Fix and refactor do_{huge_pmd_}numa_page()

[RFC PATCH 1/3] memory tiering: read last_cpupid correctly in do_huge_pmd_numa_page()

Posted by Zi Yan 1 year, 5 months ago

From: Zi Yan <ziy@nvidia.com>

last_cpupid is only available when memory tiering is off or the folio
is in toptier node. Complete the check to read last_cpupid when it is
available.

Before the fix, the default last_cpupid will be used even if memory
tiering mode is turned off at runtime instead of the actual value. This
can prevent task_numa_fault() from getting right numa fault stats, but
should not cause any crash. User might see performance changes after the
fix.

Fixes: 33024536bafd ("memory tiering: hot page selection with hint page fault latency")
Signed-off-by: Zi Yan <ziy@nvidia.com>
---
 mm/huge_memory.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/mm/huge_memory.c b/mm/huge_memory.c
index d7c84480f1a4..07d9dde4ca33 100644
--- a/mm/huge_memory.c
+++ b/mm/huge_memory.c
@@ -1705,7 +1705,8 @@ vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf)
 	 * For memory tiering mode, cpupid of slow memory page is used
 	 * to record page access time.  So use default value.
 	 */
-	if (node_is_toptier(nid))
+	if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) ||
+	    node_is_toptier(nid))
 		last_cpupid = folio_last_cpupid(folio);
 	target_nid = numa_migrate_prep(folio, vmf, haddr, nid, &flags);
 	if (target_nid == NUMA_NO_NODE)
-- 
2.43.0

Re: [RFC PATCH 1/3] memory tiering: read last_cpupid correctly in do_huge_pmd_numa_page()

Posted by David Hildenbrand 1 year, 5 months ago

On 12.07.24 04:44, Zi Yan wrote:
> From: Zi Yan <ziy@nvidia.com>
> 
> last_cpupid is only available when memory tiering is off or the folio
> is in toptier node. Complete the check to read last_cpupid when it is
> available.
> 
> Before the fix, the default last_cpupid will be used even if memory
> tiering mode is turned off at runtime instead of the actual value. This
> can prevent task_numa_fault() from getting right numa fault stats, but
> should not cause any crash. User might see performance changes after the
> fix.
> 
> Fixes: 33024536bafd ("memory tiering: hot page selection with hint page fault latency")
> Signed-off-by: Zi Yan <ziy@nvidia.com>
> ---
>   mm/huge_memory.c | 3 ++-
>   1 file changed, 2 insertions(+), 1 deletion(-)
> 
> diff --git a/mm/huge_memory.c b/mm/huge_memory.c
> index d7c84480f1a4..07d9dde4ca33 100644
> --- a/mm/huge_memory.c
> +++ b/mm/huge_memory.c
> @@ -1705,7 +1705,8 @@ vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf)
>   	 * For memory tiering mode, cpupid of slow memory page is used
>   	 * to record page access time.  So use default value.
>   	 */
> -	if (node_is_toptier(nid))
> +	if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) ||
> +	    node_is_toptier(nid))
>   		last_cpupid = folio_last_cpupid(folio);
>   	target_nid = numa_migrate_prep(folio, vmf, haddr, nid, &flags);
>   	if (target_nid == NUMA_NO_NODE)

Reported-by: ...
Closes: ...

If it applies ;)

Acked-by: David Hildenbrand <david@redhat.com>

-- 
Cheers,

David / dhildenb

Re: [RFC PATCH 1/3] memory tiering: read last_cpupid correctly in do_huge_pmd_numa_page()

Posted by Zi Yan 1 year, 5 months ago

On 12 Jul 2024, at 21:13, David Hildenbrand wrote:

> On 12.07.24 04:44, Zi Yan wrote:
>> From: Zi Yan <ziy@nvidia.com>
>>
>> last_cpupid is only available when memory tiering is off or the folio
>> is in toptier node. Complete the check to read last_cpupid when it is
>> available.
>>
>> Before the fix, the default last_cpupid will be used even if memory
>> tiering mode is turned off at runtime instead of the actual value. This
>> can prevent task_numa_fault() from getting right numa fault stats, but
>> should not cause any crash. User might see performance changes after the
>> fix.
>>
>> Fixes: 33024536bafd ("memory tiering: hot page selection with hint page fault latency")
>> Signed-off-by: Zi Yan <ziy@nvidia.com>
>> ---
>>   mm/huge_memory.c | 3 ++-
>>   1 file changed, 2 insertions(+), 1 deletion(-)
>>
>> diff --git a/mm/huge_memory.c b/mm/huge_memory.c
>> index d7c84480f1a4..07d9dde4ca33 100644
>> --- a/mm/huge_memory.c
>> +++ b/mm/huge_memory.c
>> @@ -1705,7 +1705,8 @@ vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf)
>>   	 * For memory tiering mode, cpupid of slow memory page is used
>>   	 * to record page access time.  So use default value.
>>   	 */
>> -	if (node_is_toptier(nid))
>> +	if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) ||
>> +	    node_is_toptier(nid))
>>   		last_cpupid = folio_last_cpupid(folio);
>>   	target_nid = numa_migrate_prep(folio, vmf, haddr, nid, &flags);
>>   	if (target_nid == NUMA_NO_NODE)
>
> Reported-by: ...

Reported-by: David Hildenbrand <david@redhat.com>

I suppose your email[1] reports the issue based on code inspection.

> Closes: ...

Closes: https://lore.kernel.org/linux-mm/9af34a6b-ca56-4a64-8aa6-ade65f109288@redhat.com/

Will add them in the next version.

>
> If it applies ;)
>
> Acked-by: David Hildenbrand <david@redhat.com>

Thanks.


[1] https://lore.kernel.org/linux-mm/9af34a6b-ca56-4a64-8aa6-ade65f109288@redhat.com/

--
Best Regards,
Yan, Zi

Re: [RFC PATCH 1/3] memory tiering: read last_cpupid correctly in do_huge_pmd_numa_page()

Posted by David Hildenbrand 1 year, 5 months ago

On 13.07.24 03:18, Zi Yan wrote:
> On 12 Jul 2024, at 21:13, David Hildenbrand wrote:
> 
>> On 12.07.24 04:44, Zi Yan wrote:
>>> From: Zi Yan <ziy@nvidia.com>
>>>
>>> last_cpupid is only available when memory tiering is off or the folio
>>> is in toptier node. Complete the check to read last_cpupid when it is
>>> available.
>>>
>>> Before the fix, the default last_cpupid will be used even if memory
>>> tiering mode is turned off at runtime instead of the actual value. This
>>> can prevent task_numa_fault() from getting right numa fault stats, but
>>> should not cause any crash. User might see performance changes after the
>>> fix.
>>>
>>> Fixes: 33024536bafd ("memory tiering: hot page selection with hint page fault latency")
>>> Signed-off-by: Zi Yan <ziy@nvidia.com>
>>> ---
>>>    mm/huge_memory.c | 3 ++-
>>>    1 file changed, 2 insertions(+), 1 deletion(-)
>>>
>>> diff --git a/mm/huge_memory.c b/mm/huge_memory.c
>>> index d7c84480f1a4..07d9dde4ca33 100644
>>> --- a/mm/huge_memory.c
>>> +++ b/mm/huge_memory.c
>>> @@ -1705,7 +1705,8 @@ vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf)
>>>    	 * For memory tiering mode, cpupid of slow memory page is used
>>>    	 * to record page access time.  So use default value.
>>>    	 */
>>> -	if (node_is_toptier(nid))
>>> +	if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) ||
>>> +	    node_is_toptier(nid))
>>>    		last_cpupid = folio_last_cpupid(folio);
>>>    	target_nid = numa_migrate_prep(folio, vmf, haddr, nid, &flags);
>>>    	if (target_nid == NUMA_NO_NODE)
>>
>> Reported-by: ...
> 
> Reported-by: David Hildenbrand <david@redhat.com>
> 
> I suppose your email[1] reports the issue based on code inspection.


Yes, thanks for taking care of the fix and cleanups!

-- 
Cheers,

David / dhildenb

Re: [RFC PATCH 1/3] memory tiering: read last_cpupid correctly in do_huge_pmd_numa_page()

Posted by Baolin Wang 1 year, 5 months ago


On 2024/7/12 10:44, Zi Yan wrote:
> From: Zi Yan <ziy@nvidia.com>
> 
> last_cpupid is only available when memory tiering is off or the folio
> is in toptier node. Complete the check to read last_cpupid when it is
> available.
> 
> Before the fix, the default last_cpupid will be used even if memory
> tiering mode is turned off at runtime instead of the actual value. This
> can prevent task_numa_fault() from getting right numa fault stats, but
> should not cause any crash. User might see performance changes after the
> fix.
> 
> Fixes: 33024536bafd ("memory tiering: hot page selection with hint page fault latency")
> Signed-off-by: Zi Yan <ziy@nvidia.com>

LGTM.
Reviewed-by: Baolin Wang <baolin.wang@linux.alibaba.com>

> ---
>   mm/huge_memory.c | 3 ++-
>   1 file changed, 2 insertions(+), 1 deletion(-)
> 
> diff --git a/mm/huge_memory.c b/mm/huge_memory.c
> index d7c84480f1a4..07d9dde4ca33 100644
> --- a/mm/huge_memory.c
> +++ b/mm/huge_memory.c
> @@ -1705,7 +1705,8 @@ vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf)
>   	 * For memory tiering mode, cpupid of slow memory page is used
>   	 * to record page access time.  So use default value.
>   	 */
> -	if (node_is_toptier(nid))
> +	if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) ||
> +	    node_is_toptier(nid))
>   		last_cpupid = folio_last_cpupid(folio);
>   	target_nid = numa_migrate_prep(folio, vmf, haddr, nid, &flags);
>   	if (target_nid == NUMA_NO_NODE)

Re: [RFC PATCH 1/3] memory tiering: read last_cpupid correctly in do_huge_pmd_numa_page()

Posted by Huang, Ying 1 year, 5 months ago

Zi Yan <zi.yan@sent.com> writes:

> From: Zi Yan <ziy@nvidia.com>
>
> last_cpupid is only available when memory tiering is off or the folio
> is in toptier node. Complete the check to read last_cpupid when it is
> available.
>
> Before the fix, the default last_cpupid will be used even if memory
> tiering mode is turned off at runtime instead of the actual value. This
> can prevent task_numa_fault() from getting right numa fault stats, but
> should not cause any crash. User might see performance changes after the
> fix.
>
> Fixes: 33024536bafd ("memory tiering: hot page selection with hint page fault latency")
> Signed-off-by: Zi Yan <ziy@nvidia.com>

Good catch!  Thanks!

Reviewed-by: "Huang, Ying" <ying.huang@intel.com>

> ---
>  mm/huge_memory.c | 3 ++-
>  1 file changed, 2 insertions(+), 1 deletion(-)
>
> diff --git a/mm/huge_memory.c b/mm/huge_memory.c
> index d7c84480f1a4..07d9dde4ca33 100644
> --- a/mm/huge_memory.c
> +++ b/mm/huge_memory.c
> @@ -1705,7 +1705,8 @@ vm_fault_t do_huge_pmd_numa_page(struct vm_fault *vmf)
>  	 * For memory tiering mode, cpupid of slow memory page is used
>  	 * to record page access time.  So use default value.
>  	 */
> -	if (node_is_toptier(nid))
> +	if (!(sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) ||
> +	    node_is_toptier(nid))
>  		last_cpupid = folio_last_cpupid(folio);
>  	target_nid = numa_migrate_prep(folio, vmf, haddr, nid, &flags);
>  	if (target_nid == NUMA_NO_NODE)