Messages in this thread | | | Date | Thu, 4 Jun 2015 07:59:30 +0200 | From | Ingo Molnar <> | Subject | Re: [PATCH 8/9] hrtimer: Allow hrtimer::function() to free the timer |
| |
* Peter Zijlstra <peterz@infradead.org> wrote:
> On Wed, Jun 03, 2015 at 07:41:43PM +0200, Thomas Gleixner wrote: > > On Wed, 3 Jun 2015, Peter Zijlstra wrote: > > > /** > > > * struct hrtimer - the basic hrtimer structure > > > @@ -153,6 +144,7 @@ struct hrtimer_clock_base { > > > struct timerqueue_head active; > > > ktime_t (*get_time)(void); > > > ktime_t offset; > > > + struct hrtimer *running; > > > > Aside of lacking a KernelDoc comment, it expands the struct size on > > 32bit from 32 bytes to 36 bytes which undoes some of the recent cache > > line optimizations I did. Mooo! > > > > So we might think about storing the running timer pointer in cpu_base > > instead for 32bit, which increases the foot print of the migration > > base and the extra cost for the additional indirection, but it would > > keep cache line tight for the hot pathes. > > A wee something like this then? > > --- > --- a/include/linux/hrtimer.h > +++ b/include/linux/hrtimer.h > @@ -123,8 +123,10 @@ struct hrtimer_sleeper { > > #ifdef CONFIG_64BIT > # define HRTIMER_CLOCK_BASE_ALIGN 64 > +# define __timer_base_running(timer) timer->base->running > #else > # define HRTIMER_CLOCK_BASE_ALIGN 32 > +# define __timer_base_running(timer) timer->base->cpu_base->running > #endif
Please put it into the cpu_base on 64-bit as well: the base pointer is available already on 64-bit so there should be no measurable performance difference, and readability is a primary concern with all this code.
Thanks,
Ingo
| |