Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1758606Ab3J2RbI (ORCPT ); Tue, 29 Oct 2013 13:31:08 -0400 Received: from 12.mo6.mail-out.ovh.net ([178.32.125.228]:53208 "EHLO mo6.mail-out.ovh.net" rhost-flags-OK-OK-OK-FAIL) by vger.kernel.org with ESMTP id S1758588Ab3J2RbG (ORCPT ); Tue, 29 Oct 2013 13:31:06 -0400 X-Greylist: delayed 24808 seconds by postgrey-1.27 at vger.kernel.org; Tue, 29 Oct 2013 13:31:06 EDT Message-ID: <526FEEE7.7030708@overkiz.com> Date: Tue, 29 Oct 2013 18:22:47 +0100 From: boris brezillon User-Agent: Mozilla/5.0 (X11; Linux i686; rv:24.0) Gecko/20100101 Thunderbird/24.0 MIME-Version: 1.0 To: Guenter Roeck CC: Wim Van Sebroeck , Fabio Porcedda , Nicolas Ferre , Guenter Roeck , Yang Wenyou , linux-kernel@vger.kernel.org, linux-arm-kernel@lists.infradead.org, linux-watchdog@vger.kernel.org Subject: Re: [PATCH] watchdog: at91sam9_wdt: various fixes References: <20131029075028.GF19704@spo001.leaseweb.com> <1383043053-3520-1-git-send-email-b.brezillon@overkiz.com> <20131029154519.GB9266@roeck-us.net> <526FE0DA.1020308@overkiz.com> <20131029164323.GF9266@roeck-us.net> In-Reply-To: <20131029164323.GF9266@roeck-us.net> Content-Type: text/plain; charset=ISO-8859-1; format=flowed Content-Transfer-Encoding: 7bit X-Ovh-Tracer-Id: 17204032052244215823 X-Ovh-Remote: 80.245.18.66 () X-Ovh-Local: 213.186.33.20 (ns0.ovh.net) X-OVH-SPAMSTATE: OK X-OVH-SPAMSCORE: -100 X-OVH-SPAMCAUSE: gggruggvucftvghtrhhoucdtuddrfeeiledrgeejucetufdoteggodetrfcurfhrohhfihhlvgemucfqggfjnecuuegrihhlohhuthemuceftddtnecusecvtfgvtghiphhivghnthhsucdlqddutddtmd X-Spam-Check: DONE|U 0.5/N X-VR-SPAMSTATE: OK X-VR-SPAMSCORE: -100 X-VR-SPAMCAUSE: gggruggvucftvghtrhhoucdtuddrfeeiledrgeejucetufdoteggodetrfcurfhrohhfihhlvgemucfqggfjnecuuegrihhlohhuthemuceftddtnecusecvtfgvtghiphhivghnthhsucdlqddutddtmd Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Content-Length: 10060 Lines: 263 On 29/10/2013 17:43, Guenter Roeck wrote: > On Tue, Oct 29, 2013 at 05:22:50PM +0100, boris brezillon wrote: >> On 29/10/2013 16:45, Guenter Roeck wrote: >>> On Tue, Oct 29, 2013 at 11:37:33AM +0100, Boris BREZILLON wrote: >>>> Fix the secs_to_ticks macro in case 0 is passed as an argument. >>>> >>>> Rework the heartbeat calculation to increase the security margin of the >>>> watchdog reset timer. >>>> >>>> Use the min_heartbeat value instead of the calculated heartbeat value for >>>> the first watchdog reset. >>>> >>>> Signed-off-by: Boris BREZILLON >>> Hi Boris, >>> >>> can you possibly split the three changes into separate patches ? >> Sure. My first idea was to split this in 3 patches, but, as the >> buggy at91 watchdog series was >> already applied to linux-watchdog-next, I thought it would be faster >> to only provide one >> patch to fix all the issues at once. >> >>> The first is a no-brainer. Gives my opinion of my code review capabilities >>> a slight damper ;-). >>> >>> For the other two problems, it might make sense to describe >>> the problems you are trying to solve. >>> >>> Couple of comments inline. >>> >>> Thanks, >>> Guenter >>> >>> >>>> --- >>>> drivers/watchdog/at91sam9_wdt.c | 35 ++++++++++++++++++++++++----------- >>>> 1 file changed, 24 insertions(+), 11 deletions(-) >>>> >>>> diff --git a/drivers/watchdog/at91sam9_wdt.c b/drivers/watchdog/at91sam9_wdt.c >>>> index 9bd089e..f1b59f1 100644 >>>> --- a/drivers/watchdog/at91sam9_wdt.c >>>> +++ b/drivers/watchdog/at91sam9_wdt.c >>>> @@ -51,7 +51,7 @@ >>>> #define ticks_to_hz_rounddown(t) ((((t) + 1) * HZ) >> 8) >>>> #define ticks_to_hz_roundup(t) (((((t) + 1) * HZ) + 255) >> 8) >>>> #define ticks_to_secs(t) (((t) + 1) >> 8) >>>> -#define secs_to_ticks(s) (((s) << 8) - 1) >>>> +#define secs_to_ticks(s) (s ? (((s) << 8) - 1) : 0) >>> (s) >>> >>>> #define WDT_MR_RESET 0x3FFF2FFF >>>> @@ -61,6 +61,11 @@ >>>> /* Watchdog max delta/value in secs */ >>>> #define WDT_COUNTER_MAX_SECS ticks_to_secs(WDT_COUNTER_MAX_TICKS) >>>> +/* Watchdog heartbeat shift used for security margin: >>>> + * we'll try to rshift the heartbeat value with this value to secure >>>> + * the watchdog reset. */ >>>> +#define WDT_HEARTBEAT_SHIFT 2 >>>> + >>>> /* Hardware timeout in seconds */ >>>> #define WDT_HW_TIMEOUT 2 >>>> @@ -158,7 +163,9 @@ static int at91_wdt_init(struct platform_device *pdev, struct at91wdt *wdt) >>>> int err; >>>> u32 mask = wdt->mr_mask; >>>> unsigned long min_heartbeat = 1; >>>> + unsigned long max_heartbeat; >>>> struct device *dev = &pdev->dev; >>>> + int shift; >>>> tmp = wdt_read(wdt, AT91_WDT_MR); >>>> if ((tmp & mask) != (wdt->mr & mask)) { >>>> @@ -181,23 +188,27 @@ static int at91_wdt_init(struct platform_device *pdev, struct at91wdt *wdt) >>>> if (delta < value) >>>> min_heartbeat = ticks_to_hz_roundup(value - delta); >>>> - wdt->heartbeat = ticks_to_hz_rounddown(value); >>>> - if (!wdt->heartbeat) { >>>> + max_heartbeat = ticks_to_hz_rounddown(value); >>>> + if (!max_heartbeat) { >>>> dev_err(dev, >>>> "heartbeat is too small for the system to handle it correctly\n"); >>>> return -EINVAL; >>>> } >>>> - if (wdt->heartbeat < min_heartbeat + 4) { >>>> + for (shift = WDT_HEARTBEAT_SHIFT; shift > 0; shift--) { >>>> + if ((max_heartbeat >> shift) < min_heartbeat) >>>> + continue; >>>> + >>>> + wdt->heartbeat = max_heartbeat >> shift; >>>> + break; >>>> + } >>>> + >>>> + if (!shift) >>>> wdt->heartbeat = min_heartbeat; >>> Correct me if I am wrong, but it seems to me that >>> >>> if ((max_heartbeat >> 2) >= min_heartbeat) >>> wdt->heartbeat = max_heartbeat >> 2; >>> else if ((max_heartbeat >> 1) >= min_heartbeat) >>> wdt->heartbeat = max_heartbeat >> 1; >>> else >>> wdt->heartbeat = min_heartbeat; >>> >>> would accomplish the same and be easier to understand. >> This is exactly what I'm trying to accomplish. >> I used the for loop in case we ever want to change the >> WDT_HEARTBEAT_SHIFT value >> (which is unlikely to happen). >> >> I'll move to your proposition. >> >>> However, given that, I wonder if it is really necessary to bail out above if >>> max_heartbeat is 0. After all, you set heartbeat to min_heartbeat anyway >>> in this case. >> Yes it is necessary. The max_heartbeat is a configuration that >> cannot be changed once configured. >> We have to inform the user that his max_heartbeat value is too small >> to be handled correctly by the Linux kernel. >> >> If we simply use the min_heartbeat value as heartbeat and ignore the >> wrong max_heartbeat value, >> the watchdog will reset the SoC before we can ever reset the >> watchdog counter. >> >>>> + >>>> + if (max_heartbeat < min_heartbeat + 4) >>>> dev_warn(dev, >>>> "min heartbeat and max heartbeat might be too close for the system to handle it correctly\n"); >>>> - if (wdt->heartbeat < 4) >>>> - dev_warn(dev, >>>> - "heartbeat might be too small for the system to handle it correctly\n"); >>>> - } else { >>>> - wdt->heartbeat -= 4; >>>> - } >>>> if ((tmp & AT91_WDT_WDFIEN) && wdt->irq) { >>>> err = request_irq(wdt->irq, wdt_interrupt, >>>> @@ -213,7 +224,9 @@ static int at91_wdt_init(struct platform_device *pdev, struct at91wdt *wdt) >>>> tmp & wdt->mr_mask, wdt->mr & wdt->mr_mask); >>>> setup_timer(&wdt->timer, at91_ping, (unsigned long)wdt); >>>> - mod_timer(&wdt->timer, jiffies + wdt->heartbeat); >>>> + /* Use min_heartbeat the first time because the watchdog timer might >>>> + * be running for a long time when we reach this init function. */ >>> /* >>> * Multi-line comment style >>> * >>> * Not really sure I understand what this accomplishes. What problem >>> * are you trying to solve here ? >>> */ >> Sure, I'll change the comment style. >> >> What, I'm trying to explain, is that the watchdog might (or should) >> have been resetted >> before loading the linux kernel. But loading the kernel takes some >> time (uncompressing, >> low level init, ...), and if we take the heartbeat (or max_heartbeat >> / 4 in the common case) value as >> the next trigger to reset the watchdog counter, the watchdog timer >> might have already expired. >> > But increasing anything in the watchdog driver itself won't help you there. > You can not execute any kernel code before that kernel code is running. Of course, but you can at least try to minimize the time between the watchdog driver init and the first wathdog counter reset. >> Here is an example: >> - max_heartbeat configured to 8 seconds >> - min_heartbeat configured to 1 second >> => heartbeat = 2s >> - kernel boot time (before at91 watchdog is loaded) = 6 secs >> > Guess that is where I got lost. Do you mean the time from loading the driver > to starting the watchdog application ? Because the init function is only > executed after the driver is loaded, so nothing will help you if it takes > too long for the driver to load. I think there is another bug here: the driver should not be compiled as a module. Here is why: At POR the at91 SoC configure its watchdog with these default values: - enabled - min heartbeat = 0 ticks - max heartbeat = 0xfff ticks <=> 16 secs - some reset options After a POR the watchdog can only be reconfigured once (and only once). This configuration oftenly takes place in the the bootstrap (or bootloader), but can eventually be done by the Linux kernel. If the watchdog is kept enabled by the bootstrap (or bootloader), this means the linux kernel will have to reset the watchdog counter as soon as possible to avoid a possible watchdog reset. => If we authorize the at91 sam9 watchdog to be compiled as a module, we're not sure the module will be loaded soon enough to be able to reset the watchdog counter. > > You really have two times to deal with: > - Time from ROMMON handoff to loading the driver > Nothing you can do there. If the watchdog fires before the driver is loaded, > you are lost. Only way t handle this situation is to increase the timeout > in the ROMMON. > - Time from loading driver to watchdog device open. This is really the time > you are increasing with your change. This is where it gets a bit tricky. The heartbeat I'm talking about is not the "user space" heartbeat (struct watchdog_device timeout field), it's the "hardware watchdog counter reset" heartbeat (struct at91wdt heartbeat field). Actually the at91 sam9 wdt driver does not provide a direct access to the watchdog reset counter functionnality. Instead, it periodically reset the watchdog counter (based on the timing config retrieved from the hw registers), and eventually, if the user open a the watchdog dev file and fails to reset the watchdog (using the user space API), the drivers stops resetting the hw counter, which leads to a watchdog reset. I hope I'm clear enough, cause it's quite complicated to explain. Best Regards, Boris > Thanks, > Guenter > >> If I use the heartbeat value when configuring the first expiration >> of the timer, >> it might be too late to reset the watchdog counter. >> >> I'll try to find a proper to explain this use case :-). >> >>>> + mod_timer(&wdt->timer, jiffies + min_heartbeat); >>>> /* Try to set timeout from device tree first */ >>>> if (watchdog_init_timeout(&wdt->wdd, 0, dev)) >>>> -- >>>> 1.7.9.5 >>>> >>>> -- >>>> To unsubscribe from this list: send the line "unsubscribe linux-watchdog" in >>>> the body of a message to majordomo@vger.kernel.org >>>> More majordomo info at http://vger.kernel.org/majordomo-info.html >>>> >> -- >> To unsubscribe from this list: send the line "unsubscribe linux-watchdog" in >> the body of a message to majordomo@vger.kernel.org >> More majordomo info at http://vger.kernel.org/majordomo-info.html >> -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/