Re: [RFC v7 12/22] cpu: Introduce TCGCpuOperations struct

[Eduardo Habkost <ehabkost@redhat.com>]

On Fri, Dec 04, 2020 at 07:07:09PM +0100, Claudio Fontana wrote:
> On 12/4/20 7:04 PM, Claudio Fontana wrote:
> > On 12/4/20 6:28 PM, Eduardo Habkost wrote:
> >> On Fri, Dec 04, 2020 at 06:10:49PM +0100, Philippe Mathieu-Daudé wrote:
> >>> On 11/30/20 3:35 AM, Claudio Fontana wrote:
> >>>> From: Eduardo Habkost <ehabkost@redhat.com>
> >>>>
> >>>> The TCG-specific CPU methods will be moved to a separate struct,
> >>>> to make it easier to move accel-specific code outside generic CPU
> >>>> code in the future.  Start by moving tcg_initialize().
> >>>
> >>> Good idea! One minor comment below.
> >>>
> >>>>
> >>>> The new CPUClass.tcg_opts field may eventually become a pointer,
> >>>> but keep it an embedded struct for now, to make code conversion
> >>>> easier.
> >>>>
> >>>> Signed-off-by: Eduardo Habkost <ehabkost@redhat.com>
> >>>> ---
> >>>>  MAINTAINERS                     |  1 +
> >>>>  cpu.c                           |  2 +-
> >>>>  include/hw/core/cpu.h           |  9 ++++++++-
> >>>>  include/hw/core/tcg-cpu-ops.h   | 25 +++++++++++++++++++++++++
> >>>>  target/alpha/cpu.c              |  2 +-
> >>>>  target/arm/cpu.c                |  2 +-
> >>>>  target/avr/cpu.c                |  2 +-
> >>>>  target/cris/cpu.c               | 12 ++++++------
> >>>>  target/hppa/cpu.c               |  2 +-
> >>>>  target/i386/tcg-cpu.c           |  2 +-
> >>>>  target/lm32/cpu.c               |  2 +-
> >>>>  target/m68k/cpu.c               |  2 +-
> >>>>  target/microblaze/cpu.c         |  2 +-
> >>>>  target/mips/cpu.c               |  2 +-
> >>>>  target/moxie/cpu.c              |  2 +-
> >>>>  target/nios2/cpu.c              |  2 +-
> >>>>  target/openrisc/cpu.c           |  2 +-
> >>>>  target/ppc/translate_init.c.inc |  2 +-
> >>>>  target/riscv/cpu.c              |  2 +-
> >>>>  target/rx/cpu.c                 |  2 +-
> >>>>  target/s390x/cpu.c              |  2 +-
> >>>>  target/sh4/cpu.c                |  2 +-
> >>>>  target/sparc/cpu.c              |  2 +-
> >>>>  target/tilegx/cpu.c             |  2 +-
> >>>>  target/tricore/cpu.c            |  2 +-
> >>>>  target/unicore32/cpu.c          |  2 +-
> >>>>  target/xtensa/cpu.c             |  2 +-
> >>>>  27 files changed, 63 insertions(+), 30 deletions(-)
> >>>>  create mode 100644 include/hw/core/tcg-cpu-ops.h
> >>>>
> >>>> diff --git a/MAINTAINERS b/MAINTAINERS
> >>>> index f53f2678d8..d876f504a6 100644
> >>>> --- a/MAINTAINERS
> >>>> +++ b/MAINTAINERS
> >>>> @@ -1535,6 +1535,7 @@ F: qapi/machine.json
> >>>>  F: qapi/machine-target.json
> >>>>  F: include/hw/boards.h
> >>>>  F: include/hw/core/cpu.h
> >>>> +F: include/hw/core/tcg-cpu-ops.h
> >>>>  F: include/hw/cpu/cluster.h
> >>>>  F: include/sysemu/numa.h
> >>>>  T: git https://github.com/ehabkost/qemu.git machine-next
> >>>> diff --git a/cpu.c b/cpu.c
> >>>> index 0be5dcb6f3..d02c2a17f1 100644
> >>>> --- a/cpu.c
> >>>> +++ b/cpu.c
> >>>> @@ -180,7 +180,7 @@ void cpu_exec_realizefn(CPUState *cpu, Error **errp)
> >>>>  
> >>>>      if (tcg_enabled() && !tcg_target_initialized) {
> >>>>          tcg_target_initialized = true;
> >>>> -        cc->tcg_initialize();
> >>>> +        cc->tcg_ops.initialize();
> >>>>      }
> >>>>      tlb_init(cpu);
> >>>>  
> >>>> diff --git a/include/hw/core/cpu.h b/include/hw/core/cpu.h
> >>>> index 3d92c967ff..c93b08a0fb 100644
> >>>> --- a/include/hw/core/cpu.h
> >>>> +++ b/include/hw/core/cpu.h
> >>>> @@ -76,6 +76,10 @@ typedef struct CPUWatchpoint CPUWatchpoint;
> >>>>  
> >>>>  struct TranslationBlock;
> >>>>  
> >>>> +#ifdef CONFIG_TCG
> >>>> +#include "tcg-cpu-ops.h"
> >>>> +#endif /* CONFIG_TCG */
> >>>> +
> >>>>  /**
> >>>>   * CPUClass:
> >>>>   * @class_by_name: Callback to map -cpu command line model name to an
> >>>> @@ -221,12 +225,15 @@ struct CPUClass {
> >>>>  
> >>>>      void (*disas_set_info)(CPUState *cpu, disassemble_info *info);
> >>>>      vaddr (*adjust_watchpoint_address)(CPUState *cpu, vaddr addr, int len);
> >>>> -    void (*tcg_initialize)(void);
> >>>>  
> >>>>      const char *deprecation_note;
> >>>>      /* Keep non-pointer data at the end to minimize holes.  */
> >>>>      int gdb_num_core_regs;
> >>>>      bool gdb_stop_before_watchpoint;
> >>>> +
> >>>> +#ifdef CONFIG_TCG
> >>>> +    TcgCpuOperations tcg_ops;
> >>>> +#endif /* CONFIG_TCG */
> >>>>  };
> >>
> >> I'm not a fan of #ifdefs in struct definitions (especially in
> >> generic code like hw/cpu), because there's risk the same header
> >> generate different struct layout when used by different .c files.
> >> I would prefer to gradually refactor the code so that tcg_ops is
> >> eventually removed from CPUClass.
> >>
> >> This is not a dealbreaker, because both approaches are steps in
> >> the same direction.  But the #ifdef here makes review harder and
> >> has more risks of unwanted side effects.
> >>
> >>>>  
> >>>>  /*
> >>>> diff --git a/include/hw/core/tcg-cpu-ops.h b/include/hw/core/tcg-cpu-ops.h
> >>>> new file mode 100644
> >>>> index 0000000000..4475ef0996
> >>>> --- /dev/null
> >>>> +++ b/include/hw/core/tcg-cpu-ops.h
> >>>> @@ -0,0 +1,25 @@
> >>>> +/*
> >>>> + * TCG-Specific operations that are not meaningful for hardware accelerators
> >>>> + *
> >>>> + * Copyright 2020 SUSE LLC
> >>>> + *
> >>>> + * This work is licensed under the terms of the GNU GPL, version 2 or later.
> >>>> + * See the COPYING file in the top-level directory.
> >>>> + */
> >>>> +
> >>>> +#ifndef TCG_CPU_OPS_H
> >>>> +#define TCG_CPU_OPS_H
> >>>> +
> >>>> +/**
> >>>> + * struct TcgCpuOperations: TCG operations specific to a CPU class
> >>>> + */
> >>>> +typedef struct TcgCpuOperations {
> >>>> +    /**
> >>>> +     * @initialize: Initalize TCG state
> >>>> +     *
> >>>> +     * Called when the first CPU is realized.
> >>>> +     */
> >>>> +    void (*initialize)(void);
> >>>> +} TcgCpuOperations;
> >>>> +
> >>>> +#endif /* TCG_CPU_OPS_H */
> >>> ...
> >>>
> >>>> diff --git a/target/arm/cpu.c b/target/arm/cpu.c
> >>>> index 07492e9f9a..1fa9382a7c 100644
> >>>> --- a/target/arm/cpu.c
> >>>> +++ b/target/arm/cpu.c
> >>>> @@ -2261,7 +2261,7 @@ static void arm_cpu_class_init(ObjectClass *oc, void *data)
> >>>>      cc->gdb_stop_before_watchpoint = true;
> >>>>      cc->disas_set_info = arm_disas_set_info;
> >>>>  #ifdef CONFIG_TCG
> >>>> -    cc->tcg_initialize = arm_translate_init;
> >>>> +    cc->tcg_ops.initialize = arm_translate_init;
> >>>
> >>> This one is correctly guarded by '#ifdef CONFIG_TCG'.
> >>>
> >>> For the other targets, can you either place it within
> >>> the '#ifdef CONFIG_TCG' block or if there is none, add
> >>> one please?
> >>
> >> As a new #ifdef risks having additional unwanted side effects, I
> >> would prefer to do it in separate patch, just in case.
> >>
> >> This also applies to the #ifdef Claudio added to hw/core/cpu.h
> >> above.  In case we really want to do it, I would do it in a
> >> separate patch.
> > 
> > Hi Eduardo,
> > 
> > yes, once things are settling we should dispose of as many #ifdefs are possible
> 
> 
> By that I mean, as targets are adapted (arm, s390, ...) things can be refactored in a similar was as for x86,
> so that the ifdefs disappear, and meson instead controls which pieces are built.
> 
> When it comes to the extra field, it was already very useful to have it #ifdef ed,
> because it identified quite a few problem with the existing patches, ie, code that was trying to use the field even though it was not there.
> 
> So this step helps with finding all the right pieces to refactor away,
> and then once things are in their proper place, we can take away the ifdef I think.

That's a good point.  Grepping for CONFIG_TCG will help us find
out cases where the code still has to be moved to a separate
file.

I was probably being overly cautious about the ifdef.  CONFIG_TCG
is not as tricky as target-specific macros that can't be used in
any header.

-- 
Eduardo