This largely keeps in tact what jstutte did. The initial crash was fixed by eagerly calling
LoadingFinished. The second crash is caused because we call it twice, and only in the service worker
case, where we call it once the promise rejects. Now, we check if we have cancelled, and if we have
then we don't call the scriptLoader methods from inside of the load handlers. LoadHandlers now only
use OnStreamComplete if they are "successful" -- that is, if they were not cancelled.
OnStreamComplete retains its assertion error in the case that something was cancelled and we somehow
ended up there. In a follow up, I will clean up the friend classes of the ScriptLoader so you can't
easily access these methods from the LoadHandlers.
Differential Revision: https://phabricator.services.mozilla.com/D158262
This relates to 1784482, in that it allows modules to pass through the scriptloader independently.
This was a small bug that I caught in the wpt tests. I didn't have a better segment to put it under
so it is here.
Depends on D154382
Differential Revision: https://phabricator.services.mozilla.com/D147325
This brings back the behavior to iterated over all of the load requests to cancel the cache
promise. I tried a couple of different ways of doing this, including deleting the cache, but this
seemed to be the cleanest way.
Depends on D147322
Differential Revision: https://phabricator.services.mozilla.com/D154382
This relates to 1784482, in that it allows modules to pass through the scriptloader independently.
This was a small bug that I caught in the wpt tests. I didn't have a better segment to put it under
so it is here.
Depends on D154382
Differential Revision: https://phabricator.services.mozilla.com/D147325
This brings back the behavior to iterated over all of the load requests to cancel the cache
promise. I tried a couple of different ways of doing this, including deleting the cache, but this
seemed to be the cleanest way.
Depends on D147322
Differential Revision: https://phabricator.services.mozilla.com/D154382
Previously, we had the client info for main scripts on the ScriptLoadInfo, and in the ScriptLoader
for non-main scripts (ImportScripts case). This was a bit confusing, so I moved it all to the ScriptLoader as in the
importScripts case, it is always recreated with each new load. However, for modules this is not
true. In order to make it persistant for main scripts and modules, as well as ensure that it has the
correct data, it should always live on the WorkerLoadContext.
Differential Revision: https://phabricator.services.mozilla.com/D147317
This implements the core of the change -- which is implementing an explicit IsTopLevel flag for
`WorkerLoadContexts`, and removing `mIsMainScript` from the loader. I've asked for some
clarification on this part of the spec from the whatwg editors, as I think this may be misnamed. It
is likely that what is meant here is `IsInitialScript` rather than `IsTopLevel` -- as there is a
note stating that this should be initialized with the agent cluster. Once this is clarified I may
update the name.
Differential Revision: https://phabricator.services.mozilla.com/D147319
Cleanup, optional. Repeating the same work on the NetWorkLoaderHandler as on the WorkerScriptLoader.
Previously, this was considered "safe" because the assumption is that the NetworkLoadHandler is
shorter lived than the WorkerScriptLoader. Rather than assuming this, if we end up in a situation
where this does out-live the WorkerScriptLoader, then we will end up leaking.
Differential Revision: https://phabricator.services.mozilla.com/D154384
Cleanup, optional. It seems strange to have two ways to access the WorkerPrivate, and
ThreadSafeWorkerRef seems like the more reasonable choice.
Differential Revision: https://phabricator.services.mozilla.com/D154383
We don't need the mExecutionScheduled state, as the mLoadingFinished state was only used to
determine if the promise for a given request had resolved. In fact -- we already know that it is in
a possibly resolved state when we call MaybeExecuteFinishedScripts. So, we can remove this state,
and only have the meaningful check (whether or not a promise on the service worker case hasn't
resolved yet) instead.
Differential Revision: https://phabricator.services.mozilla.com/D146183
This field is no longer necessary, as we are removing executed scripts from our list of scripts to
execute, so we cannot enter a state where something may be executed twice.
Differential Revision: https://phabricator.services.mozilla.com/D146182
This is the most substantial change in the transition from ScriptLoadInfo to ScriptLoadRequest with
regards to data representation. ScriptLoadRequests can have their data incrementally loaded,
so it is already fully decoded and ready to go by the time that we create the source buffer for
worker scripts. This simplifies some of the code, and we can add incremental loading when we are ready.
Differential Revision: https://phabricator.services.mozilla.com/D146180
Here we split ScriptLoadInfo's representation -- we are distinguishing between "what is being loaded"
and "how it is being loaded in a worker context". Requesting review from asuth for the workers side
of things, and jonco for js/loader.
Differential Revision: https://phabricator.services.mozilla.com/D146175
This section of the patch queue starts the migration from ScriptLoadInfo to
ScriptLoadRequest/LoadContext pairs.
We will be making use of the ModuleLoader, and in the future we will have a unified ScriptLoader
implementation (currently, the skeleton for this is in ScriptLoaderInterface). ScriptLoadRequest has
been rewritten to be generic to all contexts, with a companion object "LoadContext" that will handle
specialized behavior required to load the request. That is exactly the case we have with workers,
most of the fields are generic with a couple of specialized fields associated with service workers.
Quick itemization of what will happen
* Patches 1-5: Focuses on getting things into place without using them, we rely on
ScriptLoadInfo (later renamed to WorkerLoadContext).
* Patches 6-7: This is a refactoring of what will be a shared data structure for decoding scripts.
As we will be using requests, we can use this pretty much in the form it exists in on the DOM, and
this cleanup makes it a bit nicer for the DOM as well (in my opinion anyway).
* Patches 8-12: We migrate all shared data to the standard utilization used by other loaders (DOM
and component loader). The biggest pieces is `.mScriptData` transitioning from a stream of bits
that haven't been decoded, to a decoded source text that can be used directly. This was enabled by
patches 6-7.
* The final patches are small cleanups and documentation.
Differential Revision: https://phabricator.services.mozilla.com/D146173
This step is not strictly necessary, and if this is too ugly we can think of something else. The
reason for this change is that ScriptLoadRequests are initialized by a URI, and the owner of these
objects will be the worker thread. However, the current strategy requires that we create the URI
only on the main thread. We always have this information however, when we create the loader, so
there isn't any reason to defer this step until we bounce back into the main thread.
This is another spot where we can reduce reliance on the main thread, which is an eventual goal, so
it seems like this change is an improvement even outside of the goal of moving to ScriptLoadRequests
as a representation.
An alternative approach here would be to initialized the nsIURI ahead of time, and pass it with the
script url to the initialization of the WorkerScriptLoader. However, this isn't ideal as _ideally_
we would remove the text url all together, but this is necessary for errors. We can do this if we
handle workers and main thread differently (and maybe it is worthwhile) as only the main thread
needs encoding information.
Differential Revision: https://phabricator.services.mozilla.com/D146170
This follows up the cancellation requirements -- CacheCreator is a main thread only datastructure,
which must be cleaned up once the last ScriptLoadInfo is dispatched. Previously, we determined this
via iteration. Now, we manage it as part of ScriptLoadInfo, and release our reference to it when we
dispatch a given ScriptLoadInfo. Later, once we transition to ScriptLoadRequest and LoadContexts,
the cache information can live on a ServiceWorkerLoadContext[1]. This will help use differentiate
between regular worker loader environments and service worker environments.
[1]: https://searchfox.org/mozilla-central/rev/997a56b018662e2940c99bbaf57a6ac9d1aa5422/js/loader/LoadContextBase.h#22-34
Differential Revision: https://phabricator.services.mozilla.com/D145448
The joys of multi-threaded programming:
As you likely noticed, we now have list modification where previously we didn't.
We need move all list modification on the worker side, not the main thread side.
The only reason this isn't showing up on thread sanitizer now, is because the main
thread is acting as the owner of the the lists.
In order to implement Worker ownership, we need to remove all list traversal from the main thread
runnable. There are two significant cases in which we traverse lists: Cancellation, and Dispatching.
Here we handle part of the story for cancellation.
The strategy we take is fundamentally the same as in the DOM Script Loader -- We use the
"mCanceledMainThread" flag to indicate to our load handlers that the thread has been cancelled.
This is taken directly from the DOM loader [1] approach.
[1]: https://searchfox.org/mozilla-central/rev/997a56b018662e2940c99bbaf57a6ac9d1aa5422/dom/script/ScriptLoadHandler.cpp#128-132
Differential Revision: https://phabricator.services.mozilla.com/D145446
The goal of this patch stack is to transition from using the ScriptLoadInfo to represent script
requests for workers, to ScriptLoadRequest and its partner datastructure ScriptLoadRequestList.
There are a number of differences between these two datastructures, but they fundamentally represent
the same thing: a request which is either about to be loaded, or about to be executed.
To get there, in a reasonable manner, is a bit tricky. ScriptLoadRequest et all are refcounted, and
restrict which thread can delete or add them. The decision I've made in this transformation is to
make the Worker thread the home of ScriptLoadRequests. This means that we have to change how we are
doing things on the WorkerScriptLoader so that the MainThread is largely unaware of the
ScriptLoadRequestList, and it only operates in terms of ScriptLoadRequests (or, ScriptLoadInfos as
they are known for now)
The following need to be changed:
* We need to prepare everything to operate on pointers (Prep Patches 1-3)
* Cancellation must no longer iterate over a list (Prep patches 4-8, removing the mChannel reference
is a bonus)
* List management must only be done on the Worker side (Prep Patches 9-10)
This first patch transitions from using the ScriptLoadInfo directly to passing around references.
You will find a similar explaination for the actual transition in the patches that follow the prep
patches. This is bundled in the same bug as the goal here remains the same.
Differential Revision: https://phabricator.services.mozilla.com/D145441
This attempts to make a difference between "true" private classes and ones accessed by friends (even
though friends can get at private members...)
Differential Revision: https://phabricator.services.mozilla.com/D145088
Iulian Moraru
Yulia Startsev
Jens Stutte
Cosmin Sabou
Noemi Erli
Butkovits Atila
Sean Feng
Marian-Vasile Laza
Tom Schuster
Kagami Sascha Rosylight