Re: [RFC][PATCH] ring-buffer: Have nested events still record running time stamp

From: Mathieu Desnoyers
Date: Fri Jun 26 2020 - 14:13:53 EST


----- On Jun 26, 2020, at 9:58 AM, rostedt rostedt@xxxxxxxxxxx wrote:

> On Thu, 25 Jun 2020 23:35:52 -0400
> Steven Rostedt <rostedt@xxxxxxxxxxx> wrote:
>
>> I have to think about this more, as I think there's a flaw in this
>> cmpxchg algorithm.
>
> Bah, that was all wrong. I need to incorporate the seq counter into the
> time stamp. Luckily I only use 59 bits for a full time stamp (gives me
> 18 years of nanoseconds), which gives me 5 bits to pay with, and all I
> need is 4 (2 for top and 2 for bottom). Now I can make the timestamp 60
> bits with 30 bits from the top and 30 from the bottom, and using two
> bits in each of those as a sequence counter.
>
> If after reading top and bottom, if the seq bits match, then the two
> match and can be put together as a valid number. Now I have this:

[...]
>
> static inline bool __rb_time_read(rb_time_t *t, u64 *ret, unsigned long *cnt)
> {
> unsigned long top, bottom;
>
> top = local_read(&t->top);
> bottom = local_read(&t->bottom);
>
> *cnt = rb_time_cnt(top);
>
> if (*cnt != rb_time_cnt(bottom))
> return false;
>
> *ret = rb_time_val(top, bottom);
> return true;
> }

If __rb_time_read or rb_time_cmpxchg are used in an interrupt over
rb_time_set (between setting top and bottom), those will never succeed.

How is this case handled ?

Thanks,

Mathieu

--
Mathieu Desnoyers
EfficiOS Inc.
http://www.efficios.com