On 02/24/2011 07:56 AM, Michal Privoznik wrote:

--- src/conf/domain_conf.c | 2 ++ src/conf/domain_conf.h | 1 + 2 files changed, 3 insertions(+), 0 deletions(-)

diff --git a/src/conf/domain_conf.c b/src/conf/domain_conf.c index b97c1f0..454f631 100644 --- a/src/conf/domain_conf.c +++ b/src/conf/domain_conf.c @@ -2593,8 +2593,10 @@ virDomainNetDefParseXML(virCapsPtr caps, (const char *)macaddr); goto error; } + def->mac_generated = false; } else { virCapabilitiesGenerateMac(caps, def->mac); + def->mac_generated = true; }

Rather than sticking the flag into the domain definition, I'd rather pass it via the flags argument to virDomainDeviceDefParse/virDomainNetDefParseXML - that is, most callers will want: virDomainDeviceDefParse(..., 0) but qemuDomainDetachDevice (in qemu_driver.c) will want: virDomainDeviceDefParse(..., VIR_DOMAIN_PARSE_NO_GENERATE) (or maybe VIR_DOMAIN_XML_INACTIVE vs. VIR_DOMAIN_XML_INACTIVE|VIR_DOMAIN_PARSE_NO_GENERATE) That is, it's more than just detaching an interface mac addresses that has an issue. It's anywhere that parsing a device XML snippet to be used to match an existing device that we want to suppress generation of additional items in the parse (this includes generating a random mac for an interface, but may include other generated items). So adding a new flag that can be used to tell ALL parse routines to avoid generating extra data is more useful than changing the domain definition to say whether one piece of information was generated. -- Eric Blake eblake@redhat.com +1-801-349-2682 Libvirt virtualization library http://libvirt.org

(I found this mostly-composed message sitting behind a bunch of windows on my desktop when I decided it was time to reboot just so I could have an apparently clean slate...) On 03/01/2011 03:58 AM, Michal Prívozník wrote:

On 02/25/2011 05:42 PM, Eric Blake wrote:

...

On 02/24/2011 07:56 AM, Michal Privoznik wrote:

...

--- src/conf/domain_conf.c | 2 ++ src/conf/domain_conf.h | 1 + 2 files changed, 3 insertions(+), 0 deletions(-)

diff --git a/src/conf/domain_conf.c b/src/conf/domain_conf.c index b97c1f0..454f631 100644 --- a/src/conf/domain_conf.c +++ b/src/conf/domain_conf.c @@ -2593,8 +2593,10 @@ virDomainNetDefParseXML(virCapsPtr caps, (const char *)macaddr); goto error; } + def->mac_generated = false; } else { virCapabilitiesGenerateMac(caps, def->mac); + def->mac_generated = true; }

Rather than sticking the flag into the domain definition, I'd rather pass it via the flags argument to virDomainDeviceDefParse/virDomainNetDefParseXML - that is, most callers will want:

virDomainDeviceDefParse(..., 0)

but qemuDomainDetachDevice (in qemu_driver.c) will want:

virDomainDeviceDefParse(..., VIR_DOMAIN_PARSE_NO_GENERATE)

(or maybe VIR_DOMAIN_XML_INACTIVE vs. VIR_DOMAIN_XML_INACTIVE|VIR_DOMAIN_PARSE_NO_GENERATE)

That is, it's more than just detaching an interface mac addresses that has an issue. It's anywhere that parsing a device XML snippet to be used to match an existing device that we want to suppress generation of additional items in the parse (this includes generating a random mac for an interface, but may include other generated items). So adding a new flag that can be used to tell ALL parse routines to avoid generating extra data is more useful than changing the domain definition to say whether one piece of information was generated.

Personally, I think that a function called "...Parse" should do just that - parse what is provided an return it in a different format. There are several cases in libvirt of parse functions that have side effects like this and, while they are convenient, I think it would be cleaner for the parsers to stick to what they say their duty is, and for the callers to do this generation if needed.

Yes, i firstly thought in similar way. But I don't really need to suppress generation of random values. I need information whether user gave say mac in input XML or not.

Yes, you don't need to suppress them, but you also don't need to generate them. You just need to know whether or not the user supplied one.

Same apply for pci address (by the way, we don't support detaching by it now). The idea behind is: if (<mac was generated> && <pci addr was generated>) { if (domain->nnets == 1) device_clone_mac(domain->nnet[0], device) else error("don't know which detach") } else { search for device }

If i suppress generation, I wouldn't know if it was generated anyway, because mac is stored in statically allocated array. One (tricky) way out of this is to set mac before to some special mac (e.g. all 0's) and then check for change. But I'd rather avoid it. Changing array to dynamically allocated is huge amount of work to be done, because it is hard to find pieces of code touching it, and omitting even one line may lead into sigegv.

In the end, I think that the totally proper way of handling this would be to either: 1) add a flag to the def called mac_specified (or mac_is_valid or whatever), that would be 0 when the def is created, and set to 1 any time a value is given to it, or 2) as you say, make the mac a dynamically allocated piece of data, so that a NULL pointer would be an implicit "mac_specified == FALSE". The same goes for any other piece of data in any other def that is marked as optional in the RNG. But that is all a bit pie-in-the-sky, since the existing code works, and "fixing" it could possibly actually break something :-) When I recently had the same problem (adding a mac address to a virNetworkDef), I decided to make the mac address a fixed size array to simplify memory management and follow the precedent of the mac address in virDomainNetDef, and add a mac_specified boolean just after it. Consumers of virNetworkDef who really do need a mac address can generate one themselves if one isn't given, but those consumers that don't need one aren't burdened with it. (This has extra importance with objects that are saved as config, because you otherwise end up with something in the saved config that the user neither specified nor wants).

But I agree, sticking the flag into domain definition is not very clean either.

One way out of this is adding pointer to flag as (another) function parameter. Would this be acceptable?

I think that's even more obfuscated. If having the extra flag in the Def struct is confusing/unclean, then my vote would go for dynamically allocating mac when it's there (that's a potential source of null dereferences and memory leaks, though; everything has a confusion factor and a price).

On 02/24/2011 07:56 AM, Michal Privoznik wrote:

When detaching interface without <mac> specified a one is generated which leads to not found device. --- src/qemu/qemu_hotplug.c | 26 ++++++++++++++++++++------ 1 files changed, 20 insertions(+), 6 deletions(-)

diff --git a/src/qemu/qemu_hotplug.c b/src/qemu/qemu_hotplug.c index 0002af0..54c97db 100644 --- a/src/qemu/qemu_hotplug.c +++ b/src/qemu/qemu_hotplug.c @@ -1436,18 +1436,32 @@ int qemuDomainDetachNetDevice(struct qemud_driver *driver, virDomainDeviceDefPtr dev, virBitmapPtr qemuCaps) { - int i, ret = -1; + int i = 0, ret = -1; virDomainNetDefPtr detach = NULL; qemuDomainObjPrivatePtr priv = vm->privateData; int vlan; char *hostnet_name = NULL;

- for (i = 0 ; i < vm->def->nnets ; i++) { - virDomainNetDefPtr net = vm->def->nets[i]; + if (!vm->def->nnets) { + qemuReportError(VIR_ERR_OPERATION_FAILED, + _("domain has no interfaces.")); + goto cleanup; + } else if ((vm->def->nnets > 2) && (dev->data.net->mac_generated)) { + qemuReportError(VIR_ERR_OPERATION_FAILED, + _("You must specify mac address in xml file")); + goto cleanup; + }

If we refactor patch 1 to pass an additional flag to the parse routines to avoid generating the mac in the first place, then this logic needs to change a bit. It turns into this pseudocode: match = NULL; for (i = 0; i < nnets; i++) { if (nets[i] matches dev) { if (match) error: ambiguous match = nets[i] } } if (!match) error: no match This would also solve Hu Tao's point - if you have two interfaces, it is then possible to provide xml that lists just the pci address and still only matches one of the two interfaces, without requiring that the mac address be provided in the xml. -- Eric Blake eblake@redhat.com +1-801-349-2682 Libvirt virtualization library http://libvirt.org