[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

Re: [libvirt] cont command failing via JSON monitor on restore

On 01/18/2011 01:15 PM, Jim Fehlig wrote:
Laine Stump wrote:
On 01/14/2011 04:26 PM, Jim Fehlig wrote:
Laine Stump wrote:
On 01/13/2011 04:29 PM, Jim Fehlig wrote:
Daniel P. Berrange wrote:
   Yep, it wasn't really intended as a fix. It was intended to make
   the error scenario clearly detectable, which has succeeded as per
   Jim's report. The fact that QMP returned an error in this way,
   means we can now reliably detect, usleep(1000), and then retry
   the 'cont' again at which point we should succeed.

Something like the attached patch?  I'm not quite sure about the retry
bounds (currently 1 second).  In my testing, it always succeeds on the
second attempt - even with large memory vms.
In my tests, 250ms was more than enough, so I'm guessing it's okay,
although there's probably no hurt in making it a bit larger - it's not
like this is something that repeats all day every day :-)

Would it be possible for you to add the same thing into the text
monitor version of the code, so both fixes would travel together as
the patch gets cherry-picked around?
I can, but have not seen this issue with the text monitor.  And the
error reporting is not so fine grained correct?
Truthfully I haven't spent any time with the JSON code, and only
enough with the text monitor to (locally) put in a rough hack for the
same problem - I just delayed 250ms unconditionally.

Maybe it is safest for your patch to just have what you're able to
test for.
Agreed.  What error do you get back from the text monitor when cont cmd

Right now the machine that experiences this failure is stuck on qemu-0.12.5, so it doesn't fail, it "succeeds" :-) (since it doesn't have the necessary qemu patch, when "cont" is run before the migration has been able to start, it just believes that it was successful, and starts the CPU up with garbage in the guest memory.)

I'll try to get qemu-0.13 installed on this machine as soon as I can, but I'm trying to get something else fixed before Thursday.

I'd be interested in the following output from

     VIR_DEBUG("Receive command reply ret=%d errno=%d %d bytes '%s'",
               ret, msg.lastErrno, msg.rxLength, msg.rxBuffer);

Of course my system isn't setup to test exactly this fix either - the
machine that displays the problem when resuming is still running

- if (ret == 0)
-        ret = qemuMonitorJSONCheckError(cmd, reply);
+        if (ret != 0)
+            break;
+        /* If no error, we're done */
+        if ((ret = qemuMonitorJSONCheckError(cmd, reply)) == 0)
+            break;
+        /* If error class is not MigrationExpected, we're done.
+         * Otherwise try 'cont' cmd again */
+        if (!qemuMonitorJSONHasError(reply, "MigrationExpected"))
+            break;
+        virJSONValueFree(reply);
+    } while ((++i<= timeout)&&    (usleep(250000)<=0));

Doesn't this end up doing a double-free of reply if it times out?
virJSONValueFree doesn't update the pointer that's free'd like
VIR_FREE does (it can't, since it's a function call rather than a
virJSONValueFree() calls VIR_FREE() on the value passed to it, so reply
should be set to NULL when virJSONValueFree() returns.

Yes, but it's pass by value, not reference, so VIR_FREE() is NULLing
the copy of the pointer that was passed as an argument, not the
original pointer itself.
Err, right.  I'll fix this in v2 - but need to understand the text
monitor error first.

I'll do my best to get to it as soon as possible.


[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]