Re: [Qemu-devel] [PATCH v2] Changing error message of QMP 'migrate_set_downtime' to seconds

[Dr. David Alan Gilbert]

* Markus Armbruster (address@hidden) wrote:
> Daniel Henrique Barboza <address@hidden> writes:
> 
> > On 02/21/2017 06:02 AM, Markus Armbruster wrote:
> >> Daniel Henrique Barboza <address@hidden> writes:
> >>
> >>> The previous error message was displaying the values in miliseconds,
> >>> being misleading with the command that accepts the value in seconds:
> >>>
> >>> { "execute": "migrate_set_downtime", "arguments": {"value": 3000}}
> >>> {"error": {"class": "GenericError", "desc": "Parameter 'downtime_limit'
> >>> expects an integer in the range of 0 to 2000000 milliseconds"}}
> >>>
> >>> This patch changes it to '2000 seconds' to keep consistency with
> >>> the expected parameter. The macro 'QERR_INVALID_PARAMETER_VALUE'
> >>> was changed for a regular string that allows the use of the
> >>> MAX_MIGRATE_SET_DOWNTIME as a parameter, instead of hardcoding
> >>> the value in the error message.
> >>>
> >>> Signed-off-by: Daniel Henrique Barboza <address@hidden>
> >>> ---
> >>>   migration/migration.c | 12 ++++++++----
> >>>   1 file changed, 8 insertions(+), 4 deletions(-)
> >>>
> >>> diff --git a/migration/migration.c b/migration/migration.c
> >>> index c6ae69d..c05e764 100644
> >>> --- a/migration/migration.c
> >>> +++ b/migration/migration.c
> >>> @@ -49,6 +49,9 @@
> >>>    * for sending the last part */
> >>>   #define DEFAULT_MIGRATE_SET_DOWNTIME 300
> >>>   +/* Maximum migrate downtime set to 2000*1000 miliseconds */
> >>> +#define MAX_MIGRATE_SET_DOWNTIME (2000 * 1000)
> >>> +
> >>>   /* Default compression thread count */
> >>>   #define DEFAULT_MIGRATE_COMPRESS_THREAD_COUNT 8
> >>>   /* Default decompression thread count, usually decompression is at
> >>> @@ -843,10 +846,11 @@ void qmp_migrate_set_parameters(MigrationParameters 
> >>> *params, Error **errp)
> >>>           return;
> >>>       }
> >>>       if (params->has_downtime_limit &&
> >>> -        (params->downtime_limit < 0 || params->downtime_limit > 
> >>> 2000000)) {
> >>> -        error_setg(errp, QERR_INVALID_PARAMETER_VALUE,
> >>> -                   "downtime_limit",
> >>> -                   "an integer in the range of 0 to 2000000 
> >>> milliseconds");
> >>> +        (params->downtime_limit < 0 ||
> >>> +         params->downtime_limit > MAX_MIGRATE_SET_DOWNTIME)) {
> >>> +        error_setg(errp, "Parameter 'downtime_limit' expects an integer 
> >>> in "
> >>> +                         "the range of 0 to %d seconds",
> >>> +                         MAX_MIGRATE_SET_DOWNTIME / 1000);
> >>>           return;
> >>>       }
> >>>       if (params->has_x_checkpoint_delay && (params->x_checkpoint_delay < 
> >>> 0)) {
> >> Isn't this wrong for QMP migrate-set-parameters?  There, the unit is
> >> milliseconds, i.e. the new error message is as wrong as the old one is
> >> for migrate_set_downtime.
> >
> > Actually the unit for migrate-set-parameters, as seen by the caller,
> > is seconds. The underlying logic receives the input and multiplies it
> > by 1000 in qmp_migrate_set_downtime().
> 
> Yes, QMP migrate_set_downtime takes seconds, but I'm talking about QMP
> command migrate-set-parameters, which takes milliseconds.
> 
> >> I'm afraid you need to fix the error message in
> >> qmp_migrate_set_downtime().  If you assume qmp_migrate_set_parameters()
> >> fails only in one way, replace its error object by one with a better
> >> message[*].  If you'd rather not assume, you need to refactor things so
> >> that each place can set the downtime and create an appropriate error on
> >> failure.
> >
> > There is at least one similar usage of this error message just
> > above this code in max_bandwidth param. Perhaps a new
> > function/macro to deal with these cases is justified.
> >
> >>
> >> We might want to check other command wrappers that translate units.
> >>
> >> Time units are a hopeless mess in QMP.  We should've enforced uniform
> >> usage of either seconds or nanoseconds.  The latter to placate
> >> irrational fear of floating-point[**].
> >
> > I agree that my patch doesn't make it much better. I just set the
> > error message to be in seconds to be consistent with the user input,
> 
> That's reasonable!  The problem is that you fix one error message by
> breaking another one.
> 
> > but the code now feels 'awkward' when you do a verification
> > in milliseconds and deliver the error message in seconds.
> >
> > One thing that can be done is to make migrate-set-downtime to
> > accept milliseconds instead of seconds. I wasn't willing at first to
> > change the migrate-set-downtime API because of an error
> > message, however it really feels like the right thing to do here.
> 
> Changing QMP commands incompatibly is a big no-no.
> 
> > Specially when you consider that the default value of this parameter,
> > set by DEFAULT_MIGRATE_SET_DOWNTIME, is 300 - a value that
> > in theory the user shouldn't be able to set in the API.
> 
> I think that's a fine reason for deprecating the migrate_set_downtime
> command, and ask people to use migrate-set-parameters instead.

OK, but you can set it to 300 from migrate_set_downtime by doing
HMP:
   migrate_set_downtime 0.3 
QMP:
   { "execute": "migrate_set_downtime", "arguments": { "value": 0.3 } }

so it's bogus to say that's a reason to ditch it.

This has all got rather complex for a simple patch to change an error
message.  Can't we just come up with a clear error message that works in
all cases.

Dave

> >> [*] If replacing messages turns out to be a common operation, we can add
> >> a function to replace it within the same Error object.
> >>
> >> [**] If you think the rounding you can get when converting
> >> floating-point seconds to integer milli-, micro- or nanoseconds matters,
> >> I have time-keeping equipment to sell you.
> >>
--
Dr. David Alan Gilbert / address@hidden / Manchester, UK