On Wed, 23 May 2012 13:28:21 +0000
Nathan Zimmer <[email protected]> wrote:
>
> 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.
> My proposed solution is to start a file at a randomly chosen node.
>
> ...
>
> --- 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 */
> + int 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;
> diff --git a/mm/shmem.c b/mm/shmem.c
> index f99ff3e..58ef512 100644
> --- a/mm/shmem.c
> +++ b/mm/shmem.c
> @@ -819,7 +819,7 @@ 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_pgoff = index + info->node_offset;
> pvma.vm_ops = NULL;
> pvma.vm_policy = mpol_shared_policy_lookup(&info->policy, index);
>
> @@ -1153,6 +1153,7 @@ 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 = node_random(&node_online_map);
> break;
> case S_IFDIR:
> inc_nlink(inode);
The patch seems a bit arbitrary and hacky. It would have helped if you
had fully described how it works, and why this implementation was
chosen.
- Why alter (actually, lie about!) the offset-into-file? Could we
have similarly perturbed the address arg to alloc_page_vma() to do
the spreading?
- The patch is dependent upon MPOL_INTERLEAVE being in effect, isn't
it? How do we guarantee that it is in force here?
- We look up the policy via mpol_shared_policy_lookup() using the
unperturbed index. Why? Should we be using index+info->node_offset
there?
--
To unsubscribe from this list: send the line "unsubscribe stable" in
the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
