Maybe move ->anon_vma_chain ?

Maybe move ->anon_vma_chain ?

[Alexey Dobriyan]

FYI,

on x86_64, ->anon_vma_chain has offset 120 inside the structure
which means that:
* on CONFIG_NUMA=n CONFIG_USERFAULTFD=n kernels
every 4-th VMA has this list head spanning 2 cachelines and,

* on CONFIG_NUMA=y CONFIG_USERFAULTFD=y kernels
_every_ VMA has this peculiar property.

Now I don't know good benchmark for anon vmas,
but maybe you do.


struct vm_area_struct {
vm_start;             /*     0     8 */
vm_end;               /*     8     8 */
vm_next;              /*    16     8 */
vm_prev;              /*    24     8 */
vm_rb;                /*    32    24 */
rb_subtree_gap;       /*    56     8 */
        /* --- cacheline 1 boundary (64 bytes) --- */
vm_mm;                /*    64     8 */
vm_page_prot;         /*    72     8 */
vm_flags;             /*    80     8 */

rb;                   /*    88    24 */
rb_subtree_last;      /*   112     8 */
                      /*    88    32 */


===>   struct list_head anon_vma_chain; /*   120    16 */
        /* --- cacheline 2 boundary (128 bytes) was 8 bytes ago --- */


anon_vma;             /*   136     8 */
struct  * vm_ops;     /*   144     8 */
vm_pgoff;             /*   152     8 */
vm_file;              /*   160     8 */
vm_private_data;      /*   168     8 */
vm_userfaultfd_ctx;   /*   176     0 */

        /* size: 176, cachelines: 3, members: 17 */
        /* last cacheline: 48 bytes */
};

--
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: Maybe move ->anon_vma_chain ?

[Rik van Riel]

On Thu, 2016-08-18 at 17:39 +0300, Alexey Dobriyan wrote:
> FYI,
> 
> on x86_64, ->anon_vma_chain has offset 120 inside the structure
> which means that:
> * on CONFIG_NUMA=n CONFIG_USERFAULTFD=n kernels
> every 4-th VMA has this list head spanning 2 cachelines and,
> 
> * on CONFIG_NUMA=y CONFIG_USERFAULTFD=y kernels
> _every_ VMA has this peculiar property.

It may make sense to move vm_pgoff into the first
cache line, since that is likely to be used in
conjunction with the rb tree, in rb tree walks.

> Now I don't know good benchmark for anon vmas,
> but maybe you do.

I am not sure what we would use to benchmark this.

> struct vm_area_struct {
> vm_start;A A A A A A A A A A A A A /*A A A A A 0A A A A A 8 */
> vm_end;A A A A A A A A A A A A A A A /*A A A A A 8A A A A A 8 */
> vm_next;A A A A A A A A A A A A A A /*A A A A 16A A A A A 8 */
> vm_prev;A A A A A A A A A A A A A A /*A A A A 24A A A A A 8 */
> vm_rb;A A A A A A A A A A A A A A A A /*A A A A 32A A A A 24 */
> rb_subtree_gap;A A A A A A A /*A A A A 56A A A A A 8 */
> A A A A A A A A /* --- cacheline 1 boundary (64 bytes) --- */
> vm_mm;A A A A A A A A A A A A A A A A /*A A A A 64A A A A A 8 */
> vm_page_prot;A A A A A A A A A /*A A A A 72A A A A A 8 */
> vm_flags;A A A A A A A A A A A A A /*A A A A 80A A A A A 8 */
> 
> rb;A A A A A A A A A A A A A A A A A A A /*A A A A 88A A A A 24 */
> rb_subtree_last;A A A A A A /*A A A 112A A A A A 8 */
> A A A A A A A A A A A A A A A A A A A A A A /*A A A A 88A A A A 32 */
> 
> 
> ===>A A A struct list_head anon_vma_chain; /*A A A 120A A A A 16 */
> A A A A A A A A /* --- cacheline 2 boundary (128 bytes) was 8 bytes ago ---
> */
> 
> 
> anon_vma;A A A A A A A A A A A A A /*A A A 136A A A A A 8 */
> structA A * vm_ops;A A A A A /*A A A 144A A A A A 8 */
> vm_pgoff;A A A A A A A A A A A A A /*A A A 152A A A A A 8 */
> vm_file;A A A A A A A A A A A A A A /*A A A 160A A A A A 8 */
> vm_private_data;A A A A A A /*A A A 168A A A A A 8 */
> vm_userfaultfd_ctx;A A A /*A A A 176A A A A A 0 */
> 
> A A A A A A A A /* size: 176, cachelines: 3, members: 17 */
> A A A A A A A A /* last cacheline: 48 bytes */
> };

--
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>