[PATCH v4] arm64: mm: Populate vmemmap/linear at the page level for hotplugged sections

[Zhenhua Huang]

Hi Catalin,

On 2025/1/8 3:22, Catalin Marinas wrote:
> On Tue, Jan 07, 2025 at 03:42:52PM +0800, Zhenhua Huang wrote:
>> diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c
>> index e2739b69e11b..5e0f514de870 100644
>> --- a/arch/arm64/mm/mmu.c
>> +++ b/arch/arm64/mm/mmu.c
>> @@ -42,9 +42,11 @@
>>   #include <asm/pgalloc.h>
>>   #include <asm/kfence.h>
>>   
>> -#define NO_BLOCK_MAPPINGS	BIT(0)
>> -#define NO_CONT_MAPPINGS	BIT(1)
>> -#define NO_EXEC_MAPPINGS	BIT(2)	/* assumes FEAT_HPDS is not used */
>> +#define NO_PMD_BLOCK_MAPPINGS	BIT(0)
>> +#define NO_PUD_BLOCK_MAPPINGS	BIT(1)  /* Hotplug case: do not want block mapping for PUD */
>> +#define NO_BLOCK_MAPPINGS (NO_PMD_BLOCK_MAPPINGS | NO_PUD_BLOCK_MAPPINGS)
> 
> Nit: please use a tab instead of space before (NO_PMD_...)
> 
>> +#define NO_CONT_MAPPINGS	BIT(2)
>> +#define NO_EXEC_MAPPINGS	BIT(3)	/* assumes FEAT_HPDS is not used */
>>   
>>   u64 kimage_voffset __ro_after_init;
>>   EXPORT_SYMBOL(kimage_voffset);
>> @@ -254,7 +256,7 @@ static void init_pmd(pmd_t *pmdp, unsigned long addr, unsigned long end,
>>   
>>   		/* try section mapping first */
>>   		if (((addr | next | phys) & ~PMD_MASK) == 0 &&
>> -		    (flags & NO_BLOCK_MAPPINGS) == 0) {
>> +		    (flags & NO_PMD_BLOCK_MAPPINGS) == 0) {
>>   			pmd_set_huge(pmdp, phys, prot);
>>   
>>   			/*
>> @@ -356,10 +358,11 @@ static void alloc_init_pud(p4d_t *p4dp, unsigned long addr, unsigned long end,
>>   
>>   		/*
>>   		 * For 4K granule only, attempt to put down a 1GB block
>> +		 * Hotplug case: do not attempt 1GB block
>>   		 */
> 
> I don't think we need this comment added here. The hotplug case is a
> decision of the caller, so better to have the comment there.

Yeah, will remove.

> 
>>   		if (pud_sect_supported() &&
>>   		   ((addr | next | phys) & ~PUD_MASK) == 0 &&
>> -		    (flags & NO_BLOCK_MAPPINGS) == 0) {
>> +		   (flags & NO_PUD_BLOCK_MAPPINGS) == 0) {
>>   			pud_set_huge(pudp, phys, prot);
> 
> Nit: something wrong with the alignment here. I think the unmodified
> line after the 'if' one above was misaligned before your patch.

Noted and will correct in next patch.

> 
>>   
>>   			/*
>> @@ -1175,9 +1178,21 @@ int __meminit vmemmap_check_pmd(pmd_t *pmdp, int node,
>>   int __meminit vmemmap_populate(unsigned long start, unsigned long end, int node,
>>   		struct vmem_altmap *altmap)
>>   {
>> +	unsigned long start_pfn;
>> +	struct mem_section *ms;
>> +
>>   	WARN_ON((start < VMEMMAP_START) || (end > VMEMMAP_END));
>>   
>> -	if (!IS_ENABLED(CONFIG_ARM64_4K_PAGES))
>> +	start_pfn = page_to_pfn((struct page *)start);
>> +	ms = __pfn_to_section(start_pfn);
> 
> Hmm, it would have been better if the core code provided the start pfn
> as it does for vmemmap_populate_compound_pages() but I'm fine with
> deducting it from 'start'.

I found another bug, that even for early section, when vmemmap_populate 
is called, SECTION_IS_EARLY is not set. Therefore, early_section() 
always return false.

Since vmemmap_populate() occurs during section initialization, it may be 
hard to say it is a bug..
However, should we instead using SECTION_MARKED_PRESENT to check? I 
tested well in my setup.

Hot plug flow:
1. section_activate -> vmemmap_populate
2. mark PRESENT

In contrast, the early flow:
1. memblocks_present -> mark PRESENT
2. __populate_section_memmap -> vmemmap_populate

> 
>> +	/*
>> +	 * Hotplugged section does not support hugepages as
>> +	 * PMD_SIZE (hence PUD_SIZE) section mapping covers
>> +	 * struct page range that exceeds a SUBSECTION_SIZE
>> +	 * i.e 2MB - for all available base page sizes.
>> +	 */
>> +	if (!IS_ENABLED(CONFIG_ARM64_4K_PAGES) || !early_section(ms))
>>   		return vmemmap_populate_basepages(start, end, node, altmap);
>>   	else
>>   		return vmemmap_populate_hugepages(start, end, node, altmap);
>> @@ -1339,9 +1354,25 @@ int arch_add_memory(int nid, u64 start, u64 size,
>>   		    struct mhp_params *params)
>>   {
>>   	int ret, flags = NO_EXEC_MAPPINGS;
>> +	unsigned long start_pfn = page_to_pfn((struct page *)start);
>> +	struct mem_section *ms = __pfn_to_section(start_pfn);
> 
> This looks wrong. 'start' here is a physical address, you want
> PFN_DOWN() instead.

Sorry, my mistake.Thanks for catching it.

> 
>>   
>>   	VM_BUG_ON(!mhp_range_allowed(start, size, true));
>>   
>> +	/* should not be invoked by early section */
>> +	WARN_ON(early_section(ms));
>> +
>> +	/*
>> +	 * 4K base page's PMD_SIZE matches SUBSECTION_SIZE i.e 2MB. Hence
>> +	 * PMD section mapping can be allowed, but only for 4K base pages.
>> +	 * Where as PMD_SIZE (hence PUD_SIZE) for other page sizes exceed
>> +	 * SUBSECTION_SIZE.
>> +	 */
>> +	if (IS_ENABLED(CONFIG_ARM64_4K_PAGES))
>> +		flags |= NO_PUD_BLOCK_MAPPINGS | NO_CONT_MAPPINGS;
> 
> In theory we can allow contiguous PTE mappings but not PMD. You could
> probably do the same as a NO_BLOCK_MAPPINGS and split it into multiple
> components - NO_PTE_CONT_MAPPINGS and so on.
> 
>> +	else
>> +		flags |= NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS;
> 
> Similarly with 16K/64K pages we can allow contiguous PTEs as they all go
> up to 2MB blocks.

Yes!

> 
> I think we should write the flags setup in a more readable way than
> trying to do mental maths on the possible combinations, something like:
> 
> 	flags = NO_PUD_BLOCK_MAPPINGS | NO_PMD_CONT_MAPPINGS;
> 	if (SUBSECTION_SHIFT < PMD_SHIFT)
> 		flags |= NO_PMD_BLOCK_MAPPINGS;
> 	if (SUBSECTION_SHIFT < CONT_PTE_SHIFT)
> 		flags |= NO_PTE_CONT_MAPPINGS;

Good idea indeed. We no longer need to worry about PAGE SIZE CONFIG.

> 
> This way we don't care about the page size and should cover any changes
> to SUBSECTION_SHIFT making it smaller than 2MB.
>