[PATCH v2 2/2] treewide: Add the __GFP_PACKED flag to several non-DMA kmalloc() allocations

[Catalin Marinas]

On Wed, Oct 26, 2022 at 10:46:46AM -0700, Linus Torvalds wrote:
> I think we should just stop bending over backwards over this, and say
> "if your DMA isn't coherent, it's on your driver to mark its
> allocations".
[...]
> That hardware may then be one of the one-off strange cases, but those
> people with their masochistic tendencies can take the pain of "oh, now
> I need to mark my broken driver with dma_alloc()".

The driver is not necessarily broken. The same small kmalloc() in a USB
driver can work fine on a fully coherent platform but if that chip ends
up on a SoC that doesn't support coherent DMA, it needs bigger kmalloc()
alignment. The driver could check if it's coherent but that's more of an
arch detail that the driver shouldn't care about. If we define a new API
like dma_alloc() and drivers don't use it, that's when we can claim they
are broken.

A further optimisation would be for dma_alloc() to take a struct device
pointer and check dev_is_dma_coherent() before deciding to align the
size, though this doesn't work when the allocation place cannot tell the
destination device (e.g. alloc_skb(), though these buffers are
cacheline-aligned already).

Reading up on coccinelle to see if I can make this transition easier. If
not, I'll probably go back to bouncing.

-- 
Catalin