Before this commit, most of the time, components are destroyed when the
virtual dom is patched and a component node is removed. However, since
Owl is asynchronous and a component may take some time to get ready
(with onWillStart), it can happen that a component is created, then
before it is ready, it is recreated. In that case, the initial instance
has to be destroyed.
Before this commit, the destroy operation was done immediately, when we
cancel the current fibers. However, this means that we cannot have a
guarantee between micro task ticks that a component has not been
destroyed in the meantime.
For example, in Odoo, it is common to use the rpc service, which will
throw an error if called by a destroyed component. But because of the
possible destruction of a component at any time, the following code is
unsafe:
async loadSomeData() {
// guaranteed to be called when component is alive
await Promise.resolve();
// however here, component may have been destroyed
this.rpc(...)
}
So, to prevent this issue, we can slightly delay the destroy operation.
It is not entirely trivial, since we need to find a way to neutralize
the component in the meantime. But it seems like performing all that
kind of operation at the "commit" phase (so, the request animation frame
callback) makes sense to me.
So, this commit modifies the code to add a new component status
(cancelled) and use it to cancel components that are waiting to be
destroyed. These components will then be destroyed as soon as the
requestanimation frame starts, before all other dom operations.
Before this commit, in some specific situations, owl could skip updating
the DOM, even though it should. Here is what could happen:
- we have a parent component (A) and a child component (B)
- B depends on some reactive props from A
- we update some state that results in A being rendered, and B updated
- before this render is applied to the dom, we update some state again,
which causes A to be rendered again
- however, this new render is such that the B props are now identical,
so it will skip rendering B
- it will then apply the result of the render to the DOM => only A is
updated but B should also be updated!
The problem comes from the fact that Owl committed the props to the
component right after the rendering, to the new props were used in the
props comparison method to decide if B should be updated. This is
incorrect, we should always use the current props to decide what to do,
and only commit them to the component after it has been patched.
Previously, if the value of a t-component directive changed, but there
was already a scheduled render for the component before that change, the
rendering of the existing component would get delayed (as it should),
but after the parent's rendering was complete, the old component which
will be destroyed during the patch would still get rendered, despite
being stale. This can cause crashes if that component relies on state
that no longer exists.
This was caused by the fact that when rendering, we check the parent
chain for an active render, and also check that the component still
exists in the parent's fiber childrenMap (ie, we know that the upcoming
patch is not about to destroy the component). To do this, we use the
component's parentKey, but in the case of t-component, the parentKey for
both possible components is the same, causing the stale component to
incorrectly believe that it's not about to be destroyed, by finding the
next component in the childrenMap under the same key.
This commit fixes that by prepending the component's name to the
parentKey, meaning that if the value of the t-component changes, so will
the key, and the render will be further delayed until the new state is
patched, at which point the stale component will be destroyed and the
following attempt to render will be cancelled as expected.
The choice to use the component's name is to simplify the
implementation, this means that if using t-component and switching
between components that have the same class name, this protection will
not work. The alternative would be to generate a unique id for the
component class, eg with a WeakMap, but this both complicates the
implementation and adds runtime overhead, for a case that is likely
extremely rare. We are open to refine the implementation should this
problem occur in practice.
Previously, if a fiber was delayed because one of its ancestors was
rendering, and that fiber was not a root fiber, it would be rendered
after all its ancestors had finished rendering even if one of those
ancestor renderings cancelled it.
This commit fixes that by simply checking that a delayed fiber is still
its component node's current fiber before rendering it.
Before this commit, the list of all children was managed at the level of
the root fiber, but this could cause issue when subfibers would be
reused. With this commit, we use the childrenMap object that exists on
each fiber instead.
Canceling a fiber may cause user code to be run, which means that some
new renderings could be scheduled, but this could interfere with the
current renderings!
This is a breaking semantic change. With this commit, the UI is frozen
whenever owl is waiting for a parent to change
Also, this allows Owl not to render components that will be removed
later.
With this commit, component only render child
components if they have different props (shallow
equality). Otherwise, we trust the reactivity
system to make sure that all impacted components
are updated
Every use case involving some sort of key set on a component would give birth to a leak in an async context:
- If a key of a component changed, the outdated one was never destroyed.
- destroyed component were never removed from their parent's reference map.
This commit solves both issues, that are tightly linked anyway.
Have a component which does a render in its onWillPatch, onPatched, onMounted hooks.
Before this commit, the result was incorrect: the second rendering was not taken into account.
After this commit, those renderings are correctly applied at the price of a delayed render when the fiber
is in a critical state.
Have a child component on which a render is triggered.
This component delays its willUpdateProps and makes a rendering during the willUpdateProps
Before this commit, renderings of the child were inconsistent across
its parent's renderings.
After this commit, it works as expected.
When a parent and a child were rendered at the same time, it was
possible for the 2 renders to decrement the same fiber internal
counter, which meant that the render was stalled.
Aim to replace the abstraction "Context" from OWL 1 with the new primitives
"atom" and "useState":
- notification is done only after a batch of modifications.
- observers are notified at most once for a batch.
- an observer of type component is notified (and rerendered)
only if it does not have an ancestor that has to be notified for the
same batch of operations (anywhere in the web of references!).
- notification of components is done on all levels "simultaneously".
Co-authored-by: Aaron Bohy <aab@odoo.com>
Co-authored-by: Géry Debongnie <ged@odoo.com>
Co-authored-by: Mathieu Duckerts-Antoine <dam@odoo.com>
The point is to have visibility on the development of the owl2 features.
This commit reintroduces some tests keeping them skipped in order to fulfill that purpose.
There still are some missing tests though.
Géry Debongnie