Re: [PATCHv3 1/2] watchdog/softlockup: low-overhead detection of interrupt storm

[Doug Anderson]

Hi,

On Wed, Jan 31, 2024 at 9:17 AM Bitao Hu <yaoma@xxxxxxxxxxxxxxxxx> wrote:
>
> The following softlockup is caused by interrupt storm, but it cannot be
> identified from the call tree. Because the call tree is just a snapshot
> and doesn't fully capture the behavior of the CPU during the soft lockup.
>   watchdog: BUG: soft lockup - CPU#28 stuck for 23s! [fio:83921]
>   ...
>   Call trace:
>     __do_softirq+0xa0/0x37c
>     __irq_exit_rcu+0x108/0x140
>     irq_exit+0x14/0x20
>     __handle_domain_irq+0x84/0xe0
>     gic_handle_irq+0x80/0x108
>     el0_irq_naked+0x50/0x58
>
> Therefore，I think it is necessary to report CPU utilization during the
> softlockup_thresh period (report once every sample_period, for a total
> of 5 reportings), like this:
>   watchdog: BUG: soft lockup - CPU#28 stuck for 23s! [fio:83921]
>   CPU#28 Utilization every 4s during lockup:
>     #1: 0% system, 0% softirq, 100% hardirq, 0% idle
>     #2: 0% system, 0% softirq, 100% hardirq, 0% idle
>     #3: 0% system, 0% softirq, 100% hardirq, 0% idle
>     #4: 0% system, 0% softirq, 100% hardirq, 0% idle
>     #5: 0% system, 0% softirq, 100% hardirq, 0% idle
>   ...
>
> This would be helpful in determining whether an interrupt storm has
> occurred or in identifying the cause of the softlockup. The criteria for
> determination are as follows:
>   a. If the hardirq utilization is high, then interrupt storm should be
>   considered and the root cause cannot be determined from the call tree.
>   b. If the softirq utilization is high, then we could analyze the call
>   tree but it may cannot reflect the root cause.
>   c. If the system utilization is high, then we could analyze the root
>   cause from the call tree.
>
> Signed-off-by: Bitao Hu <yaoma@xxxxxxxxxxxxxxxxx>
> ---
>  kernel/watchdog.c | 84 +++++++++++++++++++++++++++++++++++++++++++++++
>  1 file changed, 84 insertions(+)

Random high-level question: I'm trying to figure out exactly when your
code will trigger. The only way it will trigger is if the timer
interrupt is a higher priority than the storming interrupt. By this I
don't mean that the timer will interrupt the storming one (it's not a
nested interrupt), but that if both interrupts are currently asserted
we'll service the timer first.

If the storming interrupt is always serviced before the timer
interrupt then the softlockup code won't trigger at all. In that case
we should detect a hard lockup and hopefully you've got the buddy
detector enabled and pseudo-NMI turned on. Then hopefully we'll have
actually interrupted the storming interrupt and it'll be on the
callstack.

I just wanted to make sure I was understanding correctly. This is why
you don't print the stats from watchdog_hardlockup_check() because
they're not useful there, right?


> diff --git a/kernel/watchdog.c b/kernel/watchdog.c
> index 81a8862295d6..046507be4eb5 100644
> --- a/kernel/watchdog.c
> +++ b/kernel/watchdog.c
> @@ -23,6 +23,8 @@
>  #include <linux/sched/debug.h>
>  #include <linux/sched/isolation.h>
>  #include <linux/stop_machine.h>
> +#include <linux/kernel_stat.h>
> +#include <linux/math64.h>

nit: instead of adding to the end, add these in sorted order. The
includes we have now are _almost_ in sorted order. I'd add these
between "init.h" and "module.h"


>  #include <asm/irq_regs.h>
>  #include <linux/kvm_para.h>
> @@ -441,6 +443,85 @@ static int is_softlockup(unsigned long touch_ts,
>         return 0;
>  }
>
> +#ifdef CONFIG_IRQ_TIME_ACCOUNTING

In v1 I think I suggested adding a new config. Even with your
optimizations you've quoted this as taking up "237,568 bytes" of
global storage when things are configured for the max number of CPUs.
It feels like someone might not want that. Adding a new Kconfig knob
shouldn't be a huge problem. Maybe you can have it default to "yes" if
the max number of CPUs is <= 64 or 128 or something?


> +#define NUM_STATS_GROUPS       5
> +enum stats_per_group {
> +       STATS_SYSTEM,
> +       STATS_SOFTIRQ,
> +       STATS_HARDIRQ,
> +       STATS_IDLE,
> +       NUM_STATS_PER_GROUP,
> +};
> +static enum cpu_usage_stat stats[NUM_STATS_PER_GROUP] = {

"static const", not just "static"

nit: maybe call this "tracked_stats" since "stats" is a bit of a
generic name for a global.


> +       CPUTIME_SYSTEM,
> +       CPUTIME_SOFTIRQ,
> +       CPUTIME_IRQ,
> +       CPUTIME_IDLE,
> +};
> +static DEFINE_PER_CPU(u16, cpustat_old[NUM_STATS_PER_GROUP]);
> +static DEFINE_PER_CPU(u8, cpustat_utilization[NUM_STATS_GROUPS][NUM_STATS_PER_GROUP]);
> +static DEFINE_PER_CPU(u8, cpustat_tail);
> +
> +/*
> + * We don't need nanosecond resolution. A granularity of 16ms is
> + * sufficient for our precision, allowing us to use u16 to store
> + * cpustats, which will roll over roughly every ~1000 seconds.
> + * 2^24 ~= 16 * 10^6
> + */
> +static u16 get_16bit_precision(u64 data)

nit: instead of "data", call it "data_ns"


> +{
> +       return data >> 24LL; /* 2^24ns ~= 16.8ms */
> +}
> +
> +static void update_cpustat(void)
> +{
> +       u8 i;

FWIW, Andrew Morton (who will likely be the one landing this patch)
was quoted in LWN [1] the other week saying that "i" should be an
integer. :-P Making it an "int" won't make the code any less
efficient.

[1] https://lwn.net/Articles/958417/


> +       u16 old;
> +       u8 utilization;
> +       u8 tail = __this_cpu_read(cpustat_tail);
> +       struct kernel_cpustat kcpustat;
> +       u64 *cpustat = kcpustat.cpustat;
> +       u16 sample_period_ms = get_16bit_precision(sample_period);

It's not really milliseconds, right? Maybe "sample_period_16"?


> +       kcpustat_cpu_fetch(&kcpustat, smp_processor_id());
> +       for (i = STATS_SYSTEM; i < NUM_STATS_PER_GROUP; i++) {

nit: start i as 0 instead of assuming that STATS_SYSTEM is 0.


> +               old = __this_cpu_read(cpustat_old[i]);
> +               cpustat[stats[i]] = get_16bit_precision(cpustat[stats[i]]);

IMO make a local called "new" and store the 16-bit precision there.
That's easier to read, gets rid of the cast below, and is probably
more efficient (the compiler doesn't need to upcast the 16-bit value
and store it in a 64-bit memory location). ...oh, or maybe "new" is a
reserved keyword? You could call them "old_stat_16" and "new_stat_16".


> +               utilization = 100 * (u16)(cpustat[stats[i]] - old) / sample_period_ms;

Maybe slightly better to round, with:

utilization = DIV_ROUND_UP(100 * (new - old), sample_period_ms);

What do you think?


> +               __this_cpu_write(cpustat_utilization[tail][i], utilization);
> +               __this_cpu_write(cpustat_old[i], cpustat[stats[i]]);
> +       }
> +       __this_cpu_write(cpustat_tail, (tail + 1) % NUM_STATS_GROUPS);
> +}
> +
> +static void print_cpustat(void)
> +{
> +       u8 i, j;
> +       u8 tail = __this_cpu_read(cpustat_tail);
> +       u64 sample_period_second = sample_period;
> +
> +       do_div(sample_period_second, NSEC_PER_SEC);
> +       /*
> +        * We do not want the "watchdog: " prefix on every line,
> +        * hence we use "printk" instead of "pr_crit".
> +        */
> +       printk(KERN_CRIT "CPU#%d Utilization every %llus during lockup:\n",
> +               smp_processor_id(), sample_period_second);
> +       for (j = STATS_SYSTEM, i = tail; j < NUM_STATS_GROUPS;

Here initting "j" to STATS_SYSTEM definitely doesn't make sense. Init to 0.

You could also make your loop easier to understand with just:

for (i = 0; i < NUM_STATS_GROUPS; i++) {
  unsigned int group = (tail + i) % NUM_STATS_GROUPS;


-Doug