On Sun, Nov 22, 2020 at 07:08:39PM -0700, Jerry Hoemann wrote:
> NMIs received during the crash path are problematic as hpwdt_pretimeout
> handling of the NMI would cause a reentry into kdump.
>
> The situation is complicated in that I/O errors can be signaled as NMI
> circumventing hpwdt_pretimeout's attempt to not claim NMI not associated
> with either the WDT or the iLO NMI switch. These NMI can additionally
> cause a secondary NMI which cause the system to hang.
>
> By disabling pretimeout and hpwdtimeout in crash path we both reduce
> the risk of receiving an NMI and simuletaneously leave the WDT running
> (if it was already in use) to allow the WDT to break the system out of
> hangs by the WDT reset.
>
> Signed-off-by: Jerry Hoemann <jerry.hoemann@xxxxxxx>
Reviewed-by: Guenter Roeck <linux@xxxxxxxxxxxx>
> ---
> drivers/watchdog/hpwdt.c | 6 ++++++
> 1 file changed, 6 insertions(+)
>
> diff --git a/drivers/watchdog/hpwdt.c b/drivers/watchdog/hpwdt.c
> index 7d34bcf..eeb4df2 100644
> --- a/drivers/watchdog/hpwdt.c
> +++ b/drivers/watchdog/hpwdt.c
> @@ -21,6 +21,7 @@
> #include <linux/types.h>
> #include <linux/watchdog.h>
> #include <asm/nmi.h>
> +#include <linux/crash_dump.h>
>
> #define HPWDT_VERSION "2.0.3"
> #define SECS_TO_TICKS(secs) ((secs) * 1000 / 128)
> @@ -334,6 +335,11 @@ static int hpwdt_init_one(struct pci_dev *dev,
> watchdog_set_nowayout(&hpwdt_dev, nowayout);
> watchdog_init_timeout(&hpwdt_dev, soft_margin, NULL);
>
> + if (is_kdump_kernel()) {
> + pretimeout = 0;
> + kdumptimeout = 0;
> + }
> +
> if (pretimeout && hpwdt_dev.timeout <= PRETIMEOUT_SEC) {
> dev_warn(&dev->dev, "timeout <= pretimeout. Setting pretimeout to zero\n");
> pretimeout = 0;
> --
> 1.8.3.1
>
