Re: [Qemu-devel] [PATCH REPOST 0/2] Add basic "detach" support for dump-guest-memory

[Markus Armbruster]

Eric Blake <address@hidden> writes:

> On 11/24/2015 04:37 AM, Fam Zheng wrote:
>
>>> I think the patch should be dropped, and periodic progress reports
>>> should be emitted from within the dump loops that do the heavy lifting.
>>>
>>> For the ELF format dumps, that loop appears to reside in dump_iterate()
>>> [dump.c].
>>>
>>> For the compressed format dumps, the loop seems to live in
>>> write_dump_pages() [dump.c].
>> 
>> This is a good idea!
>> 
>> What I'm not sure is where to report the progress. Can it be the monitor 
>> where
>> the dump-guest-memory command was issued? In other words, do we
>> support raising
>> events before the previous command returns? If yes, can libvirt handle this
>> correctly? (But the worst case is using another channel to communicate the
>> progress, it is ad-hocery but it must be better than all the risk
>> and effort to
>> enable multi-threaded dump.)
>> 
>> Eric, Markus, have any idea with the progress reporting?
>
> I'm fairly certain we support raising events prior to completion of a
> synchronous command; what I'm not sure of is whether the event hits the
> wire right away or whether it piles up waiting for the next synchronous
> command completion.  If the latter, then we need to rework it (since the
> whole point of this exercise is that we are trying to give progress of a
> long-running synchronous command that hasn't completed yet).  But we
> only have the one monitor connection for libvirt - the only way to pass
> events through a second channel is to open a second monitor connection,
> but that feels wrong to make libvirt have to track two monitors.

Short answer: events can be sent at any time.  Delay can come only from
taking the necessary lock, or rate limiting.

Longer answer: code triggers an event FOO by calling (QAPI-generated)
qapi_event_send_FOO().  This marshalls the event's arguments and calls
monitor_qapi_event_queue() (you have to peel off some obfuscating
abstraction to see that).  Assuming no rate limiting, this simply takes
the monitor_lock to call monitor_qapi_event_emit(), which immediately
sends the event to all QMP monitors.