Re: [PATCH v2] powerpc 8xx: Fixing issue with CONFIG_PIN_TLB

[leroy christophe <christophe.leroy@c-s.fr>]

Le 16/09/2013 23:02, Scott Wood a écrit :
> On Fri, 2013-09-13 at 07:04 +0200, leroy christophe wrote:
>> Le 12/09/2013 20:44, Scott Wood a écrit :
>>> On Thu, 2013-09-12 at 20:25 +0200, Christophe Leroy wrote:
>>>> This is a reorganisation of the setup of the TLB at kernel startup, in order
>>>> to handle the CONFIG_PIN_TLB case in accordance with chapter 8.10.3 of MPC866
>>>> and MPC885 reference manuals.
>>>>
>>>> Signed-off-by: Christophe Leroy <christophe.leroy@c-s.fr>
>>>>
>>>> diff -ur linux-3.11.org/arch/powerpc/kernel/head_8xx.S linux-3.11/arch/powerpc/kernel/head_8xx.S
>>>> --- linux-3.11.org/arch/powerpc/kernel/head_8xx.S	2013-09-02 22:46:10.000000000 +0200
>>>> +++ linux-3.11/arch/powerpc/kernel/head_8xx.S	2013-09-09 11:28:54.000000000 +0200
>>>> @@ -785,27 +785,24 @@
>>>>     * these mappings is mapped by page tables.
>>>>     */
>>>>    initial_mmu:
>>>> -	tlbia			/* Invalidate all TLB entries */
>>>> -/* Always pin the first 8 MB ITLB to prevent ITLB
>>>> -   misses while mucking around with SRR0/SRR1 in asm
>>>> -*/
>>>> -	lis	r8, MI_RSV4I@h
>>>> -	ori	r8, r8, 0x1c00
>>>> -
>>>> +	lis	r8, MI_RESETVAL@h
>>>>    	mtspr	SPRN_MI_CTR, r8	/* Set instruction MMU control */
>>>>    
>>>> -#ifdef CONFIG_PIN_TLB
>>>> -	lis	r10, (MD_RSV4I | MD_RESETVAL)@h
>>>> -	ori	r10, r10, 0x1c00
>>>> -	mr	r8, r10
>>>> -#else
>>>>    	lis	r10, MD_RESETVAL@h
>>>> -#endif
>>>>    #ifndef CONFIG_8xx_COPYBACK
>>>>    	oris	r10, r10, MD_WTDEF@h
>>>>    #endif
>>>>    	mtspr	SPRN_MD_CTR, r10	/* Set data TLB control */
>>>>    
>>>> +	tlbia			/* Invalidate all TLB entries */
>>> Is this change to make sure we invalidate everything even if the
>>> bootloader set RSV4I?
>> Most probably. It is step 2 of the process defined in MPC866 and MPC885
>> Reference Manuals:
>>
>> §8.10.3 Loading Locked TLB Entries:
>> The process of loading a single reserved entry in the TLB is as follows:
> To minimize code churn we should just fix actual problems, rather than
> shuffle things around to conform to a suggested sequence.  After all,
> we're not just trying to load a single entry.
Ok, I'll try again.
>
>>>> +	ori	r8, r8, 0x1c00
>>>> +	mtspr	SPRN_MI_CTR, r8	/* Set instruction MMU control */
>>>> +#ifdef CONFIG_PIN_TLB
>>>> +	ori	r10, r10, 0x1c00
>>>> +	mtspr	SPRN_MD_CTR, r10	/* Set data TLB control */
>>>> +#endif
>>> Still 0x1c00?
>> Yes, I kept the same entries in order to limit modifications:
>> * 28 = First 8Mbytes page
>> * 29 = IMMR
>> * 30 = Second 8Mbytes page
>> * 31 = Third 8Mbytes page
> If you actually want to program them in increasing order then it looks
> like you're still missing a write to CTR between the last two 8M entries
> -- thus you'll overwrite the IMMR with the last 8M entry.  That was the
> same problem that v1 fixed -- did that change get lost accidentally?
Oops, no, in fact I diffed from the version which was including it 
already. My mistake.
>
> The hardware wants to decrement; why fight it?
I see your point.
However it is not clear in the documentation if the decrement is done 
really after the update, or at xTLB interrupt. So I propose to still set 
the CTR ourself as described in the reference Manual and not assume that 
the HW decrements it.
>
>>>>    	/* Now map the lower 8 Meg into the TLBs.  For this quick hack,
>>>>    	 * we can load the instruction and data TLB registers with the
>>>>    	 * same values.
>>>> @@ -825,6 +822,12 @@
>>>>    	mtspr	SPRN_MI_AP, r8
>>>>    	mtspr	SPRN_MD_AP, r8
>>>>    
>>>> +	/* Always pin the first 8 MB ITLB to prevent ITLB
>>>> +	 * misses while mucking around with SRR0/SRR1 in asm
>>>> +	 */
>>>> +	lis	r8, (MI_RSV4I | MI_RESETVAL)@h
>>>> +	mtspr	SPRN_MI_CTR, r8	/* Set instruction MMU control */
>>> Entry 0 is not pinnable.
>> Here we are not trying to pin entry 0.
> Sorry, misread the patch.
>
>> We are at step 8, we are setting
>> MI_RSV4I. At the same time, we set MD_CTR to 0 which is off the pinned
>> range, to be sure that we won't overwrite one of the pinned entries.
>>
>> The main difference compared to the previous implementation is that
>> before, we were setting the RSV4I bit before loading the TLB entries.
>> Now, as defined in the Reference Manuals, we are doing it at the end.
> Have you seen any evidence that it matters?
Not really.

Ok, propose a new patch in a few minutes.

Christophe