KSM kernel interface

KSM kernel interface

[supercilious.dude]

Hi,

Is there a way to enable KSM globally for all eligible pages in the 
system such that applications need not call madvise() themselves? If 
not, is there a way to do so on behalf of a particular application 
without an ugly LD_PRELOAD hack?

Perhaps there is a small kernel modification that might make this 
possible or provide a sysfs flag that enables it (off by default of course)?

I read through the code very briefly, and as far as I can tell, I need 
to add the VM_MERGEABLE flag to all eligible VMAs at the time they are 
created and also __ksm_enter() the associated mm_struct at creation 
also. Would that work?

Thanks

Re: KSM kernel interface

[Hugh Dickins]

On Sat, 19 Feb 2011, supercilious.dude@gmail.com wrote:
> 
> Is there a way to enable KSM globally for all eligible pages in the system
> such that applications need not call madvise() themselves? If not, is there
> a way to do so on behalf of a particular application without an ugly
> LD_PRELOAD hack?

Sorry, there is not.

> 
> Perhaps there is a small kernel modification that might make this possible
> or provide a sysfs flag that enables it (off by default of course)?

Indeed it could be added if there were a general call for it; but ksmd
would tend to get wasteful, and the only call for it that I remember is
for testing.

I expect you've noticed the transparent_hugepage/enabled "always" option,
and you are thinking something like that could be done for KSM: yes.

> 
> I read through the code very briefly, and as far as I can tell, I need to
> add the VM_MERGEABLE flag to all eligible VMAs at the time they are created
> and also __ksm_enter() the associated mm_struct at creation also. Would that
> work?

I think that's right.

Here is the hack I use myself for testing KSM: just boot with option
"allksm" (which here disables randomize_va_space as a side-effect,
to widen the scope of sharing: cut that line out if you prefer).

You may be puzzled by the squashed comments etc: just trying not to
interfere with the line numbering in mm/mmap.c, in case I have other
patches to apply there, or hit a BUG to report.

Hugh
---

 mm/mmap.c |   27 +++++++++++++++++----------
 1 file changed, 17 insertions(+), 10 deletions(-)

--- 2.6.37/mm/mmap.c	2011-01-04 16:50:19.000000000 -0800
+++ allksm/mm/mmap.c	2011-01-04 19:47:16.000000000 -0800
@@ -960,9 +960,9 @@ void vm_stat_account(struct mm_struct *m
 #endif /* CONFIG_PROC_FS */
 
 /*
- * The caller must hold down_write(&current->mm->mmap_sem).
- */
-
+ * The caller must hold down_write(&current->mm->mmap_sem). */
+#include <linux/ksm.h>
+unsigned long vm_mergeable;
 unsigned long do_mmap_pgoff(struct file *file, unsigned long addr,
 			unsigned long len, unsigned long prot,
 			unsigned long flags, unsigned long pgoff)
@@ -1086,7 +1086,7 @@ unsigned long do_mmap_pgoff(struct file
 			/*
 			 * Set pgoff according to addr for anon_vma.
 			 */
-			pgoff = addr >> PAGE_SHIFT;
+			vm_flags |= vm_mergeable; pgoff = addr >> PAGE_SHIFT;
 			break;
 		default:
 			return -EINVAL;
@@ -1303,10 +1303,10 @@ munmap_back:
 		vma->vm_file = file;
 		get_file(file);
 		error = file->f_op->mmap(file, vma);
-		if (error)
-			goto unmap_and_free_vma;
-		if (vm_flags & VM_EXECUTABLE)
-			added_exe_file_vma(mm);
+		if (error) goto unmap_and_free_vma;
+		if (vm_flags & VM_EXECUTABLE) added_exe_file_vma(mm);
+		if (vm_mergeable)
+			ksm_madvise(vma, 0, 0, MADV_MERGEABLE,&vma->vm_flags);
 
 		/* Can addr have changed??
 		 *
@@ -2167,7 +2167,7 @@ unsigned long do_brk(unsigned long addr,
 		return error;
 
 	flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
-
+	flags |= vm_mergeable;
 	error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
 	if (error & ~PAGE_MASK)
 		return error;
@@ -2318,7 +2318,7 @@ int insert_vm_struct(struct mm_struct *
 	if (!vma->vm_file) {
 		BUG_ON(vma->anon_vma);
 		vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
-	}
+		vma->vm_flags |= vm_mergeable;	}
 	__vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent);
 	if (__vma && __vma->vm_start < vma->vm_end)
 		return -ENOMEM;
@@ -2677,3 +2677,10 @@ void __init mmap_init(void)
 	ret = percpu_counter_init(&vm_committed_as, 0);
 	VM_BUG_ON(ret);
 }
+static int __init allksm(char *s)
+{
+	randomize_va_space = 0;
+	vm_mergeable = VM_MERGEABLE;
+	return 1;
+}
+__setup("allksm", allksm);

Re: KSM kernel interface

[richard -rw- weinberger]

On Sat, Feb 19, 2011 at 7:54 PM, Hugh Dickins <hughd@google.com> wrote:
> On Sat, 19 Feb 2011, supercilious.dude@gmail.com wrote:
>>
>> Is there a way to enable KSM globally for all eligible pages in the system
>> such that applications need not call madvise() themselves? If not, is there
>> a way to do so on behalf of a particular application without an ugly
>> LD_PRELOAD hack?
>
> Sorry, there is not.
>
>>
>> Perhaps there is a small kernel modification that might make this possible
>> or provide a sysfs flag that enables it (off by default of course)?
>
> Indeed it could be added if there were a general call for it; but ksmd
> would tend to get wasteful, and the only call for it that I remember is
> for testing.
>
> I expect you've noticed the transparent_hugepage/enabled "always" option,
> and you are thinking something like that could be done for KSM: yes.
>
>>
>> I read through the code very briefly, and as far as I can tell, I need to
>> add the VM_MERGEABLE flag to all eligible VMAs at the time they are created
>> and also __ksm_enter() the associated mm_struct at creation also. Would that
>> work?
>
> I think that's right.
>
> Here is the hack I use myself for testing KSM: just boot with option
> "allksm" (which here disables randomize_va_space as a side-effect,
> to widen the scope of sharing: cut that line out if you prefer).

It would be nice to have such a "allksm" in cgroups' memory resource controller.
Especially LXC could benefit from it.

What do you think?

> You may be puzzled by the squashed comments etc: just trying not to
> interfere with the line numbering in mm/mmap.c, in case I have other
> patches to apply there, or hit a BUG to report.
>
> Hugh
> ---
>
>  mm/mmap.c |   27 +++++++++++++++++----------
>  1 file changed, 17 insertions(+), 10 deletions(-)
>
> --- 2.6.37/mm/mmap.c    2011-01-04 16:50:19.000000000 -0800
> +++ allksm/mm/mmap.c    2011-01-04 19:47:16.000000000 -0800
> @@ -960,9 +960,9 @@ void vm_stat_account(struct mm_struct *m
>  #endif /* CONFIG_PROC_FS */
>
>  /*
> - * The caller must hold down_write(&current->mm->mmap_sem).
> - */
> -
> + * The caller must hold down_write(&current->mm->mmap_sem). */
> +#include <linux/ksm.h>
> +unsigned long vm_mergeable;
>  unsigned long do_mmap_pgoff(struct file *file, unsigned long addr,
>                        unsigned long len, unsigned long prot,
>                        unsigned long flags, unsigned long pgoff)
> @@ -1086,7 +1086,7 @@ unsigned long do_mmap_pgoff(struct file
>                        /*
>                         * Set pgoff according to addr for anon_vma.
>                         */
> -                       pgoff = addr >> PAGE_SHIFT;
> +                       vm_flags |= vm_mergeable; pgoff = addr >> PAGE_SHIFT;
>                        break;
>                default:
>                        return -EINVAL;
> @@ -1303,10 +1303,10 @@ munmap_back:
>                vma->vm_file = file;
>                get_file(file);
>                error = file->f_op->mmap(file, vma);
> -               if (error)
> -                       goto unmap_and_free_vma;
> -               if (vm_flags & VM_EXECUTABLE)
> -                       added_exe_file_vma(mm);
> +               if (error) goto unmap_and_free_vma;
> +               if (vm_flags & VM_EXECUTABLE) added_exe_file_vma(mm);
> +               if (vm_mergeable)
> +                       ksm_madvise(vma, 0, 0, MADV_MERGEABLE,&vma->vm_flags);
>
>                /* Can addr have changed??
>                 *
> @@ -2167,7 +2167,7 @@ unsigned long do_brk(unsigned long addr,
>                return error;
>
>        flags = VM_DATA_DEFAULT_FLAGS | VM_ACCOUNT | mm->def_flags;
> -
> +       flags |= vm_mergeable;
>        error = get_unmapped_area(NULL, addr, len, 0, MAP_FIXED);
>        if (error & ~PAGE_MASK)
>                return error;
> @@ -2318,7 +2318,7 @@ int insert_vm_struct(struct mm_struct *
>        if (!vma->vm_file) {
>                BUG_ON(vma->anon_vma);
>                vma->vm_pgoff = vma->vm_start >> PAGE_SHIFT;
> -       }
> +               vma->vm_flags |= vm_mergeable;  }
>        __vma = find_vma_prepare(mm,vma->vm_start,&prev,&rb_link,&rb_parent);
>        if (__vma && __vma->vm_start < vma->vm_end)
>                return -ENOMEM;
> @@ -2677,3 +2677,10 @@ void __init mmap_init(void)
>        ret = percpu_counter_init(&vm_committed_as, 0);
>        VM_BUG_ON(ret);
>  }
> +static int __init allksm(char *s)
> +{
> +       randomize_va_space = 0;
> +       vm_mergeable = VM_MERGEABLE;
> +       return 1;
> +}
> +__setup("allksm", allksm);
> --
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at  http://www.tux.org/lkml/
>



-- 
Thanks,
//richard

Re: KSM kernel interface

[supercilious.dude]

Thanks for the patch. I will try it out later today.

 >
 > > Perhaps there is a small kernel modification that might make this
 > possible
 > > or provide a sysfs flag that enables it (off by default of course)?
 >
 > Indeed it could be added if there were a general call for it; but ksmd
 > would tend to get wasteful, and the only call for it that I remember is
 > for testing.
 >
 > I expect you've noticed the transparent_hugepage/enabled "always" option,
 > and you are thinking something like that could be done for KSM: yes.
 >
 >

For what its worth, I believe that the option in sysfs to enable KSM
globally would be useful to have. I'm sure i'm not the only one. Its not
like it would be a great deal of overhead, since the KSM machinery is
already there anyway, just needs a flag to enable it at runtime instead 
of at boot.

I have at least one system running mostly legacy code that is according 
to ESXI's memory report, using only 400mb per gigabyte thanks to 
transparent page sharing. I would very much like to be able to have that 
level of memory saving on arbitrary code running on bare metal. 
Administrators can make their own judgement about ksmd's CPU usage, and 
the defaults already suit most people. I don't really see how having 
that extra flag (defaulting to off of course) would hurt.

Thanks