Re: [PATCH v2] dma-coherent: add support for multi coherent rmems per dev

[Andy Shevchenko]

On Thu, Feb 08, 2024 at 03:53:37PM +0800, Howard Yen wrote:
> On Tue, Feb 6, 2024 at 11:43 PM Andy Shevchenko
> <andriy.shevchenko@xxxxxxxxxxxxxxx> wrote:
> > On Mon, Feb 05, 2024 at 02:08:00PM +0200, Andy Shevchenko wrote:
> > > On Mon, Feb 05, 2024 at 07:23:00AM +0000, Howard Yen wrote:

..

> > > > @@ -18,15 +18,9 @@ struct dma_coherent_mem {
> > > >     unsigned long   *bitmap;
> > > >     spinlock_t      spinlock;
> > > >     bool            use_dev_dma_pfn_offset;
> > > > +   struct list_head        node;
> > >
> > > Have you run `pahole`? Here I see wasted bytes for nothing.
> >
> > On top of that one may make container_of() to be no-op, by placing this member
> > to be the first one. But, double check this with bloat-o-meter (that it indeed
> > does better code generation) and on the other hand check if the current first
> > member is not performance critical and having additional pointer arithmetics is
> > okay.
> >
> > > >  };
> 
> I'm trying to re-org the members as below
> 
> from ===>
> 
> struct dma_coherent_mem {
> void *                     virt_base;            /*     0     8 */
> dma_addr_t                 device_base;          /*     8     8 */
> unsigned long              pfn_base;             /*    16     8 */
> int                        size;                 /*    24     4 */
> 
> /* XXX 4 bytes hole, try to pack */
> 
> unsigned long *            bitmap;               /*    32     8 */
> spinlock_t                 spinlock;             /*    40     4 */
> bool                       use_dev_dma_pfn_offset; /*    44     1 */
> 
> /* XXX 3 bytes hole, try to pack */
> 
> struct list_head           node;                 /*    48    16 */
> 
> /* size: 64, cachelines: 1, members: 8 */
> /* sum members: 57, holes: 2, sum holes: 7 */
> };
> 
> 
> to ===>
> 
> struct dma_coherent_mem {
> struct list_head           node;                 /*     0    16 */
> void *                     virt_base;            /*    16     8 */
> dma_addr_t                 device_base;          /*    24     8 */
> unsigned long              pfn_base;             /*    32     8 */
> int                        size;                 /*    40     4 */
> spinlock_t                 spinlock;             /*    44     4 */
> unsigned long *            bitmap;               /*    48     8 */
> bool                       use_dev_dma_pfn_offset; /*    56     1 */
> 
> /* size: 64, cachelines: 1, members: 8 */
> /* padding: 7 */

Which seems better that above, right?

> };
> 
> Looks like there is about 7 bytes padding at the end of the structure.
> Should I add __attribute__((__packed__)) to not add the padding?

No, __packed is about alignment, may give you much worse code generation.
With list_head member first you might get better code from the original,
check it with bloat-o-meter.

-- 
With Best Regards,
Andy Shevchenko