Hi Christoph,
Am Montag, 20. Juni 2022, 08:16:07 CEST schrieb Christoph Hellwig:
> On Sun, Jun 19, 2022 at 10:32:11PM +0200, Heiko Stuebner wrote:
> > +#ifdef CONFIG_RISCV_DMA_NONCOHERENT
> > +#define ARCH_DMA_MINALIGN L1_CACHE_BYTES
> > +#endif
>
> This needs to be greater or equal to riscv_cbom_block_size, but the
> core code requires a compile time constant here. So we'll need a big
> fat comment here, and panic if riscv_cbom_block_size is >
> L1_CACHE_BYTES/ARCH_DMA_MINALIGN in the code that queries
> riscv_cbom_block_size.
ARM people also had this nice WARN_TAINT to warn when the similar
case happens on ARM64 and the ARCH_DMA_MINALIGN is smaller than
the register value so I've added a similar mechanism.
I've read numerous mails from Torvalds over time that panic-ing should
only ever be the very very last resort, so that WARN_TAINT looks like
a less drastic option while still generating that big warning to users.
> Note that the arm64 folks are looking into making this variable or
> killing it off in this current form, so things might be getting better
> soon.
>
> > +void arch_sync_dma_for_device(phys_addr_t paddr, size_t size,
> > + enum dma_data_direction dir)
> > +{
> > + void *vaddr = phys_to_virt(paddr);
> > +
> > + switch (dir) {
> > + case DMA_TO_DEVICE:
> > + ALT_CMO_OP(clean, vaddr, size, riscv_cbom_block_size);
> > + break;
> > + case DMA_FROM_DEVICE:
> > + ALT_CMO_OP(inval, vaddr, size, riscv_cbom_block_size);
> > + break;
>
> For this also see:
>
> https://lore.kernel.org/all/20220606152150.GA31568@willie-the-truck/
>
> and
>
> https://lore.kernel.org/linux-arm-kernel/20220610151228.4562-1-will@xxxxxxxxxx/T/
so from that discussion, it looks like a "clean" should happen here to
prevent stale bytes (not written to by the dma transfer itself) in the
buffer area I guess.
I'll give that a spin :-)
> > +void arch_setup_dma_ops(struct device *dev, u64 dma_base, u64 size,
> > + const struct iommu_ops *iommu, bool coherent)
> > +{
> > + dev->dma_coherent = coherent;
> > +}
>
> This probably wants a sanity check warn if coherent if false without
> any support for cache flushing as that will cause data corruption.
I've added a riscv_noncoherent_supported() call that will track that
"somebody" implemented non-coherence functionality from their
setup function (zicbom_probe, thead_errata-probe) and a matching
second WARN_TAINT in arch_setup_dma_ops() when coherent value
and availability of non-coherence handling is not matched.
Heiko
