The fundamental goal of document splitting is to fix a specific performance issue. The architecture of WebRender is such that any time a content process sends a new display list to the compositor, WebRender needs to do a full scene build that includes both the content and chrome areas. Likewise if the chrome process sends a new display list to the compositor. This means that animations such as the tab loading spinner or an animated gif will cause much more WR activity than is really needed.
With document splitting, the WR scene is split into two (or more) documents based on the visual location of the elements. Everything in the "chrome area" of the window (including the tab bar, URL bar, navigation buttons, etc.) are considered part of one document, and anything below that (including content, devtools, sidebars, etc.) is in another document. Bug 1549976 introduces a third "popover" document that encompasses elements that straddle both of these visually, but that's a side-effect of the implementation rather than being driven by the fundamental problem being addressed.
With the documents split like so, an animation in the UI (such as the tab loading spinner) runs independently of the content (by virtue of being in a different document) and vice-versa, which results in better user-perceived performance.
Document splitting is so called because inside the WR code and bindings, a "document" is an independent pathway to the compositor that has its own scene, render tasks, etc.
In most of the C++ code in gfx/layers and other parts of Firefox/Gecko, the term "render root" is used instead. A "render root" is exactly equivalent to a WR document; the two terms are used interchangeably. The naming is this way because "document" has different pre-existing meanings in Gecko-land. In this documentation those other meanings are irrelevant and so "document" always refers to the same thing as "render root".
At various points there have been discussions to renaming things so that everything is more consistent and less confusing, but that hasn't happened yet.
The wr::RenderRoot data type is an enumeration of all the different documents that are expected to be created. As of this writing, the enumeration contains two entries: Default and Content. The Default document refers to the one that holds the "chrome" stuff and the Content document refers to the one that holds the "content" area stuff.
If document splitting is disabled (gfx.webrender.split-render-roots=false) then everything lives in the Default document and the Content document is always empty.
Additional data structures in the same file (e.g. wr::RenderRootArray<T>) facilitate converting pre-document-splitting code into document-splitting-aware code, usually by turning a single object into an array of objects, one per document.
The notion of having multiple documents has to be introduced at a fairly fundamental level in order to be propagated through the entire rendering pipeline. It starts in the front-end HTML/XUL code where certain elements are annotated as being the transition point between documents. For example, this code explicitly identifies an element and its descendants as being in the "content" document instead of the "default" (or "chrome") document.
These attributes are read during display list building to create the nsDisplayRenderRoot display item in the Gecko display list.
When the Gecko display list is processed to create a WebRender display list, it actually ends up creating multiple WR display lists, one for each document. This is necessary because the documents are handled independently inside WR, and so each get their own WebRenderAPI object and separate display list. The way the implementation manages this is by creating a "sub builder" for the render root that is being descended into, and using that instead of the main WR display list builder as the display list is recursed into.
Note also that clip chains and stacking contexts are per-document, so when recursing past a nsDisplayRenderRoot item, the ClipManager and StackingContextHelper being used switches to one specific to the new document. For this to work there are certain assumptions that must hold, which are described in the next section. Other things that must now be managed on a per-document basis are generally encapsulated into the RenderRootStateManager class, and the WebRenderLayerManager holds an array of these.
After the Gecko display list is converted to a set of WebRender display lists (one per document), these are sent across IPC along with any associated resources as part of the WebRender transaction. Conceptually, the parent side simply demultiplexes the data for different documents, and submits the data for each document to the corresponding WebRenderAPI instance.
One of the fundamental issues with the document splitting implementation is that we can have stuff in the UI process that's part of the "content" renderroot (e.g. a sidebar that appears to the left of the content area). The expectation for front-end authors would be that this would be affected by ancestor elements that are also in the UI process. Consider this outline of a Gecko display list:
- Root display item R - ... stuff here (call it Q) ... - display item P - display item A - display item B (flagged as being in the content renderroot) - display item C
If item P was a filter, for example, that would normally apply to all of items A, B, and C. This would mean either sharing the filter between the "chrome" renderroot and the "content" renderroot, or duplicating it such that it existed in both renderroots. The sharing is not possible as it violates the independence of WR documents. The duplication is technically possible, but could result in visual glitches as the two documents would be processed and composited separately.
In order to avoid this problem, the design of document splitting explicitly assumes that such a scenario will not happen. In particular, the only information that gets carried across the render root boundary is the positioning offset. Any filters, transforms that are not 2D axis-aligned, opacity, or mix blend mode properties do NOT get carried across the render root boundary. Similarly, a scrollframe may not contain content from multiple render roots, because that would lead to a similar problem in APZ where it would have to update the scroll position of scrollframes in multiple documents and they might get composited at separate times, resulting in visual glitches.
On the content side, all of the document splitting work happens in the UI process. In other words, content processes don't generally know what document they are part of, and don't ever split their display lists into multiple documents. Only the UI process ever sends multiple display lists to the compositor side.
There are a number of APIs on PWebRenderBridge where a wr::RenderRoot is passed across from the content side to the compositor side. And since PWebRenderBridge is a unified protocol that is used by both the UI process and content processes to communicate with the compositor, the content processes must provide some value for the wr::RenderRoot. But since it doesn't (or shouldn't) be aware of what document it's in, it must always pass wr::RenderRoot::Default.
Compositor-side code in WebRenderBridgeParent is responsible for checking that any wr::RenderRoot values provided from a content process are in fact wr::RenderRoot::Default. If this is not the case, it is either a programmer error or a hijacked content process, and appropriate handling should be used. In particular, the compositor side code should never blindly use the wr::RenderRoot value provided over the IPC channel as hijacked content processes could force the compositor into leaking information or otherwise violate the security and integrity of the browser. Instead, the compositor is responsible for determining where the content is attached in the display list of the UI process, and determine the appropriate document for that content process. This information is stored in the WebRenderBridgeParent::mRenderRoot field.
This section describes various knots of complexity in the document splitting implementation. That is, these pieces are thought to introduce higher-than-normal levels of complexity into the feature, and should be handled with care.
When a display list transaction is sent from the content side to the compositor, APZ is also notified of the update, so that it can internally update its own data structures. One of these data structures is a tree representation of the scrollable frames on the page. With document splitting, the scrollable frames may now be split across multiple documents. APZ needs to record which document each scrollable frame belongs to, so that when providing the async scroll offset to WebRender, it can send the scroll offset for a given a scrollable frame to the correct WebRender document. As one might expect, this is stored in the mRenderRoot field in the AsyncPanZoomController (there is one instance of this per scrollable frame).
Additionally, when new display list transactions and other messages are received in WebRenderBridgeParent, APZ cannot process these updates right away. Doing so would cause APZ to respond to user input based on the new display list, while the WebRender internal state still corresponds to the old display list. To ensure that APZ and WR's internal state remain in sync, APZ puts these update messages into an "updater queue" which is processed synchronously with the WebRender scene swap. This ensures that APZ's internal state is updated at the same time that WebRender swaps in the new scene, and everything stays in sync. Conceptually this is relatively simple, until we add document splitting to the mix.
Now instead of one scene swap, we have multiple scene swaps happening, one for each of the documents. In other words, even though WebRenderBridgeParent gets a single "display list transaction", the display lists for the different documents modify WR's internal state at different times. Consequently, to keep APZ in sync, we must apply a similar "splitting" to the APZ updater queue, so that messages pertaining to a particular document are applied synchronously with that document's scene swap.
(As a relevant aside: there other messages that APZ receives over other IPC channels (e.g. PAPZCTreeManager) that have ordering requirements with the PWebRenderBridge messages, and so those also normally end up in the updater queue. Consequently, these other messages are also now subjected to the splitting of the updater queue.)
Again, conceptually this is relatively simple - we just need to keep a separate queue for each document, and when an update message comes in, we decide which document a given update message is associated with, and put the message into the corresponding queue. The catch is that often these messages deal with a specific element or scrollframe on the page, and so when the message is sent from the UI process, we need to do a DOM or frame tree walk to determine which render root that element is associated with. There are some GetRenderRootForXXX helpers in gfxUtils that assist with this task.
The other catch is that an APZ message may be associated with multiple documents. A concrete example is if a user on a touch device does a multitouch action with different fingers landing on different documents, which would trigger a call to RecvSetTargetAPZC with multiple targets, each potentially belonging to a different render root. In this case, we need to ensure that the message only gets processed after the corresponding scene swaps for all the related documents. This is currently implemented by having each message in the queue associated with a set of documents rather than a single document, and only processing the message once all the documents have done their scene swap. In the example above, this is indicated by building the set of render roots here and passing that to the updater queue when queueing the message. This interaction is a source of some complexity and may have latent bugs.
Bug 1547351 provided a new and tricky problem where a content process is rendering stuff that needs to go into the "default" document because it's actually an out-of-process addon content that renders in the chrome area. Prior to this bug, the WebRenderBridgeParent instances that corresponded to content processes (hereafter referred to as "sub-WRBPs", in contrast to the "root WRBP" that corresponds to the UI process) simply assumed they were in the "Content" document, but this bug proved that this simplistic assumption does not always hold.
The solution chosen to this problem was to have the root WebRenderLayerManager (that lives in the trusted UI process) to annotate each out-of-process subpipeline with the render root it belongs in, and send that information over to the root WRBP as part of the display list transaction. The sub-WRBPs know their own pipeline ids, and therefore can find their render root by querying the root WRBP. The catch is that sub-WRBPs may receive display list transactions before the root WRBP receives the display list update that contains the render root mapping information. This happens in cases like during tab switch preload, where the user mouses over a background tab, and we pre-render it (i.e. compute and send the display list for that tab to the compositor) so that the tab switch is faster. In this scenario, that display list/subpipeline is not actually rendered, is not tied in to the display list of the UI process, and therefore doesn't get associated with a render root.
When the sub-WRBP receives a transaction in a scenario like this, it cannot actually process it (by sending it to WebRender) because it doesn't know which WR document it associated with. So it has to hold on to it in a "deferred update" queue until some later point where it does find out which WR document it is associated with, and at that point it can process the deferred update queue.
Again, conceptually this is straightforward, but the implementation produces a bunch of complexity because it needs to handle both orderings - the case where the sub-WRBP knows its render root, and the case where it doesn't yet. And the root WRBP, upon receiving a new transaction, would need to notify the sub-WRBPs of their render roots and trigger processing of the deferred updates.
Further complicating matters is Fission, because with Fission there can be pipelines nested to arbitrary depths. This results in a tree of sub-WRBPs, with each WRBP knowing what its direct children are, and only the root WRBP knowing which documents its immediate children are in. So there could be a chain of sub-WRBPs with a "missing link" (i.e. one that doesn't yet know what its children are, because it hasn't received a display list transaction yet) and upon filling in that missing link, all the descendant WRBPs from that point suddenly also know which WR document they are associated with and can process their deferred updates.
Managing all this deferred state, ensuring it is processed as soon as possible, and clearing it out when the content side is torn down (which may happen without it ever being rendered) is a source of complexity and may have latent bugs.
Transactions between the content and compositor side are throttled such that the content side doesn't go nuts pushing over display lists to the compositor when the compositor has a backlog of pending display lists. The way the throttling works is that each transaction sent has a transaction id, and after the compositor is done processing a transaction, it reports the completed transaction id back to the content side. The content side can use this information to track how many transactions are inflight at any given time and apply throttling as necessary.
With document splitting, a transaction sent from the content side gets split up and sent to multiple WR documents, each of which are operating independently of each other. If we propagate the transaction id to each of those WR documents, then the first document to complete its work would trigger the "transaction complete" message back to the content, which would unthrottle the next transaction. In this scenario, other documents may still be backlogged, so the unthrottling is undesirable.
Instead, what we want is for all documents processing a particular transaction id to finish their work and render before we send the completion message back to content. In fact, there's a bunch of work that falls into the same category as this completion message - stuff that should happen after all the WR documents are done processing their pieces of the split transaction.
The way this is managed is via a conditional in HandleFrame. This code is invoked once for each document as it advances to the rendering step, and the code in RenderThread::IncRenderingFrameCount acts as a barrier to ensure that the call chain only gets propagated once all the documents have done their processing work.
I'm listing this piece as a potential source of complexity for document splitting because it seems like a fairly important piece but the relevant code is "buried" away in a place where one might not easily stumble upon it. It's also not clear to me that the implications of this problem and solution have been fully explored. In particular, I assume that there are latent bugs here because other pieces of code were assuming a certain behaviour from the pre-document-splitting code that the post-document-splitting code may not satisfy exactly.