Re: [PATCH 3/4] hugetlb: add hugepagesnid= command-line option

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Mon, 10 Feb 2014 12:27:47 -0500 Luiz Capitulino <lcapitulino@xxxxxxxxxx> wrote:

> From: Luiz capitulino <lcapitulino@xxxxxxxxxx>
> 
> The HugeTLB command-line option hugepages= allow the user to specify
> how many huge pages should be allocated at boot-time. On NUMA systems,
> this option will try to automatically distribute the allocation equally
> among nodes. For example, if you have a 2-node NUMA system and allocates
> 200 huge pages, than hugepages= will try to allocate 100 huge pages from
> node0 and 100 from node1.
> 
> The hugepagesnid= option introduced by this commit allows the user
> to specify the nodes from which huge pages are allocated. For example,
> if you have a 2-node NUMA system and want 300 2M huge pages to be
> allocated from node1, you could do:
> 
>  hugepagesnid=1,300,2M
> 
> Or, say you want node0 to allocate 100 huge pages and node1 to
> allocate 200 huge pages, you do:
> 
>  hugepagesnid=0,100,2M hugepagesnid=1,200,1M
> 
> This commit adds support only for 2M huge pages, next commit will
> add support for 1G huge pages.
> 
> ...
>
> --- a/arch/x86/mm/hugetlbpage.c
> +++ b/arch/x86/mm/hugetlbpage.c
> @@ -188,4 +188,39 @@ static __init int setup_hugepagesz(char *opt)
>  	return 1;
>  }
>  __setup("hugepagesz=", setup_hugepagesz);
> +
> +static __init int setup_hugepagesnid(char *opt)
> +{
> +	unsigned long order, ps, nid, nr_pages;
> +	char size_str[3];
> +
> +	size_str[2] = '\0';
> +	if (sscanf(opt, "%lu,%lu,%c%c",
> +			&nid, &nr_pages, &size_str[0], &size_str[1]) < 4) {
> +		printk(KERN_ERR "hugepagesnid: failed to parse arguments\n");
> +			return 0;
> +	}

This will blow up if passed size=16M.  We can expect that powerpc (at
least) will want to copy (or generalise) this code.  It would be better
to avoid such restrictions at the outset.

> +	if (!nr_pages) {
> +		printk(KERN_ERR
> +			"hugepagesnid: zero number of pages, ignoring\n");

The code contains rather a lot of these awkward wordwraps.  Try using
pr_err() and you'll find the result quite pleasing!

> +		return 0;
> +	}
> +
> +	ps = memparse(size_str, NULL);
> +	if (ps == PMD_SIZE) {
> +		order = PMD_SHIFT - PAGE_SHIFT;
> +	} else if (ps == PUD_SIZE && cpu_has_gbpages) {
> +		order = PUD_SHIFT - PAGE_SHIFT;
> +	} else {
> +		printk(KERN_ERR "hugepagesnid: Unsupported page size %lu M\n",
> +			ps >> 20);
> +		return 0;
> +	}
> +
> +	hugetlb_add_nrpages_nid(order, nid, nr_pages);
> +	return 1;
> +}
> +__setup("hugepagesnid=", setup_hugepagesnid);
> +
>  #endif
>
> ...
>
> --- a/mm/hugetlb.c
> +++ b/mm/hugetlb.c
> @@ -46,6 +46,7 @@ __initdata LIST_HEAD(huge_boot_pages);
>  static struct hstate * __initdata parsed_hstate;
>  static unsigned long __initdata default_hstate_max_huge_pages;
>  static unsigned long __initdata default_hstate_size;
> +static unsigned long __initdata boot_alloc_nodes[HUGE_MAX_HSTATE][MAX_NUMNODES];
>  
>  /*
>   * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
> @@ -1348,6 +1349,50 @@ static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
>  	h->max_huge_pages = i;
>  }
>  
> +static unsigned long __init alloc_huge_pages_nid(struct hstate *h,
> +						int nid,
> +						unsigned long nr_pages)
> +{
> +	unsigned long i;
> +	struct page *page;
> +
> +	for (i = 0; i < nr_pages; i++) {
> +		page = alloc_fresh_huge_page_node(h, nid);
> +		if (!page) {
> +			count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
> +			break;
> +		}
> +		count_vm_event(HTLB_BUDDY_PGALLOC);
> +	}
> +
> +	return i;
> +}
> +
> +static unsigned long __init alloc_huge_pages_nodes(struct hstate *h)
> +{
> +	unsigned long i, *entry, ret = 0;
> +
> +	for (i = 0; i < MAX_NUMNODES; i++) {
> +		entry = &boot_alloc_nodes[hstate_index(h)][i];
> +		if (*entry > 0)
> +			ret += alloc_huge_pages_nid(h, i, *entry);
> +	}
> +
> +	return ret;
> +}
> +
> +static void __init hugetlb_init_hstates_nodes(void)
> +{
> +	struct hstate *h;
> +	unsigned long ret;
> +
> +	for_each_hstate(h)
> +		if (h->order < MAX_ORDER) {
> +			ret = alloc_huge_pages_nodes(h);
> +			h->max_huge_pages += ret;
> +		}
> +}

The patch adds code to mm/hugetlb.c which only x86 will use.  I guess
that's OK medium-term if we expect other architectures will use it. 
But if other architectures use it, setup_hugepagesnid() was in the wrong
directory.

Can I ask that you poke the ppc/arm/ia64/etc people, see whether they
will adpot this?  Explaining the overall justification better in [patch
0/n] would help this discussion!

>  static void __init hugetlb_init_hstates(void)
>  {
>  	struct hstate *h;
> @@ -1966,6 +2011,7 @@ static int __init hugetlb_init(void)
>  		default_hstate.max_huge_pages = default_hstate_max_huge_pages;
>  
>  	hugetlb_init_hstates();
> +	hugetlb_init_hstates_nodes();
>  	gather_bootmem_prealloc();
>  	report_hugepages();
>  
> @@ -2005,6 +2051,37 @@ void __init hugetlb_add_hstate(unsigned order)
>  	parsed_hstate = h;
>  }
>  
> +void __init hugetlb_add_nrpages_nid(unsigned order, unsigned long nid,
> +				unsigned long nr_pages)

Using an unsigned long for a NUMA node ID is overkill, surely.

> +{
> +	struct hstate *h;
> +	unsigned long *p;
> +
> +	if (parsed_hstate) {
> +		printk(KERN_WARNING
> +			"hugepagesnid: hugepagesz= specified, ignoring\n");
> +		return;
> +	}
> +
> +	for_each_hstate(h)
> +		if (h->order == order)
> +			break;
> +
> +	if (h->order != order) {
> +		hugetlb_add_hstate(order);
> +		parsed_hstate = NULL;
> +	}
> +
> +	p = &boot_alloc_nodes[hstate_index(h)][nid];
> +	if (*p != 0) {
> +		printk(KERN_WARNING
> +			"hugepagesnid: node %lu already specified, ignoring\n", nid);
> +		return;
> +	}
> +
> +	*p = nr_pages;
> +}
> +

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxx.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>




[Index of Archives]     [Linux ARM Kernel]     [Linux ARM]     [Linux Omap]     [Fedora ARM]     [IETF Annouce]     [Bugtraq]     [Linux]     [Linux OMAP]     [Linux MIPS]     [ECOS]     [Asterisk Internet PBX]     [Linux API]