From: Greg Thelen <[email protected]>
Commit 077fcf116c8c ("mm/thp: allocate transparent hugepages on local
node") restructured alloc_hugepage_vma() with the intent of only
allocating transparent hugepages locally when there was not an effective
interleave mempolicy.
alloc_pages_exact_node() does not limit the allocation to the single
node, however, but rather prefers it. This is because __GFP_THISNODE is
not set which would cause the node-local nodemask to be passed. Without
it, only a nodemask that prefers the local node is passed.
Fix this by passing __GFP_THISNODE and falling back to small pages when
the allocation fails.
Fixes: 077fcf116c8c ("mm/thp: allocate transparent hugepages on local node")
Signed-off-by: Greg Thelen <[email protected]>
Signed-off-by: David Rientjes <[email protected]>
---
mm/mempolicy.c | 3 ++-
1 file changed, 2 insertions(+), 1 deletion(-)
diff --git a/mm/mempolicy.c b/mm/mempolicy.c
--- a/mm/mempolicy.c
+++ b/mm/mempolicy.c
@@ -1985,7 +1985,8 @@ retry_cpuset:
nmask = policy_nodemask(gfp, pol);
if (!nmask || node_isset(node, *nmask)) {
mpol_cond_put(pol);
- page = alloc_pages_exact_node(node, gfp, order);
+ page = alloc_pages_exact_node(node, gfp | GFP_THISNODE,
+ order);
goto out;
}
}
From: Greg Thelen <[email protected]>
Commit 077fcf116c8c ("mm/thp: allocate transparent hugepages on local
node") restructured alloc_hugepage_vma() with the intent of only
allocating transparent hugepages locally when there was not an effective
interleave mempolicy.
alloc_pages_exact_node() does not limit the allocation to the single
node, however, but rather prefers it. This is because __GFP_THISNODE is
not set which would cause the node-local nodemask to be passed. Without
it, only a nodemask that prefers the local node is passed.
Fix this by passing __GFP_THISNODE and falling back to small pages when
the allocation fails.
Fixes: 077fcf116c8c ("mm/thp: allocate transparent hugepages on local node")
Signed-off-by: Greg Thelen <[email protected]>
Signed-off-by: David Rientjes <[email protected]>
---
v2: GFP_THISNODE actually defers compaction and reclaim entirely based on
the combination of gfp flags. We want to try compaction and reclaim,
so only set __GFP_THISNODE. We still set __GFP_NOWARN to suppress
oom warnings in the kernel log when we can simply fallback to small
pages.
mm/mempolicy.c | 5 ++++-
1 file changed, 4 insertions(+), 1 deletion(-)
diff --git a/mm/mempolicy.c b/mm/mempolicy.c
--- a/mm/mempolicy.c
+++ b/mm/mempolicy.c
@@ -1985,7 +1985,10 @@ retry_cpuset:
nmask = policy_nodemask(gfp, pol);
if (!nmask || node_isset(node, *nmask)) {
mpol_cond_put(pol);
- page = alloc_pages_exact_node(node, gfp, order);
+ page = alloc_pages_exact_node(node, gfp |
+ __GFP_THISNODE |
+ __GFP_NOWARN,
+ order);
goto out;
}
}
On 02/25/2015 12:24 AM, David Rientjes wrote:
> From: Greg Thelen <[email protected]>
>
> Commit 077fcf116c8c ("mm/thp: allocate transparent hugepages on local
> node") restructured alloc_hugepage_vma() with the intent of only
> allocating transparent hugepages locally when there was not an effective
> interleave mempolicy.
>
> alloc_pages_exact_node() does not limit the allocation to the single
> node, however, but rather prefers it. This is because __GFP_THISNODE is
> not set which would cause the node-local nodemask to be passed. Without
> it, only a nodemask that prefers the local node is passed.
Oops, good catch.
But I believe we have the same problem with khugepaged_alloc_page(),
rendering the recent node determination and zone_reclaim strictness
patches partially useless.
Then I start to wonder about other alloc_pages_exact_node() users. Some
do pass __GFP_THISNODE, others not - are they also mistaken? I guess the
function is a misnomer - when I see "exact_node", I expect the
__GFP_THISNODE behavior.
I think to avoid such hidden catches, we should create
alloc_pages_preferred_node() variant, change the exact_node() variant to
pass __GFP_THISNODE, and audit and adjust all callers accordingly.
Also, you pass __GFP_NOWARN but that should be covered by GFP_TRANSHUGE
already. Of course, nothing guarantees that hugepage == true implies
that gfp == GFP_TRANSHUGE... but current in-tree callers conform to that.
> Fix this by passing __GFP_THISNODE and falling back to small pages when
> the allocation fails.
>
> Fixes: 077fcf116c8c ("mm/thp: allocate transparent hugepages on local node")
> Signed-off-by: Greg Thelen <[email protected]>
> Signed-off-by: David Rientjes <[email protected]>
> ---
> v2: GFP_THISNODE actually defers compaction and reclaim entirely based on
> the combination of gfp flags. We want to try compaction and reclaim,
> so only set __GFP_THISNODE. We still set __GFP_NOWARN to suppress
> oom warnings in the kernel log when we can simply fallback to small
> pages.
>
> mm/mempolicy.c | 5 ++++-
> 1 file changed, 4 insertions(+), 1 deletion(-)
>
> diff --git a/mm/mempolicy.c b/mm/mempolicy.c
> --- a/mm/mempolicy.c
> +++ b/mm/mempolicy.c
> @@ -1985,7 +1985,10 @@ retry_cpuset:
> nmask = policy_nodemask(gfp, pol);
> if (!nmask || node_isset(node, *nmask)) {
> mpol_cond_put(pol);
> - page = alloc_pages_exact_node(node, gfp, order);
> + page = alloc_pages_exact_node(node, gfp |
> + __GFP_THISNODE |
> + __GFP_NOWARN,
> + order);
> goto out;
> }
> }
>
On Wed, 25 Feb 2015, Vlastimil Babka wrote:
> > Commit 077fcf116c8c ("mm/thp: allocate transparent hugepages on local
> > node") restructured alloc_hugepage_vma() with the intent of only
> > allocating transparent hugepages locally when there was not an effective
> > interleave mempolicy.
> >
> > alloc_pages_exact_node() does not limit the allocation to the single
> > node, however, but rather prefers it. This is because __GFP_THISNODE is
> > not set which would cause the node-local nodemask to be passed. Without
> > it, only a nodemask that prefers the local node is passed.
>
> Oops, good catch.
> But I believe we have the same problem with khugepaged_alloc_page(), rendering
> the recent node determination and zone_reclaim strictness patches partially
> useless.
>
Indeed.
> Then I start to wonder about other alloc_pages_exact_node() users. Some do
> pass __GFP_THISNODE, others not - are they also mistaken? I guess the function
> is a misnomer - when I see "exact_node", I expect the __GFP_THISNODE behavior.
>
I looked through these yesterday as well and could only find the
do_migrate_pages() case for page migration where __GFP_THISNODE was
missing. I proposed that separately as
http://marc.info/?l=linux-mm&m=142481989722497 -- I couldn't find any
other users that looked wrong.
> I think to avoid such hidden catches, we should create
> alloc_pages_preferred_node() variant, change the exact_node() variant to pass
> __GFP_THISNODE, and audit and adjust all callers accordingly.
>
Sounds like that should be done as part of a cleanup after the 4.0 issues
are addressed. alloc_pages_exact_node() does seem to suggest that we want
exactly that node, implying __GFP_THISNODE behavior already, so it would
be good to avoid having this come up again in the future.
> Also, you pass __GFP_NOWARN but that should be covered by GFP_TRANSHUGE
> already. Of course, nothing guarantees that hugepage == true implies that gfp
> == GFP_TRANSHUGE... but current in-tree callers conform to that.
>
Ah, good point, and it includes __GFP_NORETRY as well which means that
this patch is busted. It won't try compaction or direct reclaim in the
page allocator slowpath because of this:
/*
* GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
* __GFP_NOWARN set) should not cause reclaim since the subsystem
* (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
* using a larger set of nodes after it has established that the
* allowed per node queues are empty and that nodes are
* over allocated.
*/
if (IS_ENABLED(CONFIG_NUMA) &&
(gfp_mask & GFP_THISNODE) == GFP_THISNODE)
goto nopage;
Hmm. It would be disappointing to have to pass the nodemask of the exact
node that we want to allocate from into the page allocator to avoid using
__GFP_THISNODE.
There's a sneaky way around it by just removing __GFP_NORETRY from
GFP_TRANSHUGE so the condition above fails and since the page allocator
won't retry for such a high-order allocation, but that probably just
papers over this stuff too much already. I think what we want to do is
cause the slab allocators to not use __GFP_WAIT if they want to avoid
reclaim.
This is probably going to be a much more invasive patch than originally
thought.
On 25.2.2015 22:24, David Rientjes wrote:
>
>> alloc_pages_preferred_node() variant, change the exact_node() variant to pass
>> __GFP_THISNODE, and audit and adjust all callers accordingly.
>>
> Sounds like that should be done as part of a cleanup after the 4.0 issues
> are addressed. alloc_pages_exact_node() does seem to suggest that we want
> exactly that node, implying __GFP_THISNODE behavior already, so it would
> be good to avoid having this come up again in the future.
Oh lovely, just found out that there's alloc_pages_node which should be the
preferred-only version, but in fact does not differ from
alloc_pages_exact_node
in any relevant way. I agree we should do some larger cleanup for next
version.
>> Also, you pass __GFP_NOWARN but that should be covered by GFP_TRANSHUGE
>> already. Of course, nothing guarantees that hugepage == true implies that gfp
>> == GFP_TRANSHUGE... but current in-tree callers conform to that.
>>
> Ah, good point, and it includes __GFP_NORETRY as well which means that
> this patch is busted. It won't try compaction or direct reclaim in the
> page allocator slowpath because of this:
>
> /*
> * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
> * __GFP_NOWARN set) should not cause reclaim since the subsystem
> * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
> * using a larger set of nodes after it has established that the
> * allowed per node queues are empty and that nodes are
> * over allocated.
> */
> if (IS_ENABLED(CONFIG_NUMA) &&
> (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
> goto nopage;
>
> Hmm. It would be disappointing to have to pass the nodemask of the exact
> node that we want to allocate from into the page allocator to avoid using
> __GFP_THISNODE.
Yeah.
>
> There's a sneaky way around it by just removing __GFP_NORETRY from
> GFP_TRANSHUGE so the condition above fails and since the page allocator
> won't retry for such a high-order allocation, but that probably just
> papers over this stuff too much already. I think what we want to do is
Alternatively alloc_pages_exact_node() adds __GFP_THISNODE just to
node_zonelist() call and not to __alloc_pages() gfp_mask proper? Unless
__GFP_THISNODE
was given *also* in the incoming gfp_mask, this should give us the right
combination?
But it's also subtle....
> cause the slab allocators to not use __GFP_WAIT if they want to avoid
> reclaim.
Yes, the fewer subtle heuristics we have that include combinations of
flags (*cough*
GFP_TRANSHUGE *cough*), the better.
> This is probably going to be a much more invasive patch than originally
> thought.
Right, we might be changing behavior not just for slab allocators, but
also others using such
combination of flags.
Vlastimil Babka <[email protected]> writes:
> On 25.2.2015 22:24, David Rientjes wrote:
>>
>>> alloc_pages_preferred_node() variant, change the exact_node() variant to pass
>>> __GFP_THISNODE, and audit and adjust all callers accordingly.
>>>
>> Sounds like that should be done as part of a cleanup after the 4.0 issues
>> are addressed. alloc_pages_exact_node() does seem to suggest that we want
>> exactly that node, implying __GFP_THISNODE behavior already, so it would
>> be good to avoid having this come up again in the future.
>
> Oh lovely, just found out that there's alloc_pages_node which should be the
> preferred-only version, but in fact does not differ from
> alloc_pages_exact_node
> in any relevant way. I agree we should do some larger cleanup for next
> version.
>
>>> Also, you pass __GFP_NOWARN but that should be covered by GFP_TRANSHUGE
>>> already. Of course, nothing guarantees that hugepage == true implies that gfp
>>> == GFP_TRANSHUGE... but current in-tree callers conform to that.
>>>
>> Ah, good point, and it includes __GFP_NORETRY as well which means that
>> this patch is busted. It won't try compaction or direct reclaim in the
>> page allocator slowpath because of this:
>>
>> /*
>> * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
>> * __GFP_NOWARN set) should not cause reclaim since the subsystem
>> * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
>> * using a larger set of nodes after it has established that the
>> * allowed per node queues are empty and that nodes are
>> * over allocated.
>> */
>> if (IS_ENABLED(CONFIG_NUMA) &&
>> (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
>> goto nopage;
>>
>> Hmm. It would be disappointing to have to pass the nodemask of the exact
>> node that we want to allocate from into the page allocator to avoid using
>> __GFP_THISNODE.
>
> Yeah.
>
>>
>> There's a sneaky way around it by just removing __GFP_NORETRY from
>> GFP_TRANSHUGE so the condition above fails and since the page allocator
>> won't retry for such a high-order allocation, but that probably just
>> papers over this stuff too much already. I think what we want to do is
>
> Alternatively alloc_pages_exact_node() adds __GFP_THISNODE just to
> node_zonelist() call and not to __alloc_pages() gfp_mask proper? Unless
> __GFP_THISNODE
> was given *also* in the incoming gfp_mask, this should give us the right
> combination?
> But it's also subtle....
>
>> cause the slab allocators to not use __GFP_WAIT if they want to avoid
>> reclaim.
>
> Yes, the fewer subtle heuristics we have that include combinations of
> flags (*cough*
> GFP_TRANSHUGE *cough*), the better.
>
>> This is probably going to be a much more invasive patch than originally
>> thought.
>
> Right, we might be changing behavior not just for slab allocators, but
> also others using such
> combination of flags.
Any update on this ? Did we reach a conclusion on how to go forward here
?
-aneesh
On 04/21/2015 09:31 AM, Aneesh Kumar K.V wrote:
> Vlastimil Babka <[email protected]> writes:
>
>> On 25.2.2015 22:24, David Rientjes wrote:
>>>
>>>> alloc_pages_preferred_node() variant, change the exact_node() variant to pass
>>>> __GFP_THISNODE, and audit and adjust all callers accordingly.
>>>>
>>> Sounds like that should be done as part of a cleanup after the 4.0 issues
>>> are addressed. alloc_pages_exact_node() does seem to suggest that we want
>>> exactly that node, implying __GFP_THISNODE behavior already, so it would
>>> be good to avoid having this come up again in the future.
>>
>> Oh lovely, just found out that there's alloc_pages_node which should be the
>> preferred-only version, but in fact does not differ from
>> alloc_pages_exact_node
>> in any relevant way. I agree we should do some larger cleanup for next
>> version.
>>
>>>> Also, you pass __GFP_NOWARN but that should be covered by GFP_TRANSHUGE
>>>> already. Of course, nothing guarantees that hugepage == true implies that gfp
>>>> == GFP_TRANSHUGE... but current in-tree callers conform to that.
>>>>
>>> Ah, good point, and it includes __GFP_NORETRY as well which means that
>>> this patch is busted. It won't try compaction or direct reclaim in the
>>> page allocator slowpath because of this:
>>>
>>> /*
>>> * GFP_THISNODE (meaning __GFP_THISNODE, __GFP_NORETRY and
>>> * __GFP_NOWARN set) should not cause reclaim since the subsystem
>>> * (f.e. slab) using GFP_THISNODE may choose to trigger reclaim
>>> * using a larger set of nodes after it has established that the
>>> * allowed per node queues are empty and that nodes are
>>> * over allocated.
>>> */
>>> if (IS_ENABLED(CONFIG_NUMA) &&
>>> (gfp_mask & GFP_THISNODE) == GFP_THISNODE)
>>> goto nopage;
>>>
>>> Hmm. It would be disappointing to have to pass the nodemask of the exact
>>> node that we want to allocate from into the page allocator to avoid using
>>> __GFP_THISNODE.
>>
>> Yeah.
>>
>>>
>>> There's a sneaky way around it by just removing __GFP_NORETRY from
>>> GFP_TRANSHUGE so the condition above fails and since the page allocator
>>> won't retry for such a high-order allocation, but that probably just
>>> papers over this stuff too much already. I think what we want to do is
>>
>> Alternatively alloc_pages_exact_node() adds __GFP_THISNODE just to
>> node_zonelist() call and not to __alloc_pages() gfp_mask proper? Unless
>> __GFP_THISNODE
>> was given *also* in the incoming gfp_mask, this should give us the right
>> combination?
>> But it's also subtle....
>>
>>> cause the slab allocators to not use __GFP_WAIT if they want to avoid
>>> reclaim.
>>
>> Yes, the fewer subtle heuristics we have that include combinations of
>> flags (*cough*
>> GFP_TRANSHUGE *cough*), the better.
>>
>>> This is probably going to be a much more invasive patch than originally
>>> thought.
>>
>> Right, we might be changing behavior not just for slab allocators, but
>> also others using such
>> combination of flags.
>
> Any update on this ? Did we reach a conclusion on how to go forward here
> ?
I believe David's later version was merged already. Or what exactly are
you asking about?
> -aneesh
>
Vlastimil Babka <[email protected]> writes:
> On 04/21/2015 09:31 AM, Aneesh Kumar K.V wrote:
>> Vlastimil Babka <[email protected]> writes:
>>
>>> On 25.2.2015 22:24, David Rientjes wrote:
>>>>
>>>>> alloc_pages_preferred_node() variant, change the exact_node() variant to pass
>>>>> __GFP_THISNODE, and audit and adjust all callers accordingly.
>>>>>
....
...
>>> Right, we might be changing behavior not just for slab allocators, but
>>> also others using such
>>> combination of flags.
>>
>> Any update on this ? Did we reach a conclusion on how to go forward here
>> ?
>
> I believe David's later version was merged already. Or what exactly are
> you asking about?
When I checked last time I didn't find it. Hence I asked here. Now I
see that it got committed as 5265047ac30191ea24b16503165000c225f54feb
Thanks
-aneesh