Re: [PATCH 06/13] mm/memory: Add dax_insert_pfn

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



Jan Kara <jack@xxxxxxx> writes:

> On Thu 27-06-24 10:54:21, Alistair Popple wrote:
>> Currently to map a DAX page the DAX driver calls vmf_insert_pfn. This
>> creates a special devmap PTE entry for the pfn but does not take a
>> reference on the underlying struct page for the mapping. This is
>> because DAX page refcounts are treated specially, as indicated by the
>> presence of a devmap entry.
>> 
>> To allow DAX page refcounts to be managed the same as normal page
>> refcounts introduce dax_insert_pfn. This will take a reference on the
>> underlying page much the same as vmf_insert_page, except it also
>> permits upgrading an existing mapping to be writable if
>> requested/possible.
>> 
>> Signed-off-by: Alistair Popple <apopple@xxxxxxxxxx>
>
> Overall this looks good to me. Some comments below.
>
>> ---
>>  include/linux/mm.h |  4 ++-
>>  mm/memory.c        | 79 ++++++++++++++++++++++++++++++++++++++++++-----
>>  2 files changed, 76 insertions(+), 7 deletions(-)
>> 
>> diff --git a/include/linux/mm.h b/include/linux/mm.h
>> index 9a5652c..b84368b 100644
>> --- a/include/linux/mm.h
>> +++ b/include/linux/mm.h
>> @@ -1080,6 +1080,8 @@ int vma_is_stack_for_current(struct vm_area_struct *vma);
>>  struct mmu_gather;
>>  struct inode;
>>  
>> +extern void prep_compound_page(struct page *page, unsigned int order);
>> +
>
> You don't seem to use this function in this patch?

Thanks, bad rebase splitting this up. It belongs later in the series.

>> diff --git a/mm/memory.c b/mm/memory.c
>> index ce48a05..4f26a1f 100644
>> --- a/mm/memory.c
>> +++ b/mm/memory.c
>> @@ -1989,14 +1989,42 @@ static int validate_page_before_insert(struct page *page)
>>  }
>>  
>>  static int insert_page_into_pte_locked(struct vm_area_struct *vma, pte_t *pte,
>> -			unsigned long addr, struct page *page, pgprot_t prot)
>> +			unsigned long addr, struct page *page, pgprot_t prot, bool mkwrite)
>>  {
>>  	struct folio *folio = page_folio(page);
>> +	pte_t entry = ptep_get(pte);
>>  
>> -	if (!pte_none(ptep_get(pte)))
>> +	if (!pte_none(entry)) {
>> +		if (mkwrite) {
>> +			/*
>> +			 * For read faults on private mappings the PFN passed
>> +			 * in may not match the PFN we have mapped if the
>> +			 * mapped PFN is a writeable COW page.  In the mkwrite
>> +			 * case we are creating a writable PTE for a shared
>> +			 * mapping and we expect the PFNs to match. If they
>> +			 * don't match, we are likely racing with block
>> +			 * allocation and mapping invalidation so just skip the
>> +			 * update.
>> +			 */
>> +			if (pte_pfn(entry) != page_to_pfn(page)) {
>> +				WARN_ON_ONCE(!is_zero_pfn(pte_pfn(entry)));
>> +				return -EFAULT;
>> +			}
>> +			entry = maybe_mkwrite(entry, vma);
>> +			entry = pte_mkyoung(entry);
>> +			if (ptep_set_access_flags(vma, addr, pte, entry, 1))
>> +				update_mmu_cache(vma, addr, pte);
>> +			return 0;
>> +		}
>>  		return -EBUSY;
>
> If you do this like:
>
> 		if (!mkwrite)
> 			return -EBUSY;
>
> You can reduce indentation of the big block and also making the flow more
> obvious...

Good idea.

>> +	}
>> +
>>  	/* Ok, finally just insert the thing.. */
>>  	folio_get(folio);
>> +	if (mkwrite)
>> +		entry = maybe_mkwrite(mk_pte(page, prot), vma);
>> +	else
>> +		entry = mk_pte(page, prot);
>
> I'd prefer:
>
> 	entry = mk_pte(page, prot);
> 	if (mkwrite)
> 		entry = maybe_mkwrite(entry, vma);
>
> but I don't insist. Also insert_pfn() additionally has pte_mkyoung() and
> pte_mkdirty(). Why was it left out here?

An oversight by me, thanks for pointing it out!

> 								Honza





[Index of Archives]     [Kernel Newbies]     [Security]     [Netfilter]     [Bugtraq]     [Linux FS]     [Yosemite Forum]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Linux RAID]     [Samba]     [Video 4 Linux]     [Device Mapper]     [Linux Resources]

  Powered by Linux