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

[Qi Zheng]

On 2023/2/14 17:10, Mike Rapoport wrote:
> On Mon, Feb 13, 2023 at 09:47:43AM +0100, Vlastimil Babka wrote:
> > On 2/12/23 12:03, Qi Zheng wrote:
> > > In x86, numa_register_memblks() is only interested in
> > > those nodes which have enough memory, so it skips over
> > > all nodes with memory below NODE_MIN_SIZE (treated as
> > > a memoryless node). Later on, we will initialize these
> > > memoryless nodes (allocate pgdat in free_area_init()
> > > and build zonelist etc), and will online these nodes
> > > in init_cpu_to_node() and init_gi_nodes().
> > >
> > > After boot, these memoryless nodes are in N_ONLINE
> > > state but not in N_MEMORY state. But we can still allocate
> > > pages from these memoryless nodes.
> > >
> > > In SLUB, we only process nodes in the N_MEMORY state,
> > > such as allocating their struct kmem_cache_node. So if
> > > we allocate a page from the memoryless node above to
> > > SLUB, the struct kmem_cache_node of the node corresponding
> > > to this page is NULL, which will cause panic.
> > >
> > > 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
> > >
> > > In addition, we can also encountered this panic in the actual
> > > production environment. We set up a 2c2g container with two
> > > numa nodes, and then reserved 128M for kdump, and then we
> > > can encountered the above panic in the kdump kernel.
> > >
> > > To fix it, we can filter memoryless nodes when allocating
> > > pages.
> > >
> > > Signed-off-by: Qi Zheng <zhengqi.arch@xxxxxxxxxxxxx>
> > > Reported-by: Teng Hu <huteng.ht@xxxxxxxxxxxxx>
> >
> > Well AFAIK the key mechanism to only allocate from "good" nodes is the
> > zonelist, we shouldn't need to start putting extra checks like this. So it
> > seems to me that the code building the zonelists should take the
> > NODE_MIN_SIZE constraint in mind.
>
> Why just not drop the memory for nodes with size < NODE_MIN_SIZE from
> memblock at the first place?

In this way, it seems that no pages of size < NODE_MIN_SIZE nodes will
be released to buddy, so the pages of these nodes will not be allocated,
and the above-mentioned panic will be avoided.

But these nodes will still build zonelists for itself, which seems
unnecessary?

> Then we won't need runtime checks at all.
>   
> > > ---
> > >   mm/page_alloc.c | 5 +++++
> > >   1 file changed, 5 insertions(+)
> > >
> > > diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> > > index 588555754601..b9cce56f4e21 100644
> > > --- a/mm/page_alloc.c
> > > +++ b/mm/page_alloc.c
> > > @@ -4188,6 +4188,11 @@ get_page_from_freelist(gfp_t gfp_mask, unsigned int order, int alloc_flags,
> > >   			(alloc_flags & ALLOC_CPUSET) &&
> > >   			!__cpuset_zone_allowed(zone, gfp_mask))
> > >   				continue;
> > > +
> > > +		/* Don't allocate page from memoryless nodes. */
> > > +		if (!node_state((zone_to_nid(zone)), N_MEMORY))
> > > +			continue;
> > > +
> > >   		/*
> > >   		 * When allocating a page cache page for writing, we
> > >   		 * want to get it from a node that is within its dirty

--
Thanks,
Qi