Re: [PATCH] mm: define pte_add_end for consistency

[David Hildenbrand <david@redhat.com>]

On 01.07.20 04:11, Wei Yang wrote:
> On Tue, Jun 30, 2020 at 02:44:00PM +0200, David Hildenbrand wrote:
>> On 30.06.20 05:18, Wei Yang wrote:
>>> When walking page tables, we define several helpers to get the address of
>>> the next boundary. But we don't have one for pte level.
>>>
>>> Let's define it and consolidate the code in several places.
>>>
>>> Signed-off-by: Wei Yang <richard.weiyang@linux.alibaba.com>
>>> ---
>>>  arch/x86/mm/init_64.c   | 6 ++----
>>>  include/linux/pgtable.h | 7 +++++++
>>>  mm/kasan/init.c         | 4 +---
>>>  3 files changed, 10 insertions(+), 7 deletions(-)
>>>
>>> diff --git a/arch/x86/mm/init_64.c b/arch/x86/mm/init_64.c
>>> index dbae185511cd..f902fbd17f27 100644
>>> --- a/arch/x86/mm/init_64.c
>>> +++ b/arch/x86/mm/init_64.c
>>> @@ -973,9 +973,7 @@ remove_pte_table(pte_t *pte_start, unsigned long addr, unsigned long end,
>>>  
>>>  	pte = pte_start + pte_index(addr);
>>>  	for (; addr < end; addr = next, pte++) {
>>> -		next = (addr + PAGE_SIZE) & PAGE_MASK;
>>> -		if (next > end)
>>> -			next = end;
>>> +		next = pte_addr_end(addr, end);
>>>  
>>>  		if (!pte_present(*pte))
>>>  			continue;
>>> @@ -1558,7 +1556,7 @@ void register_page_bootmem_memmap(unsigned long section_nr,
>>>  		get_page_bootmem(section_nr, pud_page(*pud), MIX_SECTION_INFO);
>>>  
>>>  		if (!boot_cpu_has(X86_FEATURE_PSE)) {
>>> -			next = (addr + PAGE_SIZE) & PAGE_MASK;
>>> +			next = pte_addr_end(addr, end);
>>>  			pmd = pmd_offset(pud, addr);
>>>  			if (pmd_none(*pmd))
>>>  				continue;
>>> diff --git a/include/linux/pgtable.h b/include/linux/pgtable.h
>>> index 32b6c52d41b9..0de09c6c89d2 100644
>>> --- a/include/linux/pgtable.h
>>> +++ b/include/linux/pgtable.h
>>> @@ -706,6 +706,13 @@ static inline pgprot_t pgprot_modify(pgprot_t oldprot, pgprot_t newprot)
>>>  })
>>>  #endif
>>>  
>>> +#ifndef pte_addr_end
>>> +#define pte_addr_end(addr, end)						\
>>> +({	unsigned long __boundary = ((addr) + PAGE_SIZE) & PAGE_MASK;	\
>>> +	(__boundary - 1 < (end) - 1) ? __boundary : (end);		\
>>> +})
>>> +#endif
>>> +
>>>  /*
>>>   * When walking page tables, we usually want to skip any p?d_none entries;
>>>   * and any p?d_bad entries - reporting the error before resetting to none.
>>> diff --git a/mm/kasan/init.c b/mm/kasan/init.c
>>> index fe6be0be1f76..89f748601f74 100644
>>> --- a/mm/kasan/init.c
>>> +++ b/mm/kasan/init.c
>>> @@ -349,9 +349,7 @@ static void kasan_remove_pte_table(pte_t *pte, unsigned long addr,
>>>  	unsigned long next;
>>>  
>>>  	for (; addr < end; addr = next, pte++) {
>>> -		next = (addr + PAGE_SIZE) & PAGE_MASK;
>>> -		if (next > end)
>>> -			next = end;
>>> +		next = pte_addr_end(addr, end);
>>>  
>>>  		if (!pte_present(*pte))
>>>  			continue;
>>>
>>
>> I'm not really a friend of this I have to say. We're simply iterating
>> over single pages, not much magic ....
> 
> Hmm... yes, we are iterating on Page boundary, while we many have the case
> when addr or end is not PAGE_ALIGN.

I really do wonder if not having page aligned addresses actually happens
in real life. Page tables operate on page granularity, and
adding/removing unaligned parts feels wrong ... and that's also why I
dislike such a helper.

1. kasan_add_zero_shadow()/kasan_remove_zero_shadow(). If I understand
the logic (WARN_ON()) correctly, we bail out in case we would ever end
up in such a scenario, where we would want to add/remove things not
aligned to PAGE_SIZE.

2. remove_pagetable()...->remove_pte_table()

vmemmap_free() should never try to de-populate sub-pages. Even with
sub-section hot-add/remove (2MB / 512 pages), with valid struct page
sizes (56, 64, 72, 80), we always end up with full pages.

kernel_physical_mapping_remove() is only called via
arch_remove_memory(). That will never remove unaligned parts.

3. register_page_bootmem_memmap()

It operates on full pages only.


This needs in-depth analysis, but my gut feeling is that this alignment
is unnecessary.

> 
>>
>> What would definitely make sense is replacing (addr + PAGE_SIZE) &
>> PAGE_MASK; by PAGE_ALIGN() ...
>>
> 
> No, PAGE_ALIGN() is expanded to be 
> 
> 	(addr + PAGE_SIZE - 1) & PAGE_MASK;
> 
> If we change the code to PAGE_ALIGN(), we would end up with infinite loop.

Very right, it would have to be PAGE_ALIGN(addr + 1).

-- 
Thanks,

David / dhildenb