Am 23.07.2014 um 15:17 hat Luiz Capitulino geschrieben:
>
> Management software, such as OpenStack and RHEV's vdsm, wants to be able
> to allocate VM disk space on demand. The basic use case is to start a VM
> with a small disk and then the disk is enlarged when QEMU hits a ENOSPC
> condition.
>
> To this end, the management software has to be notified when QEMU
> encounters ENOSPC. The most straightforward solution is to extend QMP's
> BLOCK_IO_ERROR event with that information.
>
> This series does exactly that. The approach taken is the simplest possible:
> the BLOCK_IO_ERROR event is extended to contain a "nospace" key, which
> will be true whenever the guest runs out of space *and* werror=stop|enospc.
> Here's an example:
>
> { "event": "BLOCK_IO_ERROR",
> "data": { "device": "ide0-hd1",
> "operation": "write",
> "action": "stop",
> "nospace": true },
> "timestamp": { "seconds": 1265044230, "microseconds": 450486 } }
>
> There are three important things to observe:
>
> 1. query-block already supports querying the event by means of the
> "io-status" key. Actually, "nospace" and "io-status" keys share
> the same semantics. This is a big advantage of this approach, no
> further extension of query-block is needed
>
> 2. The event could also contain an error message key for debugging,
> But if we add it to the event, should we add it to query-block too?
I don't think it's strictly necessary, but I can imagine that it would
be a very nice feature for debugging if you could check after that fact
what caused the VM stop even if you don't have a QMP log with the event.
> 3. I'm not extending BLOCK_JOB_ERROR. The reason is that it seems
> that BLOCK_IO_ERROR is also emitted on BLOCK_JOB_ERROR
Hm, I can't see this in the code. Where do I need to look?
Or did you get both a BLOCK_JOB_ERROR and a BLOCK_IO_ERROR because the
guest tried to access the image, too, and caused a separate error?
> Now, this series is an RFC because there's an alternative solution for
> this problem: instead of extending the BLOCK_IO_ERROR event with no-space
> indicator, we could have a stringfied errno. This way management apps
> would also be able to distinguish among other errors.
I don't think sending errnos is a good approach (but if we took it,
we should use an enum rather than strings) and prefer exposing the
exact information that is actually needed.
> For example, we could have a "error-details" dict containing a
> "reason" and a "message" key:
>
> { "event": "BLOCK_IO_ERROR",
> "data": { "device": "ide0-hd1",
> "operation": "write",
> "action": "stop",
> "error-details": { "reason": "eio", "message": "I/O
> error" },
> "timestamp": { "seconds": 1265044230, "microseconds": 450486 } }
>
> And then query-block would have to be extended to contain the same
> information.
>
> IMO, this series implementation is good enough for the requirement we
> currently have but I'm open to go complex if needed.
Agreed. I would like to see the human-readable strerror() string added,
but that doesn't make this series any worse as a first step:
Acked-by: Kevin Wolf <kw...@redhat.com>
