Re: THP bug and crash on sparc64 3.8

[David Miller <davem@davemloft.net>]

From: David Rientjes <rientjes@google.com>
Date: Tue, 19 Feb 2013 11:37:30 -0800 (PST)

> On Tue, 19 Feb 2013, David Miller wrote:
> 
>> diff --git a/arch/sparc/mm/tsb.c b/arch/sparc/mm/tsb.c
>> index 7f64743..338b406 100644
>> --- a/arch/sparc/mm/tsb.c
>> +++ b/arch/sparc/mm/tsb.c
>> @@ -288,14 +288,13 @@ static unsigned long tsb_size_to_rss_limit(unsigned long new_size)
>>   * the number of entries that the current TSB can hold at once.  Currently,
>>   * we trigger when the RSS hits 3/4 of the TSB capacity.
>>   */
>> -void tsb_grow(struct mm_struct *mm, unsigned long tsb_index, unsigned long rss)
>> +void tsb_grow(struct mm_struct *mm, unsigned long tsb_index, unsigned long rss, gfp_t gfp)
>>  {
>>  	unsigned long max_tsb_size = 1 * 1024 * 1024;
>>  	unsigned long new_size, old_size, flags;
>>  	struct tsb *old_tsb, *new_tsb;
>>  	unsigned long new_cache_index, old_cache_index;
>>  	unsigned long new_rss_limit;
>> -	gfp_t gfp_flags;
>>  
>>  	if (max_tsb_size > (PAGE_SIZE << MAX_ORDER))
>>  		max_tsb_size = (PAGE_SIZE << MAX_ORDER);
>> @@ -312,12 +311,11 @@ void tsb_grow(struct mm_struct *mm, unsigned long tsb_index, unsigned long rss)
>>  		new_rss_limit = ~0UL;
>>  
>>  retry_tsb_alloc:
>> -	gfp_flags = GFP_KERNEL;
>>  	if (new_size > (PAGE_SIZE * 2))
>> -		gfp_flags = __GFP_NOWARN | __GFP_NORETRY;
>> +		gfp = __GFP_NOWARN | __GFP_NORETRY;
> 
> The existing code here already looks buggy for a different reason, it 
> should be gfp |= __GFP_NOWARN | __GFP_NORETRY;
> 
> I was wondering if you could get away with just a simple one-liner fix 
> above by doing
> 
> 	gfp_flags = rss ? GFP_KERNEL : GFP_NOWAIT;
> 
> but I agree this is more robust.

Thanks for noticing that bug.  I'll probably just fix the "=" to an "|="
for now.