[PATCHv2] mm: Don't offset memmap for flatmem

Vlastimil Babka vbabka at suse.cz
Thu Jan 29 05:13:38 PST 2015

On 01/26/2015 04:56 PM, Mel Gorman wrote:
> On Fri, Jan 23, 2015 at 10:05:48AM +0100, Vlastimil Babka wrote:
>> On 01/23/2015 01:33 AM, Laura Abbott wrote:
>>> On 1/22/2015 4:20 PM, Andrew Morton wrote:
>>>> I don't think v2 addressed Vlastimil's review comment?
>>> We're still adding the offset to node_mem_map and then subtracting it from
>>> just mem_map. Did I miss another comment somewhere?
>> Yes that was addressed, thanks. But I don't feel comfortable acking
>> it yet, as I have no idea if we are doing the right thing for
>> under the "if (page_to_pfn(mem_map) != pgdat->node_start_pfn)" will
>> probably do the right thing, but looks like a weird test for this
>> case here.
>> I have no good suggestion though, so let's CC Mel who apparently
>> wrote the ARCH_PFN_OFFSET correction?
> I don't recall introducing ARCH_PFN_OFFSET, are you sure it was me?  I'm just
> back today after been offline a week so didn't review the patch but IIRC,
> ARCH_PFN_OFFSET deals with the case where physical memory does not start
> at 0. Without the offset, virtual _PAGE_OFFSET would not physical page 0.
> I don't recall it being related to the alignment of node 0 so if there
> are crashes due to misalignment of node 0 and the fix is ARCH_PFN_OFFSET
> related then I'm surprised.

You're right that ARCH_PFN_OFFSET wasn't added by you, but by commit 
467bc461d2 which was a bugfix to your commit c713216dee, which did 
introduce the mem_map correction code, and after which the code looked like:

mem_map = NODE_DATA(0)->node_mem_map;
                if (page_to_pfn(mem_map) != pgdat->node_start_pfn)
                        mem_map -= pgdat->node_start_pfn;

It's from 2006 so I can't expect you remember the details, but I had 
some trouble finding out what this does. I assume it makes sure that 
mem_map points to struct page corresponding to pfn 0, because that's 
what translations using mem_map expect.
But pgdat->node_mem_map points to struct page corresponding to 
pgdat->node_start_pfn, which might not be 0. So it subtracts 
node_start_pfn to fix that. This is OK, as the node_mem_map is allocated 
(in this very function) with padding so that it covers a 
MAX_ORDER_NR_PAGES aligned area where node_mem_map may point to the 
middle of it.

Commit 467bc461d2 fixed this in case the first pfn is not 0, but 
ARCH_PFN_OFFSET. So mem_map points to struct page corresponding to 
pfn=ARCH_PFN_OFFSET, which is OK. But I still have few doubts:

1) The "if (page_to_pfn(mem_map) != pgdat->node_start_pfn)" sort of 
silently assumes that mem_map is allocated at the beginning of the node, 
i.e. at pgdat->node_start_pfn. And the only reason for this if-condition 
to be true, is that we haven't corrected the page_to_pfn translation, 
which uses mem_map. Is this assumption always OK to do? Shouldn't the 
if-condition be instead about pgdat->node_start_pfn not being aligned?

2) The #ifdef guard is about CONFIG_ARCH_POPULATES_NODE_MAP, which is 
nowadays called CONFIG_HAVE_MEMBLOCK_NODE_MAP. But shouldn't it be 
#ifdef FLATMEM instead? After all, we are correcting value of mem_map 
based on page_to_pfn code variant used on FLATMEM. arm doesn't define
CONFIG_ARCH_POPULATES_NODE_MAP but apparently needs this correction.

3) The node_mem_map allocation code aligns the allocation to 
MAX_ORDER_NR_PAGES, so the offset between the start of the allocated map 
and where node_mem_map points to will be up to MAX_ORDER_NR_PAGES.
However, here we subtract (in current kernel) (pgdat->node_start_pfn - 
ARCH_PFN_OFFSET). That looks like another silent assumption, that 
pgdat->node_start_pfn is always between ARCH_PFN_OFFSET and 
ARCH_PFN_OFFSET + MAX_ORDER_NR_PAGES. If it were larger, the mem_map 
correction would subtract too much and end up below what was allocated 
for node_mem_map, no? The bug report behind this patch said that first 
2MB of memory was reserved using "no-map flag using DT". Unless this 
somehow translates to ARCH_PFN_OFFSET at build time, we would underflow 
mem_map, right? Maybe I'm just overly paranoid here and of course 
ARCH_PFN_OFFSET is determined properly on arm...

If anyone can confirm my doubts or point me to what I'm missing, thanks.

More information about the linux-arm-kernel mailing list