Re: [PATCH 09/12] mm: Update vm_normal_page() callers to accept FS DAX pages

[Alistair Popple <apopple@nvidia.com>]

Dan Williams <dan.j.williams@intel.com> writes:

> Alistair Popple wrote:
>> Currently if a PTE points to a FS DAX page vm_normal_page() will
>> return NULL as these have their own special refcounting scheme. A
>> future change will allow FS DAX pages to be refcounted the same as any
>> other normal page.
>> 
>> Therefore vm_normal_page() will start returning FS DAX pages. To avoid
>> any change in behaviour callers that don't expect FS DAX pages will
>> need to explicitly check for this. As vm_normal_page() can already
>> return ZONE_DEVICE pages most callers already include a check for any
>> ZONE_DEVICE page.
>> 
>> However some callers don't, so add explicit checks where required.
>
> I would expect justification for each of these conversions, and
> hopefully with fsdax returning fully formed folios there is less need to
> sprinkle these checks around.
>
> At a minimum I think this patch needs to be broken up by file touched.

Good idea.

>> Signed-off-by: Alistair Popple <apopple@nvidia.com>
>> ---
>>  arch/x86/mm/pat/memtype.c |  4 +++-
>>  fs/proc/task_mmu.c        | 16 ++++++++++++----
>>  mm/memcontrol-v1.c        |  2 +-
>>  3 files changed, 16 insertions(+), 6 deletions(-)
>> 
>> diff --git a/arch/x86/mm/pat/memtype.c b/arch/x86/mm/pat/memtype.c
>> index 1fa0bf6..eb84593 100644
>> --- a/arch/x86/mm/pat/memtype.c
>> +++ b/arch/x86/mm/pat/memtype.c
>> @@ -951,6 +951,7 @@ static void free_pfn_range(u64 paddr, unsigned long size)
>>  static int follow_phys(struct vm_area_struct *vma, unsigned long *prot,
>>  		resource_size_t *phys)
>>  {
>> +	struct folio *folio;
>>  	pte_t *ptep, pte;
>>  	spinlock_t *ptl;
>>  
>> @@ -960,7 +961,8 @@ static int follow_phys(struct vm_area_struct *vma, unsigned long *prot,
>>  	pte = ptep_get(ptep);
>>  
>>  	/* Never return PFNs of anon folios in COW mappings. */
>> -	if (vm_normal_folio(vma, vma->vm_start, pte)) {
>> +	folio = vm_normal_folio(vma, vma->vm_start, pte);
>> +	if (folio || (folio && !folio_is_device_dax(folio))) {
>
> ...for example, I do not immediately see why follow_phys() would need to
> be careful with fsdax pages?

The intent was to maintain the original behaviour as much as
possible, partly to reduce the chance of unintended bugs/consequences
and partly to maintain my sanity by not having to dig too deeply into
all the callers.

I see I got this a little bit wrong though - it only filters FSDAX pages
and not device DAX (my intent was to filter both).

> ...but I do see why copy_page_range() (which calls follow_phys() through
> track_pfn_copy()) might care. It just turns out that vma_needs_copy(),
> afaics, bypasses dax MAP_SHARED mappings.
>
> So this touch of memtype.c looks like it can be dropped.

Ok. Although it feels safer to leave it (along with a check for device
DAX). Someone can always remove it in future if they really do want DAX
pages but this is all x86 specific so will take your guidance here.

>>  		pte_unmap_unlock(ptep, ptl);
>>  		return -EINVAL;
>>  	}
>> diff --git a/fs/proc/task_mmu.c b/fs/proc/task_mmu.c
>> index 5f171ad..456b010 100644
>> --- a/fs/proc/task_mmu.c
>> +++ b/fs/proc/task_mmu.c
>> @@ -816,6 +816,8 @@ static void smaps_pte_entry(pte_t *pte, unsigned long addr,
>>  
>>  	if (pte_present(ptent)) {
>>  		page = vm_normal_page(vma, addr, ptent);
>> +		if (page && is_device_dax_page(page))
>> +			page = NULL;
>>  		young = pte_young(ptent);
>>  		dirty = pte_dirty(ptent);
>>  		present = true;
>> @@ -864,6 +866,8 @@ static void smaps_pmd_entry(pmd_t *pmd, unsigned long addr,
>>  
>>  	if (pmd_present(*pmd)) {
>>  		page = vm_normal_page_pmd(vma, addr, *pmd);
>> +		if (page && is_device_dax_page(page))
>> +			page = NULL;
>>  		present = true;
>>  	} else if (unlikely(thp_migration_supported() && is_swap_pmd(*pmd))) {
>>  		swp_entry_t entry = pmd_to_swp_entry(*pmd);
>
> The above can be replaced with a catch like
>
>    if (folio_test_device(folio))
> 	return;
>
> ...in smaps_account() since ZONE_DEVICE pages are not suitable to
> account as they do not reflect any memory pressure on the system memory
> pool.

Sounds good.

>> @@ -1385,7 +1389,7 @@ static inline bool pte_is_pinned(struct vm_area_struct *vma, unsigned long addr,
>>  	if (likely(!test_bit(MMF_HAS_PINNED, &vma->vm_mm->flags)))
>>  		return false;
>>  	folio = vm_normal_folio(vma, addr, pte);
>> -	if (!folio)
>> +	if (!folio || folio_is_device_dax(folio))
>>  		return false;
>>  	return folio_maybe_dma_pinned(folio);
>
> The whole point of ZONE_DEVICE is to account for DMA so I see no reason
> for pte_is_pinned() to special case dax. The caller of pte_is_pinned()
> is doing it for soft_dirty reasons, and I believe soft_dirty is already
> disabled for vma_is_dax(). I assume MEMORY_DEVICE_PRIVATE also does not
> support soft-dirty, so I expect all ZONE_DEVICE already opt-out of this.

Actually soft-dirty is theoretically supported on DEVICE_PRIVATE pages
in the sense that the soft-dirty bits are copied around. Whether or not
it actually works is a different question though, I've certainly never
tried it.

Again, was just trying to maintain previous behaviour but can drop this
check.

>>  }
>> @@ -1710,6 +1714,8 @@ static pagemap_entry_t pte_to_pagemap_entry(struct pagemapread *pm,
>>  			frame = pte_pfn(pte);
>>  		flags |= PM_PRESENT;
>>  		page = vm_normal_page(vma, addr, pte);
>> +		if (page && is_device_dax_page(page))
>> +			page = NULL;
>>  		if (pte_soft_dirty(pte))
>>  			flags |= PM_SOFT_DIRTY;
>>  		if (pte_uffd_wp(pte))
>> @@ -2096,7 +2102,8 @@ static unsigned long pagemap_page_category(struct pagemap_scan_private *p,
>>  
>>  		if (p->masks_of_interest & PAGE_IS_FILE) {
>>  			page = vm_normal_page(vma, addr, pte);
>> -			if (page && !PageAnon(page))
>> +			if (page && !PageAnon(page) &&
>> +			    !is_device_dax_page(page))
>>  				categories |= PAGE_IS_FILE;
>>  		}
>>  
>> @@ -2158,7 +2165,8 @@ static unsigned long pagemap_thp_category(struct pagemap_scan_private *p,
>>  
>>  		if (p->masks_of_interest & PAGE_IS_FILE) {
>>  			page = vm_normal_page_pmd(vma, addr, pmd);
>> -			if (page && !PageAnon(page))
>> +			if (page && !PageAnon(page) &&
>> +			    !is_device_dax_page(page))
>>  				categories |= PAGE_IS_FILE;
>>  		}
>>  
>> @@ -2919,7 +2927,7 @@ static struct page *can_gather_numa_stats_pmd(pmd_t pmd,
>>  		return NULL;
>>  
>>  	page = vm_normal_page_pmd(vma, addr, pmd);
>> -	if (!page)
>> +	if (!page || is_device_dax_page(page))
>>  		return NULL;
>
> I am not immediately seeing a reason to block pagemap_read() from
> interrogating dax-backed virtual mappings. I think these protections can
> be dropped.

Ok.

>>  
>>  	if (PageReserved(page))
>> diff --git a/mm/memcontrol-v1.c b/mm/memcontrol-v1.c
>> index b37c0d8..e16053c 100644
>> --- a/mm/memcontrol-v1.c
>> +++ b/mm/memcontrol-v1.c
>> @@ -667,7 +667,7 @@ static struct page *mc_handle_present_pte(struct vm_area_struct *vma,
>>  {
>>  	struct page *page = vm_normal_page(vma, addr, ptent);
>>  
>> -	if (!page)
>> +	if (!page || is_device_dax_page(page))
>>  		return NULL;
>>  	if (PageAnon(page)) {
>>  		if (!(mc.flags & MOVE_ANON))
>
> I think this better handled with something like this to disable all
> memcg accounting for ZONE_DEVICE pages:

Ok, thanks for the review.

> diff --git a/mm/memcontrol-v1.c b/mm/memcontrol-v1.c
> index b37c0d870816..cfc43e8c59fe 100644
> --- a/mm/memcontrol-v1.c
> +++ b/mm/memcontrol-v1.c
> @@ -940,8 +940,7 @@ static enum mc_target_type get_mctgt_type(struct vm_area_struct *vma,
>                  */
>                 if (folio_memcg(folio) == mc.from) {
>                         ret = MC_TARGET_PAGE;
> -                       if (folio_is_device_private(folio) ||
> -                           folio_is_device_coherent(folio))
> +                       if (folio_is_device(folio))
>                                 ret = MC_TARGET_DEVICE;
>                         if (target)
>                                 target->folio = folio;