Re: [PATCH 0/3] hw/scsi: Cleanup around scsi_bus_legacy_handle_cmdline()

[Paolo Bonzini <pbonzini@redhat.com>]

On 11/22/24 13:37, Thomas Huth wrote:
> On 22/11/2024 12.19, Philippe Mathieu-Daudé wrote:
>> When a device model requires legacy command line handling,
>> call scsi_bus_legacy_handle_cmdline() in its realize handler
>> instead of having each user call it.
>>
>> This applies to:
>>   - spapr_vscsi
>>   - lsi53c810 / lsi53c895a
>>   - sysbus_esp
>>
>> Note, scsi_bus_legacy_handle_cmdline() prototype could be
>> made private to hw/scsi/ to restrict its use to scsi device
>> implementations.
> 
> Not sure whether this is the right way to go ... shouldn't the handling 
> of the legacy command line be rather part of the machine than being part 
> of the SCSI controller device? Imagine for example a machine that has 
> multiple, different SCSI controllers - I think you'd rather want to let 
> the machine decide where the legacy devices should be grabbed from than 
> having the SCSI controller devices fight for them...?

I agree that it should be done in the machines generally:

1) if the machine creates a SCSI controller, it should then call
scsi_bus_legacy_handle_cmdline().  This is the case for esp and for
spapr-vscsi (so spapr_vscsi_create() could be inlined in its caller).

2) lsi53c* is the odd one out because it was "the" way to add "-drive
if=scsi" to the PC machine.

For case (2) it's okay to call it in the realize function, I guess.
However, let's only process -drive if=scsi for devices added on the
command-line.

The LSI HBA should not call  scsi_bus_legacy_handle_cmdline() if
dev->hotplugged.  I think we can do it without a deprecation period,
and in fact assert that !phase_check(PHASE_MACHINE_READY) in
scsi_bus_legacy_handle_cmdline().

Paolo