On Sat, 25 Oct 2025, Philippe Mathieu-Daudé wrote:
On 25/10/25 16:49, BALATON Zoltan wrote:Hello,Added a few more people to cc hoping to get some opinion to clear this up. This is brought up by my patch trying to simplify hw/pci-host/ raven.c part of this series:https://patchew.org/QEMU/[email protected]/ (First submitted in May here: https://patchew.org/QEMU/[email protected]/ but that went relatively unnoticed and missed the previous release.) Find discussion below the patch. On Sat, 25 Oct 2025, BALATON Zoltan wrote:On Fri, 24 Oct 2025, Mark Cave-Ayland wrote:On 23/10/2025 16:26, BALATON Zoltan wrote:Export memory regions as sysbus mmio regions and let the board code map them similar to how it is done in grackle. While at it rename raven_pcihost_realizefn to raven_pcihost_realize. Signed-off-by: BALATON Zoltan <[email protected]> Reviewed-by: Philippe Mathieu-Daudé <[email protected]> --- hw/pci-host/raven.c | 38 +++++++++++++------------------------- hw/ppc/prep.c | 10 ++++++++-- 2 files changed, 21 insertions(+), 27 deletions(-)@@ -180,7 +178,18 @@ static void raven_pcihost_realizefn(DeviceState *d, Error **errp)qdev_init_gpio_in(d, raven_change_gpio, 1); + memory_region_init(&s->pci_io, o, "pci-io", 0x3f800000); + memory_region_init_io(&s->pci_discontiguous_io, o, + &raven_io_ops, &s->pci_io, + "pci-discontiguous-io", 8 * MiB); + memory_region_set_enabled(&s->pci_discontiguous_io, false); + memory_region_init(&s->pci_memory, o, "pci-memory", 0x3f000000); + + sysbus_init_mmio(dev, &s->pci_io); + sysbus_init_mmio(dev, &s->pci_discontiguous_io); + sysbus_init_mmio(dev, &s->pci_memory); sysbus_init_irq(dev, &s->irq); +h->bus = pci_register_root_bus(d, NULL, raven_set_irq, raven_map_irq, &s->irq, &s->pci_memory, &s- >pci_io, 0, 1,TYPE_PCI_BUS);@@ -219,32 +228,12 @@ static void raven_pcihost_realizefn(DeviceState *d, Error **errp)pci_setup_iommu(h->bus, &raven_iommu_ops, s); } -static void raven_pcihost_initfn(Object *obj) -{ - PREPPCIState *s = RAVEN_PCI_HOST_BRIDGE(obj); - MemoryRegion *address_space_mem = get_system_memory(); - - memory_region_init(&s->pci_io, obj, "pci-io", 0x3f800000); - memory_region_init_io(&s->pci_discontiguous_io, obj, - &raven_io_ops, &s->pci_io, - "pci-discontiguous-io", 8 * MiB); - memory_region_init(&s->pci_memory, obj, "pci-memory", 0x3f000000); - - /* CPU address space */ - memory_region_add_subregion(address_space_mem, PCI_IO_BASE_ADDR, - &s->pci_io);- memory_region_add_subregion_overlap(address_space_mem, PCI_IO_BASE_ADDR,- &s->pci_discontiguous_io, 1); - memory_region_set_enabled(&s->pci_discontiguous_io, false);- memory_region_add_subregion(address_space_mem, 0xc0000000, &s- >pci_memory);-} -@@ -293,6 +296,9 @@ static void ibm_40p_init(MachineState *machine) pcihost = SYS_BUS_DEVICE(dev);object_property_add_child(qdev_get_machine(), "raven", OBJECT(dev));sysbus_realize_and_unref(pcihost, &error_fatal); + sysbus_mmio_map(pcihost, 0, PCI_IO_BASE_ADDR); + sysbus_mmio_map_overlap(pcihost, 1, PCI_IO_BASE_ADDR, 1); + sysbus_mmio_map(pcihost, 2, PCI_MEM_BASE_ADDR); pci_bus = PCI_BUS(qdev_get_child_bus(dev, "pci.0")); if (!pci_bus) { error_report("could not create PCI host controller");In general the expectation is that memory regions should be initialised in the _init() function, unless they depend upon a property in which case they should be initialised in the _realize() function. Why do you feel this needs to be different?Is any of it needed before realize? If not why have an init method at all? As shown here this works perfectly without one and is more comprehensible that way for people reading it without deep knowledge about Qdev. In general I think simple devices only need a realize method and the init method is rarely needed, e.g. if there are some child objects that need to be init for passing properties that can be set before realize or similar unusual cases but for most classes init is not needed at all. I only want to keep what's necessary and remove everything that's not needed. I think that makes the device model easier to understand.I've checked documentation here: https://www.qemu.org/docs/master/devel/qdev-api.htmlbut it's not really clear on what's the preferred way of using init and realize. It's not even very clear on when to use which to me. So becuase that did not help I did a quick survey on what other pci-host models do. Of the 32 .c files in hw/pci-host 16 have an init method:aspeed_pcie.c, astro.c, designware.c, gpex.c, grackle.c, i440fx.c, pnv_phb3.c, pnv_phb3_msi.c, pnv_phb3_pbcq.c, pnv_phb4.c, q35.c, raven.c, sabre.c, uninorth.c, versatile.c, xilinx-pcie.cOf these astro.c has an empty init function that should be removed; grackle.c, sabre.c and uninorth.c are maintained by you so I'll ignore them here; we're discussing raven.c now and i440fx.c has two memory_region_init_io calls in init that could be in realize where all others are and otherwise all other models do this in realize and only init child objects and add properties in init methods when that's needed because they need to be available before realize. The other 16 device models don't have an init method at all and do all in the realize like I proposed in this patch for raven. Since only device models that you maintain do it differently I think what you say is not following the preferred way so you should not block this patch.I'd be interested if there is a consensus on this or can we cone to one that we can document to avoid this repeating every time.I've been told to stop arguing about QDev on the mailing list, and instead spend my time and energy in documenting QDev, so we'll discuss the documentation patches :) Also we'll try to provide a QDev meaningful state machine, which will help to enforce doing in the correct places. Meanwhile... .instance_init is actually QOM layer, it is called once, and cannot fail. What is allocated here has to be de-allocated in .instance_finalize..realize is QDev where we check the device properties, reporting error. What is allocated/configured there has to be de-allocated in .unrealize. The big difference is for hot-pluggable devices, where unplug calls unrealize(), keeping the device initialized. Re-plug calls .realize() again, and we should be able to do that multiple times. With that in mind, IMO it is better to allocate all we can once in .instance_init().
I'd really like if simple devices could be implemented without further complication. I don't mind if the method is called init or realize and when it's called but most devices should only need one of them and not have to care about where each step in their init phase should go when most of the time they are just created once and never finalized like this one. For these classes having two init methods just distracts from the actual functionality of the class and makes it more difficult to write them and understand what they do. The way that a lot of classes follow currently that they only have realize and only need init if some properties or child objects providing properties need to be created before realize to be possible to set them is easy to understand and keeps devices simple. So is there an advantage for hot-pluggable devices are not deinit when they are unplugged or is it a big performance issue to rerun realize on re-plug? Given that such classes are not common I'd trade that for keeping most other classes simple and more easily understandable and only require multiple methods where really needed.
Regards, BALATON Zoltan
