On Mon, 2009-09-14 at 14:33 +0100, Mel Gorman wrote:
> On Fri, Sep 11, 2009 at 03:27:30PM -0700, David Rientjes wrote:
> > On Thu, 10 Sep 2009, Mel Gorman wrote:
> >
> > > > Would you explain why introducing a new mempolicy flag, MPOL_F_HUGEPAGES,
> > > > and only using the new behavior when this is set would be inconsistent or
> > > > inadvisible?
> > >
> > > I already explained this. The interface in numactl would look weird. There
> > > would be an --interleave switch and a --hugepages-interleave that only
> > > applies to nr_hugepages. The smarts could be in hugeadm to apply the mask
> > > when --pool-pages-min is specified but that wouldn't help scripts that are
> > > still using echo.
> > >
> >
> > I don't think we need to address the scripts that are currently using echo
> > since they're (hopefully) written to the kernel implementation, i.e. no
> > mempolicy restriction on writing to nr_hugepages.
> >
>
> Ok.
>
> > > I hate to have to do this, but how about nr_hugepages which acts
> > > system-wide as it did traditionally and nr_hugepages_mempolicy that obeys
> > > policies? Something like the following untested patch. It would be fairly
> > > trivial for me to implement a --obey-mempolicies switch for hugeadm which
> > > works in conjunction with --pool--pages-min and less likely to cause confusion
> > > than --hugepages-interleave in numactl.
> > >
> >
> > I like it.
> >
>
> Ok, when I get this tested, I'll sent it as a follow-on patch to Lee's
> for proper incorporation.
>
> > > Sorry the patch is untested. I can't hold of a NUMA machine at the moment
> > > and fake NUMA support sucks far worse than I expected it to.
> > >
> >
> > Hmm, I rewrote most of fake NUMA a couple years ago. What problems are
> > you having with it?
> >
>
> On PPC64, the parameters behave differently. I couldn't convince it to
> create more than one NUMA node. On x86-64, the NUMA nodes appeared to
> exist and would be visible on /proc/buddyinfo for example but the sysfs
> directories for the fake nodes were not created so nr_hugepages couldn't
> be examined on a per-node basis for example.
>
> > > ==== BEGIN PATCH ====
> > >
> > > [PATCH] Optionally use a memory policy when tuning the size of the static hugepage pool
> > >
> > > Patch "derive huge pages nodes allowed from task mempolicy" brought
> > > huge page support more in line with the core VM in that tuning the size
> > > of the static huge page pool would obey memory policies. Using this,
> > > administrators could interleave allocation of huge pages from a subset
> > > of nodes. This is consistent with how dynamic hugepage pool resizing
> > > works and how hugepages get allocated to applications at run-time.
> > >
> > > However, it was pointed out that scripts may exist that depend on being
> > > able to drain all hugepages via /proc/sys/vm/nr_hugepages from processes
> > > that are running within a memory policy. This patch adds
> > > /proc/sys/vm/nr_hugepages_mempolicy which when written to will obey
> > > memory policies. /proc/sys/vm/nr_hugepages continues then to be a
> > > system-wide tunable regardless of memory policy.
> > >
> > > Signed-off-by: Mel Gorman <mel@xxxxxxxxx>
> > > ---
> > > include/linux/hugetlb.h | 1 +
> > > kernel/sysctl.c | 11 +++++++++++
> > > mm/hugetlb.c | 35 ++++++++++++++++++++++++++++++++---
> > > 3 files changed, 44 insertions(+), 3 deletions(-)
> > >
> >
> > It'll need an update to Documentation/vm/hugetlb.txt, but this can
> > probably be done in one of Lee's patches that edits the same file when he
> > reposts.
> >
>
> Agreed.
So, I'm respinning V7 today. Shall I add this in as a separate patch?
Also, see below:
>
> > > diff --git a/include/linux/hugetlb.h b/include/linux/hugetlb.h
> > > index fcb1677..fc3a659 100644
> > > --- a/include/linux/hugetlb.h
> > > +++ b/include/linux/hugetlb.h
> > > @@ -21,6 +21,7 @@ static inline int is_vm_hugetlb_page(struct vm_area_struct *vma)
> > >
> > > void reset_vma_resv_huge_pages(struct vm_area_struct *vma);
> > > int hugetlb_sysctl_handler(struct ctl_table *, int, void __user *, size_t *, loff_t *);
> > > +int hugetlb_mempolicy_sysctl_handler(struct ctl_table *, int, void __user *, size_t *, loff_t *);
> > > int hugetlb_overcommit_handler(struct ctl_table *, int, void __user *, size_t *, loff_t *);
> > > int hugetlb_treat_movable_handler(struct ctl_table *, int, void __user *, size_t *, loff_t *);
> > > int copy_hugetlb_page_range(struct mm_struct *, struct mm_struct *, struct vm_area_struct *);
> > > diff --git a/kernel/sysctl.c b/kernel/sysctl.c
> > > index 8bac3f5..0637655 100644
> > > --- a/kernel/sysctl.c
> > > +++ b/kernel/sysctl.c
> > > @@ -1171,6 +1171,17 @@ static struct ctl_table vm_table[] = {
> > > .extra1 = (void *)&hugetlb_zero,
> > > .extra2 = (void *)&hugetlb_infinity,
> > > },
> > > +#ifdef CONFIG_NUMA
> > > + {
> > > + .procname = "nr_hugepages_mempolicy",
> > > + .data = NULL,
> > > + .maxlen = sizeof(unsigned long),
> > > + .mode = 0644,
> > > + .proc_handler = &hugetlb_mempolicy_sysctl_handler,
> > > + .extra1 = (void *)&hugetlb_zero,
> > > + .extra2 = (void *)&hugetlb_infinity,
> > > + },
> > > +#endif
> > > {
> > > .ctl_name = VM_HUGETLB_GROUP,
> > > .procname = "hugetlb_shm_group",
> > > diff --git a/mm/hugetlb.c b/mm/hugetlb.c
> > > index 83decd6..68abef0 100644
> > > --- a/mm/hugetlb.c
> > > +++ b/mm/hugetlb.c
> > > @@ -1244,6 +1244,7 @@ static int adjust_pool_surplus(struct hstate *h, nodemask_t *nodes_allowed,
> > > return ret;
> > > }
> > >
> > > +#define NUMA_NO_NODE_OBEY_MEMPOLICY (-2)
How about defining NUMA_NO_NODE_OBEY_MEMPOLICY as (NUMA_NO_NODE - 1)
just to ensure that it's different. Not sure it's worth an enum at this
point. NUMA_NO_NODE_OBEY_MEMPOLICY is private to hugetlb at this time.
> > > #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
> > > static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
> > > int nid)
> >
> > I think it would be possible to avoid adding NUMA_NO_NODE_OBEY_MEMPOLICY
> > if the nodemask was allocated in the sysctl handler instead and passing it
> > into set_max_huge_pages() instead of a nid. Lee, what do you think?
> >
> > Other than that, I like this approach because it avoids the potential for
> > userspace breakage while adding the new feature in way that avoids
> > confusion.
> >
>
> Indeed. While the addition of another proc tunable sucks, it seems like
> the only available compromise.
And, I supposed we need to replicate this under the global sysfs
hstates?
Lee
>
--
To unsubscribe from this list: send the line "unsubscribe linux-numa" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
