may hang. >>> >>> Doing a free_irq() in that case is fine, as long as the rate of spurious >>> interrupts is low. >> >> Yep. >> >>> >>>>> The real solution to this is to prevent the device itself from >>>>> generating interrupts (or to forbid interrupt sharing if it isn't >>>>> possible). >>>> >>>> I fully agree that the device should be configured to not deliver >>>> interrupt, this is the first and most important step a driver should >>>> take. For example it should mask its device's IRQ register bits. >>>> >>>> However, this isn't sufficient, because of shared IRQs and buggy HWs >>>> delivering spurious IRQs. >>> >>> It *is* sufficient for shared IRQs. Actually, it is the only way to >>> sanely implement shared IRQs (you must gate the interrupt upstream of >>> the summing interrupt controller). Buggy HW is another story (and that's >>> probably the case here). >>> >>> Now: can we please get this patch merged? ;-) >> >> Right, I have applied it for fixes! > > Thanks a lot for that. > >> Thanks for the discussion! However it would be nice to reach a >> conclusion for the problem generically. > > The only thing I can come up with is to have a requester-specific > callback that would get called when doing a requester-specific > disable_irq(). This callback would have to disable the interrupt at the > source level, instead of doing it at the irqchip level (and would only > make sense for shared interrupts). > > You'd need a per-action refcount so that enable/disable can nest, and > some new APIs to request, enable and disable specific actions. > > I could look into it if there would be more than one user... Great! If you start hacking on something, please keep me in the loop. For mmc, we have all the mmc sdhci variant drivers/devices which likely suffers from similar problems, as there have been a couple of different reports. I think the link I gave was for sdhci, reported several years ago. I have also received reports for the dw_mmc host driver [1], pointing to similar problems that the irq handler can be called when the device is in a low power state and thus hanging on register accesses. At that point the reporter used CONFIG_DEBUG_SHIRQ to trigger the problems. [1] https://patchwork.kernel.org/patch/9898377/ Kind regards Uffe From mboxrd@z Thu Jan 1 00:00:00 1970 From: ulf.hansson@linaro.org (Ulf Hansson) Date: Thu, 5 Jul 2018 17:45:29 +0200 Subject: [PATCH v2 1/1] mmc: sunxi: Disable irq during pm_suspend In-Reply-To: <90c9c35b-f900-4d75-b813-1b6ba3ee2fc9@arm.com> References: <1530685741-20604-1-git-send-email-stefan@olimex.com> <9b8f30fd-12aa-46ba-ced9-aed38ada0059@arm.com> <20180704212951.43c6a62a@why.wild-wind.fr.eu.org> <79f3c720-cc1d-f480-8e6b-8dbc3232837a@arm.com> <90c9c35b-f900-4d75-b813-1b6ba3ee2fc9@arm.com> Message-ID: To: linux-arm-kernel@lists.infradead.org List-Id: linux-arm-kernel.lists.infradead.org On 5 July 2018 at 15:56, Marc Zyngier wrote: > On 05/07/18 13:07, Ulf Hansson wrote: >> On 5 July 2018 at 13:40, Marc Zyngier wrote: >>> On 05/07/18 12:12, Ulf Hansson wrote: >>>> On 4 July 2018 at 22:29, Marc Zyngier wrote: >>>>> On Wed, 4 Jul 2018 15:34:36 +0200 >>>>> Ulf Hansson wrote: >>>>> >>>>>> On 4 July 2018 at 13:34, Marc Zyngier wrote: >>>>>>> On 04/07/18 11:50, Ulf Hansson wrote: >>>>>>>> + Marc >>>>>>>> >>>>>>>> On 4 July 2018 at 08:28, Stefan Mavrodiev wrote: >>>>>>>>> When mmc host controller enters suspend state, the clocks are >>>>>>>>> disabled, but irqs are not. For some reason the irqchip emits >>>>>>>>> false interrupts, which causes system lock loop. >>>>>>>>> >>>>>>>>> Debug log is: >>>>>>>>> ... >>>>>>>>> sunxi-mmc 1c11000.mmc: setting clk to 52000000, rounded 51200000 >>>>>>>>> sunxi-mmc 1c11000.mmc: enabling the clock >>>>>>>>> sunxi-mmc 1c11000.mmc: cmd 13(8000014d) arg 10000 ie 0x0000bbc6 len 0 >>>>>>>>> sunxi-mmc 1c11000.mmc: irq: rq (ptrval) mi 00000004 idi 00000000 >>>>>>>>> sunxi-mmc 1c11000.mmc: cmd 6(80000146) arg 3210101 ie 0x0000bbc6 len 0 >>>>>>>>> sunxi-mmc 1c11000.mmc: irq: rq (ptrval) mi 00000004 idi 00000000 >>>>>>>>> sunxi-mmc 1c11000.mmc: cmd 13(8000014d) arg 10000 ie 0x0000bbc6 len 0 >>>>>>>>> sunxi-mmc 1c11000.mmc: irq: rq (ptrval) mi 00000004 idi 00000000 >>>>>>>>> mmc1: new DDR MMC card at address 0001 >>>>>>>>> mmcblk1: mmc1:0001 AGND3R 14.6 GiB >>>>>>>>> mmcblk1boot0: mmc1:0001 AGND3R partition 1 4.00 MiB >>>>>>>>> mmcblk1boot1: mmc1:0001 AGND3R partition 2 4.00 MiB >>>>>>>>> sunxi-mmc 1c11000.mmc: cmd 18(80003352) arg 0 ie 0x0000fbc2 len 409 >>>>>>>>> sunxi-mmc 1c11000.mmc: irq: rq (ptrval) mi 00004000 idi 00000002 >>>>>>>>> mmcblk1: p1 >>>>>>>>> sunxi-mmc 1c11000.mmc: irq: rq (null) mi 00000000 idi 00000000 >>>>>>>>> sunxi-mmc 1c11000.mmc: irq: rq (null) mi 00000000 idi 00000000 >>>>>>>>> sunxi-mmc 1c11000.mmc: irq: rq (null) mi 00000000 idi 00000000 >>>>>>>>> sunxi-mmc 1c11000.mmc: irq: rq (null) mi 00000000 idi 00000000 >>>>>>>>> and so on... >>>>>>>>> >>>>>>>>> This issue apears on eMMC cards, routed on MMC2 slot. The patch is >>>>>>>>> tested with A20-OLinuXino-MICRO/LIME/LIME2 boards. >>>>>>>>> >>>>>>>>> Fixes: 9a8e1e8cc2c0 ("mmc: sunxi: Add runtime_pm support") >>>>>>>>> Signed-off-by: Stefan Mavrodiev >>>>>>>>> --- >>>>>>>>> Changes in v2: >>>>>>>>> - Add comment why disable_irq() is necessary >>>>>>>>> >>>>>>>>> --- >>>>>>>>> drivers/mmc/host/sunxi-mmc.c | 7 +++++++ >>>>>>>>> 1 file changed, 7 insertions(+) >>>>>>>>> >>>>>>>>> diff --git a/drivers/mmc/host/sunxi-mmc.c b/drivers/mmc/host/sunxi-mmc.c >>>>>>>>> index e747259..8e7f3e3 100644 >>>>>>>>> --- a/drivers/mmc/host/sunxi-mmc.c >>>>>>>>> +++ b/drivers/mmc/host/sunxi-mmc.c >>>>>>>>> @@ -1446,6 +1446,7 @@ static int sunxi_mmc_runtime_resume(struct device *dev) >>>>>>>>> sunxi_mmc_init_host(host); >>>>>>>>> sunxi_mmc_set_bus_width(host, mmc->ios.bus_width); >>>>>>>>> sunxi_mmc_set_clk(host, &mmc->ios); >>>>>>>>> + enable_irq(host->irq); >>>>>>>>> >>>>>>>>> return 0; >>>>>>>>> } >>>>>>>>> @@ -1455,6 +1456,12 @@ static int sunxi_mmc_runtime_suspend(struct device *dev) >>>>>>>>> struct mmc_host *mmc = dev_get_drvdata(dev); >>>>>>>>> struct sunxi_mmc_host *host = mmc_priv(mmc); >>>>>>>>> >>>>>>>>> + /* >>>>>>>>> + * When clocks are off, it's possible receiving >>>>>>>>> + * fake interrupts, which will stall the system. >>>>>>>>> + * Disabling the irq will prevent this. >>>>>>>>> + */ >>>>>>>>> + disable_irq(host->irq); >>>>>>>> >>>>>>>> No, this doesn't work for shared IRQs. >>>>>>> >>>>>>> Well, in this case, it does work, because that interrupt line cannot be >>>>>>> shared with anything else, if I understand how the SoC is wired: each >>>>>>> MMC controller has a dedicated interrupt line to the GIC, and it isn't >>>>>>> shared with anything (that's on the A20 though, and I don't know about >>>>>>> other SoCs integrating the same IP). >>>>>> >>>>>> That's the problem. This may work on some SoCs but not on others. >>>>>> >>>>>>> >>>>>>>> >>>>>>>>> sunxi_mmc_reset_host(host); >>>>>>>>> sunxi_mmc_disable(host); >>>>>>>>> >>>>>>>>> -- >>>>>>>>> 2.7.4 >>>>>>>>> >>>>>>>> >>>>>>>> The only option today is to use free_irq() in runtime suspend and then >>>>>>>> re-request the irq to re-install the handler at runtime resume. >>>>>>>> >>>>>>>> That's not an optimal solution, which is pointed out in the below >>>>>>>> discussion as well. Moreover, it has also turned out using free_irq() >>>>>>>> is also problematic in cases threaded handlers are used. >>>>>>>> >>>>>>>> Here's the link to the discussion, it's not the only one I know of, so >>>>>>>> this is common problem. >>>>>>>> https://lkml.org/lkml/2016/1/28/213 >>>>>>>> >>>>>>>> Care to have a hack on the "common" solution, which in principle means >>>>>>>> adding APIs to genirq that can disable/enable handlers from being >>>>>>>> called, rather than the entire IRQ line. >>>>>>> >>>>>>> That doesn't work. You still end-up with a screaming interrupt, and you >>>>>>> will still spend 100% of your time in interrupt context for nothing. >>>>>>> >>>>>>> Eventually, the kernel will have enough (the /other/ shared handlers >>>>>>> returning IRQ_NONE all the time), and will forcefully kill that >>>>>>> particular interrupt interrupt line, meaning you end-up in the same >>>>>>> situation of having the line disabled for all the users of that >>>>>>> interrupt line. Except that now, it is disabled forever. >>>>>> >>>>>> Ahh, correct! >>>>>> >>>>>> Sounds like free_irq() is what we need. Only that it's bit heavy >>>>>> weight as we need to re-install handlers. >>>>> >>>>> BTW, free_irq() doesn't help you either in the case of a shared >>>>> handler. You'll end-up in the exact same scenario as above. >>>> >>>> In regards to the spurious interrupt storm issue, yes, I fully agree. >>>> >>>> On the other hand, in case of a shared IRQ, don't we want the genirq >>>> core to deal with disabling the IRQ, rather than the driver? >>> >>> How do you propose we do that? You have an OR gate between two device, >>> and the result of that gate is directly plugged in the interrupt controller. >>> >>> The only thing the genirq subsystem can do is take the interrupt. If >>> nobody cares, the whole interrupt *line* will eventually get disabled. >> >> Yep, something like that. That would work, right? >> >>> >>>> Also, don't forget the other related issue, which is when the IRQ >>>> handler gets invoked (not as a storm, but once is enough), either >>>> because of a spurious IRQ or because of a shared IRQ - while the >>>> device is in a low power state (runtime suspended with clock gated for >>>> example). If that happens and the handler accesses a register the >>>> handler may hang. >>> >>> Doing a free_irq() in that case is fine, as long as the rate of spurious >>> interrupts is low. >> >> Yep. >> >>> >>>>> The real solution to this is to prevent the device itself from >>>>> generating interrupts (or to forbid interrupt sharing if it isn't >>>>> possible). >>>> >>>> I fully agree that the device should be configured to not deliver >>>> interrupt, this is the first and most important step a driver should >>>> take. For example it should mask its device's IRQ register bits. >>>> >>>> However, this isn't sufficient, because of shared IRQs and buggy HWs >>>> delivering spurious IRQs. >>> >>> It *is* sufficient for shared IRQs. Actually, it is the only way to >>> sanely implement shared IRQs (you must gate the interrupt upstream of >>> the summing interrupt controller). Buggy HW is another story (and that's >>> probably the case here). >>> >>> Now: can we please get this patch merged? ;-) >> >> Right, I have applied it for fixes! > > Thanks a lot for that. > >> Thanks for the discussion! However it would be nice to reach a >> conclusion for the problem generically. > > The only thing I can come up with is to have a requester-specific > callback that would get called when doing a requester-specific > disable_irq(). This callback would have to disable the interrupt at the > source level, instead of doing it at the irqchip level (and would only > make sense for shared interrupts). > > You'd need a per-action refcount so that enable/disable can nest, and > some new APIs to request, enable and disable specific actions. > > I could look into it if there would be more than one user... Great! If you start hacking on something, please keep me in the loop. For mmc, we have all the mmc sdhci variant drivers/devices which likely suffers from similar problems, as there have been a couple of different reports. I think the link I gave was for sdhci, reported several years ago. I have also received reports for the dw_mmc host driver [1], pointing to similar problems that the irq handler can be called when the device is in a low power state and thus hanging on register accesses. At that point the reporter used CONFIG_DEBUG_SHIRQ to trigger the problems. [1] https://patchwork.kernel.org/patch/9898377/ Kind regards Uffe