Re: [PATCH 02/12] pci/p2pdma: Don't initialise page refcount to one

[Dan Williams <dan.j.williams@intel.com>]

Alistair Popple wrote:
> The reference counts for ZONE_DEVICE private pages should be
> initialised by the driver when the page is actually allocated by the
> driver allocator, not when they are first created. This is currently
> the case for MEMORY_DEVICE_PRIVATE and MEMORY_DEVICE_COHERENT pages
> but not MEMORY_DEVICE_PCI_P2PDMA pages so fix that up.
> 
> Signed-off-by: Alistair Popple <apopple@nvidia.com>
> ---
>  drivers/pci/p2pdma.c |  6 ++++++
>  mm/memremap.c        | 17 +++++++++++++----
>  mm/mm_init.c         | 22 ++++++++++++++++++----
>  3 files changed, 37 insertions(+), 8 deletions(-)
> 
> diff --git a/drivers/pci/p2pdma.c b/drivers/pci/p2pdma.c
> index 4f47a13..210b9f4 100644
> --- a/drivers/pci/p2pdma.c
> +++ b/drivers/pci/p2pdma.c
> @@ -129,6 +129,12 @@ static int p2pmem_alloc_mmap(struct file *filp, struct kobject *kobj,
>  	}
>  
>  	/*
> +	 * Initialise the refcount for the freshly allocated page. As we have
> +	 * just allocated the page no one else should be using it.
> +	 */
> +	set_page_count(virt_to_page(kaddr), 1);

Perhaps VM_WARN_ONCE to back up that assumption?

I also notice that there are multiple virt_to_page() lookups in this
routine, so maybe time for a local @page variable.

> +
> +	/*
>  	 * vm_insert_page() can sleep, so a reference is taken to mapping
>  	 * such that rcu_read_unlock() can be done before inserting the
>  	 * pages
> diff --git a/mm/memremap.c b/mm/memremap.c
> index 40d4547..07bbe0e 100644
> --- a/mm/memremap.c
> +++ b/mm/memremap.c
> @@ -488,15 +488,24 @@ void free_zone_device_folio(struct folio *folio)
>  	folio->mapping = NULL;
>  	folio->page.pgmap->ops->page_free(folio_page(folio, 0));
>  
> -	if (folio->page.pgmap->type != MEMORY_DEVICE_PRIVATE &&
> -	    folio->page.pgmap->type != MEMORY_DEVICE_COHERENT)
> +	switch (folio->page.pgmap->type) {
> +	case MEMORY_DEVICE_PRIVATE:
> +	case MEMORY_DEVICE_COHERENT:
> +		put_dev_pagemap(folio->page.pgmap);
> +		break;
> +
> +	case MEMORY_DEVICE_FS_DAX:
> +	case MEMORY_DEVICE_GENERIC:
>  		/*
>  		 * Reset the refcount to 1 to prepare for handing out the page
>  		 * again.
>  		 */
>  		folio_set_count(folio, 1);
> -	else
> -		put_dev_pagemap(folio->page.pgmap);
> +		break;
> +
> +	case MEMORY_DEVICE_PCI_P2PDMA:
> +		break;

A follow on cleanup is that either all implementations should be
put_dev_pagemap(), or none of them. Put the onus on the implementation
to track how many pages it has handed out in the implementation
allocator.

> +	}
>  }
>  
>  void zone_device_page_init(struct page *page)
> diff --git a/mm/mm_init.c b/mm/mm_init.c
> index 4ba5607..0489820 100644
> --- a/mm/mm_init.c
> +++ b/mm/mm_init.c
> @@ -1015,12 +1015,26 @@ static void __ref __init_zone_device_page(struct page *page, unsigned long pfn,
>  	}
>  
>  	/*
> -	 * ZONE_DEVICE pages are released directly to the driver page allocator
> -	 * which will set the page count to 1 when allocating the page.
> +	 * ZONE_DEVICE pages other than MEMORY_TYPE_GENERIC and
> +	 * MEMORY_TYPE_FS_DAX pages are released directly to the driver page
> +	 * allocator which will set the page count to 1 when allocating the
> +	 * page.
> +	 *
> +	 * MEMORY_TYPE_GENERIC and MEMORY_TYPE_FS_DAX pages automatically have
> +	 * their refcount reset to one whenever they are freed (ie. after
> +	 * their refcount drops to 0).

I'll send some follow on patches to clean up device-dax.

For this one:

Reviewed-by: Dan Williams <dan.j.williams@intel.com>