On Tue, 6 May 2025, Markus Armbruster wrote:
Akihiko Odaki <akihiko.od...@daynix.com> writes:
On 2025/02/06 18:48, Markus Armbruster wrote:
Akihiko Odaki <akihiko.od...@daynix.com> writes:
[...]
I understand we have something like this:
* true: on if possible, else off
* false: off (always possible)
Which one is the default?
It depends. Some properties have true by default. The others have false.
There is no way to reliably configure "on", i.e. fail if it's not
possible. I agree that's a problem.
This problem can be solved
using an existing mechanism, OnOffAuto, which differentiates the "auto"
state and explicit the "on" state.
I guess you're proposing something like this:
* auto: on if possible, else off
* on: on if possible, else error
* off: off (always possible)
Which one is the default?
I converted on to auto and off to false in a following patch.
However, converting bool to OnOffAuto surfaces another problem: they
disagree how "on" and "off" should be written. Please note that the
disagreement already exists and so it is nice to solve anyway.
Yes, converting bool to OnOffAuto is an incompatible change.
Not just about conversion, but this inconsistency require users to know
whether a property is bool or OnOffAuto and change how the values are
written in JSON accordingly. This somewhat hurts usability.
This patch tries to solve it by tolerating bool values for OnOffAuto. As
you pointed out, this approach has a downside: it makes OnOffAuto more
complicated by having multiple ways to express the same thing.
It also affects existing uses of OnOffAuto, where such a change is
unnecessary and undesirable.
To be clear: this is pretty much a deal-breaker for me.
We established above that you need certain boolean properties to take a
third state. I'm willing to discuss patches that change exactly these
properties. I'm going to reject patches that affect properties that do
not need such a change.
Even if the change is backwards compatible not breaking existing usage
just adding additional ways to specify same value?
I may not understand this problem completely and may have forgotten what I
understood but I think the proposed solution only added new alternative
names for existing values so the old values still work. Then the only way
this could break if somebody using new names want them to work on older
QEMU versions but that's not reasonable. Also JSON is not primarily used
by humans but by management apps and scripts which won't change the names
they use when we add new names that will likely only be used on the
command line so this should not break it even if those apps or script are
used with older QEMU.
Another approach is to have one unified way to express "on"/"off" for
bool and OnOffAuto. This will give three options in total:
1. Let OnOffAuto accept JSON bool and "on"/"off" (what this patch does)
The parenthesis is inaccurate. This patch only affects qdev properties.
It does not affect use of OnOffAuto elsewhere, e.g. QOM object
"sev-guest" property "legacy-vm-type", or QMP command blockdev-add
argument "locking" with driver "file".
2. Let OnOffAuto and bool accept JSON bool and deprecate "on"/"off"
3. Let OnOffAuto and bool accept "on"/"off" and deprecate JSON bool
For each of these options:
(a) Change exactly the uses of OnOffAuto that need to become tri-state
(b) Change all qdev properties (currently a superset of (a); what this
patch does)
(c) Change all uses of OnOffAuto
I dislike (c) and especially (b).
I'm fine with either of these approaches; they are at least better than
the current situation where users need to care if the value is OnOffAuto
or bool when they just want to express on/off. Please tell me what you
prefer.
We managed to maneuver ourselves into a bit of a corner in just a few
simple steps:
* The obvious type for a flag is bool.
* The obvious type for a small set of values is enum.
* Thus, the obvious type for a tri-state is enum.
* But this prevents growing a flag into a tri-state compatibly. Which
is what you want to do.
However, we actually have a second way to do a tri-state: optional bool,
i.e. present and true, present and false, absent.
Permit me a digression... I'm not a fan of assigning "absent" a meaning
different from any present value. But it's a design choice QAPI made.
It's a new insight I didn't know. Properties in qdev have a default
value instead of special "absent". But if QAPI does have special
"absent", perhaps qdev may be modified to align with.
Nothing stops you from creating qdev properties with a special "absent"
value. All you need is a special value that cannot be set.
The problem is that the default value is 0 and that means false. You can't
change it without either changing the default or requiring to set the
default explicitly which may need a lot of changes and error prone so we
should keep the current default to avoid this.
In fact, the humble "str" property already works that way: it's a char *
where null means "absent".
This case works as the default value matches the uninitialised value but
would not work for bool.
Code can recognize "absent" and do whatever needs doing then. For
instance, consider device "ide-cd". It has three such properties:
"ver", "serial", and "model". "ver" defaults to "2.5+", "serial" to
some unique string, but "model" defaults to NULL. Since you cannot set
such a value, it effectively means "absent". The code responsible for
this is in ide_dev_initfn():
if (!dev->version) {
dev->version = g_strdup(s->version);
}
if (!dev->serial) {
dev->serial = g_strdup(s->drive_serial_str);
}
Note it leaves a null dev->model null.
Using optional that way can occasionally lead to trouble. Consider
migrate-set-parameters. Its arguments are all optional. For each
argument present, the respective migration parameter is set to the
argument value. You cannot use this to reset a migration parameter from
present to absent. Matters for parameters where "absent" has a meaning
different from any "present" value.
End of digression.
Start of next digression :)
Note that qdev properties are generally optional. The only way to make
them mandatory is to reject their default value in .realize(). When
users set this default value explicitly, the error message will almost
certainly be confusing.
End of digression.
Optional bool may enable a fourth solution:
4. Make "absent" mean on if possible, else off, "present and true" mean
on if possible, else error, and "present and false" mean off (always
possible).
This changes the meaning of "present and true", but it's precisely
the change you want, isn't it?
We have "false by default" properties so it unfortunately does not work.
Then make the code make "absent" mean what you need it to mean. Just
like the code from ide_dev_initfn() I quoted above.
This would change the value for false as 0 is now taken to mean auto so
either all places that assume false need to be updated or if we use
different value for auto then all places need to explicitly set that and
you can't easily change a bool into OnOffAuto.
What is the core of the issue that you don't like about allowing on/off
for bool? If it's just some cases are not covered and still would not
accept alternative names would it be easier to fix those?
Regards,
BALATON Zoltan
Yet another solutions:
5. Alternate of bool and an enum with a single value "auto".
Falls apart with the keyval visitor used for the command line.
Fixable, I believe, but a good chunk of work and complexity.
I may have missed something, but I think that will break JSON string
literals "on" and "off".
Unbreaking it will be a good chunk of work and complexity, I believe.
My gut feeling: explore 4. first.
