[PATCH] mm/hugetlb: split hugetlb_cma in nodes with memory

[Song Bao Hua (Barry Song)]

> -----Original Message-----
> From: Anshuman Khandual [mailto:anshuman.khandual at arm.com]
> Sent: Wednesday, July 8, 2020 4:18 PM
> To: Song Bao Hua (Barry Song) <song.bao.hua at hisilicon.com>;
> akpm at linux-foundation.org
> Cc: x86 at kernel.org; linux-mm at kvack.org; linux-kernel at vger.kernel.org;
> Linuxarm <linuxarm at huawei.com>; linux-arm-kernel at lists.infradead.org;
> Roman Gushchin <guro at fb.com>; Catalin Marinas
> <catalin.marinas at arm.com>; Will Deacon <will at kernel.org>; Thomas Gleixner
> <tglx at linutronix.de>; Ingo Molnar <mingo at redhat.com>; Borislav Petkov
> <bp at alien8.de>; H . Peter Anvin <hpa at zytor.com>; Mike Kravetz
> <mike.kravetz at oracle.com>; Mike Rapoport <rppt at linux.ibm.com>; Jonathan
> Cameron <jonathan.cameron at huawei.com>
> Subject: Re: [PATCH] mm/hugetlb: split hugetlb_cma in nodes with memory
> 
> Hello Barry,
> 
> On 07/08/2020 05:53 AM, Barry Song wrote:
> > Rather than splitting huge_cma in online nodes, it is better to do it in
> > nodes with memory.
> 
> Right, it makes sense to avoid nodes without memory, hence loosing portions
> of CMA reservation intended for HugeTLB. N_MEMORY is better than
> N_ONLINE
> and will help avoid this situation.

Thanks for taking a look, Anshuman.

> 
> > For an ARM64 server with four numa nodes and only node0 has memory. If I
> > set hugetlb_cma=4G in bootargs,
> >
> > without this patch, I got the below printk:
> > hugetlb_cma: reserve 4096 MiB, up to 1024 MiB per node
> > hugetlb_cma: reserved 1024 MiB on node 0
> > hugetlb_cma: reservation failed: err -12, node 1
> > hugetlb_cma: reservation failed: err -12, node 2
> > hugetlb_cma: reservation failed: err -12, node 3
> 
> As expected.
> 
> >
> > hugetlb_cma size is broken once the system has nodes without memory.
> 
> I would not say that it is 'broken'. It is just not optimal but still works
> as designed.
> 
> >
> > With this patch, I got the below printk:
> > hugetlb_cma: reserve 4096 MiB, up to 4096 MiB per node
> > hugetlb_cma: reserved 4096 MiB on node 0
> 
> As expected, the per node CMA reservation quota has changed from
> N_ONLINE
> to N_MEMORY.
> 
> >
> > So this patch fixes the broken hugetlb_cma size on arm64.
> 
> There is nothing arm64 specific here. A platform where N_ONLINE !=
> N_MEMORY
> i.e with some nodes without memory when CMA reservation gets called, will
> have this problem.

Agreed. one fact is that right now only x86 and arm64 are calling hugetlb_cma_reserve().
So I don't know how eager other platforms need this function.

> 
> >
> > Jonathan Cameron tested this patch on x86 platform. Jonathan figured out
> x86
> > is much different with arm64. hugetlb_cma size has never broken on x86.
> > On arm64 all nodes are marked online at the same time. On x86, only
> > nodes with memory are initially marked as online:
> > initmem_init()->x86_numa_init()->numa_init()->
> > numa_register_memblks()->alloc_node_data()->node_set_online()
> > So at time of the existing cma setup call only the memory containing nodes
> > are online. The other nodes are brought up much later.
> 
> The problem is always there if N_ONLINE != N_MEMORY but in this case, it
> is just hidden because N_ONLINE happen to match N_MEMORY during the
> boot
> process when hugetlb_cma_reserve() gets called.

Yes. Exactly.

> 
> >
> > Thus, the change is simply to fix ARM64.  A change is needed to x86 only
> > because the inherent assumptions in cma_hugetlb_reserve() have changed.
> 
> cma_hugetlb_reserve() will now scan over N_MEMORY and hence expects all
> platforms to have N_MEMORY initialized properly before calling it. This
> needs to be well documented for the hugetlb_cma_reserve() function along
> with it's call sites.
> 

Yep. will document this.

> >
> > Fixes: cf11e85fc08c ("mm: hugetlb: optionally allocate gigantic hugepages
> using cma")
> 
> I would not call this a "Fix". The current code still works, though in
> a sub optimal manner.

Do you think it is worth linux-stable? For example, is it better for this optimal manner
to be in 5.7 and 5.8? or we have this patch in 5.9-rc1?

To me, I would prefer 5.7 and 5.8 users can still have a hugetlb cma size which is consistent
with the bootargs.

> 
> > Cc: Roman Gushchin <guro at fb.com>
> > Cc: Catalin Marinas <catalin.marinas at arm.com>
> > Cc: Will Deacon <will at kernel.org>
> > Cc: Thomas Gleixner <tglx at linutronix.de>
> > Cc: Ingo Molnar <mingo at redhat.com>
> > Cc: Borislav Petkov <bp at alien8.de>
> > Cc: H. Peter Anvin <hpa at zytor.com>
> > Cc: Mike Kravetz <mike.kravetz at oracle.com>
> > Cc: Mike Rapoport <rppt at linux.ibm.com>
> > Cc: Andrew Morton <akpm at linux-foundation.org>
> > Cc: Anshuman Khandual <anshuman.khandual at arm.com>
> > Cc: Jonathan Cameron <jonathan.cameron at huawei.com>
> > Signed-off-by: Barry Song <song.bao.hua at hisilicon.com>
> > ---
> >  arch/arm64/mm/init.c    | 18 +++++++++---------
> >  arch/x86/kernel/setup.c | 13 ++++++++++---
> >  mm/hugetlb.c            |  4 ++--
> >  3 files changed, 21 insertions(+), 14 deletions(-)
> >
> > diff --git a/arch/arm64/mm/init.c b/arch/arm64/mm/init.c
> > index 1e93cfc7c47a..f6090ef6812b 100644
> > --- a/arch/arm64/mm/init.c
> > +++ b/arch/arm64/mm/init.c
> > @@ -420,15 +420,6 @@ void __init bootmem_init(void)
> >
> >  	arm64_numa_init();
> >
> > -	/*
> > -	 * must be done after arm64_numa_init() which calls numa_init() to
> > -	 * initialize node_online_map that gets used in hugetlb_cma_reserve()
> > -	 * while allocating required CMA size across online nodes.
> > -	 */
> > -#ifdef CONFIG_ARM64_4K_PAGES
> > -	hugetlb_cma_reserve(PUD_SHIFT - PAGE_SHIFT);
> > -#endif
> > -
> >  	/*
> >  	 * Sparsemem tries to allocate bootmem in memory_present(), so must
> be
> >  	 * done after the fixed reservations.
> > @@ -438,6 +429,15 @@ void __init bootmem_init(void)
> >  	sparse_init();
> >  	zone_sizes_init(min, max);
> >
> > +	/*
> > +	 * must be done after zone_sizes_init() which calls node_set_state() to
> > +	 * setup node_states[N_MEMORY] that gets used in
> hugetlb_cma_reserve()
> > +	 * while allocating required CMA size across nodes with memory.
> > +	 */
> 
> Needs better wording here, in particular a reference to free_area_init()
> that updates N_MEMORY via node_set_state(). Also mention the fact that
> now hugetlb_cma_reserve() scans over N_MEMORY nodemask and hence
> expects
> the platforms to have a properly initialized one.

Ok. free_area_init() needs to be highlighted.

> 
> > +#ifdef CONFIG_ARM64_4K_PAGES
> > +	hugetlb_cma_reserve(PUD_SHIFT - PAGE_SHIFT);
> > +#endif
> > +
> >  	memblock_dump_all();
> >  }
> >
> > diff --git a/arch/x86/kernel/setup.c b/arch/x86/kernel/setup.c
> > index a3767e74c758..fdb3a934b6c6 100644
> > --- a/arch/x86/kernel/setup.c
> > +++ b/arch/x86/kernel/setup.c
> > @@ -1164,9 +1164,6 @@ void __init setup_arch(char **cmdline_p)
> >  	initmem_init();
> >  	dma_contiguous_reserve(max_pfn_mapped << PAGE_SHIFT);
> >
> > -	if (boot_cpu_has(X86_FEATURE_GBPAGES))
> > -		hugetlb_cma_reserve(PUD_SHIFT - PAGE_SHIFT);
> > -
> >  	/*
> >  	 * Reserve memory for crash kernel after SRAT is parsed so that it
> >  	 * won't consume hotpluggable memory.
> > @@ -1180,6 +1177,16 @@ void __init setup_arch(char **cmdline_p)
> >
> >  	x86_init.paging.pagetable_init();
> >
> > +	/*
> > +	 * must be done after zone_sizes_init() which calls node_set_state() to
> > +	 * setup node_states[N_MEMORY] that gets used in
> hugetlb_cma_reserve()
> > +	 * while allocating required CMA size across nodes with memory.
> > +	 * And zone_sizes_init() is done in x86_init.paging.pagetable_init()
> > +	 * which is typically paging_init().
> > +	 */
> 
> Drop the last sentence here. Should have just the same comment as arm64.

Do we need something to explain why x86_init.paging.pagetable_init() can do 
free_area_init()?

> 
> > +	if (boot_cpu_has(X86_FEATURE_GBPAGES))
> > +		hugetlb_cma_reserve(PUD_SHIFT - PAGE_SHIFT);
> > +
> >  	kasan_init();
> >
> >  	/*
> > diff --git a/mm/hugetlb.c b/mm/hugetlb.c
> > index d293c823121e..3a0ad49187e4 100644
> > --- a/mm/hugetlb.c
> > +++ b/mm/hugetlb.c
> > @@ -5699,12 +5699,12 @@ void __init hugetlb_cma_reserve(int order)
> >  	 * If 3 GB area is requested on a machine with 4 numa nodes,
> >  	 * let's allocate 1 GB on first three nodes and ignore the last one.
> >  	 */
> > -	per_node = DIV_ROUND_UP(hugetlb_cma_size, nr_online_nodes);
> > +	per_node = DIV_ROUND_UP(hugetlb_cma_size,
> num_node_state(N_MEMORY));
> >  	pr_info("hugetlb_cma: reserve %lu MiB, up to %lu MiB per node\n",
> >  		hugetlb_cma_size / SZ_1M, per_node / SZ_1M);
> >
> >  	reserved = 0;
> > -	for_each_node_state(nid, N_ONLINE) {
> > +	for_each_node_state(nid, N_MEMORY) {
> >  		int res;
> >
> >  		size = min(per_node, hugetlb_cma_size - reserved);
> >
> 
> The patch makes sense. But it needs better articulation of the problem in
> the commit message, specifically pointing out the fact that it originates
> primarily from a scenario where N_ONLINE != N_MEMORY because the
> presence
> of memory less online nodes. It manifests itself on arm64 because of how
> N_ONLINE and N_MEMORY gets initialized during boot but remains hidden on
> x86 for the very same reason.

Thanks
Barry