Re: [PATCH 11/15] mm, dax, pmem: introduce __pfn_t

[Dave Hansen <dave@sr71.net>]

On 09/22/2015 09:42 PM, Dan Williams wrote:
>  /*
> + * __pfn_t: encapsulates a page-frame number that is optionally backed
> + * by memmap (struct page).  Whether a __pfn_t has a 'struct page'
> + * backing is indicated by flags in the low bits of the value;
> + */
> +typedef struct {
> +	unsigned long val;
> +} __pfn_t;
> +
> +/*
> + * PFN_SG_CHAIN - pfn is a pointer to the next scatterlist entry
> + * PFN_SG_LAST - pfn references a page and is the last scatterlist entry
> + * PFN_DEV - pfn is not covered by system memmap by default
> + * PFN_MAP - pfn has a dynamic page mapping established by a device driver
> + */
> +enum {
> +	PFN_SHIFT = 4,
> +	PFN_MASK = (1UL << PFN_SHIFT) - 1,
> +	PFN_SG_CHAIN = (1UL << 0),
> +	PFN_SG_LAST = (1UL << 1),
> +	PFN_DEV = (1UL << 2),
> +	PFN_MAP = (1UL << 3),
> +};

Please forgive a little bikeshedding here...

Why __pfn_t?  Because the KVM code has a pfn_t?  If so, I think we
should rescue pfn_t from KVM and give them a kvm_pfn_t.

I think you should do one of two things:  Make PFN_SHIFT 12 so that a
physical addr can be stored in a __pfn_t with no work.  Or, use the
*high* 12 bits of __pfn_t.val.

If you use the high bits, *and* make it store a plain pfn when all the
bits are 0, then you get a zero-cost pfn<->__pfn_t conversion which will
hopefully generate the exact same code which is there today.

The one disadvantage here is that it makes it more likely that somebody
that's just setting __pfn_t.val=foo will get things subtly wrong
somehow, but that it will work most of the time.

Also, about naming...  PFN_SHIFT is pretty awful name for this.  It
probably needs to be __PFN_T_SOMETHING.  We don't want folks doing
craziness like:

	unsigned long phys_addr = pfn << PFN_SHIFT.

Which *looks* OK.