Re: [PATCH 00/17] powernv: introduce pnv-phb unified devices

[Mark Cave-Ayland]

On 07/05/2022 20:06, Daniel Henrique Barboza wrote:

> Hi,
>
> Since the 7.0.0 release cycle we have a desire to use the powernv
> emulation with libvirt. To do that we need to enable user creatable
> pnv-phb devices to allow more user customization an to avoid spamming
> multiple default devices in the domain XML. In the middle of the
> previous cycle we experimented with user created
> pnv-phb3/pnv-phb3-root-port/pnv-phb4/pnv-phb4-root-port/pnv-phb5. The
> end result, although functional, is that the user needs to deal with a
> lot of versioned devices that, from the user perspective, does the same
> thing. In a way we outsourced the implementation details of the PHBs
> (e.g. pnv8 uses phb3, pnv9 uses phb4) to the command line. Having more
> devices also puts an extra burden in the libvirt support we want to
> have.
>
> To solve this, Cedric and Frederic gave the idea of adding a common
> virtual pnv-phb device that the user can add in the command line, and
> QEMU handles the details internally. Unfortunatelly this idea turned out
> to be way harder than anticipated. Between creating a device that would
> just forward the callbacks to the existing devices internally, creating
> a PnvPHB device with the minimal attributes and making the other devices
> inherit from it, and making an 'abstract' device that could be casted
> for both phb3 and phb4 PHBs,

This bit sounds all good...

> all sorts of very obscure problems occured:
> PHBs not being detected, interrupts not being delivered and memory
> regions not being able to read/write registers. My initial impression is
> that there are assumptions made both in ppc/pnv and hw/pci-host that
> requires the memory region and the bus being in the same device. Even
> if we somehow figure all this out, the resulting code is hacky and
> annoying to maitain.

But this seems really surprising since there are many examples of similar QOM 
patterns within the codebase, and in my experience they work really well. Do you have 
a particular example that you can share to demonstrate why the result of the QOM 
mapping is making things more difficult?

> This brings us to this series. The cleaner way I found to accomplish
> what we want to do is to create real, unified pnv-phb/phb-phb-root-port
> devices, and get rid of all the versioned ones. This is done by
> merging all the PHB3/PHB4 attributes in unified devices. pnv_phb3* and pnv_phb4*
> files end up using the same pnv-phb and phb-phb-root-port unified devices,
> with the difference that pnv_phb3* only cares about version 3 attributes
> and pnv_phb4* only cares about PHB4 attributes. Introducing new PHB
> versions in the future will be a matter of adding any non-existent
> attributes in the unified pnv-phb* devices and using them in separated
> pnv_phbN* files.
>
> The pnv-phb implementation per se is just a router for either phb3 or
> phb4 logic, done in init() and realize() time, after we check which powernv
> machine we're running. If running with powernv8 we redirect control to
> pnv_phb3_realize(), otherwise we redirect to pnv_phb4_realize(). The
> unified device does not do anything per se other than handling control
> to the right version.
>
> After this series, this same powernv8 command line that boots a powernv8
> machine with some phbs and root ports and with network:
>
> ./qemu-system-ppc64 -m 4G \
> -machine powernv8 -smp 2,sockets=2,cores=1,threads=1  \
> -accel tcg,thread=multi -bios skiboot.lid  \
> -kernel vmlinux -initrd buildroot.rootfs.cpio -append 'console=hvc0 ro xmon=on' 
>  \
> -nodefaults  -serial mon:stdio -nographic  \
> -device pnv-phb,chip-id=0,index=0,id=pcie.0 \
> -device pnv-phb,chip-id=0,index=1,id=pcie.1 \
> -device pnv-phb,chip-id=1,index=2,id=pcie.2 \
> -device pnv-phb-root-port,id=root0,bus=pcie.2 \
> -device pnv-phb-root-port,id=root1,bus=pcie.1 \
> -device pcie-pci-bridge,id=bridge1,bus=root0,addr=0x0  \
> -device nvme,bus=bridge1,addr=0x1,drive=drive0,serial=1234  \
> -drive file=./simics-disk.raw,if=none,id=drive0,format=raw,cache=none  \
> -device e1000e,netdev=net0,mac=C0:ff:EE:00:01:04,bus=bridge1,addr=0x3 \
> -netdev bridge,helper=/usr/libexec/qemu-bridge-helper,br=virbr0,id=net0 \
> -device nec-usb-xhci,bus=bridge1,addr=0x2
>
>
> Can be used to boot powernv9 and powernv10 machines with the same attributes
> just by changing the machine type.
>
>
> Daniel Henrique Barboza (17):
>    ppc/pnv: rename PnvPHB3.ioda* to PnvPHB3.ioda2*
>    ppc/pnv: rename PnvPHB3.regs[] to PnvPHB3.regs3[]
>    ppc/pnv: rename PnvPHB3.dma_spaces to PnvPHB3.v3_dma_spaces
>    ppc/pnv: add unified pnv-phb header
>    ppc/pnv: add pnv-phb device
>    ppc/pnv: remove PnvPHB3
>    ppc/pnv: user created pnv-phb for powernv8
>    ppc/pnv: remove PnvPHB4
>    ppc/pnv: user creatable pnv-phb for powernv9
>    ppc/pnv: use PnvPHB.version
>    ppc/pnv: change pnv_phb4_get_pec() to also retrieve chip10->pecs
>    ppc/pnv: user creatable pnv-phb for powernv10
>    ppc/pnv: add pnv_phb_get_current_machine()
>    ppc/pnv: add pnv-phb-root-port device
>    ppc/pnv: remove pnv-phb3-root-port
>    ppc/pnv: remove pnv-phb4-root-port
>    ppc/pnv: remove pecc->rp_model
>
>   hw/pci-host/meson.build        |   3 +-
>   hw/pci-host/pnv_phb.c          | 258 ++++++++++++++++++++++++++++
>   hw/pci-host/pnv_phb3.c         | 256 +++++++++++-----------------
>   hw/pci-host/pnv_phb3_msi.c     |  12 +-
>   hw/pci-host/pnv_phb3_pbcq.c    |   8 +-
>   hw/pci-host/pnv_phb4.c         | 298 ++++++++++++---------------------
>   hw/pci-host/pnv_phb4_pec.c     |  14 +-
>   hw/ppc/pnv.c                   |  41 ++++-
>   include/hw/pci-host/pnv_phb.h  | 224 +++++++++++++++++++++++++
>   include/hw/pci-host/pnv_phb3.h | 118 +------------
>   include/hw/pci-host/pnv_phb4.h | 108 ++----------
>   include/hw/ppc/pnv.h           |   3 +-
>   12 files changed, 757 insertions(+), 586 deletions(-)
>   create mode 100644 hw/pci-host/pnv_phb.c
>   create mode 100644 include/hw/pci-host/pnv_phb.h

I'm completely on-board with having a proxy-like PHB device that maps to the correct 
underlying PHB version, but to me it feels like combining multiple separate devices 
into a single one is going to make things more complicated to maintain in the long term.


ATB,

Mark.