Accessibility/CacheTheWorld

From MozillaWiki
Jump to: navigation, search

What?

Firefox's current architecture for multi-process accessibility suffers from severe performance issues and is costly and difficult to maintain due to the massively different and specialised approaches necessary on different operating systems. In addition, it is currently impossible to support builtin Windows accessibility tools such as Narrator and Windows Speech Recognition. This project aims to re-architect our multi-process accessibility support to cache the entire accessibility trees for all content processes within the parent process.

Why?

This will allow us to address several problems that are difficult or impossible to fix with the current architecture:

  1. Performance, especially on Windows. While performance is acceptable for daily usage in most cases, there are many use cases which are far from delightful and some which are still completely unusable. Performance with the JAWS screen reader, which is used heavily in enterprise, is sluggish at best. A great deal of work has been done over the past few years to improve this, but we're approaching a point where we will not be able to improve this any further with the current architecture. Because software other than assistive technology uses accessibility APIs (e.g. Windows touch, East Asian input methods, enterprise SSO tools), this can impact even users without disabilities. The proposed new architecture will allow us to significantly improve performance across all operating systems. See bug 1737192 for a list of performance bugs which we expect will be fixed (or at least improved) by Cache the World.
  2. Stability. Because the current architecture makes heavy use of synchronous IPC, there is a high risk of deadlocks between accessibility and other Firefox components. While all known cases have been addressed as they have been discovered, the underlying cause remains and future instances of this problem are very likely. In addition, our use of obscure COM features on Windows has resulted in stability problems which are extremely difficult to diagnose and fix. One of these remains a problem today, despite months of investigation, and forces some NVDA screen reader users to forcibly kill Firefox (or worse, forcibly power off their computers) every few hours. The proposed new architecture will not suffer from these inherent stability risks and known problems. See bug 1737193 for a list of stability/reliability bugs which we expect will be fixed (or at least improved) by Cache the World.
  3. Complexity and cost. Our existing architecture is necessarily very different on each operating system, making it extremely complex and difficult to maintain. For example, the IPC layer on Windows (~8000 lines of code) has an entirely different architecture to other platforms. This also means that maintenance is very costly, especially when implementing support for new operating systems (e.g. Android, Mac) or major Gecko architectural initiatives (e.g. Fission). Furthermore, our use of esoteric operating system specific features, especially on Windows (where we depend on a whole separate ~11000 line module), makes it very difficult for this work to be distributed across the team because of the highly specific expertise required. In the proposed new architecture, most of the "heavy lifting" will be done in cross-platform code, decreasing complexity and maintenance cost.
  4. Support for other builtin accessibility tools. It is impossible to support Windows Narrator and Windows Speech Recognition with our current architecture. The proposed new architecture will allow for this. As these builtin tools rise in popularity, we do not want Firefox to be left behind.

Meeting Notes

Jamie, Morgan, and Eitan meet weekly to discuss this project. You can find the meeting notes in this google doc.

Roadmap

Given the large scope of the project, we are breaking the project down into quarterly milestones. Each milestone will aim to support a set of user scenarios. This roadmap is in the early stages and subject to significant change. It will be updated as milestones become clearer, with future milestones being less well defined than earlier ones.

Milestone 0: December 2021

Initial proof of concept.

All testing will be performed with the NVDA screen reader, for two reasons:

  1. Cache the World is all or nothing on Windows. That makes it easy to determine where we're at regarding real world usage.
  2. The performance benefits are most necessary and noticeable on Windows.

In milestone 0, the following capabilities will be provided:

  1. Reading and navigating a page with text, links and headings.
  2. Plain text editing: reading the line of text when focused; backspacing; moving the caret by line, word and character.
  3. Access to formatting information: font, bold/italic, etc.
  4. Access to screen coordinates on simple pages.
  5. Loading very large pages will be at least 10x faster with the cache than without.

Test Scenarios

  1. Do a Google search, navigate the results using heading navigation and follow a result link.
  2. Fill out the form for a Google Advanced Search.
  3. Go to https://www.reaper.fm. Check the formatting of the “This is REAPER.” text (which isn’t a heading even though it should be) and confirm that the font size is reported as bigger than the paragraph of text below it.
  4. Build Gecko in the background. Do a Google search. Verify that the browser does not become unresponsive.
  5. Load https://searchfox.org/mozilla-central/source/layout/base/nsCSSFrameConstructor.cpp. Page should take < 10 sec to be usable.
  6. Open Gmail, find a message in the inbox, open it, read it, return to the inbox.
  7. Compose a message in Gmail containing text, a link, a bulleted list and a block quote. Read back through the message.
  8. Open Slack, use the quick switcher to switch to a channel, read some messages, write a message.

Bugzilla

Note that the roadmap wasn't created until late in milestone 0, so many bugs are missing below.

Full Query
ID Summary Assigned to Status
1735955 Cached bounds all 0s for many (most?) elements on Google search James Teh [:Jamie] RESOLVED
1739050 If the focused Accessible is moved, the RemoteAccessible is recreated but focus isn't fired on it (AKA broken Google Search box on Windows + CTW) James Teh [:Jamie] RESOLVED
1741792 Cache the caret James Teh [:Jamie] RESOLVED
1742902 Fix window emulation when the cache is enabled James Teh [:Jamie] RESOLVED
1742915 Cache tag object attribute James Teh [:Jamie] RESOLVED
1742917 Implement StartOffset for RemoteAccessible and LinkIndexAtOffset for HyperTextAccessibleBase James Teh [:Jamie] RESOLVED
1746827 Crash in [@ PLDHashTable::Search | mozilla::a11y::RemoteAccessibleBase<T>::MinValue] James Teh [:Jamie] RESOLVED

7 Total; 0 Open (0%); 7 Resolved (100%); 0 Verified (0%);


Milestone 1: March 2022

Android.

The primary focus of this milestone is getting the cache working for Android. Mozilla aims to implement Fission for Android in 2022h1. Modifying the existing multi-process architecture to support Fission on Android will require significant engineering effort. Rather than investing in a solution which we will be throwing away once the cache is implemented, we will instead switch Android to use the cache and extend the cache to include functionality required by Android.

In milestone 1, the following capabilities will be provided:

  1. The cache will support GroupPosition.
  2. TextLeafRange will support word end and line end boundaries, which are needed for Android text navigation.
  3. Pivot will support navigating text using TextLeafRange, which is needed for Android text navigation.
  4. Cached screen bounds will be updated appropriately when scrolling.
  5. The cache will support tables. Only a very small subset of table functionality is needed for Android, but full table functionality will be implemented regardless.
  6. Android will use the cache for all functionality except hit testing.
  7. As an interim solution, Android will use the existing async IPDL mechanism for hit testing, updated to target the call at the correct document to handle OOP iframes. (Synchronous hit testing in the core cache will take longer to implement and will be done in a future milestone.)

Test Scenarios

  1. Do a Google search, navigate the results using heading navigation and follow a result link.
  2. Sign up for facebook, enter information including alternative gender. Don’t submit password (that is when it gets real).
  3. Go to en.wikipedia.org and navigate by char/word/line.
  4. Navigate en.wikipedia.org with explore by touch.
  5. With Talkback, load https://www.nvaccess.org/. Navigate to the embedded video in four different ways: item navigation, explore by touch, controls navigation and Talkback search. Activate the Play button to play the video.
  6. On Windows with NVDA, load https://developer.mozilla.org/en-US/docs/Web/API/Window/getSelection.
    1. Press control+home to ensure you're at the top of the document, then t twice to move to the Browser Compatibility table.
    2. Press control+alt+rightArrow twice to move to the Mobile column. Ensure NVDA says "column 8 through 13".
    3. Press control+alt+downArrow once. Ensure NVDA says "WebView Android".
    4. Press control+alt+downArrow again. Ensure NVDA says "getSelection" and "Full support".
    5. Press control+alt+leftArrow. Ensure NVDA says "Safari", "Desktop" and "Full support".

Bugzilla

Full Query
ID Summary Assigned to Status
1730085 TextLeafPoint: Implement BOUNDARY_WORD_END and BOUNDARY_LINE_END James Teh [:Jamie] RESOLVED
1730086 TextLeafPoint: Implement HyperTextAccessible::TextBefore/AfterOffset using TextLeafPoint James Teh [:Jamie] RESOLVED
1730090 TextLeafPoint: Implement BOUNDARY_PARAGRAPH James Teh [:Jamie] RESOLVED
1733268 Cache OPAQUE1 state (no state change event) Morgan Reschenberg [:morgan] RESOLVED
1735970 New table implementation which can work with the cache James Teh [:Jamie] RESOLVED
1737944 Implement GroupPosition for cached RemoteAccessibles Eitan Isaacson [:eeejay] RESOLVED
1741793 Cache text selection James Teh [:Jamie] RESOLVED
1746833 Implement textual value for cached RemoteAccessibles Eitan Isaacson [:eeejay] RESOLVED
1747172 Make Pivot text functions use HyperTextAccessibleBase James Teh [:Jamie] RESOLVED
1747835 Cache text input type Eitan Isaacson [:eeejay] RESOLVED
1748749 Cache ActionCount/ActionNameAt Eitan Isaacson [:eeejay] RESOLVED
1749828 Support cached grouppos in windows and xpcom Eitan Isaacson [:eeejay] RESOLVED
1752380 Support actions in linkable accessible's indirect actions Eitan Isaacson [:eeejay] RESOLVED
1754360 Move EndOffset to base Accessible James Teh [:Jamie] RESOLVED
1755420 Clean up HyperTextAccessible methods now that all boundaries are supported by TextLeafPoint James Teh [:Jamie] RESOLVED
1755817 Implement cached version of LandmarkRole Eitan Isaacson [:eeejay] RESOLVED
1756229 Cache scroll offset Morgan Reschenberg [:morgan] RESOLVED
1756528 TextRange::Crop doesn't crop siblings James Teh [:Jamie] RESOLVED
1757812 Make android pivot traversal rules remote friendly. Eitan Isaacson [:eeejay] RESOLVED
1758009 Description is not proxied in RemoteAccessibleWrap Eitan Isaacson [:eeejay] RESOLVED
1758540 Pref on a11y.cache.enabled in Android Eitan Isaacson [:eeejay] REOPENED
1758559 Add cross domain iframe tests to Android a11y tests Eitan Isaacson [:eeejay] RESOLVED
1758592 Use pivot result to determine boundaries in parent-side traversal Eitan Isaacson [:eeejay] RESOLVED
1758811 Stop using java cache when gecko a11y cache is enabled Eitan Isaacson [:eeejay] RESOLVED
1760735 Support TableAccessibleBase in AccGroupInfo James Teh [:Jamie] RESOLVED
1765430 Remove RemoteAccessibleWrap and map accessibles to their IDs in SessionAccessibility. Eitan Isaacson [:eeejay] RESOLVED
1765433 Make GetNodeInfo and Pivot methods run safely in Android UI thread Eitan Isaacson [:eeejay] RESOLVED
1768726 Make RemoteAccessibleBase::BoundsWithOffset thread safe for Android Eitan Isaacson [:eeejay] RESOLVED
1768913 Don't send WINDOW_CONTENT_CHANGED when about:blank loads Eitan Isaacson [:eeejay] RESOLVED
1768927 Don't use StringBundle off main thread Eitan Isaacson [:eeejay] RESOLVED
1768972 DocAccessibleParent::AddChildDoc calls RecvEvent causing monitor lock reentry Eitan Isaacson [:eeejay] RESOLVED
1769640 Cache AppUnitsPerDevPixel in top DocAccessibleParents Eitan Isaacson [:eeejay] NEW

32 Total; 2 Open (6.25%); 30 Resolved (93.75%); 0 Verified (0%);


Milestone 2: June 2022

Opt-in user preview.

This milestone will focus on the work required to get most daily usage working well on Windows and Mac. Before the end of this milestone, we will invite interested users to enable the cache in Firefox Nightly, test it and provide any feedback.

In milestone 2, the following capabilities will be provided:

  1. Text implementation on Mac which utilises the cache.
  2. Caching of additional properties, including document URL, access key/keyboard shortcut and ARIA attributes such as aria-current.
  3. Missing cache updates; e.g. coalesced selection events, name change events for selection changes in Gmail message lists.
  4. Caching of spelling errors.
  5. Caching of relations (labelledBy/labelFor, etc.).
  6. Support for live regions.
  7. Caching of text screen coordinates.
  8. Support for interaction triggered by clients, including scrolling.
  9. Hit testing.
  10. Any fixes required to get JAWS working correctly.

Test Scenarios

  1. Relevant tests from m0 (which was NVDA only), but with JAWS and VoiceOver:
    1. Do a Google search, navigate the results using heading navigation and follow a result link.
    2. Fill out the form for a Google Advanced Search.
    3. Go to https://www.reaper.fm. Check the formatting of the “This is REAPER.” text (which isn’t a heading even though it should be) and confirm that the font size is reported as bigger than the paragraph of text below it.
    4. Open Gmail, find a message in the inbox, open it, read it, return to the inbox.
    5. Compose a message in Gmail containing text, a link, a bulleted list and a block quote. Read back through the message.
    6. Open Slack, use the quick switcher to switch to a channel, read some messages, write a message.
  2. NVDA, JAWS: Open https://www.mozilla.org/en-US/about/manifesto/. Press h several times to move to the "Principle 2" heading. In the same tab, open https://google.com/. Press alt+leftArrow to go back. Ensure the screen reader cursor is still near "Principle 2". (Tests document URL caching.)
  3. NVDA, JAWS, VoiceOver: Open https://wiki.mozilla.org/. Move to the navigation landmark. Ensure that Edit and View history report access keys (on Windows, alt+shift+e and alt+shift+h, respectively).
  4. TBD: Real world aria-current test case.
  5. NVDA, JAWS, VoiceOver: Open the Gmail inbox. Focus a message in the list. Press x to select it. Ensure the screen reader reports that the message is selected.
  6. NVDA, JAWS, VoiceOver: Open Slack. Focus the message box. Type "tset" and press space. Ensure the screen reader reports a spelling error.
  7. TBD: Real world aria-labelledby test case where the relation matters.
  8. TBD: JAWS: Real world aria-controls test case.
  9. NVDA, JAWS, VoiceOver: Open a document in Google Docs. Press control+alt+a, control+alt+c. Ensure the screen reader says "Not on a comment".
  10. NVDA, JAWS (is this possible with VoiceOver?): Open https://google.com/. In the search box, type "asdf". Move the screen reader's review cursor to "s", route the mouse and simulate a left click. Verify that the caret is now on "s".
  11. NVDA, JAWS, VoiceOver: Open https://www.nvaccess.org/. Move the screen reader's cursor to the bottom of the page. Verify that the bottom of the page is scrolled into view and that the top of the page is scrolled out of view.
  12. NVDA, JAWS, VoiceOver: Open https://google.com/. Move the mouse to About, Store, the Search button, I'm Feeling Lucky, Privacy, Terms and Settings. Ensure that the screen reader reports these as you move the mouse to them.
  13. NVDA, JAWS, VoiceOver: Open a blank document in Google Docs. Add a heading, some text wrapping across two lines, a bulleted list and a link. Read back through the text and verify that it is read correctly. Go to the start of the first line of wrapped text. Type two additional words so that the line wrapping changes. Read the first line and verify that it is reported correctly, accounting for the changed wrapping.

Bugzilla

Full Query
ID Summary Assigned to Status
1699339 Fire name/description change event when text in hidden aria-labelled/describedby subtree changes NEW
1730095 Implement Mac text navigation using TextLeafPoint/Range NEW
1733417 [meta] Cache relations NEW
1735101 Cache text screen bounds Morgan Reschenberg [:morgan] RESOLVED
1737919 TextLeafPoint: Implement support for spelling errors James Teh [:Jamie] NEW
1739560 Update cache when line boundaries change James Teh [:Jamie] RESOLVED
1743966 Implement ScrollTo for cached RemoteAccessible on Windows James Teh [:Jamie] RESOLVED
1748393 Update cache when title of document changes James Teh [:Jamie] RESOLVED
1748755 Support live regions with the cache enabled on Windows James Teh [:Jamie] NEW
1755377 Coalesced selection events drop SELECTED state change events NEW
1756728 Allow a11y clients to query URL for cached remote documents NEW
1757121 Cache live region attributes NEW
1757126 Cache AccessKey and KeyboardShortcut NEW
1758689 Implement CTW hit testing Morgan Reschenberg [:morgan] NEW
1760739 Expose ARIA object attributes for cached RemoteAccessibles NEW
1760939 Move LinkCount to HyperTextAccessibleBase James Teh [:Jamie] RESOLVED
1763211 Cache text bounds for horizontal (LtR and RtL) text Morgan Reschenberg [:morgan] RESOLVED
1764249 With the cache enabled, document incorrectly reports text selection when there is a caret James Teh [:Jamie] VERIFIED
1764499 [CTW] JAWS virtual cursor won't report formatting info NEW
1764500 [CTW] JAWS treats a document as a read only edit if it contains a contentEditable James Teh [:Jamie] RESOLVED
1765460 [CTW] Bounds reported incorrectly when some pages are scrolled James Teh [:Jamie] RESOLVED
1765944 [CTW] Transform cache does not get updated when setting a transform James Teh [:Jamie] RESOLVED
1766142 [CTW] Expose explicit-name object attribute on cached RemoteAccessible James Teh [:Jamie] RESOLVED
1766206 [CTW] Line offsets incorrect for line feed at end of wrapped line James Teh [:Jamie] RESOLVED
1766546 [CTW] Ability to set the caret for cached RemoteAccessibles NEW
1766547 Support setting the caret in TextLeafPoint NEW
1766560 [CTW] First line start of a leaf is ignored when fetching the next line start James Teh [:Jamie] RESOLVED
1766794 [CTW] Terrible performance in NVDA vbuf loading using View Source James Teh [:Jamie] RESOLVED
1767887 [CTW] Paragraph boundary incorrect inside list items James Teh [:Jamie] RESOLVED
1768391 [CTW] Crash in [@ mozilla::detail::InvalidArrayIndex_CRASH | mozilla::a11y::HyperTextAccessibleBase::GetChildOffset] James Teh [:Jamie] RESOLVED
1769165 Make CharBounds handle embedded characters NEW
1769166 Push updates to TextBounds James Teh [:Jamie] NEW
1770162 [CTW] Assertion in nsITimer::SetDelay when MSAA ids released during XPCOM shutdown James Teh [:Jamie] NEW

33 Total; 17 Open (51.52%); 15 Resolved (45.45%); 1 Verified (3.03%);


Milestone 3: September 2022

TBD. Beta experiment? Release experiment?

Bugzilla

No results.

0 Total; 0 Open (0%); 0 Resolved (0%); 0 Verified (0%);


Backlog

Full Query
ID Summary Assigned to Status
1234121 Cache active item from child process in parent and remove sync call for FocusedChild NEW
1694573 Make AccEvent use new Accessible base class Morgan Reschenberg [:morgan] NEW
1730868 TextLeafPoint: Better handling for inline-block for line boundaries NEW
1733188 [meta] Notify of state changes for caching purposes NEW
1733486 [meta] Cache text NEW
1739559 Update cache when text attributes change NEW
1743749 Clean up use of atoms for cache keys NEW
1754905 Reimagine XPCOM testing for Pivot NEW
1757127 Cache Language NEW
1757260 Cache initial focus, caret and text selection NEW
1759971 CTW support for MathML on Windows NEW
1764750 [CTW] Document sometimes keeps stale state NEW
1765217 Cleanup: Add OPAQUE1 to kRemoteCalculatedStates NEW
1765968 [CTW] Use QueueCacheUpdate instead of SendCacheUpdate wherever possible NEW
1766147 Get rid of Accessible::IsProxy and eProxyType AccType NEW
1768396 [CTW] Consider caching RemoteAccessibleBase::IndexInParent NEW
1768951 Some text bounds are incorrectly calculated within iframes NEW
1769688 [CTW] Include info about the parent process a11y cache in about:memory NEW
1769824 Move some of HyperTextAccessibleBase::TextBounds into TextLeafRange::TextBounds NEW

19 Total; 19 Open (100%); 0 Resolved (0%); 0 Verified (0%);