On Wed, Feb 22, 2023 at 04:20:16PM +0000, Jiaxun Yang wrote:

quoted

2023年2月22日 16:02，Conor Dooley [off-list ref] 写道：
On Wed, Feb 22, 2023 at 03:55:19PM +0000, Jiaxun Yang wrote:

quoted

quoted

2023年2月22日 14:50，Conor Dooley [off-list ref] 写道：
On Wed, Feb 22, 2023 at 01:37:11PM +0000, Jiaxun Yang wrote:

quoted

For riscv our assumption is unless a device states it is non-coherent,
we take it to be DMA coherent.

For devicetree probed devices that have been true since very begining
with OF_DMA_DEFAULT_COHERENT selected.

Signed-off-by: Jiaxun Yang <jiaxun.yang@flygoat.com>
---
arch/riscv/kernel/setup.c | 3 +++
1 file changed, 3 insertions(+)

diff --git a/arch/riscv/kernel/setup.c b/arch/riscv/kernel/setup.c
index 376d2827e736..34b371180976 100644
--- a/arch/riscv/kernel/setup.c
+++ b/arch/riscv/kernel/setup.c

@@ -300,6 +300,9 @@ void __init setup_arch(char **cmdline_p)

riscv_init_cbom_blocksize();
riscv_fill_hwcap();
apply_boot_alternatives();
+#ifdef CONFIG_RISCV_DMA_NONCOHERENT
+ dma_default_coherent = true;
+#endif

Do we really need to add ifdeffery for this here?
It's always coherent by default, so why do we need to say set it in
setup_arch() when we know that, regardless of options, it is true?

Because this symbol is only a variable when:

defined(CONFIG_ARCH_HAS_SYNC_DMA_FOR_DEVICE) || \
defined(CONFIG_ARCH_HAS_SYNC_DMA_FOR_CPU) || \
defined(CONFIG_ARCH_HAS_SYNC_DMA_FOR_CPU_ALL)

Which is only true if  CONFIG_RISCV_DMA_NONCOHERENT is selected.

Otherwise this symbol is defined to true and we can’t make a assignment to it.

quoted

Maybe I am just slow today, but I don't get why you need to add
ifdeffery to setup_arch() to do something that is always true.
Why can't you just set this in riscv/mm/dma-noncoherent.c? What am I
missing?

Hmm that sounds like a good idea but I was unable to find a place for this.

riscv/mm/dma-noncoherent.c are just a bunch of callbacks, there is no initialisation
function that will always run on every boot. riscv_noncoherent_supported is only
called conditionally.

Right, that's fair. riscv_noncoherent_supported() is for when we know we
can support non-coherent DMA, not to detect whether we can, hence the
conditional nature.
Apologies for sending you on a wild goose chase there.

Perhaps I can add a initcall in dma-noncoherent.c but it seems to be a little bit
overkilling.

Yah, probably would be.

Actually I’d prefer to have a config option for the default value but it seems like
it’s not in Christoph’s flavour.

We already have a config option that sets things up nicely for RISC-V
that you're getting rid of! ;)