Re: Perf loses events without reporting

[Jiri Olsa <jolsa@redhat.com>]

On Fri, Jul 02, 2021 at 06:22:10PM +0200, Jiri Olsa wrote:
> On Thu, Jul 01, 2021 at 11:00:24AM -0700, Stephen Brennan wrote:
> 
> SNIP
> 
> > 
> > static ssize_t irqsoff_write(struct file *f, const char __user *data, size_t amt, loff_t *off)
> > {
> > 	char buf[32];
> > 	int rv;
> > 	unsigned long millis = 0;
> > 
> > 	if (amt > sizeof(buf) - 1)
> > 		return -EFBIG;
> > 
> > 	if ((rv = copy_from_user(buf, data, amt)) != 0)
> > 		return -EFAULT;
> > 
> > 	buf[amt] = '\0';
> > 
> > 	if (sscanf(buf, "%lu", &millis) != 1)
> > 		return -EINVAL;
> > 
> > 	/* Threshold it at 5 minutes for safety. */
> > 	if (millis > 5 * 60 * 1000)
> > 		return -EINVAL;
> > 
> > 	pr_info("[irqoff] lock for %lu millis\n", millis);
> > 	spin_lock_irq(&irqoff_lock);
> > 	irqsoff_inirq_delay(millis);
> > 	spin_unlock_irq(&irqoff_lock);
> 
> yea, I can see that as well.. if I took the spin_lock_irq it works as
> expected.. will check

ok, so with disabling irqs you disable task tick which
kicks the unthrotling

throtling is started when you cross the limit of samples
per tick, which you do real fast with busy waiting

then when event is throttled (and no samples are generated),
the next task tick restores it.. but with disabled irqs,
there's no task tick, so the event is throttled all the
time when irqs are disabled

is there any use case you have to have irqs disabled for so long?

cc-ing peterz
initial email: https://lore.kernel.org/linux-perf-users/YN89LwSzTKc8wBar@krava/T/#m616481aa95da10e17d9b0769717f006b9ff38637

jirka