Re: [REGRESSION] RLIMIT_DATA crashes named

From: Konstantin Khlebnikov
Date: Sat Sep 17 2016 - 08:09:23 EST


On Sat, Sep 17, 2016 at 12:09 PM, Cyrill Gorcunov <gorcunov@xxxxxxxxx> wrote:
> On Sat, Sep 17, 2016 at 11:33:56AM +0300, Konstantin Khlebnikov wrote:
>> >
>> > do_just_once just isn't a good name for a global
>> > rate limited mechanism that does something very
>> > different than the name.
>> >
>> > Maybe allow_once_per_ratelimit or the like
>> >
>> > There could be an equivalent do_once
>> >
>> > https://lkml.org/lkml/2009/5/22/3
>> >
>>
>> What about this printk_reriodic() and pr_warn_once_per_minute()?
>>
>> It simply remembers next jiffies to print rather than using that
>> complicated ratelimiting engine.
>
> +#define printk_periodic(period, fmt, ...) \
> +({ \
> + static unsigned long __print_next __read_mostly = INITIAL_JIFFIES; \
> + bool __do_print = time_after_eq(jiffies, __print_next); \
> + \
> + if (__do_print) { \
> + __print_next = jiffies + (period); \
> + printk(fmt, ##__VA_ARGS__); \
> + } \
> + unlikely(__do_print); \
> +})
>
> Seems I don't understand the bottom unlikely...

This is gcc extrension: https://gcc.gnu.org/onlinedocs/gcc/Statement-Exprs.html
Here macro works as a function which returns bool

After second though macro should update __print_next if it's too far
if first warning happens too late here will long period of silence
untill next jiffies overlap.

something like

#define printk_periodic(period, fmt, ...)
({
static unsigned long __print_next = INITIAL_JIFFIES;
unsigned long __print_jiffies = jiffies;
bool __do_print = time_after_eq(__print_jiffies, __print_next);

if (__do_print) {
__print_next = __print_jiffies + (period);
printk(fmt, ##__VA_ARGS__);
} else if (time_after(__print_next, __print_jiffies + (period))
__print_next = __print_jiffies + (period);
unlikely(__do_print);
})