Re: [PATCH 5/5] mm: clear N_HIGH_MEMORY map before se set it again -v2

[Yinghai Lu <yinghai@kernel.org>]

Andrew Morton wrote:
> On Thu, 14 May 2009 10:05:43 -0700
> Yinghai Lu <yinghai@kernel.org> wrote:
> 
>> Andrew Morton wrote:
>>> On Thu, 14 May 2009 09:43:22 -0700
>>> Yinghai Lu <yinghai@kernel.org> wrote:
>>>
>>>> incase some system strange SRAT table. some kind of small range.
>>>> or with mem= etc
>>>>
>>> That description is very hard to understand.  Please provide more details.
>> if the wrong SRAT table, have small range for some node. that node will not be onlined.
>> In the early checking, the bit in node_states[N_HIGH_MEMORY] for the node is set even 
>> that node has less RAM like 1M, and it is not cleared before the bit is set again in 
>> the following loop according online nodes.
> 
> Where in the kernel does this setting of the bit in node_states[]
> occur?  early_calculate_totalpages()?

yes.

> 
> Where in the kernel is it later decided to _not_ use these pages in
> that node?  Perhaps that's the place where the problem should be fixed.

in free_area_init_nodes()

        /* Initialise every node */
        mminit_verify_pageflags_layout();
        setup_nr_node_ids();
        for_each_online_node(nid) {
                pg_data_t *pgdat = NODE_DATA(nid);
                free_area_init_node(nid, NULL,
                                find_min_pfn_for_node(nid), NULL);

                /* Any memory on that node */
                if (pgdat->node_present_pages)
                        node_set_state(nid, N_HIGH_MEMORY);
                check_for_regular_memory(pgdat);
        }

so that patch clear that node_mask before set those bits according if that node is online
and node_present_pages is there.

> 
> 
>>>> Signed-off-by: Yinghai Lu <Yinghai@kernel.org>
>>>> Tested-by: Jack Steiner <steiner@sgi.com>
>>> What reason did Jack have to test this?  Perhaps he hit some bug? 
>>> If so, please fully describe that bug in the changelog.
>> for some memmoryless node and strange memmap.
> 
> That's not a very good problem description.
> 
> Put yourself in the position of a distro engineer whose customer
> reports a 2.6.26 problem.  He's going to look at your patch wondering
> whether it might fix his customer's problem.  We should provide him
> with sufficient information to be able to determine this.
> 
>>>
>>>> Index: linux-2.6/mm/page_alloc.c
>>>> ===================================================================
>>>> --- linux-2.6.orig/mm/page_alloc.c
>>>> +++ linux-2.6/mm/page_alloc.c
>>>> @@ -4041,6 +4047,11 @@ void __init free_area_init_nodes(unsigne
>>>>  						early_node_map[i].start_pfn,
>>>>  						early_node_map[i].end_pfn);
>>>>  
>>>> +	/*
>>>> +	 * find_zone_movable_pfns_for_nodes/early_calculate_totalpages init
>>>> +	 * that node_mask, clear it at first
>>>> +	 */
>>>> +	nodes_clear(node_states[N_HIGH_MEMORY]);
>>>>  	/* Initialise every node */
>>>>  	mminit_verify_pageflags_layout();
>>>>  	setup_nr_node_ids();
>>> If CONFIG_HIGHMEM=n, this will clear the N_NORMAL_MEMORY entry in
>>> node_states[].  Why is this correct and desirable?
>> then N_NORMAL_MEMORY == N_HIGH_MEMORY
> 
> I know.
> 
> But it's unobvious that this change is correct and desirable with both
> CONFIG_HIGHMEM=n and CONFIG_HIGHMEM=y.

use ifdef ?

YH