Re: [PATCH V8 2/2] arm64/mm: Enable memory hot remove

[Catalin Marinas]

On Mon, Sep 23, 2019 at 11:13:45AM +0530, Anshuman Khandual wrote:
> The arch code for hot-remove must tear down portions of the linear map and
> vmemmap corresponding to memory being removed. In both cases the page
> tables mapping these regions must be freed, and when sparse vmemmap is in
> use the memory backing the vmemmap must also be freed.
> 
> This patch adds unmap_hotplug_range() and free_empty_tables() helpers which
> can be used to tear down either region and calls it from vmemmap_free() and
> ___remove_pgd_mapping(). The sparse_vmap argument determines whether the
> backing memory will be freed.

Can you change the 'sparse_vmap' name to something more meaningful which
would suggest freeing of the backing memory?

> It makes two distinct passes over the kernel page table. In the first pass
> with unmap_hotplug_range() it unmaps, invalidates applicable TLB cache and
> frees backing memory if required (vmemmap) for each mapped leaf entry. In
> the second pass with free_empty_tables() it looks for empty page table
> sections whose page table page can be unmapped, TLB invalidated and freed.
> 
> While freeing intermediate level page table pages bail out if any of its
> entries are still valid. This can happen for partially filled kernel page
> table either from a previously attempted failed memory hot add or while
> removing an address range which does not span the entire page table page
> range.
> 
> The vmemmap region may share levels of table with the vmalloc region.
> There can be conflicts between hot remove freeing page table pages with
> a concurrent vmalloc() walking the kernel page table. This conflict can
> not just be solved by taking the init_mm ptl because of existing locking
> scheme in vmalloc(). So free_empty_tables() implements a floor and ceiling
> method which is borrowed from user page table tear with free_pgd_range()
> which skips freeing page table pages if intermediate address range is not
> aligned or maximum floor-ceiling might not own the entire page table page.
> 
> While here update arch_add_memory() to handle __add_pages() failures by
> just unmapping recently added kernel linear mapping. Now enable memory hot
> remove on arm64 platforms by default with ARCH_ENABLE_MEMORY_HOTREMOVE.
> 
> This implementation is overall inspired from kernel page table tear down
> procedure on X86 architecture and user page table tear down method.
> 
> Acked-by: Steve Capper <steve.capper@xxxxxxx>
> Acked-by: David Hildenbrand <david@xxxxxxxxxx>
> Signed-off-by: Anshuman Khandual <anshuman.khandual@xxxxxxx>

Given the amount of changes since version 7, do the acks still stand?

[...]
> +static void free_pte_table(pmd_t *pmdp, unsigned long addr, unsigned long end,
> +			   unsigned long floor, unsigned long ceiling)
> +{
> +	struct page *page;
> +	pte_t *ptep;
> +	int i;
> +
> +	if (!pgtable_range_aligned(addr, end, floor, ceiling, PMD_MASK))
> +		return;
> +
> +	ptep = pte_offset_kernel(pmdp, 0UL);
> +	for (i = 0; i < PTRS_PER_PTE; i++) {
> +		if (!pte_none(READ_ONCE(ptep[i])))
> +			return;
> +	}
> +
> +	page = pmd_page(READ_ONCE(*pmdp));

Arguably, that's not the pmd page we are freeing here. Even if you get
the same result, pmd_page() is normally used for huge pages pointed at
by the pmd entry. Since you have the ptep already, why not use
virt_to_page(ptep)?

> +	pmd_clear(pmdp);
> +	__flush_tlb_kernel_pgtable(addr);
> +	free_hotplug_pgtable_page(page);
> +}
> +
> +static void free_pmd_table(pud_t *pudp, unsigned long addr, unsigned long end,
> +			   unsigned long floor, unsigned long ceiling)
> +{
> +	struct page *page;
> +	pmd_t *pmdp;
> +	int i;
> +
> +	if (CONFIG_PGTABLE_LEVELS <= 2)
> +		return;
> +
> +	if (!pgtable_range_aligned(addr, end, floor, ceiling, PUD_MASK))
> +		return;
> +
> +	pmdp = pmd_offset(pudp, 0UL);
> +	for (i = 0; i < PTRS_PER_PMD; i++) {
> +		if (!pmd_none(READ_ONCE(pmdp[i])))
> +			return;
> +	}
> +
> +	page = pud_page(READ_ONCE(*pudp));

Same here, virt_to_page(pmdp).

> +	pud_clear(pudp);
> +	__flush_tlb_kernel_pgtable(addr);
> +	free_hotplug_pgtable_page(page);
> +}
> +
> +static void free_pud_table(pgd_t *pgdp, unsigned long addr, unsigned  long end,
> +			   unsigned long floor, unsigned long ceiling)
> +{
> +	struct page *page;
> +	pud_t *pudp;
> +	int i;
> +
> +	if (CONFIG_PGTABLE_LEVELS <= 3)
> +		return;
> +
> +	if (!pgtable_range_aligned(addr, end, floor, ceiling, PGDIR_MASK))
> +		return;
> +
> +	pudp = pud_offset(pgdp, 0UL);
> +	for (i = 0; i < PTRS_PER_PUD; i++) {
> +		if (!pud_none(READ_ONCE(pudp[i])))
> +			return;
> +	}
> +
> +	page = pgd_page(READ_ONCE(*pgdp));

As above.

> +	pgd_clear(pgdp);
> +	__flush_tlb_kernel_pgtable(addr);
> +	free_hotplug_pgtable_page(page);
> +}
> +
> +static void unmap_hotplug_pte_range(pmd_t *pmdp, unsigned long addr,
> +				    unsigned long end, bool sparse_vmap)
> +{
> +	struct page *page;
> +	pte_t *ptep, pte;
> +
> +	do {
> +		ptep = pte_offset_kernel(pmdp, addr);
> +		pte = READ_ONCE(*ptep);
> +		if (pte_none(pte))
> +			continue;
> +
> +		WARN_ON(!pte_present(pte));
> +		page = sparse_vmap ? pte_page(pte) : NULL;
> +		pte_clear(&init_mm, addr, ptep);
> +		flush_tlb_kernel_range(addr, addr + PAGE_SIZE);
> +		if (sparse_vmap)
> +			free_hotplug_page_range(page, PAGE_SIZE);

You could only set 'page' if sparse_vmap (or even drop 'page' entirely).
The compiler is probably smart enough to optimise it but using a
pointless ternary operator just makes the code harder to follow.

> +	} while (addr += PAGE_SIZE, addr < end);
> +}
[...]
> +static void free_empty_pte_table(pmd_t *pmdp, unsigned long addr,
> +				 unsigned long end)
> +{
> +	pte_t *ptep, pte;
> +
> +	do {
> +		ptep = pte_offset_kernel(pmdp, addr);
> +		pte = READ_ONCE(*ptep);
> +		WARN_ON(!pte_none(pte));
> +	} while (addr += PAGE_SIZE, addr < end);
> +}
> +
> +static void free_empty_pmd_table(pud_t *pudp, unsigned long addr,
> +				 unsigned long end, unsigned long floor,
> +				 unsigned long ceiling)
> +{
> +	unsigned long next;
> +	pmd_t *pmdp, pmd;
> +
> +	do {
> +		next = pmd_addr_end(addr, end);
> +		pmdp = pmd_offset(pudp, addr);
> +		pmd = READ_ONCE(*pmdp);
> +		if (pmd_none(pmd))
> +			continue;
> +
> +		WARN_ON(!pmd_present(pmd) || !pmd_table(pmd) || pmd_sect(pmd));
> +		free_empty_pte_table(pmdp, addr, next);
> +		free_pte_table(pmdp, addr, next, floor, ceiling);

Do we need two closely named functions here? Can you not collapse
free_empty_pud_table() and free_pte_table() into a single one? The same
comment for the pmd/pud variants. I just find this confusing.

> +	} while (addr = next, addr < end);

You could make these function in two steps: first, as above, invoke the
next level recursively; second, after the do..while loop, check whether
it's empty and free the pmd page as in free_pmd_table().

> +}
[...]

-- 
Catalin