Re: [PATCH] mm: fix a regression with HIGHMEM introduced by changeset 7f1290f2f2a4d

[Wen Congyang <wencongyang@gmail.com>]

At 2012/11/15 19:28, Bob Liu Wrote:
> On Thu, Nov 15, 2012 at 5:22 PM, Wen Congyang<wency@cn.fujitsu.com>  wrote:
>> Hi, Liu Jiang
>>
>> At 11/14/2012 10:52 PM, Jiang Liu Wrote:
>>> On 11/07/2012 04:43 AM, Andrew Morton wrote:
>>>> On Tue, 6 Nov 2012 09:31:57 +0800
>>>> Jiang Liu<jiang.liu@huawei.com>  wrote:
>>>>
>>>>> Changeset 7f1290f2f2 tries to fix a issue when calculating
>>>>> zone->present_pages, but it causes a regression to 32bit systems with
>>>>> HIGHMEM. With that changeset, function reset_zone_present_pages()
>>>>> resets all zone->present_pages to zero, and fixup_zone_present_pages()
>>>>> is called to recalculate zone->present_pages when boot allocator frees
>>>>> core memory pages into buddy allocator. Because highmem pages are not
>>>>> freed by bootmem allocator, all highmem zones' present_pages becomes
>>>>> zero.
>>>>>
>>>>> Actually there's no need to recalculate present_pages for highmem zone
>>>>> because bootmem allocator never allocates pages from them. So fix the
>>>>> regression by skipping highmem in function reset_zone_present_pages()
>>>>> and fixup_zone_present_pages().
>>>>>
>>>>> ...
>>>>>
>>>>> --- a/mm/page_alloc.c
>>>>> +++ b/mm/page_alloc.c
>>>>> @@ -6108,7 +6108,8 @@ void reset_zone_present_pages(void)
>>>>>      for_each_node_state(nid, N_HIGH_MEMORY) {
>>>>>              for (i = 0; i<  MAX_NR_ZONES; i++) {
>>>>>                      z = NODE_DATA(nid)->node_zones + i;
>>>>> -                   z->present_pages = 0;
>>>>> +                   if (!is_highmem(z))
>>>>> +                           z->present_pages = 0;
>>>>>              }
>>>>>      }
>>>>>   }
>>>>> @@ -6123,10 +6124,11 @@ void fixup_zone_present_pages(int nid, unsigned long start_pfn,
>>>>>
>>>>>      for (i = 0; i<  MAX_NR_ZONES; i++) {
>>>>>              z = NODE_DATA(nid)->node_zones + i;
>>>>> +           if (is_highmem(z))
>>>>> +                   continue;
>>>>> +
>>>>>              zone_start_pfn = z->zone_start_pfn;
>>>>>              zone_end_pfn = zone_start_pfn + z->spanned_pages;
>>>>> -
>>>>> -           /* if the two regions intersect */
>>>>>              if (!(zone_start_pfn>= end_pfn || zone_end_pfn<= start_pfn))
>>>>>                      z->present_pages += min(end_pfn, zone_end_pfn) -
>>>>>                                          max(start_pfn, zone_start_pfn);
>>>>
>>>> This ...  isn't very nice.  It is embeds within
>>>> reset_zone_present_pages() and fixup_zone_present_pages() knowledge
>>>> about their caller's state.  Or, more specifically, it is emebedding
>>>> knowledge about the overall state of the system when these functions
>>>> are called.
>>>>
>>>> I mean, a function called "reset_zone_present_pages" should reset
>>>> ->present_pages!
>>>>
>>>> The fact that fixup_zone_present_page() has multiple call sites makes
>>>> this all even more risky.  And what are the interactions between this
>>>> and memory hotplug?
>>>>
>>>> Can we find a cleaner fix?
>>>>
>>>> Please tell us more about what's happening here.  Is it the case that
>>>> reset_zone_present_pages() is being called *after* highmem has been
>>>> populated?  If so, then fixup_zone_present_pages() should work
>>>> correctly for highmem?  Or is it the case that highmem hasn't yet been
>>>> setup?  IOW, what is the sequence of operations here?
>>>>
>>>> Is the problem that we're *missing* a call to
>>>> fixup_zone_present_pages(), perhaps?  If we call
>>>> fixup_zone_present_pages() after highmem has been populated,
>>>> fixup_zone_present_pages() should correctly fill in the highmem zone's
>>>> ->present_pages?
>>> Hi Andrew,
>>>        Sorry for the late response:(
>>>        I have done more investigations according to your suggestions. Currently
>>> we have only called fixup_zone_present_pages() for memory freed by bootmem
>>> allocator and missed HIGHMEM pages. We could also call fixup_zone_present_pages()
>>> for HIGHMEM pages, but that will need to change arch specific code for x86, powerpc,
>>> sparc, microblaze, arm, mips, um and tile etc. Seems a little overhead.
>>>        And sadly enough, I found the quick fix is still incomplete. The original
>>> patch still have another issue that, reset_zone_present_pages() is only called
>>> for IA64, so it will cause trouble for other arches which make use of "bootmem.c".
>>>        Then I feel a little guilty and tried to find a cleaner solution without
>>> touching arch specific code. But things are more complex than my expectation and
>>> I'm still working on that.
>>>        So how about totally reverting the changeset 7f1290f2f2a4d2c3f1b7ce8e87256e052ca23125
>>> and I will post another version once I found a cleaner way?
>>
>> I think fixup_zone_present_pages() are very useful for memory hotplug.
>>
>
> I might miss something, but if memory hotplug is the only user depends on
> fixup_zone_present_pages().

IIRC, water_mask depends on zone->present_pages. But I don't meet any 
problem
even if zone->present_pages is wrong.

> Why not reverting the changeset 7f1290f2f2a4d2c3f1b7ce8e87256e052ca23125
> And add checking to offline_pages() like:
> if (zone->present_pages>= offlined_page)
>          zone->present_pages -= offlined_pages;
> else
>          zone->present_pages = 0;
>
> It's more simple and can minimize the effect to other parts of kernel.

Hmm, zone->present_pages may be 0 when there is memory in this zone which is
onlined and in use. If zone->present_pages becomes to 0, we will free pcp
list for this zone. It will cause some unexpected error.

>
>> We calculate zone->present_pages in free_area_init_core(), but its value is wrong.
>> So it is why we fix it in fixup_zone_present_pages().
>>
>> What about this:
>> 1. init zone->present_pages to the present pages in this zone(include bootmem)
>> 2. don't reset zone->present_pages for HIGHMEM pages
>>
>> We don't allocate bootmem from HIGHMEM. So its present pages is inited in step1
>> and there is no need to fix it in step2.
>>
>> Is it OK?
>>
>> If it is OK, I will resend the patch for step1(the patch is from laijs).
>>
>

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