[PATCH 0/2 v5][resend] tmpfs not interleaving properly

[PATCH 0/2 v5][resend] tmpfs not interleaving properly

[Nathan Zimmer]

When tmpfs has the memory policy interleaved it always starts allocating at each
file at node 0.  When there are many small files the lower nodes fill up
disproportionately.
This patch spreads out node usage by starting files at nodes other then 0.
The tmpfs superblock grants an offset for each inode as they are created. Each
then uses that offset to proved a prefered first node for its interleave in
the shmem_interleave.

v2: passed preferred node via addr.
v3: using current->cpuset_mem_spread_rotor instead of random_node.
v4: Switching the rotor and attempting to provide an interleave function.
Also splitting the patch into two sections.
v5: Corrected unsigned to long.

Nathan Zimmer (2):
  shmem: provide vm_ops when also providing a mem policy
  tmpfs: interleave the starting node of /dev/shmem

 include/linux/mm.h       |    7 +++++++
 include/linux/shmem_fs.h |    3 +++
 mm/mempolicy.c           |    4 ++++
 mm/shmem.c               |   35 ++++++++++++++++++++++++++++++++---
 4 files changed, 46 insertions(+), 3 deletions(-)

--
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>

[PATCH 1/2 v5][resend] shmem: provide vm_ops when also providing a mem policy

[Nathan Zimmer]

Updating shmem_get_policy to use the vma_policy if provided.
This is to allows us to safely provide shmem_vm_ops to the vma when the vm_file
has not been setup which is the case on the pseudo vmas.

Cc: Christoph Lameter <cl@linux.com>
Cc: Nick Piggin <npiggin@gmail.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Rik van Riel <riel@redhat.com>
Signed-off-by: Nathan Zimmer <nzimmer@sgi.com>
---
 mm/shmem.c |   18 +++++++++++++++---
 1 files changed, 15 insertions(+), 3 deletions(-)

diff --git a/mm/shmem.c b/mm/shmem.c
index a15a466..d073252 100644
--- a/mm/shmem.c
+++ b/mm/shmem.c
@@ -921,8 +921,11 @@ static struct page *shmem_swapin(swp_entry_t swap, gfp_t gfp,
 	/* Create a pseudo vma that just contains the policy */
 	pvma.vm_start = 0;
 	pvma.vm_pgoff = index;
-	pvma.vm_ops = NULL;
 	pvma.vm_policy = spol;
+	if (pvma.vm_policy)
+		pvma.vm_ops = &shmem_vm_ops;
+	else
+		pvma.vm_ops = NULL;
 	return swapin_readahead(swap, gfp, &pvma, 0);
 }
 
@@ -934,8 +937,11 @@ static struct page *shmem_alloc_page(gfp_t gfp,
 	/* Create a pseudo vma that just contains the policy */
 	pvma.vm_start = 0;
 	pvma.vm_pgoff = index;
-	pvma.vm_ops = NULL;
 	pvma.vm_policy = mpol_shared_policy_lookup(&info->policy, index);
+	if (pvma.vm_policy)
+		pvma.vm_ops = &shmem_vm_ops;
+	else
+		pvma.vm_ops = NULL;
 
 	/*
 	 * alloc_page_vma() will drop the shared policy reference
@@ -1296,8 +1302,14 @@ static int shmem_set_policy(struct vm_area_struct *vma, struct mempolicy *mpol)
 static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
 					  unsigned long addr)
 {
-	struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
 	pgoff_t index;
+	struct inode *inode;
+
+	/* If the vma knows what policy it wants use that one. */
+	if (vma->vm_policy)
+		return vma->vm_policy;
+
+	inode = vma->vm_file->f_path.dentry->d_inode;
 
 	index = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
 	return mpol_shared_policy_lookup(&SHMEM_I(inode)->policy, index);
-- 
1.6.0.2

--
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>

[PATCH 2/2 v5][resend] tmpfs: interleave the starting node of /dev/shmem

[Nathan Zimmer]

The tmpfs superblock grants an offset for each inode as they are created. Each
inode then uses that offset to provide a preferred first node for its interleave
in the newly provided shmem_interleave.

Cc: Christoph Lameter <cl@linux.com>
Cc: Nick Piggin <npiggin@gmail.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Rik van Riel <riel@redhat.com>
Signed-off-by: Nathan Zimmer <nzimmer@sgi.com>
---
 include/linux/mm.h       |    7 +++++++
 include/linux/shmem_fs.h |    3 +++
 mm/mempolicy.c           |    4 ++++
 mm/shmem.c               |   17 +++++++++++++++++
 4 files changed, 31 insertions(+), 0 deletions(-)

diff --git a/include/linux/mm.h b/include/linux/mm.h
index b36d08c..651109e 100644
--- a/include/linux/mm.h
+++ b/include/linux/mm.h
@@ -238,6 +238,13 @@ struct vm_operations_struct {
 	 */
 	struct mempolicy *(*get_policy)(struct vm_area_struct *vma,
 					unsigned long addr);
+
+	/*
+	 * If the policy is interleave allow the vma to suggest a node.
+	 */
+	unsigned long (*interleave)(struct vm_area_struct *vma,
+					unsigned long addr);
+
 	int (*migrate)(struct vm_area_struct *vma, const nodemask_t *from,
 		const nodemask_t *to, unsigned long flags);
 #endif
diff --git a/include/linux/shmem_fs.h b/include/linux/shmem_fs.h
index bef2cf0..6995556 100644
--- a/include/linux/shmem_fs.h
+++ b/include/linux/shmem_fs.h
@@ -17,6 +17,7 @@ struct shmem_inode_info {
 		char		*symlink;	/* unswappable short symlink */
 	};
 	struct shared_policy	policy;		/* NUMA memory alloc policy */
+	unsigned long           node_offset;	/* bias for interleaved nodes */
 	struct list_head	swaplist;	/* chain of maybes on swap */
 	struct list_head	xattr_list;	/* list of shmem_xattr */
 	struct inode		vfs_inode;
@@ -32,6 +33,8 @@ struct shmem_sb_info {
 	kgid_t gid;		    /* Mount gid for root directory */
 	umode_t mode;		    /* Mount mode for root directory */
 	struct mempolicy *mpol;     /* default memory policy for mappings */
+	unsigned long next_pref_node;
+			 /* next interleave bias to suggest for inodes */
 };
 
 static inline struct shmem_inode_info *SHMEM_I(struct inode *inode)
diff --git a/mm/mempolicy.c b/mm/mempolicy.c
index 1d771e4..e2cbe9e 100644
--- a/mm/mempolicy.c
+++ b/mm/mempolicy.c
@@ -1663,6 +1663,10 @@ static inline unsigned interleave_nid(struct mempolicy *pol,
 {
 	if (vma) {
 		unsigned long off;
+		if (vma->vm_ops && vma->vm_ops->interleave) {
+			off = vma->vm_ops->interleave(vma, addr);
+			return offset_il_node(pol, vma, off);
+		}
 
 		/*
 		 * for small pages, there is no difference between
diff --git a/mm/shmem.c b/mm/shmem.c
index d073252..e569338 100644
--- a/mm/shmem.c
+++ b/mm/shmem.c
@@ -922,6 +922,7 @@ static struct page *shmem_swapin(swp_entry_t swap, gfp_t gfp,
 	pvma.vm_start = 0;
 	pvma.vm_pgoff = index;
 	pvma.vm_policy = spol;
+	pvma.vm_private_data = (void *) info->node_offset;
 	if (pvma.vm_policy)
 		pvma.vm_ops = &shmem_vm_ops;
 	else
@@ -938,6 +939,7 @@ static struct page *shmem_alloc_page(gfp_t gfp,
 	pvma.vm_start = 0;
 	pvma.vm_pgoff = index;
 	pvma.vm_policy = mpol_shared_policy_lookup(&info->policy, index);
+	pvma.vm_private_data = (void *) info->node_offset;
 	if (pvma.vm_policy)
 		pvma.vm_ops = &shmem_vm_ops;
 	else
@@ -1314,6 +1316,19 @@ static struct mempolicy *shmem_get_policy(struct vm_area_struct *vma,
 	index = ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
 	return mpol_shared_policy_lookup(&SHMEM_I(inode)->policy, index);
 }
+
+static unsigned long shmem_interleave(struct vm_area_struct *vma,
+					unsigned long addr)
+{
+	unsigned long offset;
+
+	/* Use the vm_files prefered node as the initial offset. */
+	offset = (unsigned long *) vma->vm_private_data;
+
+	offset += ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
+
+	return offset;
+}
 #endif
 
 int shmem_lock(struct file *file, int lock, struct user_struct *user)
@@ -1386,6 +1401,7 @@ static struct inode *shmem_get_inode(struct super_block *sb, const struct inode
 			inode->i_fop = &shmem_file_operations;
 			mpol_shared_policy_init(&info->policy,
 						 shmem_get_sbmpol(sbinfo));
+			info->node_offset = ++(sbinfo->next_pref_node);
 			break;
 		case S_IFDIR:
 			inc_nlink(inode);
@@ -2871,6 +2887,7 @@ static const struct super_operations shmem_ops = {
 static const struct vm_operations_struct shmem_vm_ops = {
 	.fault		= shmem_fault,
 #ifdef CONFIG_NUMA
+	.interleave	= shmem_interleave,
 	.set_policy     = shmem_set_policy,
 	.get_policy     = shmem_get_policy,
 #endif
-- 
1.6.0.2

--
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>

Re: [PATCH 2/2 v5][resend] tmpfs: interleave the starting node of /dev/shmem

[Cong Wang]

On Mon, 09 Jul 2012 at 14:46 GMT, Nathan Zimmer <nzimmer@sgi.com> wrote:
> +static unsigned long shmem_interleave(struct vm_area_struct *vma,
> +					unsigned long addr)
> +{
> +	unsigned long offset;
> +
> +	/* Use the vm_files prefered node as the initial offset. */
> +	offset = (unsigned long *) vma->vm_private_data;
> +

offset is 'unsigned long', but here you cast ->private_data to 'unsigned
long *'?? Please test your patches before posting.

Thanks.


--
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>

Re: [PATCH 2/2 v5][resend] tmpfs: interleave the starting node of /dev/shmem

[Dan Carpenter]

On Mon, Jul 09, 2012 at 09:46:39AM -0500, Nathan Zimmer wrote:
> +static unsigned long shmem_interleave(struct vm_area_struct *vma,
> +					unsigned long addr)
> +{
> +	unsigned long offset;
> +
> +	/* Use the vm_files prefered node as the initial offset. */
> +	offset = (unsigned long *) vma->vm_private_data;

Should this be?:
	offset = (unsigned long)vma->vm_private_data;

offset is an unsigned long, not a pointer.  ->vm_private_data is a
void pointer.

It causes a GCC warning:
mm/shmem.c: In function a??shmem_interleavea??:
mm/shmem.c:1341:9: warning: assignment makes integer from pointer without a cast [enabled by default]

> +
> +	offset += ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
> +
> +	return offset;
> +}
>  #endif

regards,
dan carpenter

--
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>

Re: [PATCH 2/2 v5][resend] tmpfs: interleave the starting node of /dev/shmem

[Nathan Zimmer]

Yes I had failed to notice that.
I'll send a fix shortly.


On 07/23/2012 05:58 AM, Dan Carpenter wrote:
> On Mon, Jul 09, 2012 at 09:46:39AM -0500, Nathan Zimmer wrote:
>> +static unsigned long shmem_interleave(struct vm_area_struct *vma,
>> +					unsigned long addr)
>> +{
>> +	unsigned long offset;
>> +
>> +	/* Use the vm_files prefered node as the initial offset. */
>> +	offset = (unsigned long *) vma->vm_private_data;
> Should this be?:
> 	offset = (unsigned long)vma->vm_private_data;
>
> offset is an unsigned long, not a pointer.  ->vm_private_data is a
> void pointer.
>
> It causes a GCC warning:
> mm/shmem.c: In function a??shmem_interleavea??:
> mm/shmem.c:1341:9: warning: assignment makes integer from pointer without a cast [enabled by default]
>
>> +
>> +	offset += ((addr - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
>> +
>> +	return offset;
>> +}
>>   #endif
> regards,
> dan carpenter
>
> --
> 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>

--
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>

Re: [PATCH 2/2 v5][resend] tmpfs: interleave the starting node of /dev/shmem

[Hugh Dickins]

Nathan, Kosaki-san,

I have, at long last, reached the point of looking at this patchset.
And I'm puzzled as to why it has grown more complicated than what you
first sent out.

I've read through the various threads, and some of the changes I like.

I'm glad Andrew took out the stable Cc: obviously the interleave policy
was never intended for a filesystem of many small files, and it could
be that some usages with larger files have actually optimized to the
current node layout, and will regress with this change.  Let's keep it
simple and assume not; but if there are complaints, then we shall have
to make the new behaviour dependent on a mount option.

And I'm glad you switched from random number to rotor: I'm probably
missing the mark by orders of magnitude, but I always think of random
numbers as a precious resource, and was unsure if this deserved them.

But other changes just seem unnecessary to me.  And I don't see how
we can accuse you of being hackish, so long as we have that horrid
business of pseudo-vma on the shmem stack.  I believe the mempolicy
work was designed around vmas, then at the last moment had shmem
grafted on, and the quick way to shoehorn it in was the pseudo-vma.
It's just a way of massaging the info into a format that mempolicy.c
expects, and the arguments about addresses and offsets mystified me.

I did set out to replace the pseudo-vma by adding an alloc_page_mpol()
three years ago; but, no surprise, I got stuck when it came to
understanding the mpol reference counting, and had to move away.
Maybe we can revisit that once Kosaki-san has the refcounting fixed.

Please, what's wrong with the patch below, to replace the current
two or three?  I don't have real NUMA myself: does it work?
If it doesn't work, can you see why not?

Nathan, I've presumptuously put in your signoff, because
you generally seemed happy to incorporate suggestions made.
Kosaki-san, I'm sorry if this version annoys you, but I've not
seen an actual explanation as to why anything more is needed.

Hugh

From: Nathan Zimmer <nzimmer@sgi.com>
Subject: tmpfs: distribute interleave better across nodes

When tmpfs has the interleave memory policy, it always starts allocating
for each file from node 0 at offset 0.  When there are many small files,
the lower nodes fill up disproportionately.

This patch spreads out node usage by starting files at nodes other than
0, by using the inode number to bias the starting node for interleave.

Signed-off-by: Nathan Zimmer <nzimmer@sgi.com>
Signed-off-by: Hugh Dickins <hughd@google.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: Nick Piggin <npiggin@gmail.com>
Cc: Lee Schermerhorn <lee.schermerhorn@hp.com>
Cc: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
Cc: Rik van Riel <riel@redhat.com>
Cc: Andi Kleen <andi@firstfloor.org>
Cc: Andrew Morton <akpm@linux-foundation.org>
---

 mm/shmem.c |    6 ++++--
 1 file changed, 4 insertions(+), 2 deletions(-)

--- v3.5/mm/shmem.c	2012-07-21 13:58:29.000000000 -0700
+++ linux/mm/shmem.c	2012-07-24 20:13:58.468797969 -0700
@@ -929,7 +929,8 @@ static struct page *shmem_swapin(swp_ent
 
 	/* Create a pseudo vma that just contains the policy */
 	pvma.vm_start = 0;
-	pvma.vm_pgoff = index;
+	/* Bias interleave by inode number to distribute better across nodes */
+	pvma.vm_pgoff = index + info->vfs_inode.i_ino;
 	pvma.vm_ops = NULL;
 	pvma.vm_policy = spol;
 	return swapin_readahead(swap, gfp, &pvma, 0);
@@ -942,7 +943,8 @@ static struct page *shmem_alloc_page(gfp
 
 	/* Create a pseudo vma that just contains the policy */
 	pvma.vm_start = 0;
-	pvma.vm_pgoff = index;
+	/* Bias interleave by inode number to distribute better across nodes */
+	pvma.vm_pgoff = index + info->vfs_inode.i_ino;
 	pvma.vm_ops = NULL;
 	pvma.vm_policy = mpol_shared_policy_lookup(&info->policy, index);
 

--
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>

Re: [PATCH 2/2 v5][resend] tmpfs: interleave the starting node of /dev/shmem

[KOSAKI Motohiro]

> Please, what's wrong with the patch below, to replace the current
> two or three?  I don't have real NUMA myself: does it work?
> If it doesn't work, can you see why not?

It works. It doesn't match my preference. but I don't want block your way.
this area is maintained you. please go ahead.

at least, inode bias is better than random.


--
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>

Re: [PATCH 2/2 v5][resend] tmpfs: interleave the starting node of /dev/shmem

[Nathan Zimmer]

On Tue, Jul 24, 2012 at 09:38:21PM -0700, Hugh Dickins wrote:
> 
> I'm glad Andrew took out the stable Cc: 
Actually I did that.  I have a habit of thinking about performance issues as
bugs and that is not always the case.

> Please, what's wrong with the patch below, to replace the current
> two or three?  I don't have real NUMA myself: does it work?
Yes it works and spreads quite nicely. 

> Nathan, I've presumptuously put in your signoff, because
> you generally seemed happy to incorporate suggestions made.
I am always grateful for suggestions, advise, and help.

Nate

--
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>