Re: [PATCH] MM: Support more pagesizes for MAP_HUGETLB/SHM_HUGETLBv3

From: Andrew Morton
Date: Tue Oct 09 2012 - 18:18:59 EST


On Wed, 3 Oct 2012 15:24:23 -0700
Andi Kleen <andi@xxxxxxxxxxxxxx> wrote:

> From: Andi Kleen <ak@xxxxxxxxxxxxxxx>
>
> There was some desire in large applications using MAP_HUGETLB/SHM_HUGETLB
> to use 1GB huge pages on some mappings, and stay with 2MB on others. This
> is useful together with NUMA policy: use 2MB interleaving on some mappings,
> but 1GB on local mappings.
>
> This patch extends the IPC/SHM syscall interfaces slightly to allow specifying
> the page size.
>
> It borrows some upper bits in the existing flag arguments and allows encoding
> the log of the desired page size in addition to the *_HUGETLB flag.
> When 0 is specified the default size is used, this makes the change fully
> compatible.
>
> Extending the internal hugetlb code to handle this is straight forward. Instead
> of a single mount it just keeps an array of them and selects the right
> mount based on the specified page size.
>
> I also exported the new flags to the user headers
> (they were previously under __KERNEL__). Right now only symbols
> for x86 and some other architecture for 1GB and 2MB are defined.
> The interface should already work for all other architectures
> though.

So some manpages need updating. I'm not sure which - mmap(2) surely,
but which for the IPC change?

> v2: Port to new tree. Fix unmount.
> v3: Ported to latest tree.
> Acked-by: Rik van Riel <riel@xxxxxxxxxx>
> Acked-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@xxxxxxxxxxxxxx>
> Signed-off-by: Andi Kleen <ak@xxxxxxxxxxxxxxx>
> ---
> arch/x86/include/asm/mman.h | 3 ++
> fs/hugetlbfs/inode.c | 63 ++++++++++++++++++++++++++++++++++---------
> include/asm-generic/mman.h | 13 +++++++++
> include/linux/hugetlb.h | 12 +++++++-
> include/linux/shm.h | 19 +++++++++++++
> ipc/shm.c | 3 +-
> mm/mmap.c | 5 ++-

Alas, include/asm-generic/mman.h doesn't exist now.

Does this change touch all the hugetlb-capable architectures?

z:/usr/src/linux-3.6> grep -rl MAP_HUGETLB arch
arch/alpha/include/asm/mman.h
arch/xtensa/include/asm/mman.h
arch/parisc/include/asm/mman.h
arch/tile/include/asm/mman.h
arch/sparc/include/asm/mman.h
arch/powerpc/include/asm/mman.h
arch/mips/include/asm/mman.h

>
> ...
>
> @@ -933,9 +933,22 @@ static int can_do_hugetlb_shm(void)
> return capable(CAP_IPC_LOCK) || in_group_p(shm_group);
> }
>
> +static int get_hstate_idx(int page_size_log)

nitlet: "page_size_order" would be more kernely. Or just "page_order".

> +{
> + struct hstate *h;
> +
> + if (!page_size_log)
> + return default_hstate_idx;
> + h = size_to_hstate(1 << page_size_log);
> + if (!h)
> + return -1;
> + return h - hstates;
> +}
>
> ...
>
> static int __init init_hugetlbfs_fs(void)
> {
> + struct hstate *h;
> int error;
> - struct vfsmount *vfsmount;
> + int i;
>
> error = bdi_init(&hugetlbfs_backing_dev_info);
> if (error)
> @@ -1030,14 +1049,26 @@ static int __init init_hugetlbfs_fs(void)
> if (error)
> goto out;
>
> - vfsmount = kern_mount(&hugetlbfs_fs_type);
> + i = 0;
> + for_each_hstate (h) {
> + char buf[50];
> + unsigned ps_kb = 1U << (h->order + PAGE_SHIFT - 10);
>
> - if (!IS_ERR(vfsmount)) {
> - hugetlbfs_vfsmount = vfsmount;
> - return 0;
> - }
> + snprintf(buf, sizeof buf, "pagesize=%uK", ps_kb);
> + hugetlbfs_vfsmount[i] = kern_mount_data(&hugetlbfs_fs_type,
> + buf);
>
> - error = PTR_ERR(vfsmount);
> + if (IS_ERR(hugetlbfs_vfsmount[i])) {
> + pr_err(
> + "hugetlb: Cannot mount internal hugetlbfs for page size %uK",
> + ps_kb);
> + error = PTR_ERR(hugetlbfs_vfsmount[i]);
> + }
> + i++;
> + }
> + /* Non default hstates are optional */
> + if (hugetlbfs_vfsmount[default_hstate_idx])
> + return 0;

hm, so if I'm understanding this, the patch mounts hugetlbfs N times,
once for each page size. And presumably the shm code somehow selects
one of these mounts, based on incoming flags. And presumably if those
flags are all-zero, the behaviour is unaltered.

Please update the changelog to describe all this - the overview of how
the patch actually operates.

Also, all this affects the /proc/mounts contents, yes? Let's changelog
that very-slightly-non-back-compatible user-visible change as well.

There's some overhead to doing all those additional mounts. Can we
quantify it?

>
> ...
>

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/