Re: [Qemu-devel] [PATCH v4 0/5] Add a valid_cpu_types property

[Eduardo Habkost]

On Thu, Jan 11, 2018 at 11:25:08AM +0100, Igor Mammedov wrote:
> On Wed, 10 Jan 2018 19:48:00 -0200
> Eduardo Habkost <address@hidden> wrote:
> 
> > On Wed, Jan 10, 2018 at 01:30:29PM -0800, Alistair Francis wrote:
> > > On Thu, Dec 28, 2017 at 6:59 AM, Eduardo Habkost <address@hidden> wrote:  
> > > > On Thu, Dec 28, 2017 at 02:39:31PM +0100, Igor Mammedov wrote:  
> > > >> On Fri, 22 Dec 2017 11:47:00 -0800
> > > >> Alistair Francis <address@hidden> wrote:
> > > >>  
> > > >> > On Fri, Dec 22, 2017 at 10:45 AM, Alistair Francis
> > > >> > <address@hidden> wrote:  
> > > >> > > On Wed, Dec 20, 2017 at 2:06 PM, Eduardo Habkost <address@hidden> 
> > > >> > > wrote:  
> > > >> > >> On Tue, Dec 19, 2017 at 05:03:59PM -0800, Alistair Francis wrote: 
> > > >> > >>  
> > > >> > >>> On Tue, Dec 19, 2017 at 4:55 PM, Alistair Francis
> > > >> > >>> <address@hidden> wrote:  
> > > >> > >>> > On Tue, Dec 19, 2017 at 4:43 PM, Peter Maydell 
> > > >> > >>> > <address@hidden> wrote:  
> > > >> > >>> >> On 20 December 2017 at 00:27, Alistair Francis
> > > >> > >>> >> <address@hidden> wrote:  
> > > >> > >>> >>> There are numorous QEMU machines that only have a single or 
> > > >> > >>> >>> a handful of
> > > >> > >>> >>> valid CPU options. To simplyfy the management of 
> > > >> > >>> >>> specificying which CPU
> > > >> > >>> >>> is/isn't valid let's create a property that can be set in 
> > > >> > >>> >>> the machine
> > > >> > >>> >>> init. We can then check to see if the user supplied CPU is 
> > > >> > >>> >>> in that list
> > > >> > >>> >>> or not.
> > > >> > >>> >>>
> > > >> > >>> >>> I have added the valid_cpu_types for some ARM machines only 
> > > >> > >>> >>> at the
> > > >> > >>> >>> moment.
> > > >> > >>> >>>
> > > >> > >>> >>> Here is what specifying the CPUs looks like now:
> > > >> > >>> >>>
> > > >> > >>> >>> $ aarch64-softmmu/qemu-system-aarch64 -M netduino2 -kernel 
> > > >> > >>> >>> ./u-boot.elf -nographic -cpu "cortex-m3" -S
> > > >> > >>> >>> QEMU 2.10.50 monitor - type 'help' for more information
> > > >> > >>> >>> (qemu) info cpus
> > > >> > >>> >>> * CPU #0: thread_id=24175
> > > >> > >>> >>> (qemu) q
> > > >> > >>> >>>
> > > >> > >>> >>> $ aarch64-softmmu/qemu-system-aarch64 -M netduino2 -kernel 
> > > >> > >>> >>> ./u-boot.elf -nographic -cpu "cortex-m4" -S
> > > >> > >>> >>> QEMU 2.10.50 monitor - type 'help' for more information
> > > >> > >>> >>> (qemu) q
> > > >> > >>> >>>
> > > >> > >>> >>> $ aarch64-softmmu/qemu-system-aarch64 -M netduino2 -kernel 
> > > >> > >>> >>> ./u-boot.elf -nographic -cpu "cortex-m5" -S
> > > >> > >>> >>> qemu-system-aarch64: unable to find CPU model 'cortex-m5'
> > > >> > >>> >>>
> > > >> > >>> >>> $ aarch64-softmmu/qemu-system-aarch64 -M netduino2 -kernel 
> > > >> > >>> >>> ./u-boot.elf -nographic -cpu "cortex-a9" -S
> > > >> > >>> >>> qemu-system-aarch64: Invalid CPU type: cortex-a9-arm-cpu
> > > >> > >>> >>> The valid types are: cortex-m3-arm-cpu, cortex-m4-arm-cpu  
> > > >> > >>> >>
> > > >> > >>> >> Thanks for this; we really should be more strict about
> > > >> > >>> >> forbidding "won't work" combinations than we have
> > > >> > >>> >> been in the past.
> > > >> > >>> >>
> > > >> > >>> >> In the last of these cases, I think that when we
> > > >> > >>> >> list the invalid CPU type and the valid types
> > > >> > >>> >> we should use the same names we want the user to
> > > >> > >>> >> use on the command line, without the "-arm-cpu"
> > > >> > >>> >> suffixes.  
> > > >> > >>> >
> > > >> > >>> > Hmm... That is a good point, it is confusing that they don't 
> > > >> > >>> > line up.  
> > > >> > >>
> > > >> > >> Agreed.
> > > >> > >>  
> > > >> > >>> >
> > > >> > >>> > The problem is that we are just doing a simple
> > > >> > >>> > object_class_dynamic_cast() in hw/core/machine.c which I think
> > > >> > >>> > (untested) requires us to have the full name in the valid cpu 
> > > >> > >>> > array.  
> > > >> > >> [...]  
> > > >> > >>>
> > > >> > >>> I think an earlier version of my previous series adding the 
> > > >> > >>> support to
> > > >> > >>> machine.c did string comparison, but it was decided to utilise 
> > > >> > >>> objects
> > > >> > >>> instead. One option is to make the array 2 wide and have the 
> > > >> > >>> second
> > > >> > >>> string be user friendly?  
> > > >> > >>
> > > >> > >> Making the array 2-column will duplicate information that we can
> > > >> > >> already find out using other methods, and it won't solve the
> > > >> > >> problem if an entry has a parent class with multiple subclasses
> > > >> > >> (the original reason I suggested object_class_dynamic_cast()).
> > > >> > >>
> > > >> > >> The main obstacle to fix this easily is that we do have a common
> > > >> > >>   ObjectClass *cpu_class_by_name(const char *cpu_model)
> > > >> > >> function, but not a common method to get the model name from a
> > > >> > >> CPUClass.  Implementing this is possible, but probably better to
> > > >> > >> do it after moving the existing arch-specific CPU model
> > > >> > >> enumeration hooks to common code (currently we duplicate lots of
> > > >> > >> CPU enumeration/lookup boilerplate code that we shouldn't have
> > > >> > >> to).
> > > >> > >>
> > > >> > >> Listing only the human-friendly names in the array like in the
> > > >> > >> original patch could be a reasonable temporary solution.  It
> > > >> > >> won't allow us to use a single entry for all subclasses of a
> > > >> > >> given type by now (e.g. listing only TYPE_X86_CPU on PC), but at
> > > >> > >> least we can address this issue without waiting for a refactor of
> > > >> > >> the CPU model enumeration code.  
> > > >> >
> > > >> > Ah, I just re-read this. Do you mean go back to the original RFC and
> > > >> > just use strcmp() to compare the human readable cpu_model?  
> > > >> It's sort of going backwards but I won't object to this as far as you
> > > >> won't use machine->cpu_model (which is in process of being removed)  
> > > 
> > > Wait, machine->cpu_model is the human readable name. Without using
> > > that we can't use just human readable strings for the valid cpu types.  
> > 
> > Well, if we want to deprecate machine->cpu_model we need to offer
> > an alternative first, otherwise we can't prevent people from
> > using it.
> > 
> > Igor, do you see an (existing) alternative to machine->cpu_model
> > that would allow us to avoid using it in
> > machine_run_board_init()?
> In recently merged refactoring machine->cpu_model is being replaced
> by machine->cpu_type. So currently we don't need machine->cpu_model
> anywhere except machine('none'), and once I refactor that it could
> be dropped completely and after some work on *-user targets we can
> practically get rid of cpu_model notion completely
> (excluding of -cpu option parser). 
> 
> My dislike of idea is that it's adding back cpumodel strings
> in boards code again (which I've just got rid of).
> 
> I hate to say that but it looks like we need more refactoring
> for this series to print cpumodels back to user.
> 
> We already have FOO_cpu_list()/FOO_query_cpu_definitions()
> which already do cpu type => cpumodel conversion (and even
> have some code duplication within a target), I'd suggest
> generalizing that across targets and then using generic
> helper for printing back to user converted cpu types from
> mc->valid_cpu_types which this series introduces.

I agree with the long term goal of making cpu type => cpu model
conversion generic.  But until we refactor the arch-specific code
and implement that, we have 2 options:

1) Keep printing a confusing error message until we implement cpu
   type => cpu model conversion;
2) Keep the MachineState::cpu_model field until we implement cpu
   type => cpu model conversion.

I don't see any reason to pick (1) instead of (2).

-- 
Eduardo