1:31 p.m.
On Thu, Dec 15, 2016 at 01:39:30PM +0100, Markus Armbruster wrote:
Eduardo Habkost <ehabkost(a)redhat.com> writes:
> This adds a new command to QMP: query-device-slots. It will allow
> management software to query possible slots where devices can be
> plugged.
>
> This implementation of the command will return:
>
> * Multiple PCI slots per bus, in the case of PCI buses;
> * One slot per bus for the other buses (that don't
> implement slot enumeration yet);
> * One slot for each entry from query-hotpluggable-cpus.
>
> Git tree
> --------
>
> This patch needs the previous query-machines series I am working
> on. The full tree can be found on the git tree at:
>
> git://github.com/ehabkost/qemu-hacks.git work/query-machines-bus-info
>
> Example output
> --------------
>
> The following output was returned by QEMU when running it as:
>
> $ qemu-system-x86_64 -machine q35 \
> -readconfig docs/q35-chipset.cfg \
> -smp 4,maxcpus=8,sockets=2,cores=2,threads=2
>
> {
> "return": [
> {
> "available": false,
> "hotpluggable": false,
> "props": {
> "bus": "i2c"
Missing: bus "i2c-bus" device address, property name "address",
value
0..255.
> },
> "type": "non-slot",
I guess "non-slot" means "props" is incomplete, but the rest should
be
okay.
Yes, non-slot is explicitly documented as incomplete.
> "accepted-device-types": [
> "i2c-slave"
> ]
> },
> {
> "available": false,
> "hotpluggable": false,
> "props": {
> "bus": "ide.4"
Bus "IDE" device address property "unit", value 0 (because this is
SATA,
for PATA it would be 0..1).
> },
> "type": "non-slot",
> "accepted-device-types": [
> "ide-device"
> ]
> },
[...]
> {
> "available": false,
> "hotpluggable": false,
> "props": {
> "bus": "main-system-bus"
> },
> "type": "non-slot",
> "accepted-device-types": [
> "sys-bus-device"
> ]
> },
Can't give guidance on this one, I'm afraid.
So, the good news here seems to be: very few machines-type accept
sysbus devices on -device. This means we can investigate why
exactly we set has_dynamic_sysbus=true on some machine-types
(maybe this will be obvious from the git log, maybe it won't, but
I'm feeling optimistic today), manually whitelist the (hopefully)
few device types that can be created dynamically, create specific
slots for them, and hide sysbus completely from clients.
> {
> "available": false,
> "hotpluggable": false,
> "props": {
> "bus": "isa.0"
> },
> "type": "non-slot",
> "accepted-device-types": [
> "isa-device"
> ]
> },
Bus "ISA" is iffy, too.
> {
> "available": false,
> "hotpluggable": false,
> "props": {
> "bus": "pcie.0",
> "device-number": 0
Really? "device_add e1000e,help" shows now property
"device-number".
Oh, I see, you're adding it in this patch, along with "function".
Yes. But I am still not sure about how to model/describe
"function". Should we say that (device-number=0, function=0) and
(device-number=0, function=1) are part of the same "slot", or two
different slots?
If we implement something to easily represent slot sets, it will
be easier to represent each PCI function as a separate slot.
> },
> "type": "pci",
> "accepted-device-types": [
> "legacy-pci-device",
> "pci-express-device"
> ]
> },
> {
> "available": false,
> "hotpluggable": false,
> "props": {
> "bus": "pcie.0",
> "device-number": 1
> },
> "type": "pci",
> "accepted-device-types": [
> "legacy-pci-device",
> "pci-express-device"
> ]
> },
[...]
> {
> "available": false,
> "hotpluggable": false,
> "props": {
> "bus": "pcie.0",
> "device-number": 31
> },
> "type": "pci",
> "accepted-device-types": [
> "legacy-pci-device",
> "pci-express-device"
> ]
> },
Design question: do we need to describe the bus with 32 separate
objects, or would a single object do? The objects only differ in the
value of props.device-number...
For a hypothetical bus with uint32_t device addresses, separate objects
would be impossible. So we probably need to describe slot sets, anyway.
My question is: do we need to describe slots sets in the first
version, or can we do it as an extension later?
> {
> "available": true,
> "hotpluggable": true,
> "props": {
> "bus": "ich9-pcie-port-1",
> "device-number": 0
> },
> "type": "pci",
> "accepted-device-types": [
> "pci-express-device"
> ]
> },
> {
> "available": true,
> "hotpluggable": true,
> "props": {
> "bus": "ich9-pcie-port-2",
> "device-number": 0
> },
> "type": "pci",
> "accepted-device-types": [
> "pci-express-device"
> ]
> },
> {
> "available": true,
> "hotpluggable": true,
> "props": {
> "bus": "ich9-pcie-port-3",
> "device-number": 0
> },
> "type": "pci",
> "accepted-device-types": [
> "pci-express-device"
> ]
> },
> {
> "available": true,
> "hotpluggable": true,
> "props": {
> "bus": "ich9-pcie-port-4",
> "device-number": 0
> },
> "type": "pci",
> "accepted-device-types": [
> "pci-express-device"
> ]
> },
> {
> "available": false,
> "hotpluggable": false,
> "props": {
> "bus": "ich9-pci-bridge",
> "device-number": 0
> },
> "type": "pci",
> "accepted-device-types": [
> "legacy-pci-device"
> ]
> },
[...]
> {
> "available": false,
> "hotpluggable": false,
> "props": {
> "bus": "ich9-pci-bridge",
> "device-number": 31
> },
> "type": "pci",
> "accepted-device-types": [
> "legacy-pci-device"
> ]
> },
Likewise.
> {
> "available": true,
> "hotpluggable": true,
> "props": {
> "bus": "ich9-ehci-1.0"
> },
> "type": "non-slot",
> "accepted-device-types": [
> "usb-device"
> ]
> },
Okay, more bad news: bus "usb-bus" device address property is
"port",
value is a sequence of port numbers separated by '.'. The reason is
"convenience"
A real USB host controller has a fixed number of ports, numbered 1..N.
This one has six.
A real USB hub device plugs into a port and in turn provides a fixed
number of ports, numbered 1..N. The USB hub device we model provides
eight.
An obvious way to model this would be to have the host controller
provide a bus with a fixed number of slots, and the usb-hub provide a
separate bus with eight. This is how we do PCI bridges. But it's not
how we do USB hubs.
Host software can treat the tree formed by USB host controller, hubs and
other devices as a single bus, and that's how we model it.
A "port" value that is a single number addresses that port on the host
controller.
A value containing '.' addresses something behind a hub. The leftmost
number addresses a host controller port. A hub must be plugged in
there. Recurse for the rest of the port string.
If you device_add an USB device without specifying a port, it picks an
unused one. When it has to pick the last port, it tries to add a
usb-hub device there, which provides eight ports, then picks a port for
your device. If adding the usb-hub succeeds, it'll get the new hub's
port 1. Else, it gets the last port. There's a hard-coded limit of
five on the depth of the hub chain. So, if you keep adding USB devices
to the bus provided by ich9-ehci, the first five go into ports "1", ...,
"5", the sixth goes into "6.1" (with usb-hub in port "6"),
and so forth,
until the 41st goes into "6.8.8.8.8.8". Adding hubs manually permits
more devices.
I assume we won't need to represent the automagic port creation
rules in the query. Only the next case:
If you do specify "port", then that port must already exist.
So, hubs are _not_ created automatically at ports "A", "A.B" and
"A.B.C" if I explicitly try to plug something to port "A.B.C.D"?
That's good news, because then we can simply represent only the
available ports.
I think the sanest way to represent the existing ports is
enumerating
them, like this:
{
"available": true,
"hotpluggable": true,
"props": {
"bus": "ich9-ehci-1.0", "port":
"1"
},
"type": "usb",
"accepted-device-types": [
"usb-device"
]
},
[...]
{
"available": true,
"hotpluggable": true,
"props": {
"bus": "ich9-ehci-1.0", "port":
"6"
},
"type": "usb",
"accepted-device-types": [
"usb-device"
]
},
Plugging usb-hub into a port with port: "P" then flips that port's
"available" to false, and adds eight new ports with "port":
"P.1",
... "P.8".
So, plugging a device can not only add buses provided by the device, but
also, and less obviously, add slots to the bus it plugs into! Likewise
for unplug.
Interesting.
Perhaps we should emit events when slots come and go.
Perhaps. Should we leave this for later?
FWIW, the problem I am trying to solve right now is querying for
machine-type capabilities before creating the VM. The runtime
querying command in this patch is just the first step. Then I
plan to also return DeviceSlotInfo for machine-types just like
the "query-machines"/"available-buses" series I sent.
> {
> "available": true,
> "hotpluggable": true,
> "props": {
> "bus": "ich9-ehci-2.0"
> },
> "type": "non-slot",
> "accepted-device-types": [
> "usb-device"
> ]
> },
> {
> "available": false,
> "hotpluggable": false,
> "props": {
> "bus": "ich9-hda-audio.0"
Bus "HDA" device address property is "cad", value 0..14.
> },
> "type": "non-slot",
> "accepted-device-types": [
> "hda-codec"
> ]
> },
> {
> "available": true,
> "hotpluggable": true,
> "props": {
> "socket-id": 1,
> "core-id": 1,
> "thread-id": 1
> },
> "type": "cpu",
> "accepted-device-types": [
> "qemu64-x86_64-cpu"
> ]
> },
You know CPUs better than me.
Fortunately CPU slots are already modelled, I am just translating
the query-hotpluggable-cpus data.
[More of them...]
> ]
> }
There's still more...
Bus "SCSI" device address property is "scsi-id", value 0..N, where N
depends on the SCSI HBA.
Bus "virtio-serial-bus" device address property is "nr", value 0..N,
where N is configurable, currenrly up to 511.
Sounds like this is large enough to justify representing
slot-sets since the first version of this interface.
ccid-bus device address property is "slot", value 0..N, where N depends
on the device providing the bus (I think).
Buses not yet covered by docs/qdev-device-use.txt: "floppy-bus",
"apple-desktop-bus", "IndustryPack", "aux-bus",
"spapr-vio-bus",
"virtual-css-bus", "s390-sclp-events-bus", "s390-pcibus",
"sd-bus",
"SSI", "virtio-bus", "xen-sysbus". Grep for ".parent
= TYPE_BUS".
This will be fun...
Thanks a lot for the information and ideas about the cases above.
To find all buses in a certain build, use:
{ "execute": "qom-list-types", "arguments": {
"implements": "bus" } }
[No code review, yet]
--
Eduardo