2024-03-27 15:09:52

by David Hildenbrand

Subject: [PATCH v1] mm: optimize CONFIG_PER_VMA_LOCK member placement in vm_area_struct

Currently, we end up wasting some memory in each vm_area_struct. Pahole
states that:
[...]
int vm_lock_seq; /* 40 4 */

/* XXX 4 bytes hole, try to pack */

struct vma_lock * vm_lock; /* 48 8 */
bool detached; /* 56 1 */

/* XXX 7 bytes hole, try to pack */
[...]

Let's reduce the holes and memory wastage by moving the bool:
[...]
bool detached; /* 40 1 */

/* XXX 3 bytes hole, try to pack */

int vm_lock_seq; /* 44 4 */
struct vma_lock * vm_lock; /* 48 8 */
[...]

Effectively shrinking the vm_area_struct with CONFIG_PER_VMA_LOCK by
8 byte.

Likely, we could place "detached" in the lowest bit of vm_lock, but at
least on 64bit that won't really make a difference, so keep it simple.

Cc: Andrew Morton <[email protected]>
Cc: Suren Baghdasaryan <[email protected]>
Signed-off-by: David Hildenbrand <[email protected]>
---
include/linux/mm_types.h | 6 +++---
1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/include/linux/mm_types.h b/include/linux/mm_types.h
index 4ae4684d1add..f56739dece7a 100644
--- a/include/linux/mm_types.h
+++ b/include/linux/mm_types.h
@@ -671,6 +671,9 @@ struct vm_area_struct {
};

#ifdef CONFIG_PER_VMA_LOCK
+ /* Flag to indicate areas detached from the mm->mm_mt tree */
+ bool detached;
+
/*
* Can only be written (using WRITE_ONCE()) while holding both:
* - mmap_lock (in write mode)
@@ -687,9 +690,6 @@ struct vm_area_struct {
*/
int vm_lock_seq;
struct vma_lock *vm_lock;
-
- /* Flag to indicate areas detached from the mm->mm_mt tree */
- bool detached;
#endif

/*
--
2.43.2

2024-03-27 20:13:25

by Suren Baghdasaryan

Subject: Re: [PATCH v1] mm: optimize CONFIG_PER_VMA_LOCK member placement in vm_area_struct

On Wed, Mar 27, 2024 at 7:36 AM David Hildenbrand <[email protected]> wrote:
>
> Currently, we end up wasting some memory in each vm_area_struct. Pahole
> states that:
> [...]
> int vm_lock_seq; /* 40 4 */
>
> /* XXX 4 bytes hole, try to pack */
>
> struct vma_lock * vm_lock; /* 48 8 */
> bool detached; /* 56 1 */
>
> /* XXX 7 bytes hole, try to pack */
> [...]
>
> Let's reduce the holes and memory wastage by moving the bool:
> [...]
> bool detached; /* 40 1 */
>
> /* XXX 3 bytes hole, try to pack */
>
> int vm_lock_seq; /* 44 4 */
> struct vma_lock * vm_lock; /* 48 8 */
> [...]
>
> Effectively shrinking the vm_area_struct with CONFIG_PER_VMA_LOCK by
> 8 byte.
>
> Likely, we could place "detached" in the lowest bit of vm_lock, but at
> least on 64bit that won't really make a difference, so keep it simple.
>
> Cc: Andrew Morton <[email protected]>
> Cc: Suren Baghdasaryan <[email protected]>
> Signed-off-by: David Hildenbrand <[email protected]>

Looks reasonable to me. Thanks!

> ---
> include/linux/mm_types.h | 6 +++---
> 1 file changed, 3 insertions(+), 3 deletions(-)
>
> diff --git a/include/linux/mm_types.h b/include/linux/mm_types.h
> index 4ae4684d1add..f56739dece7a 100644
> --- a/include/linux/mm_types.h
> +++ b/include/linux/mm_types.h
> @@ -671,6 +671,9 @@ struct vm_area_struct {
> };
>
> #ifdef CONFIG_PER_VMA_LOCK
> + /* Flag to indicate areas detached from the mm->mm_mt tree */
> + bool detached;
> +
> /*
> * Can only be written (using WRITE_ONCE()) while holding both:
> * - mmap_lock (in write mode)
> @@ -687,9 +690,6 @@ struct vm_area_struct {
> */
> int vm_lock_seq;
> struct vma_lock *vm_lock;
> -
> - /* Flag to indicate areas detached from the mm->mm_mt tree */
> - bool detached;
> #endif
>
> /*
> --
> 2.43.2
>