Re: [Qemu-devel] [PATCH 2/3] ioh3420: Provide a unique bus name and an interrupt mapping function

[Markus Armbruster]

"Michael S. Tsirkin" <address@hidden> writes:

> On Wed, Aug 20, 2014 at 02:06:38PM +0200, Markus Armbruster wrote:
>> Knut Omang <address@hidden> writes:
>> 
>> > On Wed, 2014-08-20 at 10:52 +0200, Paolo Bonzini wrote:
>> >> Il 20/08/2014 08:53, Knut Omang ha scritto:
>> >> > A unique bus name is necessary to be able to refer to each instance
>> >> > from the command line and monitors.
>> >> 
>> >> Is it needed?  Can't you just add id= to the -device option?
>> >
>> > Yes, as far as I understand the problem is that the id= would work on
>> > the ioh3420 device itself, while what is needed here is to name the
>> > secondary bus of the ioh3420, which I haven't found a way to name from
>> > the command line.
>> 
>> Bus names in qdev are a mess.  Here are the rules:
>> 
>> 1. If code provides a name, that's the name.
>> 
>> 2. Else, if the device has an ID, the name is ID.N, where N counts the
>> device's buses from zero.
>> 
>> 3. Else, the name is BUS-TYPE-NAME.N, where N counts the buses of this
>> type from zero.
>> 
>> This results in a usable bus name unless device IDs collide with bus
>> type names, or the code provides names that collide.
>> 
>> The user needs to take care to use IDs that don't collide with bus
>> names.  Adding new bus names may screw some users.
>> 
>> The user needs to take further care to use IDs whenever the code
>> provides a bus name that collides.  Adding code that provides bus names
>> may screw some users.
>> 
>> Broken by design.
>> 
>> The problem here is "code provides names that collide": device
>> q35-pcihost provides the name "pcie.0".  Bound to collide with the first
>> PCIE bus named under rule 3.  For instance, if you next add an ioh3420
>> without ID, its bus is also named "pcie.0".
>> 
>> Rule 1 should be taken out and shot.  Unfortunately, that'll break ABI
>> left & right.  Instead, we can try to reduce its use.  The appended does
>> exactly that for q35-pcihost.  With it applied, the bus provided by
>> q35-pcihost still gets the same name "pcie.0", but under rule 3 instead
>> of rule 1.  Rule 3 then names further PCIE buses "pcie.1", "pcie.2", ...
>> instead of "pcie.0", "pcie.1", ...  Better, but it's still an ABI break.
>> 
>> > Maybe an even better solution would be to have default names for
>> > everything, if not specified, from a user friendliness perspective? 
>> 
>> Buses *have* a default name!  You're confusing this with device IDs,
>> which exist only when the user sets one.
>> 
>> Changes in this area are difficult, because the names are all ABI.
>> Names that cannot be used are fair game, of course.
>> 
>> > I suppose this is a more general issue of sensible default values
>> > though, but the fact that it is easy to create devices which cannot be
>> > referred has caused me some confusion from time to time.
>> 
>> Picking default qdev IDs risks collisions with the user's IDs.  We
>> shouldn't do that.  We do it anyway in a few places, for historical
>> reasons.
>> 
>> QOM paths might be a sane way to let users refer to devices without IDs.
>> 
>> 
>> While writing the above, I stumbled another rule 1 screwup: pci_bridge.c
>> attempts to "improve" the boring standard bus names chosen via rule 2 or
>> 3.
>> 
>> pci_bridge_initfn() provides a bus name of its own (commit 8a3d80f
>> pci_bridge: user-friendly default bus name):
>> a. If pci_bridge_map_irq() set a bus name, that's the name.
>> 
>> b. Else, if the device has an ID, that's the name.  Thus, ID.N is
>> "improved" to just ID, at the cost of a special case: now users have to
>> avoid not just IDs of the form BUS-TYPE-NAME.N, but also plain
>> BUS-TYPE-NAME.
>> 
>> Callers of pci_bridge_map_irq() generally provide a name.  Some names
>> contain spaces, thus can't collide (but would be bloody inconvenient on
>> the command line or in the monitor).  Others don't, but thankfully the
>> ones I checked are in dead code.  Craptastic.
>> 
>> 
>> diff --git a/hw/pci-host/q35.c b/hw/pci-host/q35.c
>> index 37f228e..469aafd 100644
>> --- a/hw/pci-host/q35.c
>> +++ b/hw/pci-host/q35.c
>> @@ -47,7 +47,7 @@ static void q35_host_realize(DeviceState *dev, Error 
>> **errp)
>>      sysbus_add_io(sbd, MCH_HOST_BRIDGE_CONFIG_DATA, &pci->data_mem);
>>      sysbus_init_ioports(sbd, MCH_HOST_BRIDGE_CONFIG_DATA, 4);
>>  
>> -    pci->bus = pci_bus_new(DEVICE(s), "pcie.0",
>> +    pci->bus = pci_bus_new(DEVICE(s), NULL,
>>                             s->mch.pci_address_space, 
>> s->mch.address_space_io,
>>                             0, TYPE_PCIE_BUS);
>>      qdev_set_parent_bus(DEVICE(&s->mch), BUS(pci->bus));
>
> This is for the root bus, I think it won't help Knut who's trying to
> add devices behind root ports.

Read again, more slowly :)

Yes, I null the name of the root bus.  That makes the qdev machinery
derive the very same "pcie.0" name via rule 3 instead of rule 1, with
the side effect that future (non-root) PCIE buses get different names.
In particular, the next one named via rule 3 will be called "pcie.1"
instead of "pcie.0", making it actually accessible.