v7_flush_kern_cache_louis flushes up to L2?

v7_flush_kern_cache_louis flushes up to L2?

[Bastian Hecht]

Hello,

I've got a Cortex-A9 UP with a L2 and want to submit some PM code I've
written. Just to make sure I've made no mistake, it would be very
helpful if you can confirm a hypothesis I use in my code:

v7_flush_kern_cache_louis: Flush the data cache up to Level of
Unification Inner Shareable

This flushes the data out up to the L2, right? The ARM docs say that
the Point of Unification would be my L2. I'm a bit confused by the
term "Level of Unification Inner Shareable" (that states that in an
SMP system L1 coherency is guaranteed and all is flushed to the L2?).

Thanks,

 Bastian

v7_flush_kern_cache_louis flushes up to L2?

[Jonathan Austin]

Hi Bastian,

On 10/04/13 11:43, Bastian Hecht wrote:
> Hello,
>
> I've got a Cortex-A9 UP with a L2 and want to submit some PM code I've

To clarify, is this an MPCore with a single core, or a genuine UP? This 
can be established from the 'U' bit of the MPIDR.

> written. Just to make sure I've made no mistake, it would be very
> helpful if you can confirm a hypothesis I use in my code:
>
> v7_flush_kern_cache_louis: Flush the data cache up to Level of
> Unification Inner Shareable
>

Depending on whether you're SMP or UP (bearing in mind that you can be 
SMP, but still only have one processor!) the IS is ignored in the 
v7_flush_dcache_louis operation:
(from cache-v7.S)

         mrc     p15, 1, r0, c0, c0, 1           @ read clidr, r0 = clidr
         ALT_SMP(ands    r3, r0, #(7 << 21))     @ extract LoUIS from clidr
         ALT_UP(ands     r3, r0, #(7 << 27))     @ extract LoUU from clidr
         ALT_SMP(mov     r3, r3, lsr #20)        @ r3 = LoUIS * 2
         ALT_UP(mov      r3, r3, lsr #26)        @ r3 = LoUU * 2
	...
	flush levels based on value in r3

> This flushes the data out up to the L2, right? The ARM docs say that
> the Point of Unification would be my L2. I'm a bit confused by the
> term "Level of Unification Inner Shareable" (that states that in an
> SMP system L1 coherency is guaranteed and all is flushed to the L2?).
>

As you say, for the A9 (from the TRM) the CLIDR reports LoUIS is the 
same as LoUU and both specify L2.

Does that make things clearer, or are you still unsure about something?

Jonny

v7_flush_kern_cache_louis flushes up to L2?

[Bastian Hecht]

Hi Jonny!

2013/4/10 Jonathan Austin <jonathan.austin@arm.com>:
> Hi Bastian,
>
>
> On 10/04/13 11:43, Bastian Hecht wrote:
>>
>> Hello,
>>
>> I've got a Cortex-A9 UP with a L2 and want to submit some PM code I've
>
>
> To clarify, is this an MPCore with a single core, or a genuine UP? This can
> be established from the 'U' bit of the MPIDR.
>

I didn't actually read out the U bit, but I'm sure I've got no SCU, so
I bet high that it's a genuine UP system.

>> written. Just to make sure I've made no mistake, it would be very
>> helpful if you can confirm a hypothesis I use in my code:
>>
>> v7_flush_kern_cache_louis: Flush the data cache up to Level of
>> Unification Inner Shareable
>>
>
> Depending on whether you're SMP or UP (bearing in mind that you can be SMP,
> but still only have one processor!) the IS is ignored in the
> v7_flush_dcache_louis operation:
> (from cache-v7.S)
>
>         mrc     p15, 1, r0, c0, c0, 1           @ read clidr, r0 = clidr
>         ALT_SMP(ands    r3, r0, #(7 << 21))     @ extract LoUIS from clidr
>         ALT_UP(ands     r3, r0, #(7 << 27))     @ extract LoUU from clidr
>         ALT_SMP(mov     r3, r3, lsr #20)        @ r3 = LoUIS * 2
>         ALT_UP(mov      r3, r3, lsr #26)        @ r3 = LoUU * 2
>         ...
>         flush levels based on value in r3
>
>
>> This flushes the data out up to the L2, right? The ARM docs say that
>> the Point of Unification would be my L2. I'm a bit confused by the
>> term "Level of Unification Inner Shareable" (that states that in an
>> SMP system L1 coherency is guaranteed and all is flushed to the L2?).
>>
>
> As you say, for the A9 (from the TRM) the CLIDR reports LoUIS is the same as
> LoUU and both specify L2.

Ok, this is the golden info I was looking for. So after cpu_suspend()
I am good with the following sequence?
flush L2 (outer_flush_all)
disable L2 (outer_disable)
Clear the SCTLR.C bit and issue an "isb"
flush L1 (v7_flush_dcache_all)
cpu_do_idle

and for resume:
invalidate L1
(trust cpu_resume to resume the L1 and enable the SCTLR.C bit)
resume L2 (outer_resume)

> Does that make things clearer, or are you still unsure about something?

If you could confirm the above sequence, I'm perfectly fine. Thanks
for the quick and exhaustive support.

Cheers,

 Bastian

v7_flush_kern_cache_louis flushes up to L2?

[Jonathan Austin]

Hi again Bastian,

On 10/04/13 13:16, Bastian Hecht wrote:
[...]
>>> This flushes the data out up to the L2, right? The ARM docs say that
>>> the Point of Unification would be my L2. I'm a bit confused by the
>>> term "Level of Unification Inner Shareable" (that states that in an
>>> SMP system L1 coherency is guaranteed and all is flushed to the L2?).
>>>
>>
>> As you say, for the A9 (from the TRM) the CLIDR reports LoUIS is the same as
>> LoUU and both specify L2.
>
> Ok, this is the golden info I was looking for. So after cpu_suspend()
> I am good with the following sequence?
> flush L2 (outer_flush_all)
> disable L2 (outer_disable)
> Clear the SCTLR.C bit and issue an "isb"
> flush L1 (v7_flush_dcache_all)

This looks good but I think there's a minor modification to be made to 
the sequence here:

After you turn of the caches with the SCTLR.C bit, you can no longer hit 
in the L1 cache:
SCTLR.C = 0:
"The Cortex-A9 L1 Data Cache is not enabled. All memory accesses to 
Normal Memory Cacheable regions are treated as Normal Memory 
Non-Cacheable, without lookup and without allocation in the L1 Data Cache."

If you had dirty data in the L1 cache you will lose it at that point.

So, you should flush L1 before turning off caches.

Note that on SMP, you would also need to clean/flush again after turning 
off the caches, as read-speculation could have sucked dirty data from 
another cache in to your cache, which would need to be written back 
before sleeping.

> cpu_do_idle
>
> and for resume:
> invalidate L1

How do you plan to do this? I notice that there's an increasing tendency 
to call v7_invalidate_l1 which isn't actually part of the cacheflush API 
(see arch/arm/include/asm/cacheflush.h arch/arm/include/asm/glue-cache.h...)

Does your cache definitely come up in an undefined state? If not, could 
you use v7_flush_dcache_all? (see the comment above v7_invalidate_l1 in 
arch/arm/mm/cache-v7.S)

Hope that helps,

Jonny

v7_flush_kern_cache_louis flushes up to L2?

[Bastian Hecht]

Heyho!

2013/4/10 Jonathan Austin <jonathan.austin@arm.com>:
> Hi again Bastian,
>
> On 10/04/13 13:16, Bastian Hecht wrote:
> [...]
>
>>>> This flushes the data out up to the L2, right? The ARM docs say that
>>>> the Point of Unification would be my L2. I'm a bit confused by the
>>>> term "Level of Unification Inner Shareable" (that states that in an
>>>> SMP system L1 coherency is guaranteed and all is flushed to the L2?).
>>>>
>>>
>>> As you say, for the A9 (from the TRM) the CLIDR reports LoUIS is the same
>>> as
>>> LoUU and both specify L2.
>>
>>
>> Ok, this is the golden info I was looking for. So after cpu_suspend()
>> I am good with the following sequence?
>> flush L2 (outer_flush_all)
>> disable L2 (outer_disable)
>> Clear the SCTLR.C bit and issue an "isb"
>> flush L1 (v7_flush_dcache_all)
>
>
> This looks good but I think there's a minor modification to be made to the
> sequence here:
>
> After you turn of the caches with the SCTLR.C bit, you can no longer hit in
> the L1 cache:
> SCTLR.C = 0:
> "The Cortex-A9 L1 Data Cache is not enabled. All memory accesses to Normal
> Memory Cacheable regions are treated as Normal Memory Non-Cacheable, without
> lookup and without allocation in the L1 Data Cache."
>
> If you had dirty data in the L1 cache you will lose it at that point.
>
> So, you should flush L1 before turning off caches.

Hmm, I wondered if I need that as cpu_suspend flushes all out to the
L2 as we discussed. So if I don't write anything to RAM after
cpu_suspend why do I need the extra flush before turning off
L1-caching?

> Note that on SMP, you would also need to clean/flush again after turning off
> the caches, as read-speculation could have sucked dirty data from another
> cache in to your cache, which would need to be written back before sleeping.
>

Aaaaaah, thanks! I always wondered why this is needed, but that makes sense.

>> cpu_do_idle
>>
>> and for resume:
>> invalidate L1
>
>
> How do you plan to do this? I notice that there's an increasing tendency to
> call v7_invalidate_l1 which isn't actually part of the cacheflush API (see
> arch/arm/include/asm/cacheflush.h arch/arm/include/asm/glue-cache.h...)

I used to have my gazillionth version of v7_invalidate_l1 in my
assembly code but as it is in cache-v7.S now, I use this.

> Does your cache definitely come up in an undefined state? If not, could you
> use v7_flush_dcache_all? (see the comment above v7_invalidate_l1 in
> arch/arm/mm/cache-v7.S)

To be honest, I have no clue if my cache comes up in an undefined
state. But the comment about random data getting flushed to RAM scared
me so much that I blindly go with what other Cortex A9 users do.

> Hope that helps,

It does!

Thanks,

 Bastian

v7_flush_kern_cache_louis flushes up to L2?

[Russell King - ARM Linux]

On Wed, Apr 10, 2013 at 02:35:05PM +0100, Jonathan Austin wrote:
> Note that on SMP, you would also need to clean/flush again after turning  
> off the caches, as read-speculation could have sucked dirty data from  
> another cache in to your cache, which would need to be written back  
> before sleeping.

Err, no.  If you read data into a cache, it is always populated in a clean
state.  You never suck dirty data out of one cache, remove it there and
place it immediately in a dirty state in an upper level cache.

So, if we start from the idea that reads will populate the upper level
cache with clean data, that cache is populated merely with a _copy_ of
the data held elsewhere in the system.

In that case, there is no need to clean or flush that cache after turning
it off - if the cache only contains _copies_ of data elsewhere in the
system, then it's perfectly fine for that data to be discarded in any
way - be that by turning the power off to the cache or invalidating it.

There is a step further in this: if the cache is being dirtied by the
local CPU, and that data will only ever be used by the local CPU, again,
in the CPU shutdown path, you really do not need to write that data out -
because the only CPU which cares about it is the local CPU which is going
to be shutdown.  So, loss of that data is not a problem.

In other words, shutdown of a CPU is a process which involves pushing out
data which the rest of the system requires (that's done by an initial
flush).  Once that flush is done, provided the dying CPU does not _write_
to any shared data (this includes spinlocks and such like too), there is
no need to flush the cache anymore.

v7_flush_kern_cache_louis flushes up to L2?

[Russell King - ARM Linux]

On Fri, Apr 19, 2013 at 11:21:42AM +0100, Russell King - ARM Linux wrote:
> On Wed, Apr 10, 2013 at 02:35:05PM +0100, Jonathan Austin wrote:
> > Note that on SMP, you would also need to clean/flush again after turning  
> > off the caches, as read-speculation could have sucked dirty data from  
> > another cache in to your cache, which would need to be written back  
> > before sleeping.
> 
> Err, no.  If you read data into a cache, it is always populated in a clean
> state.  You never suck dirty data out of one cache, remove it there and
> place it immediately in a dirty state in an upper level cache.
> 
> So, if we start from the idea that reads will populate the upper level
> cache with clean data, that cache is populated merely with a _copy_ of
> the data held elsewhere in the system.

Hmm, Will tells me that later ARM cores can migrate dirty cache lines up
the cache tree, which means that we're potentially into deep problems
with CPU hotplug and suspend paths, because it means there's potentially
no way to ensure that we cleanly push data out the caches and shut the
caches down before the CPU goes down.

Combine this cache behaviour with the implementation specific details
about SCTRL.C (whether caches are just prevented from being allocated,
or whether it also affects the caches being searched) and you have a
recipe for multiple chunks of complex platform specific and CPU
specific code - especially when combined with things like removal of
CPU power and synchronisation of that in the hotplug case.

I do hope that ARM Ltd have thought about this, and have a way that we
can shut CPU caches down sanely, but I'm fearing that they haven't
really considered that to be something anyone really wants to do in a
generic way.

v7_flush_kern_cache_louis flushes up to L2?

[Lorenzo Pieralisi]

On Wed, Apr 10, 2013 at 01:16:03PM +0100, Bastian Hecht wrote:
> Hi Jonny!
> 
> 2013/4/10 Jonathan Austin <jonathan.austin@arm.com>:
> > Hi Bastian,
> >
> >
> > On 10/04/13 11:43, Bastian Hecht wrote:
> >>
> >> Hello,
> >>
> >> I've got a Cortex-A9 UP with a L2 and want to submit some PM code I've
> >
> >
> > To clarify, is this an MPCore with a single core, or a genuine UP? This can
> > be established from the 'U' bit of the MPIDR.
> >
> 
> I didn't actually read out the U bit, but I'm sure I've got no SCU, so
> I bet high that it's a genuine UP system.
> 
> >> written. Just to make sure I've made no mistake, it would be very
> >> helpful if you can confirm a hypothesis I use in my code:
> >>
> >> v7_flush_kern_cache_louis: Flush the data cache up to Level of
> >> Unification Inner Shareable
> >>
> >
> > Depending on whether you're SMP or UP (bearing in mind that you can be SMP,
> > but still only have one processor!) the IS is ignored in the
> > v7_flush_dcache_louis operation:
> > (from cache-v7.S)
> >
> >         mrc     p15, 1, r0, c0, c0, 1           @ read clidr, r0 = clidr
> >         ALT_SMP(ands    r3, r0, #(7 << 21))     @ extract LoUIS from clidr
> >         ALT_UP(ands     r3, r0, #(7 << 27))     @ extract LoUU from clidr
> >         ALT_SMP(mov     r3, r3, lsr #20)        @ r3 = LoUIS * 2
> >         ALT_UP(mov      r3, r3, lsr #26)        @ r3 = LoUU * 2
> >         ...
> >         flush levels based on value in r3
> >
> >
> >> This flushes the data out up to the L2, right? The ARM docs say that
> >> the Point of Unification would be my L2. I'm a bit confused by the
> >> term "Level of Unification Inner Shareable" (that states that in an
> >> SMP system L1 coherency is guaranteed and all is flushed to the L2?).
> >>
> >
> > As you say, for the A9 (from the TRM) the CLIDR reports LoUIS is the same as
> > LoUU and both specify L2.
> 
> Ok, this is the golden info I was looking for. So after cpu_suspend()
> I am good with the following sequence?

Is L2 RAM retained on power down ?

> flush L2 (outer_flush_all)
> disable L2 (outer_disable)

Disabling L2 is not mandatory. And the code above (if L2 RAM is not
retained) can be simply

outer_disable

the cleaning is done in the PL310 disable function properly.

> Clear the SCTLR.C bit and issue an "isb"
> flush L1 (v7_flush_dcache_all)

Two steps above are ok, as long as flush L1 does not push data on the
stack on entry and L1 clean routine does not need any dirty data present
in L1. Clearing C bit on A9 stops the core from searching the
cache so data writes should not be executed with C bit cleared and a
data cache still containing dirty lines. For the same reason the cache
cleaning routine should not require any data to execute since the data
can be sitting in the cache that is not searched anymore when the C bit
is cleared.

The sequence:

clear C bit
bl v7_flush_dcache_all

is better coded in assembly (in the cpu_suspend finisher, ie the
function you pass as cpu_suspend 2 argument) to control what you are doing.

The cache cleaning routine (v7_flush_dcache_all) does not require any
data to execute, so running it with C bit cleared is fine.

> cpu_do_idle
> 
> and for resume:
> invalidate L1

Use v7_invalidate_l1 here.

> (trust cpu_resume to resume the L1 and enable the SCTLR.C bit)
> resume L2 (outer_resume)

Again, it depends on L2 behaviour on shutdown, if it is retained or not.
If L2 RAM content is lost on power down the sequence above seems ok.

Post the code, happy to have a look.

Lorenzo

v7_flush_kern_cache_louis flushes up to L2?

[Bastian Hecht]

Hi Lorenzo,

2013/4/10 Lorenzo Pieralisi <lorenzo.pieralisi@arm.com>:
> On Wed, Apr 10, 2013 at 01:16:03PM +0100, Bastian Hecht wrote:
>> Hi Jonny!
>>
>> 2013/4/10 Jonathan Austin <jonathan.austin@arm.com>:
>> > Hi Bastian,
>> >
>> >
>> > On 10/04/13 11:43, Bastian Hecht wrote:
>> >>
>> >> Hello,
>> >>
>> >> I've got a Cortex-A9 UP with a L2 and want to submit some PM code I've
>> >
>> >
>> > To clarify, is this an MPCore with a single core, or a genuine UP? This can
>> > be established from the 'U' bit of the MPIDR.
>> >
>>
>> I didn't actually read out the U bit, but I'm sure I've got no SCU, so
>> I bet high that it's a genuine UP system.
>>
>> >> written. Just to make sure I've made no mistake, it would be very
>> >> helpful if you can confirm a hypothesis I use in my code:
>> >>
>> >> v7_flush_kern_cache_louis: Flush the data cache up to Level of
>> >> Unification Inner Shareable
>> >>
>> >
>> > Depending on whether you're SMP or UP (bearing in mind that you can be SMP,
>> > but still only have one processor!) the IS is ignored in the
>> > v7_flush_dcache_louis operation:
>> > (from cache-v7.S)
>> >
>> >         mrc     p15, 1, r0, c0, c0, 1           @ read clidr, r0 = clidr
>> >         ALT_SMP(ands    r3, r0, #(7 << 21))     @ extract LoUIS from clidr
>> >         ALT_UP(ands     r3, r0, #(7 << 27))     @ extract LoUU from clidr
>> >         ALT_SMP(mov     r3, r3, lsr #20)        @ r3 = LoUIS * 2
>> >         ALT_UP(mov      r3, r3, lsr #26)        @ r3 = LoUU * 2
>> >         ...
>> >         flush levels based on value in r3
>> >
>> >
>> >> This flushes the data out up to the L2, right? The ARM docs say that
>> >> the Point of Unification would be my L2. I'm a bit confused by the
>> >> term "Level of Unification Inner Shareable" (that states that in an
>> >> SMP system L1 coherency is guaranteed and all is flushed to the L2?).
>> >>
>> >
>> > As you say, for the A9 (from the TRM) the CLIDR reports LoUIS is the same as
>> > LoUU and both specify L2.
>>
>> Ok, this is the golden info I was looking for. So after cpu_suspend()
>> I am good with the following sequence?
>
> Is L2 RAM retained on power down ?

I have two different versions of powering down the SoC. Currently I'm
focusing on a shutdown mode that contains the powering off of the L2.

>> flush L2 (outer_flush_all)
>> disable L2 (outer_disable)
>
> Disabling L2 is not mandatory. And the code above (if L2 RAM is not
> retained) can be simply
>
> outer_disable
>
> the cleaning is done in the PL310 disable function properly.

The problem I see with this approach is:
What advantage do we get at all if we have to flush the L2 (which is
done in the PL310 disable routine)? Isn't this exactly the part we
want to save? Not to have to flush the L2.

>> Clear the SCTLR.C bit and issue an "isb"
>> flush L1 (v7_flush_dcache_all)
>
> Two steps above are ok, as long as flush L1 does not push data on the
> stack on entry and L1 clean routine does not need any dirty data present
> in L1. Clearing C bit on A9 stops the core from searching the
> cache so data writes should not be executed with C bit cleared and a
> data cache still containing dirty lines. For the same reason the cache
> cleaning routine should not require any data to execute since the data
> can be sitting in the cache that is not searched anymore when the C bit
> is cleared.

Ah true! When I enter my assembly code there are of course stack modifications.
And here comes a point I've just realized recently: Sometimes the WFI
command doesn't enter the low power state mode I've requested. I
haven't observed this when using Suspend-To-Ram but when using
CPUIdle.
I've seen code from the OMAP people that check for the case that WFI
doesn't succeed. Probably I need to do the same. And for this I need
the stack to be synced.

> The sequence:
>
> clear C bit
> bl v7_flush_dcache_all
>
> is better coded in assembly (in the cpu_suspend finisher, ie the
> function you pass as cpu_suspend 2 argument) to control what you are doing.
>
> The cache cleaning routine (v7_flush_dcache_all) does not require any
> data to execute, so running it with C bit cleared is fine.
>
>> cpu_do_idle
>>
>> and for resume:
>> invalidate L1
>
> Use v7_invalidate_l1 here.
>
>> (trust cpu_resume to resume the L1 and enable the SCTLR.C bit)
>> resume L2 (outer_resume)
>
> Again, it depends on L2 behaviour on shutdown, if it is retained or not.
> If L2 RAM content is lost on power down the sequence above seems ok.
>
> Post the code, happy to have a look.
>

Ok great, I'm quite blown away by the support here, should have
contacted you earlier!

Bastian