Turns out it's currently the fastest among the three currently
implemented (Cache, browser.storage.session, indexedDB). Possibly
because indexedDB can natively persist structure-cloneable data,
something uBO can now benefit with the work on abstracting away
the limitations of various storages being limited to persist only
text or JSON data.
Related issue:
https://github.com/uBlockOrigin/uBlock-issues/issues/2969
In uBO, the "cache storage" is used to save resources which can
be safely discarded, though at the cost of having to fetch or
recompute them again.
Extension storage (browser.storage.local) is now always used as
cache storage backend. This has always been the default for
Chromium-based browsers.
For Firefox-based browsers, IndexedDB was used as backend for
cache storage, with fallback to extension storage when using
Firefox in private mode by default.
Extension storage is reliable since it works in all contexts,
though it may not be the most performant one.
To speed-up loading of resources from extension storage, uBO will
now make use of Cache API storage, which will mirror content of
key assets saved to extension storage. Typically loading resources
from Cache API is faster than loading the same resources from
the extension storage.
Only resources which must be loaded in memory as fast as possible
will make use of the Cache API storage layered on top of the
extension storage.
Compiled filter lists and memory snapshot of filtering engines
(aka "selfies") will be mirrored to the Cache API storage, since
these must be loaded into memory as fast as possible, and reloading
filter lists from their compiled counterpart is a common
operation.
This new design makes it now seamless to work in permanent private
mode for Firefox-based browsers, since extension storage now
always contains cache-related assets.
Support for IndexedDB is removed for the time being, except to
support migration of cached assets the first time uBO runs with
the new cache storage design.
In order to easily support all choices of storage, a new serializer
has been introduced, which is capable of serializing/deserializing
structure-cloneable data to/from a JS string.
Because of this new serializer, JS data structures can be stored
directly from their native representation, and deserialized
directly to their native representation from uBO's point of view,
since the serialization occurs (if needed) only at the storage
interface level.
This new serializer simplifies many code paths where data
structures such as Set, Map, TypedArray, RegExp, etc. had to be
converted in a disparate manner to be able to persist them to
extension storage.
The new serializer supports workers and LZ4 compression. These
can be configured through advanced settings.
With this new layered design, it's possible to introduce more
storage layers if measured as beneficial (i.e. maybe
browser.storage.session)
References:
- https://developer.mozilla.org/en-US/docs/Mozilla/Add-ons/WebExtensions/API/storage/local
- https://developer.mozilla.org/en-US/docs/Web/API/Cache
- https://developer.mozilla.org/en-US/docs/Web/API/Web_Workers_API/Structured_clone_algorithm
As reported internally to ubo-security by https://github.com/distinctmondaylila
One issue is a regression from the rewriting of the static filtering
parser in version 1.47.0, specifically the following commit:
https://github.com/gorhill/uBlock/commit/8ea3b0f64c
The existing regex was no longer suitable to properly detect
some usage of `report-xxx` in the rwritten parser.
Another issue which predates 1.47.0 is that the regex used for
validation was case-sensititive, while the `report-uri` directive
can be written using uppercase letters, i.e. `Report-uri`.
Avoid race conditions between isolated world-side broadcast channel
and main-side broadcast channel, so as to not lose logging
information if the isolated world-side is not yet ready to
receive through its broadcast channel.
Additionally, added new scriptlet: `trusted-replace-argument`.
[...]##+js(trusted-replace-argument, fn, argpos, argval [,condition, pattern])
Where:
- `fn` is the function we want to proxy through an `apply` handler.
This can also be a class, in which case the scriptlet will proxy
through `construct` handler. At the moment, `fn` must exist at the
time the scriptlet executes.
- `argpos` is the 0-based position of the argument we want to change
- `argval` is the value we want to have for the argument -- the value
is interpreted the same way the value for `set-constant` is
interpreted.
- `condition, pattern` is a vararg which tells the scriptlet to act
only if `pattern` is found in the argument to overwrite.
Example of usage:
alliptvlinks.com##+js(trusted-replace-argument, MutationObserver, 0, noopFunc)
This commit brings the following changes to the logger:
All logging output generated by injected scriptlets are now sent to
the logger, the developer console will no longer be used to log
scriptlet logging information.
When the logger is not opened, the scriplets will not output any
logging information.
The goal with this new approach is to allow filter authors to
more easily assess the working of scriptlets without having to
go through scriptlet parameters to enable logging.
Consequently all the previous ways to tell scriptlets to log
information are now obsolete: if the logger is opened, the
scriptlets will log information to the logger.
Another benefit of this approach is that the dev tools do not
need to be open to obtain scriptlets logging information.
Accordingly, new filter expressions have been added to the logger:
"info" and "error". Selecting the "scriptlet" expression will also
keep the logging information from scriptlets.
A new button has been added to the logger (not yet i18n-ed): a
"volume" icon, which allows to enable verbose mode. When verbose
mode is enabled, the scriptlets may choose to output more
information regarding their inner working.
The entries in the logger will automatically expand on mouse hover.
This allows to scroll through entries which text does not fit into
a single row.
Clicking anywhere on an entry in the logger will open the detailed
view when applicable.
Generic information/errors will now be rendered regardless of which
tab is currently selected in the logger (similar to how tabless
entries are already being rendered).
Procedural filters with `:xpath` operator were silently rejected
at conversion time because the parser was failing to evaluate the
xpath expression due to the absence of a `document` object in
nodejs.
If `document` object is not present, the parser will assume the
xpath expression is valid.
Scriptlets parameters which are quoted must be re-quoted when
output to the logger to be sure they can be properly looked up
in the list, and that they can be used through copy-paste
operations.
Related issue:
https://github.com/uBlockOrigin/uBlock-issues/issues/3083
This will not completely eliminate the issue but it should
lower the likelihood it will occur -- so at least uBO can
still benefit from reliable scriptlet execution in Firefox.
The idea is to remove as many dependencies as possible for
low-level ScriptletFilteringEngine in order to make it easier
to reuse the module outside uBO itself.
The high-level derived class takes care of caching and
injection of scriptlets into documents, which requires
more knowledge about the environment in which scriptlets
are to be used.
Also improve scriptlet cache usage to minimize overhead of
retrieving scriptlets.
If the `manual` parameter is assigned a date in the form of
`YYMMDD`, this will tell uBO to update lists from origin sources
when the current time is within the range of the specified date,
otherwise lists will be updated from CDNs. Updating from CDNs
is always strongly recommended since this enables differential
updates.
For the time being, `manual=1` will always cause to update lists
from origin, but this form will be deprecated once next stable
release is widespread. The idea is to not leave behind stale
and obsolete links which would be detrimental to differential
updates should someone click on one of these old links left
behind.
Purging all the lists from cache storage is detrimental to
differential update, and cause filter lists to be updated less
often and consequently to be less up to date then when letting
differential updater do its work.
Possibly related issue:
https://github.com/uBlockOrigin/uBlock-issues/issues/2136
Also reported internally, steps to reproduce the issue fixed here:
- Open uBO's dashboard through 3-dot > Add-ons > uBO > Settings
- Bring forth "Filter lists" pane
We want the tab to be already opened at next launch
- Quit Firefox for Android
- Launch Firefox for Android
Result:
Very long launch time, lists marked as out of date.
In Support pane's troubleshooting information, a delta sign will
be added to list update information when the list was last updated
through differential update.
Related commit:
https://github.com/uBlockOrigin/uAssets/commit/20312c2178
To work around incompatibity with `urltransform` in 1.53.0.
With this commit support for `urltransform` is dropped in
favor of `uritransform`. Since `uritransform` won't be parsed
by older versions of uBO, the change in syntax for `urltransform`
between 1.53.0 and 1.54.0 will no longer cause issue.
Related issue:
https://github.com/uBlockOrigin/uBlock-issues/issues/2969
Changes:
Use browser.alarms to trigger selfie creation. Presence of a selfie
improve markedly time to readiness when uBO is unsuspended.
Mirror content of storage.local to (in-memory) storage.session for
faster load to readiness when uBO is ususpended.
Broadcast channels are more suited to uBO than DOM events to dispatch
notifications to different parts of uBO.
DOM events can only be dispatched to local context, broadcast channels
dispatch to all contexts (i.e. background process, workers, auxiliary
pages) -- this last behavior is better suited to uBO to communicate
internal changes to all potential listeners, not just those in the local
context.
Additionally, broadcasting to content scripts is now done through
tabs.sendMessage() instead of through potentially opened message
ports, this simplifies broadcasting to content scripts, and this
doesn't require to have long-lived message ports in content
scripts.
Fortunately, getAvailableLists() filters out non-filter
list resources thus preventing those mistakenly selected
resources from being loaded as filter lists.
There is a lot asynchronicity in the auto-update code, and
the fix here is to detect then fix instances of out of sync
state between a cached filter list and its metadata stored
separately.
