Re: [PATCH v6 11/15] software node: move small properties inline when copying

[Dmitry Torokhov]

On Fri, Nov 08, 2019 at 02:34:48AM +0100, Rafael J. Wysocki wrote:
> On Friday, November 8, 2019 1:49:46 AM CET Dmitry Torokhov wrote:
> > On Fri, Nov 08, 2019 at 01:45:03AM +0100, Rafael J. Wysocki wrote:
> > > On Friday, November 8, 2019 1:28:44 AM CET Dmitry Torokhov wrote:
> > > > On Fri, Nov 08, 2019 at 01:04:31AM +0100, Rafael J. Wysocki wrote:
> > > > > On Wednesday, November 6, 2019 12:56:56 AM CET Dmitry Torokhov wrote:
> > > > > > Hi Rafael,
> > > > > > 
> > > > > > On Wed, Nov 06, 2019 at 12:42:02AM +0100, Rafael J. Wysocki wrote:
> > > > > > > On Wednesday, October 23, 2019 10:02:29 PM CET Dmitry Torokhov wrote:
> > > > > > > > When copying/duplicating set of properties, move smaller properties that
> > > > > > > > were stored separately directly inside property entry structures. We can
> > > > > > > > move:
> > > > > > > > 
> > > > > > > > - up to 8 bytes from U8 arrays
> > > > > > > > - up to 4 words
> > > > > > > > - up to 2 double words
> > > > > > > > - one U64 value
> > > > > > > > - one or 2 strings.
> > > > > > > 
> > > > > > > Yes, we can do that, but how much of a difference does this really make?
> > > > > > 
> > > > > > Arguably not much I think, but it was pretty cheap to do.
> > > > > > 
> > > > > > > 
> > > > > > > Also, how can one distinguish between a single-value property and an inline
> > > > > > > array which this change?  By looking at the length?
> > > > > > 
> > > > > > We do not really need to distinguish between the 2. The device
> > > > > > properties API is typically wrap single values around arrays (i.e. it is
> > > > > > perfectly fine to use scalar API to fetch first element of array and use
> > > > > > array API to fetch a scalar). So we have property of certain type with
> > > > > > certain number of elements, and it can either be stored inside
> > > > > > property_entry structure, or outside of it. They are 2 orthogonal
> > > > > > concepts.
> > > > > > 
> > > > > > > 
> > > > > > > > Signed-off-by: Dmitry Torokhov <dmitry.torokhov@xxxxxxxxx>
> > > > > > > > ---
> > > > > > > >  drivers/base/swnode.c | 10 ++++++++++
> > > > > > > >  1 file changed, 10 insertions(+)
> > > > > > > > 
> > > > > > > > diff --git a/drivers/base/swnode.c b/drivers/base/swnode.c
> > > > > > > > index 18a30fb3cc58..49e1108aa4b7 100644
> > > > > > > > --- a/drivers/base/swnode.c
> > > > > > > > +++ b/drivers/base/swnode.c
> > > > > > > > @@ -280,6 +280,16 @@ static int property_entry_copy_data(struct property_entry *dst,
> > > > > > > >  	if (!dst->name)
> > > > > > > >  		goto out_free_data;
> > > > > > > >  
> > > > > > > > +	if (!dst->is_inline && dst->length <= sizeof(dst->value)) {
> > > > > > > > +		/* We have an opportunity to move the data inline */
> > > > > > > > +		const void *tmp = dst->pointer;
> > > > > > > > +
> > > > > > > > +		memcpy(&dst->value, tmp, dst->length);
> > > > > > > > +		dst->is_inline = true;
> > > > > > > > +
> > > > > > > > +		kfree(tmp);
> > > > > > > 
> > > > > > > This would have been more useful if we had been able to avoid making the
> > > > > > > allocation altogether.
> > > > > > 
> > > > > > OK, I can do that and re-send this patch and the one with the tests.
> > > > > 
> > > > > But if you do that, IMO it would be prudent to extend the definition of
> > > > > struct property_entry like this:
> > > > > 
> > > > >  struct property_entry {
> > > > >  	const char *name;
> > > > >  	size_t length;
> > > > >  	bool is_array;
> > > > >  	enum dev_prop_type type;
> > > > >  	union {
> > > > >  		union {
> > > > >  			const u8 *u8_data;
> > > > >  			const u16 *u16_data;
> > > > >  			const u32 *u32_data;
> > > > >  			const u64 *u64_data;
> > > > >  			const char * const *str;
> > > > >  		} pointer;
> > > > >  		union {
> > > > >  			u8 u8_data;
> > > > >  			u16 u16_data;
> > > > >  			u32 u32_data;
> > > > >  			u64 u64_data;
> > > > >  			const char *str;
> > > > > +			u8 u8_buf[sizeof(u64)];
> > > > > +			u16 u16_buf[sizeof(u64)/sizeof(u16)];
> > > > > +			u32 u32_buf[sizeof(u64)/sizeof(u32)];
> > > > > +			char char_buf[sizeof(u64)];
> > > > >  		} value;
> > > > >  	};
> > > > >  };
> > > > > 
> > > > > to make it clear that the value field is going to be used as an array in
> > > > > some cases.
> > > > 
> > > > Sorry, just sent out updated series before receiving your email. I can
> > > > cook up new patch cleaning this.
> > > 
> > > I'd prefer a new version of the series, honestly.
> > 
> > OK, sure.
> > 
> > > 
> > > > I think we can drop scalars and only have arrays and have initializers use
> > > > <type>_data[0] to create initial property entries.
> > > 
> > > Why [0]?  IMO it is better to use the exact size (which is known) in this
> > > particular case.
> > 
> > diff --git a/include/linux/property.h b/include/linux/property.h
> > index b315fdc0ec28d..b28c81af7bb68 100644
> > --- a/include/linux/property.h
> > +++ b/include/linux/property.h
> > @@ -257,11 +257,11 @@ struct property_entry {
> >         union {
> >                 const void *pointer;
> >                 union {
> > -                       u8 u8_data;
> > -                       u16 u16_data;
> > -                       u32 u32_data;
> > -                       u64 u64_data;
> > -                       const char *str;
> > +                       u8 u8_data[sizeof(u64) / sizeof(u8)];
> > +                       u16 u16_data[sizeof(u64) / sizeof(u16)];
> > +                       u32 u32_data[sizeof(u64) / sizeof(u32)];
> > +                       u64 u64_data[sizeof(u64) / sizeof(u64)];
> 
> IMO with a scalar u64 this kind of would explain itself, but with a u64 array
> it becomes somewhat confusing.
> 
> > +                       const char *str[sizeof(u64) / sizeof(char *)];
> >                 } value;
> >         };
> >  };
> > @@ -273,7 +273,7 @@ struct property_entry {
> >   */
> > 
> >  #define __PROPERTY_ENTRY_ELEMENT_SIZE(_elem_)                          \
> > -       sizeof(((struct property_entry *)NULL)->value._elem_)
> > +       sizeof(((struct property_entry *)NULL)->value._elem_[0])
> > 
> >  #define __PROPERTY_ENTRY_ARRAY_ELSIZE_LEN(_name_, _elsize_, _Type_,    \
> >                                           _val_, _len_)                 \
> > @@ -323,7 +323,7 @@ struct property_entry {
> >         .length = __PROPERTY_ENTRY_ELEMENT_SIZE(_elem_),                \
> >         .is_inline = true,                                              \
> >         .type = DEV_PROP_##_Type_,                                      \
> > -       { .value = { ._elem_ = _val_ } },                               \
> > +       { .value = { ._elem_[0] = _val_ } },                            \
> >  }
> > 
> >  #define PROPERTY_ENTRY_U8(_name_, _val_)                               \
> > 
> > > 
> > > Also note that u64 is naturally a scalar only.
> > 
> > It still can be expressed as array of 1 element.
> 
> It can, but for what purpose?

Just so we do not have to special-case handling of U64 in
PROPERTY_ENTRY_Unnn() macros, as I shown in the snippet above.

-- 
Dmitry