On 07/27/2010 11:11 AM, Ryota Ozaki wrote:
On Mon, Jul 26, 2010 at 11:37 PM, Laine Stump<laine(a)laine.org>
wrote:
> On 07/26/2010 08:31 AM, Ryota Ozaki wrote:
>> On Mon, Jul 26, 2010 at 6:51 PM, Daniel P. Berrange<berrange(a)redhat.com>
>> wrote:
>>> You could just change
>>>
>>> return cmdResult
>>>
>>> to
>>>
>>> return -cmdResult;
>>>
>>> That would still let you give the error code, while also keeping the
>>> value
>>> < 0
>> It looks better than mine ;-) I'll rewrite my patch in such a way.
>>
>> Laine, is it ok for you too?
>>
> Doing that is fine when all possible failures in the function have an
> associated errno. In the case of virRun'ing an external program that could
> return a non-zero exit code, this is unfortunately not the case, so those
> functions that call virRun will need to report the error themselves and
> return -1.
For veth.c all functions match the latter case while bridge.c has both.
If we don't take care about the consistency between veth.c and bridge.c,
we can focus on how to treat the latter case. (Ideally keeping the consistency
is better, but it seems difficult...)
We can ignore bridge.c for now - there is a lot of code in libvirt that
doesn't follow the "-errno on failure" convention, so we should try to
be as narrow as possible. For this cleanup, I like the scope of just
veth.c and its direct callers.
Another thing I didn't notice when I wrote my first comment on this
thread is that most of the error conditions in the involved functions
are due to an external program failing, which will likely be due to the
program returning a non-0 code. However, that's not *always* the case,
so we can't really say that the function will always return a valid
"-exitcode", nor can we say it will always return a valid "-errno"
(and
as you point out, they can conflict).
> When non-0 exits from the called program are all failures, the
simplest way
> to do it is, as you say, to just not pass a pointer to a resultcode to
> virRun (as long as the error message reported by virRun is useful enough -
Yes.
> remember that it gives the name of the program being run, and "virRun",
but
> not the name of the calling function).
Agreed. That'll lose useful information for debugging. One option is
to re-report
an error in the calling function. It will lead reporting two messages,
but it should
be more useful than less reports.
One concern aobut virRun's error reporting is that it shows standard
errors through
DEBUG so by default it shows nothing while they are important for ifconfig
and ip commands because their error messages may be according to errno.
I also recall once in some other code letting virRun display the error
and the result was that someone else who got the error (which was really
unimportant) was handed a giant scary looking (and uninformative) error
message that took their attention away from the real problem, which was
elsewhere (I forget the details now, but you get the idea). It's
convenient, but if you want meaningful errors, I would recommend against
letting virRun report them.
I went through all the calls to all the functions listed in veth.h, and
I see that they are called from relatively few places, and the error
logs are almost always to the veth.c function that's being called, not
to the caller. So I think it could work to have:
1) all calls to virRun in veth.c request cmdResult.
2) all functions in veth.c log the errors themselves (with one possible
exception - vethDelete() - because it is often called during error
recovery).
3) all functions in veth.c return -1 to their caller
4) all functions called by those functions also return -1 (this is
mostly the case already, once you change veth.c, and has no effect on
those functions' callers).
Logging errors in veth.c functions you have all the information you need
from virRun, but also know the context in which the program is being
called (see my notes below) so you can avoid the "always big and scary"
messages from virRun, and you can consistently return -1.
(The only downside is that you don't have access to the stderr of the
executed program, but that will be solved when we switch to Dan
Berrange's new exec library ;-)
Does this sound reasonable and doable? If you agree with it, but think
it's not doable in time for the release on Friday, but the original bug
has real consequences (ie it makes something potentially fail), then I
think it would be okay to push your original patch, with intent to do a
further cleanup after the release.
(sorry for turning this 2 line bugfix into an odyssey of refactoring -
I'm just paying it forward ;-)
*****************************
Here are my notes from looking at the functions in veth.c and their callers:
vethCreate
called from:
lxcSetupInterfaces - on error, it logs "Failed to create device
pair %d", rc
solution - log that error in vethCreate, and have vethCreate return
-1 on failure
vethDelete
called from:
lxcControllerCleanupInterfaces - logs, but doesn't display code
lxcVmCleanup - ignores, no log
lxcVmStart - ignores, no log
Both of the cases where there is no log on failure are cases where there
was already an error, and this is just cleaning up. The other case is, I
think, as the container is shutting down. There's a good reason for not
logging an error if you're already in error recovery - the newer error
message will overwrite the old one (eg, virsh only prints out the last
error that was logged during any given API call).
suggestion - since the existing error log doesn't reference the
return code, just have vethDelete return -1, and continue logging the
error in the caller (when desired).
vethInterfaceUpOrDown
called from:
lxcContainerRenameAndEnableInterfaces - enabling, log error
lxcVmCleanup - disabling, don't log error
lxcSetupInterfaces - enabling, log error
This one doesn't log an error if upOrDown is 0, and in the two cases
where it does log an error, the text is nearly identical.
suggestion - log errors in vethInterfaceUpOrDown if upOrDown != 0,
and return -1 on failure. Don't log in the caller.
moveInterfaceToNetNs
called from: lxcControllerMoveInterfaces - error is logged, but
doesn't give code.
suggestion - do the logging in moveInterfaceToNetNs instead, and add
in the code, then return -1
setMacAddr
called from: lxcSetupInterfaces - error is logged, but doesn't give code
suggestion - do the logging in setMacAddr, and add in the code, then
return -1
setInterfaceName
called from: lxcContainerRenameAndEnableInterfaces - code is saved,
then used to log error
suggestion - log the same error in setInterfaceName that was
previously logged in the caller, then return -1.