On Wed, Apr 14, 2021 at 10:10:44AM +0200, Jesper Dangaard Brouer wrote:

quoted

Yes, indeed! - And very frustrating.  It's keeping me up at night.
I'm dreaming about 32 vs 64 bit data structures. My fitbit stats tell
me that I don't sleep well with these kind of dreams ;-)

Then you're going to love this ... even with the latest patch, there's
still a problem.  Because dma_addr_t is still 64-bit aligned _as a type_,
that forces the union to be 64-bit aligned (as we already knew and worked
around), but what I'd forgotten is that forces the entirety of struct
page to be 64-bit aligned.  Which means ...

        /* size: 40, cachelines: 1, members: 4 */
        /* padding: 4 */
        /* forced alignments: 1 */
        /* last cacheline: 40 bytes */
} __attribute__((__aligned__(8)));

.. that we still have a hole!  It's just moved from being at offset 4
to being at offset 36.

quoted

That said, I think we need to have a quicker fix for the immediate
issue with 64-bit bit dma_addr on 32-bit arch and the misalignment hole
it leaves[3] in struct page.  In[3] you mention ppc32, does it only
happens on certain 32-bit archs?

AFAICT it happens on mips32, ppc32, arm32 and arc.  It doesn't happen
on x86-32 because dma_addr_t is 32-bit aligned.

Doing this fixes it:

+++ b/include/linux/types.h

@@ -140,7 +140,7 @@ typedef u64 blkcnt_t;
  * so they don't care about the size of the actual bus addresses.
  */
 #ifdef CONFIG_ARCH_DMA_ADDR_T_64BIT
-typedef u64 dma_addr_t;
+typedef u64 __attribute__((aligned(sizeof(void *)))) dma_addr_t;
 #else
 typedef u32 dma_addr_t;
 #endif

quoted

I'm seriously considering removing page_pool's support for doing/keeping
DMA-mappings on 32-bit arch's.  AFAIK only a single driver use this.

... if you're going to do that, then we don't need to do this.