Re: [PATCH 9/9] um: Add initial SMP support

[Johannes Berg <johannes@sipsolutions.net>]

On Tue, 2025-07-29 at 23:06 +0800, Tiwei Bie wrote:
> On Mon, 28 Jul 2025 18:27:53 +0200, Johannes Berg wrote:
> > On Tue, 2025-07-29 at 00:04 +0800, Tiwei Bie wrote:
> > > > > +++ b/arch/um/include/asm/spinlock.h
> > > > > @@ -0,0 +1,8 @@
> > > > > +/* SPDX-License-Identifier: GPL-2.0 */
> > > > > +#ifndef __ASM_UM_SPINLOCK_H
> > > > > +#define __ASM_UM_SPINLOCK_H
> > > > > +
> > > > > +#include <asm/processor.h>
> > > > > +#include <asm-generic/spinlock.h>
> > > > > +
> > > > > +#endif /* __ASM_UM_SPINLOCK_H */
> > > > 
> > > > Do we need this file? Maybe asm-generic should be including the right
> > > > things it needs?
> > > 
> > > I added this file to include asm/processor.h; otherwise, there would be
> > > a lot of compilation errors. Other architectures seem to do the same:
> > > 
> > > $ grep -r asm/processor.h arch/ | grep asm/spinlock.h
> > > arch/arm/include/asm/spinlock.h:#include <asm/processor.h>
> > > arch/alpha/include/asm/spinlock.h:#include <asm/processor.h>
> > > arch/arc/include/asm/spinlock.h:#include <asm/processor.h>
> > > arch/hexagon/include/asm/spinlock.h:#include <asm/processor.h>
> > > arch/parisc/include/asm/spinlock.h:#include <asm/processor.h>
> > > arch/x86/include/asm/spinlock.h:#include <asm/processor.h>
> > > arch/s390/include/asm/spinlock.h:#include <asm/processor.h>
> > > arch/mips/include/asm/spinlock.h:#include <asm/processor.h>
> > > arch/loongarch/include/asm/spinlock.h:#include <asm/processor.h>
> > 
> > Except for loongarch they all do something else too though. Feels to me
> > um (and loongarch) really shouldn't need that file.
> 
> Sorry for the confusion. My point is that since other architectures
> also do this, it seems common practice to include asm/processor.h in
> asm/spinlock.h when necessary.

Yeah, I understand.

> 
> The reason we need to include asm/processor.h in asm/spinlock.h on UML
> is because:
> 
> ticket_spin_lock() (which is an inline function indirectly provided by
> asm-generic/spinlock.h) relies on atomic_cond_read_acquire(), which
> is defined as smp_cond_load_acquire().

Right, but that's not the architecture's "fault".

It seems to me that either spinlock.h should include asm/processor.h for
it, or (at least, but I think less appropriate) asm-generic/spinlock.h
should be doing this.

> On UML, smp_cond_load_acquire() is provided by asm-generic/barrier.h,
> and it relies on smp_cond_load_relaxed(), which is also provided by
> asm-generic/barrier.h on UML. And smp_cond_load_relaxed() is a macro
> that relies on cpu_relax(), which is provided by asm/processor.h.

In general though, there ought to be some definition of which header
file(s) is/are expected to provide smp_cond_load_acquire() and/or
atomic_cond_read_acquire(). And that header file/those header files
should be included by the files that use the functions/macros.


IOW, I think you've stumbled across an inconsistency in the generic
files, and hence we should fix that, rather than having each
architecture paper over it.


johannes