Re: [PATCH] Don't touch single threaded PTEs which are on the right node
On Thu, Oct 13, 2016 at 11:04:02AM -0700, Andi Kleen wrote:
> > > do {
> > > oldpte = *pte;
> > > @@ -94,6 +100,13 @@ static unsigned long change_pte_range(struct
> > > vm_area_struct *vma, pmd_t *pmd,
> > > /* Avoid TLB flush if possible */
> > > if (pte_protnone(oldpte))
> > > continue;
> > > +
> > > + /*
> > > + * Don't mess with PTEs if page is already on
> > > the node
> > > + * a single-threaded process is running on.
> > > + */
> > > + if (target_node == page_to_nid(page))
> > > + continue;
> > > }
> > >
> >
> > Check target_node != NUMA_NODE && target_node == page_to_nid(page) to
> > avoid unnecessary page->flag masking and shifts?
>
> I didn't do this last change because I expect a potentially mispredicted
> check is more expensive than some shifting/masking.
>
Ok, that's fair enough. For something that minor I expect it to be a
case of "you win some you lose some" depending on workload, CPU and
phase of the moon.
Thanks.
--
Mel Gorman
SUSE Labs
[PATCH] Don't touch single threaded PTEs which are on the right node
From: Andi Kleen
We had some problems with pages getting unmapped in single threaded
affinitized processes. It was tracked down to NUMA scanning.
In this case it doesn't make any sense to unmap pages if the
process is single threaded and the page is already on the
node the process is running on.
Add a check for this case into the numa protection code,
and skip unmapping if true.
In theory the process could be migrated later, but we
will eventually rescan and unmap and migrate then.
In theory this could be made more fancy: remembering this
state per process or even whole mm. However that would
need extra tracking and be more complicated, and the
simple check seems to work fine so far.
v2: Only do it for private VMAs. Move most of check out of
loop.
v3: Minor updates from Mel. Change code layout.
Acked-by: Mel Gorman
Signed-off-by: Andi Kleen
---
mm/mprotect.c | 13 +
1 file changed, 13 insertions(+)
diff --git a/mm/mprotect.c b/mm/mprotect.c
index a4830f0325fe..11b8857c3437 100644
--- a/mm/mprotect.c
+++ b/mm/mprotect.c
@@ -68,11 +68,17 @@ static unsigned long change_pte_range(struct vm_area_struct
*vma, pmd_t *pmd,
pte_t *pte, oldpte;
spinlock_t *ptl;
unsigned long pages = 0;
+ int target_node = NUMA_NO_NODE;
pte = lock_pte_protection(vma, pmd, addr, prot_numa, &ptl);
if (!pte)
return 0;
+ /* Get target node for single threaded private VMAs */
+ if (prot_numa && !(vma->vm_flags & VM_SHARED) &&
+ atomic_read(&vma->vm_mm->mm_users) == 1)
+ target_node = numa_node_id();
+
arch_enter_lazy_mmu_mode();
do {
oldpte = *pte;
@@ -94,6 +100,13 @@ static unsigned long change_pte_range(struct vm_area_struct
*vma, pmd_t *pmd,
/* Avoid TLB flush if possible */
if (pte_protnone(oldpte))
continue;
+
+ /*
+* Don't mess with PTEs if page is already on
the node
+* a single-threaded process is running on.
+*/
+ if (target_node == page_to_nid(page))
+ continue;
}
ptent = ptep_modify_prot_start(mm, addr, pte);
--
2.5.5
Re: [PATCH] Don't touch single threaded PTEs which are on the right node
> > do {
> > oldpte = *pte;
> > @@ -94,6 +100,13 @@ static unsigned long change_pte_range(struct
> > vm_area_struct *vma, pmd_t *pmd,
> > /* Avoid TLB flush if possible */
> > if (pte_protnone(oldpte))
> > continue;
> > +
> > + /*
> > +* Don't mess with PTEs if page is already on
> > the node
> > +* a single-threaded process is running on.
> > +*/
> > + if (target_node == page_to_nid(page))
> > + continue;
> > }
> >
>
> Check target_node != NUMA_NODE && target_node == page_to_nid(page) to
> avoid unnecessary page->flag masking and shifts?
I didn't do this last change because I expect a potentially mispredicted
check is more expensive than some shifting/masking.
-Andi
Re: [PATCH] Don't touch single threaded PTEs which are on the right node
On Wed, Oct 12, 2016 at 09:15:49AM -0700, Andi Kleen wrote:
> From: Andi Kleen
>
> We had some problems with pages getting unmapped in single threaded
> affinitized processes. It was tracked down to NUMA scanning.
>
> In this case it doesn't make any sense to unmap pages if the
> process is single threaded and the page is already on the
> node the process is running on.
>
> Add a check for this case into the numa protection code,
> and skip unmapping if true.
>
> In theory the process could be migrated later, but we
> will eventually rescan and unmap and migrate then.
>
> In theory this could be made more fancy: remembering this
> state per process or even whole mm. However that would
> need extra tracking and be more complicated, and the
> simple check seems to work fine so far.
>
> v2: Only do it for private VMAs. Move most of check out of
> loop.
> Signed-off-by: Andi Kleen
Minor comments
> ---
> mm/mprotect.c | 13 +
> 1 file changed, 13 insertions(+)
>
> diff --git a/mm/mprotect.c b/mm/mprotect.c
> index a4830f0325fe..e9473e7e1468 100644
> --- a/mm/mprotect.c
> +++ b/mm/mprotect.c
> @@ -68,11 +68,17 @@ static unsigned long change_pte_range(struct
> vm_area_struct *vma, pmd_t *pmd,
> pte_t *pte, oldpte;
> spinlock_t *ptl;
> unsigned long pages = 0;
> + int target_node = -1;
>
Proper convention is to use NUMA_NO_NODE instead of -1 although it's not
always adhered to.
> pte = lock_pte_protection(vma, pmd, addr, prot_numa, &ptl);
> if (!pte)
> return 0;
>
> + if (prot_numa &&
> + !(vma->vm_flags & VM_SHARED) &&
> + atomic_read(&vma->vm_mm->mm_users) == 1)
> + target_node = cpu_to_node(raw_smp_processor_id());
> +
Use numa_node_id() instead of open-coding this. A short comment probably
would not hurt even if git blame should make it obvious.
> arch_enter_lazy_mmu_mode();
> do {
> oldpte = *pte;
> @@ -94,6 +100,13 @@ static unsigned long change_pte_range(struct
> vm_area_struct *vma, pmd_t *pmd,
> /* Avoid TLB flush if possible */
> if (pte_protnone(oldpte))
> continue;
> +
> + /*
> + * Don't mess with PTEs if page is already on
> the node
> + * a single-threaded process is running on.
> + */
> + if (target_node == page_to_nid(page))
> + continue;
> }
>
Check target_node != NUMA_NODE && target_node == page_to_nid(page) to
avoid unnecessary page->flag masking and shifts?
The last one will be fairly marginal, the others are taste so whether
you spin a v3 with the corrections or not;
Acked-by: Mel Gorman
Thanks.
--
Mel Gorman
SUSE Labs
[PATCH] Don't touch single threaded PTEs which are on the right node
From: Andi Kleen
We had some problems with pages getting unmapped in single threaded
affinitized processes. It was tracked down to NUMA scanning.
In this case it doesn't make any sense to unmap pages if the
process is single threaded and the page is already on the
node the process is running on.
Add a check for this case into the numa protection code,
and skip unmapping if true.
In theory the process could be migrated later, but we
will eventually rescan and unmap and migrate then.
In theory this could be made more fancy: remembering this
state per process or even whole mm. However that would
need extra tracking and be more complicated, and the
simple check seems to work fine so far.
v2: Only do it for private VMAs. Move most of check out of
loop.
Signed-off-by: Andi Kleen
---
mm/mprotect.c | 13 +
1 file changed, 13 insertions(+)
diff --git a/mm/mprotect.c b/mm/mprotect.c
index a4830f0325fe..e9473e7e1468 100644
--- a/mm/mprotect.c
+++ b/mm/mprotect.c
@@ -68,11 +68,17 @@ static unsigned long change_pte_range(struct vm_area_struct
*vma, pmd_t *pmd,
pte_t *pte, oldpte;
spinlock_t *ptl;
unsigned long pages = 0;
+ int target_node = -1;
pte = lock_pte_protection(vma, pmd, addr, prot_numa, &ptl);
if (!pte)
return 0;
+ if (prot_numa &&
+ !(vma->vm_flags & VM_SHARED) &&
+ atomic_read(&vma->vm_mm->mm_users) == 1)
+ target_node = cpu_to_node(raw_smp_processor_id());
+
arch_enter_lazy_mmu_mode();
do {
oldpte = *pte;
@@ -94,6 +100,13 @@ static unsigned long change_pte_range(struct vm_area_struct
*vma, pmd_t *pmd,
/* Avoid TLB flush if possible */
if (pte_protnone(oldpte))
continue;
+
+ /*
+* Don't mess with PTEs if page is already on
the node
+* a single-threaded process is running on.
+*/
+ if (target_node == page_to_nid(page))
+ continue;
}
ptent = ptep_modify_prot_start(mm, addr, pte);
--
2.5.5
Re: [PATCH] Don't touch single threaded PTEs which are on the right node
> You shouldn't need to check the number of mm_users and the node the task
> is running on for every PTE being scanned.
Ok.
>
> A more important corner case is if the VMA is shared with a task running on
> another node. By avoiding the NUMA hinting faults here, the hinting faults
> trapped by the remote process will appear exclusive and allow migration of
> the page. This will happen even if the single-threade task is continually
> using the pages.
>
> When you said "we had some problems", you didn't describe the workload or
> what the problems were (I'm assuming latency/jitter). Would restricting
> this check to private VMAs be sufficient?
The problem we ran into was that prefetches were not working, but
yes it would also cause extra latencies and jitter and in general
is unnecessary overhead.
It is super easy to reproduce. Just run main() {for(;;);}
It will eventually get some of its pages unmapped.
Yes doing it for private only would be fine. I'll add a check
for that.
-Andi
Re: [PATCH] Don't touch single threaded PTEs which are on the right node
On Tue, Oct 11, 2016 at 01:28:58PM -0700, Andi Kleen wrote: > From: Andi Kleen > > We had some problems with pages getting unmapped in single threaded > affinitized processes. It was tracked down to NUMA scanning. > > In this case it doesn't make any sense to unmap pages if the > process is single threaded and the page is already on the > node the process is running on. > > Add a check for this case into the numa protection code, > and skip unmapping if true. > > In theory the process could be migrated later, but we > will eventually rescan and unmap and migrate then. > > In theory this could be made more fancy: remembering this > state per process or even whole mm. However that would > need extra tracking and be more complicated, and the > simple check seems to work fine so far. > > Signed-off-by: Andi Kleen > --- > mm/mprotect.c | 8 > 1 file changed, 8 insertions(+) > > diff --git a/mm/mprotect.c b/mm/mprotect.c > index a4830f0325fe..e8028658e817 100644 > --- a/mm/mprotect.c > +++ b/mm/mprotect.c > @@ -94,6 +94,14 @@ static unsigned long change_pte_range(struct > vm_area_struct *vma, pmd_t *pmd, > /* Avoid TLB flush if possible */ > if (pte_protnone(oldpte)) > continue; > + > + /* > + * Don't mess with PTEs if page is already on > the node > + * a single-threaded process is running on. > + */ > + if (atomic_read(&vma->vm_mm->mm_users) == 1 && > + cpu_to_node(raw_smp_processor_id()) == > page_to_nid(page)) > + continue; > } You shouldn't need to check the number of mm_users and the node the task is running on for every PTE being scanned. A more important corner case is if the VMA is shared with a task running on another node. By avoiding the NUMA hinting faults here, the hinting faults trapped by the remote process will appear exclusive and allow migration of the page. This will happen even if the single-threade task is continually using the pages. When you said "we had some problems", you didn't describe the workload or what the problems were (I'm assuming latency/jitter). Would restricting this check to private VMAs be sufficient? -- Mel Gorman SUSE Labs
[PATCH] Don't touch single threaded PTEs which are on the right node
From: Andi Kleen We had some problems with pages getting unmapped in single threaded affinitized processes. It was tracked down to NUMA scanning. In this case it doesn't make any sense to unmap pages if the process is single threaded and the page is already on the node the process is running on. Add a check for this case into the numa protection code, and skip unmapping if true. In theory the process could be migrated later, but we will eventually rescan and unmap and migrate then. In theory this could be made more fancy: remembering this state per process or even whole mm. However that would need extra tracking and be more complicated, and the simple check seems to work fine so far. Signed-off-by: Andi Kleen --- mm/mprotect.c | 8 1 file changed, 8 insertions(+) diff --git a/mm/mprotect.c b/mm/mprotect.c index a4830f0325fe..e8028658e817 100644 --- a/mm/mprotect.c +++ b/mm/mprotect.c @@ -94,6 +94,14 @@ static unsigned long change_pte_range(struct vm_area_struct *vma, pmd_t *pmd, /* Avoid TLB flush if possible */ if (pte_protnone(oldpte)) continue; + + /* +* Don't mess with PTEs if page is already on the node +* a single-threaded process is running on. +*/ + if (atomic_read(&vma->vm_mm->mm_users) == 1 && + cpu_to_node(raw_smp_processor_id()) == page_to_nid(page)) + continue; } ptent = ptep_modify_prot_start(mm, addr, pte); -- 2.5.5
