Re: [PATCH 1/3] ARM: at91: flush the L2 cache before entering cpu idle

[Jean-Jacques Hiblot]

2017-01-11 12:18 GMT+01:00 Russell King - ARM Linux <linux@xxxxxxxxxxxxxxx>:
> On Wed, Jan 11, 2017 at 12:05:05PM +0100, Jean-Jacques Hiblot wrote:
>> 2017-01-11 9:15 GMT+01:00  <Wenyou.Yang@xxxxxxxxxxxxx>:
>> > Hi Jean-Jacques,
>> >
>> >> -----Original Message-----
>> >> From: Jean-Jacques Hiblot [mailto:jjhiblot@xxxxxxxxx]
>> >> Sent: 2017å1æ11æ 0:51
>> >> To: Alexandre Belloni <alexandre.belloni@xxxxxxxxxxxxxxxxxx>
>> >> Cc: Wenyou Yang - A41535 <Wenyou.Yang@xxxxxxxxxxxxx>; Mark Rutland
>> >> <mark.rutland@xxxxxxx>; devicetree <devicetree@xxxxxxxxxxxxxxx>; Russell
>> >> King <linux@xxxxxxxxxxxxxxxx>; Wenyou Yang - A41535
>> >> <Wenyou.Yang@xxxxxxxxxxxxx>; Nicolas Ferre <nicolas.ferre@xxxxxxxxx>;
>> >> Linux Kernel Mailing List <linux-kernel@xxxxxxxxxxxxxxx>; Rob Herring
>> >> <robh+dt@xxxxxxxxxx>; linux-arm-kernel@xxxxxxxxxxxxxxxxxxx
>> >> Subject: Re: [PATCH 1/3] ARM: at91: flush the L2 cache before entering cpu idle
>> >>
>> >> 2017-01-10 17:18 GMT+01:00 Alexandre Belloni
>> >> <alexandre.belloni@xxxxxxxxxxxxxxxxxx>:
>> >> > I though a bit more about it, and I don't really like the new
>> >> > compatible string. I don't feel this should be necessary.
>> >> >
>> >> > What about the following:
>> >> >
>> >> > diff --git a/arch/arm/mach-at91/pm.c b/arch/arm/mach-at91/pm.c index
>> >> > b4332b727e9c..0333aca63e44 100644
>> >> > --- a/arch/arm/mach-at91/pm.c
>> >> > +++ b/arch/arm/mach-at91/pm.c
>> >> > @@ -53,6 +53,7 @@ extern void at91_pinctrl_gpio_resume(void);  static
>> >> > struct {
>> >> >         unsigned long uhp_udp_mask;
>> >> >         int memctrl;
>> >> > +       bool has_l2_cache;
>> >> >  } at91_pm_data;
>> >> >
>> >> >  void __iomem *at91_ramc_base[2];
>> >> > @@ -267,6 +268,11 @@ static void at91_ddr_standby(void)
>> >> >         u32 lpr0, lpr1 = 0;
>> >> >         u32 saved_lpr0, saved_lpr1 = 0;
>> >> >
>> >>
>> >> > +       if (at91_pm_data.has_l2_cache) {
>> >> > +               flush_cache_all();
>> >> what is the point of calling flush_cache_all() here ? Do we really care that dirty
>> >> data in L1 is written to DDR ? I may be missing something but to me it's just extra
>> >> latency.
>> >
>> > Are you mean use outer_flush_all() to flush all cache lines in the outer cache only?
>>
>> Yes that's what I meant. You see, you don't flush the cache for
>> sama5d3 so it shouldn't be required either for sam5d4. You should be
>> able to test it quickly and see if L1 flush is indeed required by
>> replacing  flush_cache_all() with outer_flush_all(). BTW is highly
>> probable that L2 cache flush is done in outer_disable() so calling
>> outer_flush_all() is probably no required.
>
> Please don't.  Read the comments in the code, and understand the APIs
> that you're suggesting people use _before_ making the suggestion:
>
> /**
>  * outer_flush_all - clean and invalidate all cache lines in the outer cache
>  *
>  * Note: depending on implementation, this may not be atomic - it must
>  * only be called with interrupts disabled and no other active outer
>  * cache masters.
>  *
>  * It is intended that this function is only used by implementations
>  * needing to override the outer_cache.disable() method due to security.
>  * (Some implementations perform this as a clean followed by an invalidate.)
>  */
>
> So, outer_flush_all() should not be called except from L2 cache code
> implementing the outer_disable() function - it's not intended for
> platforms to use.

OK. My bad. I didn't understand the comments.

>
> There are, however, sadly three users of outer_flush_all() which have
> crept in through arm-soc, that should be outer_disable() instead.
>
> --
> RMK's Patch system: http://www.armlinux.org.uk/developer/patches/
> FTTC broadband for 0.8mile line: currently at 9.6Mbps down 400kbps up
> according to speedtest.net.