When transparent_hugepage is set to never by cmdline or echo, defrag
still show what it used to be and can be modified which makes user
confusing whether defrag would take effect.
Actually if transparent_hugepage is set to never, defrag will not take
effect. Just Display never and remain unchangeable to for defrag when
transparent_hugepage is set to never.
Suggested-by: Liu Song <[email protected]>
Signed-off-by: Yuanhe Shu <[email protected]>
---
mm/huge_memory.c | 38 ++++++++++++++++++++++++--------------
1 file changed, 24 insertions(+), 14 deletions(-)
diff --git a/mm/huge_memory.c b/mm/huge_memory.c
index 94c958f7ebb5..03cd22a87b93 100644
--- a/mm/huge_memory.c
+++ b/mm/huge_memory.c
@@ -349,21 +349,25 @@ static ssize_t defrag_show(struct kobject *kobj,
{
const char *output;
- if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
- &transparent_hugepage_flags))
- output = "[always] defer defer+madvise madvise never";
- else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
- &transparent_hugepage_flags))
- output = "always [defer] defer+madvise madvise never";
- else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG,
- &transparent_hugepage_flags))
- output = "always defer [defer+madvise] madvise never";
- else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG,
- &transparent_hugepage_flags))
- output = "always defer defer+madvise [madvise] never";
- else
+ if (!test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags) &&
+ !test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags))
output = "always defer defer+madvise madvise [never]";
-
+ else {
+ if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_DIRECT_FLAG,
+ &transparent_hugepage_flags))
+ output = "[always] defer defer+madvise madvise never";
+ else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG,
+ &transparent_hugepage_flags))
+ output = "always [defer] defer+madvise madvise never";
+ else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG,
+ &transparent_hugepage_flags))
+ output = "always defer [defer+madvise] madvise never";
+ else if (test_bit(TRANSPARENT_HUGEPAGE_DEFRAG_REQ_MADV_FLAG,
+ &transparent_hugepage_flags))
+ output = "always defer defer+madvise [madvise] never";
+ else
+ output = "always defer defer+madvise madvise [never]";
+ }
return sysfs_emit(buf, "%s\n", output);
}
@@ -371,6 +375,12 @@ static ssize_t defrag_store(struct kobject *kobj,
struct kobj_attribute *attr,
const char *buf, size_t count)
{
+ if (!test_bit(TRANSPARENT_HUGEPAGE_FLAG, &transparent_hugepage_flags) &&
+ !test_bit(TRANSPARENT_HUGEPAGE_REQ_MADV_FLAG, &transparent_hugepage_flags)) {
+ pr_warn("transparent_hugepage not enabled, defrag takes no effect\n");
+ return -EINVAL;
+ }
+
if (sysfs_streq(buf, "always")) {
clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_FLAG, &transparent_hugepage_flags);
clear_bit(TRANSPARENT_HUGEPAGE_DEFRAG_KSWAPD_OR_MADV_FLAG, &transparent_hugepage_flags);
--
2.39.3
On 22.02.24 12:53, Yuanhe Shu wrote:
> When transparent_hugepage is set to never by cmdline or echo, defrag
> still show what it used to be and can be modified which makes user
> confusing whether defrag would take effect.
>
> Actually if transparent_hugepage is set to never, defrag will not take
> effect. Just Display never and remain unchangeable to for defrag when
> transparent_hugepage is set to never.
>
> Suggested-by: Liu Song <[email protected]>
> Signed-off-by: Yuanhe Shu <[email protected]>
> ---
No, I don't think we want such a dependency between both options.
You might just end up breaking existing scripts (enable defrag before
enabling THP) for no good reason.
--
Cheers,
David / dhildenb
在 2024/2/22 20:14, David Hildenbrand 写道:
> On 22.02.24 12:53, Yuanhe Shu wrote:
>> When transparent_hugepage is set to never by cmdline or echo, defrag
>> still show what it used to be and can be modified which makes user
>> confusing whether defrag would take effect.
>>
>> Actually if transparent_hugepage is set to never, defrag will not take
>> effect. Just Display never and remain unchangeable to for defrag when
>> transparent_hugepage is set to never.
>>
>> Suggested-by: Liu Song <[email protected]>
>> Signed-off-by: Yuanhe Shu <[email protected]>
>> ---
>
> No, I don't think we want such a dependency between both options.
>
> You might just end up breaking existing scripts (enable defrag before
> enabling THP) for no good reason.
>
In certain situations where khugepaged_thread is NULL, it would be more
reasonable for the value of
/sys/kernel/mm/transparent_hugepage/khugepaged/defrag to be 0. The patch
should include a fix for this case.
Thanks
On 23.02.24 12:01, Liu Song wrote:
>
> 在 2024/2/22 20:14, David Hildenbrand 写道:
>> On 22.02.24 12:53, Yuanhe Shu wrote:
>>> When transparent_hugepage is set to never by cmdline or echo, defrag
>>> still show what it used to be and can be modified which makes user
>>> confusing whether defrag would take effect.
>>>
>>> Actually if transparent_hugepage is set to never, defrag will not take
>>> effect. Just Display never and remain unchangeable to for defrag when
>>> transparent_hugepage is set to never.
>>>
>>> Suggested-by: Liu Song <[email protected]>
>>> Signed-off-by: Yuanhe Shu <[email protected]>
>>> ---
>>
>> No, I don't think we want such a dependency between both options.
>>
>> You might just end up breaking existing scripts (enable defrag before
>> enabling THP) for no good reason.
>>
> In certain situations where khugepaged_thread is NULL, it would be more
> reasonable for the value of
> /sys/kernel/mm/transparent_hugepage/khugepaged/defrag to be 0. The patch
> should include a fix for this case.
Why?
We have a bunch of THP toggles. They reside in
"/sys/kernel/mm/transparent_hugepage/", indicating that they are THP
specific.
Some of them are only in effect if some other toggles are set.
That is very common practice.
If you think something could be confusing, maybe clarify the doc? I
don't immediately see why any code changes are required, really.
--
Cheers,
David / dhildenb
在 2024/2/26 16:54, David Hildenbrand 写道:
> On 23.02.24 12:01, Liu Song wrote:
>>
>> 在 2024/2/22 20:14, David Hildenbrand 写道:
>>> On 22.02.24 12:53, Yuanhe Shu wrote:
>>>> When transparent_hugepage is set to never by cmdline or echo, defrag
>>>> still show what it used to be and can be modified which makes user
>>>> confusing whether defrag would take effect.
>>>>
>>>> Actually if transparent_hugepage is set to never, defrag will not take
>>>> effect. Just Display never and remain unchangeable to for defrag when
>>>> transparent_hugepage is set to never.
>>>>
>>>> Suggested-by: Liu Song <[email protected]>
>>>> Signed-off-by: Yuanhe Shu <[email protected]>
>>>> ---
>>>
>>> No, I don't think we want such a dependency between both options.
>>>
>>> You might just end up breaking existing scripts (enable defrag before
>>> enabling THP) for no good reason.
>>>
>> In certain situations where khugepaged_thread is NULL, it would be more
>> reasonable for the value of
>> /sys/kernel/mm/transparent_hugepage/khugepaged/defrag to be 0. The patch
>> should include a fix for this case.
>
> Why?
>
> We have a bunch of THP toggles. They reside in
> "/sys/kernel/mm/transparent_hugepage/", indicating that they are THP
> specific.
>
> Some of them are only in effect if some other toggles are set.
>
> That is very common practice.
>
> If you think something could be confusing, maybe clarify the doc? I
> don't immediately see why any code changes are required, really.
We should explain this in the documentation, but to be honest, many
people don't read the documentation, and even after we explicitly
disable THP with transparent_hugepage=never, khugepaged/defrag is still
set to 1, and users come asking why it's still defragging. We can't
expect all users to have technical expertise, or to diligently read
through the documentation; it would obviously be best if we could avoid
user confusion altogether.
Thanks
On 26.02.24 12:48, David Hildenbrand wrote:
> On 26.02.24 12:22, Liu Song wrote:
>>
>> 在 2024/2/26 16:54, David Hildenbrand 写道:
>>> On 23.02.24 12:01, Liu Song wrote:
>>>>
>>>> 在 2024/2/22 20:14, David Hildenbrand 写道:
>>>>> On 22.02.24 12:53, Yuanhe Shu wrote:
>>>>>> When transparent_hugepage is set to never by cmdline or echo, defrag
>>>>>> still show what it used to be and can be modified which makes user
>>>>>> confusing whether defrag would take effect.
>>>>>>
>>>>>> Actually if transparent_hugepage is set to never, defrag will not take
>>>>>> effect. Just Display never and remain unchangeable to for defrag when
>>>>>> transparent_hugepage is set to never.
>>>>>>
>>>>>> Suggested-by: Liu Song <[email protected]>
>>>>>> Signed-off-by: Yuanhe Shu <[email protected]>
>>>>>> ---
>>>>>
>>>>> No, I don't think we want such a dependency between both options.
>>>>>
>>>>> You might just end up breaking existing scripts (enable defrag before
>>>>> enabling THP) for no good reason.
>>>>>
>>>> In certain situations where khugepaged_thread is NULL, it would be more
>>>> reasonable for the value of
>>>> /sys/kernel/mm/transparent_hugepage/khugepaged/defrag to be 0. The patch
>>>> should include a fix for this case.
>>>
>>> Why?
>>>
>>> We have a bunch of THP toggles. They reside in
>>> "/sys/kernel/mm/transparent_hugepage/", indicating that they are THP
>>> specific.
>>>
>>> Some of them are only in effect if some other toggles are set.
>>>
>>> That is very common practice.
>>>
>>> If you think something could be confusing, maybe clarify the doc? I
>>> don't immediately see why any code changes are required, really.
>>
>> We should explain this in the documentation, but to be honest, many
>> people don't read the documentation, and even after we explicitly
>> disable THP with transparent_hugepage=never, khugepaged/defrag is still
>> set to 1, and users come asking why it's still defragging. We can't
>> expect all users to have technical expertise, or to diligently read
>> through the documentation; it would obviously be best if we could avoid
>> user confusion altogether.
>
> Then tell your users if a feature is enabled
s/enabled/disabled/
s/1on1/101/
Time for a break :D
Anyhow, there has to be a pretty good reason to change the semantics of
these toggles that have been around forever. "users could be confused"
is a not sufficient.
--
Cheers,
David / dhildenb
On 26.02.24 12:22, Liu Song wrote:
>
> 在 2024/2/26 16:54, David Hildenbrand 写道:
>> On 23.02.24 12:01, Liu Song wrote:
>>>
>>> 在 2024/2/22 20:14, David Hildenbrand 写道:
>>>> On 22.02.24 12:53, Yuanhe Shu wrote:
>>>>> When transparent_hugepage is set to never by cmdline or echo, defrag
>>>>> still show what it used to be and can be modified which makes user
>>>>> confusing whether defrag would take effect.
>>>>>
>>>>> Actually if transparent_hugepage is set to never, defrag will not take
>>>>> effect. Just Display never and remain unchangeable to for defrag when
>>>>> transparent_hugepage is set to never.
>>>>>
>>>>> Suggested-by: Liu Song <[email protected]>
>>>>> Signed-off-by: Yuanhe Shu <[email protected]>
>>>>> ---
>>>>
>>>> No, I don't think we want such a dependency between both options.
>>>>
>>>> You might just end up breaking existing scripts (enable defrag before
>>>> enabling THP) for no good reason.
>>>>
>>> In certain situations where khugepaged_thread is NULL, it would be more
>>> reasonable for the value of
>>> /sys/kernel/mm/transparent_hugepage/khugepaged/defrag to be 0. The patch
>>> should include a fix for this case.
>>
>> Why?
>>
>> We have a bunch of THP toggles. They reside in
>> "/sys/kernel/mm/transparent_hugepage/", indicating that they are THP
>> specific.
>>
>> Some of them are only in effect if some other toggles are set.
>>
>> That is very common practice.
>>
>> If you think something could be confusing, maybe clarify the doc? I
>> don't immediately see why any code changes are required, really.
>
> We should explain this in the documentation, but to be honest, many
> people don't read the documentation, and even after we explicitly
> disable THP with transparent_hugepage=never, khugepaged/defrag is still
> set to 1, and users come asking why it's still defragging. We can't
> expect all users to have technical expertise, or to diligently read
> through the documentation; it would obviously be best if we could avoid
> user confusion altogether.
Then tell your users if a feature is enabled, other toggles that affect
that feature are ignored. Basic 1on1.
If my care engine is off but it's not in neutral, it's not suddenly
driving. And most users understand that reasoning.
--
Cheers,
David / dhildenb