Re: [PATCH] mm: Fix the TLB range flushed when __tlb_remove_page() runs out of slots

[Vineet Gupta <Vineet.Gupta1@synopsys.com>]

On 05/29/2013 07:59 PM, Catalin Marinas wrote:
> On Wed, May 29, 2013 at 03:08:37PM +0100, Vineet Gupta wrote:
>> On 05/29/2013 07:33 PM, Catalin Marinas wrote:
>>> On Wed, May 29, 2013 at 01:56:13PM +0100, Vineet Gupta wrote:
>>>> zap_pte_range loops from @addr to @end. In the middle, if it runs out of
>>>> batching slots, TLB entries needs to be flushed for @start to @interim,
>>>> NOT @interim to @end.
>>>>
>>>> Since ARC port doesn't use page free batching I can't test it myself but
>>>> this seems like the right thing to do.
>>>> Observed this when working on a fix for the issue at thread:
>>>> 	http://www.spinics.net/lists/linux-arch/msg21736.html
>>>>
>>>> Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
>>>> Cc: Andrew Morton <akpm@linux-foundation.org>
>>>> Cc: Mel Gorman <mgorman@suse.de>
>>>> Cc: Hugh Dickins <hughd@google.com>
>>>> Cc: Rik van Riel <riel@redhat.com>
>>>> Cc: David Rientjes <rientjes@google.com>
>>>> Cc: Peter Zijlstra <peterz@infradead.org>
>>>> Cc: linux-mm@kvack.org
>>>> Cc: linux-arch@vger.kernel.org <linux-arch@vger.kernel.org>
>>>> Cc: Catalin Marinas <catalin.marinas@arm.com>
>>>> Cc: Max Filippov <jcmvbkbc@gmail.com>
>>>> ---
>>>>  mm/memory.c |    9 ++++++---
>>>>  1 file changed, 6 insertions(+), 3 deletions(-)
>>>>
>>>> diff --git a/mm/memory.c b/mm/memory.c
>>>> index 6dc1882..d9d5fd9 100644
>>>> --- a/mm/memory.c
>>>> +++ b/mm/memory.c
>>>> @@ -1110,6 +1110,7 @@ static unsigned long zap_pte_range(struct mmu_gather *tlb,
>>>>  	spinlock_t *ptl;
>>>>  	pte_t *start_pte;
>>>>  	pte_t *pte;
>>>> +	unsigned long range_start = addr;
>>>>  
>>>>  again:
>>>>  	init_rss_vec(rss);
>>>> @@ -1215,12 +1216,14 @@ again:
>>>>  		force_flush = 0;
>>>>  
>>>>  #ifdef HAVE_GENERIC_MMU_GATHER
>>>> -		tlb->start = addr;
>>>> -		tlb->end = end;
>>>> +		tlb->start = range_start;
>>>> +		tlb->end = addr;
>>>>  #endif
>>>>  		tlb_flush_mmu(tlb);
>>>> -		if (addr != end)
>>>> +		if (addr != end) {
>>>> +			range_start = addr;
>>>>  			goto again;
>>>> +		}
>>>>  	}
>>> Isn't this code only run if force_flush != 0? force_flush is set to
>>> !__tlb_remove_page() and this function always returns 1 on (generic TLB)
>>> UP since tlb_fast_mode() is 1. There is no batching on UP with the
>>> generic TLB code.
>> Correct ! That's why the changelog says I couldn't test it on ARC port itself :-)
>>
>> However based on the other discussion (Max's TLB/PTE inconsistency), as I started
>> writing code to reuse this block to flush the TLB even for non forced case, I
>> realized that what this is doing is incorrect and won't work for the general flushing.
> An alternative would be to make sure the above block is always called
> when tlb_fast_mode():
>
> diff --git a/mm/memory.c b/mm/memory.c
> index 6dc1882..f8b1f30 100644
> --- a/mm/memory.c
> +++ b/mm/memory.c
> @@ -1211,7 +1211,7 @@ again:
>  	 * the PTE lock to avoid doing the potential expensive TLB invalidate
>  	 * and page-free while holding it.
>  	 */
> -	if (force_flush) {
> +	if (force_flush || tlb_fast_mode(tlb)) {
>  		force_flush = 0;

I agree with tlb_fast_mode() addition (to solve Max's issue). The problem however
is that when we hit this at the end of loop - @addr is already pointing to @end so
range flush gets start = end - not what we really intended.

>> Ignoring all other threads, do we agree that the exiting code - if used in any
>> situations is incorrect semantically ?
> It is incorrect unless there are requirements for
> arch_leave_lazy_mmu_mode() to handle the TLB invalidation (it doesn't
> look like it's widely implemented though).

This patch is preparatory - independent of Max's issue. It is fixing just the
forced flush case - whoever uses it right now (ofcourse UP + generic TLB doesn't).

Thx,
-Vineet

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>