Re: [PATCH] mm: page_alloc: don't allocate page from memoryless nodes

[Qi Zheng <zhengqi.arch@bytedance.com>]

On 2023/2/15 18:08, Mike Rapoport wrote:
> On Wed, Feb 15, 2023 at 05:41:52PM +0800, Qi Zheng wrote:
>>
>>
>> On 2023/2/15 17:30, Mike Rapoport wrote:
>>> On Tue, Feb 14, 2023 at 02:38:44PM +0100, Michal Hocko wrote:
>>>> On Tue 14-02-23 12:58:39, David Hildenbrand wrote:
>>>>> On 14.02.23 12:48, David Hildenbrand wrote:
>>>>>> On 14.02.23 12:44, Mike Rapoport wrote:
>>>>>>> (added x86 folks)
>>>>>>>
>>>>>>> On Tue, Feb 14, 2023 at 12:29:42PM +0100, David Hildenbrand wrote:
>>>>>>>> On 14.02.23 12:26, Qi Zheng wrote:
>>>>>>>>> On 2023/2/14 19:22, David Hildenbrand wrote:
>>>>>>>>>>
>>>>>>>>>> TBH, this is the first time I hear of NODE_MIN_SIZE and it seems to be a
>>>>>>>>>> pretty x86 specific thing.
>>>>>>>>>>
>>>>>>>>>> Are we sure we want to get NODE_MIN_SIZE involved?
>>>>>>>>>
>>>>>>>>> Maybe add an arch_xxx() to handle it?
>>>>>>>>
>>>>>>>> I still haven't figured out what we want to achieve with NODE_MIN_SIZE at
>>>>>>>> all. It smells like an arch-specific hack looking at
>>>>>>>>
>>>>>>>> "Don't confuse VM with a node that doesn't have the minimum amount of
>>>>>>>> memory"
>>>>>>>>
>>>>>>>> Why shouldn't mm-core deal with that?
>>>>>>>
>>>>>>> Well, a node with <4M RAM is not very useful and bears all the overhead of
>>>>>>> an extra live node.
>>>>>>
>>>>>> And totally not with 4.1M, haha.
>>>>>>
>>>>>> I really like the "Might fix boot" in the commit description.
>>>>>>
>>>>>>>
>>>>>>> But, hey, why won't we just drop that '< NODE_MIN_SIZE' and let people with
>>>>>>> weird HW configurations just live with this?
>>>>>>
>>>>>>
>>>>>> ;)
>>>>>>
>>>>>
>>>>> Actually, remembering 09f49dca570a ("mm: handle uninitialized numa nodes
>>>>> gracefully"), this might be the right thing to do. That commit assumes that
>>>>> all offline nodes would get the pgdat allocated in free_area_init(). So that
>>>>> we end up with an allocated pgdat for all possible nodes. The reasoning IIRC
>>>>> was that we don't care about wasting memory in weird VM setups.
>>>>
>>>> Yes, that is the case indeed. I suspect the NODE_MIN_SIZE is a relict of
>>>> the past when some PXM entries were incorrect or fishy. I would just
>>>> drop the check and see whether something breaks. Or make those involved
>>>> back then remember whether this is addressing something that is relevant
>>>> these days. Even 5MB node makes (as the memmap is allocated for the
>>>> whole memory section anyway and that is 128MB) a very little sense if you ask me.
>>>
>>> How about we try this:
>>
>> I'm curious how we can test this? I guess no one remembers the
>> historical background of NODE_MIN_SIZE. :(
>   
> At the very least we can verify that your setup works fine with this ;-)
> 
> Of course we cannot test the exact same configuration that NODE_MIN_SIZE
> was supposed to fix, but there was a lot of effort to make core mm
> initialization robust to cope with weird memory layouts and I'm quite
> confident this won't break anything.

Got it. And for the following patch:
Reviewed-by: Qi Zheng <zhengqi.arch@bytedance.com>

In addition, with this patch, although the crash problem I mentioned
will not exist, I still think it is necessary to modify the
find_next_best_node() and offline_pages() mentioned in the summary in
my other reply. What do you think?

>   
>>>   From b670120bcacd3fe34a40d7179c70ca2ab69279e0 Mon Sep 17 00:00:00 2001
>>> From: "Mike Rapoport (IBM)" <rppt@kernel.org>
>>> Date: Wed, 15 Feb 2023 11:12:18 +0200
>>> Subject: [PATCH] x86/mm: drop 4MB restriction on minimal NUMA node size
>>>
>>> Qi Zheng reports crashes in a production environment and provides a
>>> simplified example as a reproducer:
>>>
>>>     For example, if we use qemu to start a two NUMA node kernel,
>>>     one of the nodes has 2M memory (less than NODE_MIN_SIZE),
>>>     and the other node has 2G, then we will encounter the
>>>     following panic:
>>>
>>>     [    0.149844] BUG: kernel NULL pointer dereference, address: 0000000000000000
>>>     [    0.150783] #PF: supervisor write access in kernel mode
>>>     [    0.151488] #PF: error_code(0x0002) - not-present page
>>>     <...>
>>>     [    0.156056] RIP: 0010:_raw_spin_lock_irqsave+0x22/0x40
>>>     <...>
>>>     [    0.169781] Call Trace:
>>>     [    0.170159]  <TASK>
>>>     [    0.170448]  deactivate_slab+0x187/0x3c0
>>>     [    0.171031]  ? bootstrap+0x1b/0x10e
>>>     [    0.171559]  ? preempt_count_sub+0x9/0xa0
>>>     [    0.172145]  ? kmem_cache_alloc+0x12c/0x440
>>>     [    0.172735]  ? bootstrap+0x1b/0x10e
>>>     [    0.173236]  bootstrap+0x6b/0x10e
>>>     [    0.173720]  kmem_cache_init+0x10a/0x188
>>>     [    0.174240]  start_kernel+0x415/0x6ac
>>>     [    0.174738]  secondary_startup_64_no_verify+0xe0/0xeb
>>>     [    0.175417]  </TASK>
>>>     [    0.175713] Modules linked in:
>>>     [    0.176117] CR2: 0000000000000000
>>>
>>> The crashes happen because of inconsistency between nodemask that has
>>> nodes with less than 4MB as memoryless and the actual memory fed into
>>> core mm.
>>>
>>> The commit 9391a3f9c7f1 ("[PATCH] x86_64: Clear more state when ignoring
>>> empty node in SRAT parsing") that introduced minimal size of a NUMA node
>>> does not explain why a node size cannot be less than 4MB and what boot
>>> failures this restriction might fix.
>>>
>>> Since then a lot has changed and core mm won't confuse badly about small
>>> node sizes.
>>>
>>> Drop the limitation for the minimal node size.
>>>
>>> Link: https://lore.kernel.org/all/20230212110305.93670-1-zhengqi.arch@bytedance.com/
>>> Signed-off-by: Mike Rapoport (IBM) <rppt@kernel.org>
>>> ---
>>>    arch/x86/include/asm/numa.h | 7 -------
>>>    arch/x86/mm/numa.c          | 7 -------
>>>    2 files changed, 14 deletions(-)
>>>
>>> diff --git a/arch/x86/include/asm/numa.h b/arch/x86/include/asm/numa.h
>>> index e3bae2b60a0d..ef2844d69173 100644
>>> --- a/arch/x86/include/asm/numa.h
>>> +++ b/arch/x86/include/asm/numa.h
>>> @@ -12,13 +12,6 @@
>>>    #define NR_NODE_MEMBLKS		(MAX_NUMNODES*2)
>>> -/*
>>> - * Too small node sizes may confuse the VM badly. Usually they
>>> - * result from BIOS bugs. So dont recognize nodes as standalone
>>> - * NUMA entities that have less than this amount of RAM listed:
>>> - */
>>> -#define NODE_MIN_SIZE (4*1024*1024)
>>> -
>>>    extern int numa_off;
>>>    /*
>>> diff --git a/arch/x86/mm/numa.c b/arch/x86/mm/numa.c
>>> index 2aadb2019b4f..55e3d895f15c 100644
>>> --- a/arch/x86/mm/numa.c
>>> +++ b/arch/x86/mm/numa.c
>>> @@ -601,13 +601,6 @@ static int __init numa_register_memblks(struct numa_meminfo *mi)
>>>    		if (start >= end)
>>>    			continue;
>>> -		/*
>>> -		 * Don't confuse VM with a node that doesn't have the
>>> -		 * minimum amount of memory:
>>> -		 */
>>> -		if (end && (end - start) < NODE_MIN_SIZE)
>>> -			continue;
>>> -
>>>    		alloc_node_data(nid);
>>>    	}
>>
>> -- 
>> Thanks,
>> Qi
> 

-- 
Thanks,
Qi