Re: [rcu:from-joel.2019.08.16a 143/172] kernel/rcu/tree.c:2808:6: note: in expansion of macro 'xchg'

[Paul Walmsley <paul.walmsley@sifive.com>]

On Fri, 16 Aug 2019, Joel Fernandes wrote:

> On Sat, Aug 17, 2019 at 05:10:59AM +0800, kbuild test robot wrote:
> > tree:   https://kernel.googlesource.com/pub/scm/linux/kernel/git/paulmck/linux-rcu.git from-joel.2019.08.16a
> > head:   01b0e4d3e0ac279b295bc06a3591f0b810b9908f
> > commit: bda80ba9decc7a32413e88d2f070de180c4b76ab [143/172] rcu/tree: Add basic support for kfree_rcu() batching
> > config: riscv-defconfig (attached as .config)
> > compiler: riscv64-linux-gcc (GCC) 7.4.0
> > reproduce:
> >         wget https://raw.githubusercontent.com/intel/lkp-tests/master/sbin/make.cross -O ~/bin/make.cross
> >         chmod +x ~/bin/make.cross
> >         git checkout bda80ba9decc7a32413e88d2f070de180c4b76ab
> >         # save the attached .config to linux build tree
> >         GCC_VERSION=7.4.0 make.cross ARCH=riscv 
> 
> This seems to be a riscv issue:
> 
> A call to '__compiletime_assert_2792' declared with attribute error:
> BUILD_BUG failed
> 
> Could riscv folks take a look at it? Why is using xchg() causing issues? The
> xchg() is being done on a bool.

sizeof(bool) = 1, and the only xchg() implementation that's currently 
present on all RISC-V in Linux is usable only on 32-bit types.  SPARC, 
Microblaze, C-SKY, and Hexagon all look like they have the same 
limitation.  You'll probably see similar build breakage for those 
platforms too.

MIPS, OpenRISC, Xtensa, and SuperH have compatibility functions for the 
8-bit and 16-bit cases.  We could add those for RISC-V.  However, they'll 
be slower than the 4- and 8-byte cases.  This is both because there's more 
code involved, and because there's an increased risk of false-sharing.

Can you use a u32 instead for xchg/cmpxchg operations like struct 
kfree_rcu_cpu.monitor_todo that are intended to be cross-platform?  Based 
on a cursory glance through the tree, it looks like there's only one 
cross-platform driver that uses 'true' or 'false' with the kernel 
xchg/cmpxchg, and it uses an int as the underlying data type.

I suppose one could typedef "xchg_bool" on a per-architecture basis, to be 
smallest type that makes sense to use in an xchg() or cmpxchg().  That 
might save some memory on architectures that have a fast 1-byte 
xchg()/cmpxchg().  But given the lack of users in the tree, I'm not sure 
it's worth the effort.


- Paul