On Tue, 5 May 2020 at 16:04, Arnd Bergmann [off-list ref] wrote:

With the kcsan changes, __read_once_size() is not inlined, but
clang can decide to emit a version that hardcodes the address, which
in turn triggers a warning for dss_conv_list being __initdata but
__read_once_size() not being __init:

WARNING: modpost: vmlinux.o(.text+0x6e4d7a): Section mismatch in
reference from the function __read_once_size() to the variable
.init.data:dss_conv_list
The function __read_once_size() references
the variable __initdata dss_conv_list.
This is often because __read_once_size lacks a __initdata
annotation or the annotation of dss_conv_list is wrong.

This is clearly a false positive warning, but it's hard to tell
who is to blame for it. Work around it by removing the __initdata
annotation, wasting the space of two pointers in return for getting
rid of the warning.

Fixes: dfd402a4c4ba ("kcsan: Add Kernel Concurrency Sanitizer infrastructure")
Signed-off-by: Arnd Bergmann <arnd@arndb.de>

[+Cc Will]

I think Will is working on a series that completely overhauls
READ_ONCE/WRITE_ONCE, also getting rid of __read_once_size() in the
process, which would make this patch redundant. If we can live with
this warning until the new READ_ONCE/WRITE_ONCE gets merged, we can
probably keep things as-is for now.

Thanks,
-- Marco