[4.4-rc6-rt1 PATCH 0/2] ARM: am437x: boot test report

[4.4-rc6-rt1 PATCH 0/2] ARM: am437x: boot test report

[Grygorii Strashko]

Hi Sebastian,

On 01/13/2016 10:14 PM, Grygorii Strashko wrote:
> On 01/13/2016 01:41 PM, Grygorii Strashko wrote:
>> On 01/12/2016 07:15 PM, Sebastian Andrzej Siewior wrote:
>>> * Grygorii Strashko | 2015-12-28 20:00:09 [+0200]:
>>>
>>>
>>>> 3) Seems there are some problem with HIGHMEM :(. System crashes if I'm trying
>>>>      to run hackbench (see log 1). I can run hackbench only if HIGHMEM is disabled.
>>>>      I've also tried to disable CONFIG_CPU_SW_DOMAIN_PAN, CONFIG_ARM_PATCH_PHYS_VIRT,
>>>>      and CONFIG_CPUSETS, before disabling HIGHMEM.
>>>>      Sorry, I'll not able to investigate it further because of Christmas holidays here.
>>>
>>> I see. It seems not to trigger anything on x86.
>>
>> Yah. I saw other reports - it seems this issue is not triggered on ARM SMP also
>> (as per report from Tim Sander). Will try to dig deeper.
>>
> 
> I did more tests, but still don't know what is wrong.
> 1) vanilla 4.4 - no issue
> 2) 4.4-rt: CONFIG_PREEMPT_RT_BASE=n and CONFIG_PREEMPT_RT_FULL=n -- no issue
> 3) 4.4-rt: CONFIG_DEBUG_HIGHMEM=y -- issue is present log1
> 4) 4.4-rt: this is interesting ;)
> I can run hackbench few times without issues If I run hackbench with RT priority (log2)
> But I got crash immediately once I run it using non-RT priorities.


Seems I've found root cause of the issue - FIX_KMAP_BEGIN offset need to be taken into
account in switch_kmaps() and below diff fixes this issue for me.
What I'm not fully understand is why I'm the only one who hit this issue :(
New functionality was added by commit:
commit a5f4c561b3b19a9bc43a81da6382b0098ebbc1fb
    ARM: 8415/1: early fixmap support for earlycon

and it is enabled by default.

Luckiness :(

diff --git a/arch/arm/mm/highmem.c b/arch/arm/mm/highmem.c
index deabc36..b9b4f9c 100644
--- a/arch/arm/mm/highmem.c
+++ b/arch/arm/mm/highmem.c
@@ -167,7 +167,7 @@ void switch_kmaps(struct task_struct *prev_p, struct task_struct *next_p)
         * Clear @prev's kmap_atomic mappings
         */
        for (i = 0; i < prev_p->kmap_idx; i++) {
-               int idx = i + KM_TYPE_NR * smp_processor_id();
+               int idx = FIX_KMAP_BEGIN + i + KM_TYPE_NR * smp_processor_id();
 
                set_fixmap_pte(idx, __pte(0));
        }
@@ -175,7 +175,7 @@ void switch_kmaps(struct task_struct *prev_p, struct task_struct *next_p)
         * Restore @next_p's kmap_atomic mappings
         */
        for (i = 0; i < next_p->kmap_idx; i++) {
-               int idx = i + KM_TYPE_NR * smp_processor_id();
+               int idx = FIX_KMAP_BEGIN + i + KM_TYPE_NR * smp_processor_id();
 
                if (!pte_none(next_p->kmap_pte[i]))
                        set_fixmap_pte(idx, next_p->kmap_pte[i]);

-- 
regards,
-grygorii

[4.4-rc6-rt1 PATCH 0/2] ARM: am437x: boot test report

[Russell King - ARM Linux]

On Thu, Jan 14, 2016 at 10:11:09PM +0200, Grygorii Strashko wrote:
> diff --git a/arch/arm/mm/highmem.c b/arch/arm/mm/highmem.c
> index deabc36..b9b4f9c 100644
> --- a/arch/arm/mm/highmem.c
> +++ b/arch/arm/mm/highmem.c
> @@ -167,7 +167,7 @@ void switch_kmaps(struct task_struct *prev_p, struct task_struct *next_p)
>          * Clear @prev's kmap_atomic mappings
>          */
>         for (i = 0; i < prev_p->kmap_idx; i++) {
> -               int idx = i + KM_TYPE_NR * smp_processor_id();
> +               int idx = FIX_KMAP_BEGIN + i + KM_TYPE_NR * smp_processor_id();
>  
>                 set_fixmap_pte(idx, __pte(0));
>         }
> @@ -175,7 +175,7 @@ void switch_kmaps(struct task_struct *prev_p, struct task_struct *next_p)
>          * Restore @next_p's kmap_atomic mappings
>          */
>         for (i = 0; i < next_p->kmap_idx; i++) {
> -               int idx = i + KM_TYPE_NR * smp_processor_id();
> +               int idx = FIX_KMAP_BEGIN + i + KM_TYPE_NR * smp_processor_id();
>  
>                 if (!pte_none(next_p->kmap_pte[i]))
>                         set_fixmap_pte(idx, next_p->kmap_pte[i]);

This looks like it introduces the 4th and 5th copies of the same
calcuation, so can I suggest that we do this to prevent this kind
of error?

diff --git a/arch/arm/mm/highmem.c b/arch/arm/mm/highmem.c
index d02f8187b1cc..61f0d5941116 100644
--- a/arch/arm/mm/highmem.c
+++ b/arch/arm/mm/highmem.c
@@ -34,6 +34,11 @@ static inline pte_t get_fixmap_pte(unsigned long vaddr)
 	return *ptep;
 }
 
+static unsigned int fixmap_idx(int type)
+{
+	return FIX_KMAP_BEGIN + type + KM_TYPE_NR * smp_processor_id();
+}
+
 void *kmap(struct page *page)
 {
 	might_sleep();
@@ -79,7 +84,7 @@ void *kmap_atomic(struct page *page)
 
 	type = kmap_atomic_idx_push();
 
-	idx = FIX_KMAP_BEGIN + type + KM_TYPE_NR * smp_processor_id();
+	idx = fixmap_idx(type);
 	vaddr = __fix_to_virt(idx);
 #ifdef CONFIG_DEBUG_HIGHMEM
 	/*
@@ -106,7 +111,7 @@ void __kunmap_atomic(void *kvaddr)
 
 	if (kvaddr >= (void *)FIXADDR_START) {
 		type = kmap_atomic_idx();
-		idx = FIX_KMAP_BEGIN + type + KM_TYPE_NR * smp_processor_id();
+		idx = fixmap_idx(type);
 
 		if (cache_is_vivt())
 			__cpuc_flush_dcache_area((void *)vaddr, PAGE_SIZE);
@@ -138,7 +143,7 @@ void *kmap_atomic_pfn(unsigned long pfn)
 		return page_address(page);
 
 	type = kmap_atomic_idx_push();
-	idx = FIX_KMAP_BEGIN + type + KM_TYPE_NR * smp_processor_id();
+	idx = fixmap_idx(type);
 	vaddr = __fix_to_virt(idx);
 #ifdef CONFIG_DEBUG_HIGHMEM
 	BUG_ON(!pte_none(get_fixmap_pte(vaddr)));


-- 
RMK's Patch system: http://www.arm.linux.org.uk/developer/patches/
FTTC broadband for 0.8mile line: currently at 9.6Mbps down 400kbps up
according to speedtest.net.

[4.4-rc6-rt1 PATCH 0/2] ARM: am437x: boot test report

[Grygorii Strashko]

Hi Russell,

On 01/14/2016 11:30 PM, Russell King - ARM Linux wrote:
> On Thu, Jan 14, 2016 at 10:11:09PM +0200, Grygorii Strashko wrote:
>> diff --git a/arch/arm/mm/highmem.c b/arch/arm/mm/highmem.c
>> index deabc36..b9b4f9c 100644
>> --- a/arch/arm/mm/highmem.c
>> +++ b/arch/arm/mm/highmem.c
>> @@ -167,7 +167,7 @@ void switch_kmaps(struct task_struct *prev_p, struct task_struct *next_p)
>>           * Clear @prev's kmap_atomic mappings
>>           */
>>          for (i = 0; i < prev_p->kmap_idx; i++) {
>> -               int idx = i + KM_TYPE_NR * smp_processor_id();
>> +               int idx = FIX_KMAP_BEGIN + i + KM_TYPE_NR * smp_processor_id();
>>   
>>                  set_fixmap_pte(idx, __pte(0));
>>          }
>> @@ -175,7 +175,7 @@ void switch_kmaps(struct task_struct *prev_p, struct task_struct *next_p)
>>           * Restore @next_p's kmap_atomic mappings
>>           */
>>          for (i = 0; i < next_p->kmap_idx; i++) {
>> -               int idx = i + KM_TYPE_NR * smp_processor_id();
>> +               int idx = FIX_KMAP_BEGIN + i + KM_TYPE_NR * smp_processor_id();
>>   
>>                  if (!pte_none(next_p->kmap_pte[i]))
>>                          set_fixmap_pte(idx, next_p->kmap_pte[i]);
> 
> This looks like it introduces the 4th and 5th copies of the same
> calcuation, so can I suggest that we do this to prevent this kind
> of error?
> 
> diff --git a/arch/arm/mm/highmem.c b/arch/arm/mm/highmem.c
> index d02f8187b1cc..61f0d5941116 100644
> --- a/arch/arm/mm/highmem.c
> +++ b/arch/arm/mm/highmem.c
> @@ -34,6 +34,11 @@ static inline pte_t get_fixmap_pte(unsigned long vaddr)
>   	return *ptep;
>   }
>   
> +static unsigned int fixmap_idx(int type)
> +{
> +	return FIX_KMAP_BEGIN + type + KM_TYPE_NR * smp_processor_id();
> +}
> +


This looks very reasonable - I'll updated and re-send.
Would you agree if I'll add your Signed-off-by: in final patch?

Thanks for review.


[...]


-- 
regards,
-grygorii

[4.4-rc6-rt1 PATCH 0/2] ARM: am437x: boot test report

[Russell King - ARM Linux]

On Fri, Jan 15, 2016 at 01:35:33PM +0200, Grygorii Strashko wrote:
> This looks very reasonable - I'll updated and re-send.
> Would you agree if I'll add your Signed-off-by: in final patch?

Yes, that's fine.  Please use:

Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>

Thanks.

-- 
RMK's Patch system: http://www.arm.linux.org.uk/developer/patches/
FTTC broadband for 0.8mile line: currently at 9.6Mbps down 400kbps up
according to speedtest.net.