Re: [PATCH] Fix the j-core SOC build.

Hello,

On Wed, Nov 10, 2021 at 07:53:13PM -0500, Rich Felker wrote:

On Wed, Nov 10, 2021 at 04:33:46PM -0500, Rich Felker wrote:

quoted

On Sat, Oct 02, 2021 at 10:18:51PM -0400, Rich Felker wrote:

quoted

On Sat, Oct 02, 2021 at 02:32:15PM -0500, Rob Landley wrote:

quoted

From: Rob Landley <redacted>

Commit b67177ecd956 broke the j-core SOC build with a link failure, because
mm/percpu.c function pcpu_post_unmap_tlb_flush() calls flush_tlb_kernel_range()
which is defined under #ifdef CONFIG_MMU.

Signed-off-by: Rob Landley <redacted>
---

 arch/sh/kernel/smp.c |    5 +++++
 1 file changed, 5 insertions(+)

diff --git a/arch/sh/kernel/smp.c b/arch/sh/kernel/smp.c
index 65924d9ec245..3ec8f32aad85 100644
--- a/arch/sh/kernel/smp.c
+++ b/arch/sh/kernel/smp.c

@@ -468,4 +468,9 @@ void flush_tlb_one(unsigned long asid, unsigned long vaddr)
 	local_flush_tlb_one(asid, vaddr);
 }

+#else
+void flush_tlb_kernel_range(unsigned long start, unsigned long end)
+{
+	local_flush_tlb_all();
+}
 #endif

local_flush_tlb_all() is defined in arch/sh/mm/nommu.c as BUG(); so
this is most likely wrong unless it just doesn't get called. I think
there should probably be something at a very general level dummying
out these functions/macros on nommu but I don't know where it should
be.

I've looked into this some more, and while arch/arm does dummy its
equivalent functions out on nommu, arch/sh has always had them as
BUG(), and indeed it makes some sense to catch erroneous usage.
pcpu_post_unmap_tlb_flush should probably have the flush under #ifdef
CONFIG_MMU or something.

I've added the author of the commit that broke this to Cc in case he
has any thoughts.

I think we actually have the wrong commit to blame; it looks like
93274f1dd6b0 ("percpu: flush tlb in pcpu_reclaim_populated()") added
the TLB flush. However, this rabbit hole seems to go a lot deeper.
According to the comments in the files, percpu-km.c, not percpu-vm.c,
is supposed to be used on nommu archs. However, mm/Kconfig has
NEED_PER_CPU_KM depending on !SMP rather than !SMP || !MMU, which
mismatches the comments in the source. So should we be trying to fix
the Kconfig constraints to use -km? Or fixing -vm to work on nommu
like it used to?

Rich

I'm surprised I haven't heard any reports of this until now. Thanks for
ccing me. Fwiw, it's generally good to cc the author / subsystem
maintainer when there is an issue.

I would need to think about it a little bit more, but I think
percpu-km.c was written more in the mindset of !MMU == !SMP. It sounds
like for superh that's not true.

Given that, if sh has no MMU, then it really should be using percpu-km.c
not percpu-vm.c as there is no notion of unpopulated pages for that
configuration. I think we should fix the Kconfig constraints and give
that a shot. I didn't have any known users of !mmu/percpu-km.c, so I
didn't want to make the changes without someone verifying it or needing
it.

Thanks,
Dennis