Related issues:
- https://github.com/uBlockOrigin/uBlock-issues/issues/761
- https://github.com/uBlockOrigin/uBlock-issues/issues/528
The previous bidi-trie code could only hold filters which
are plain pattern, i.e. no wildcard characters, and which
had no origin option (`domain=`), right and/or left anchor,
and no `csp=` option.
Example of filters that could be moved into a bidi-trie
data structure:
&ad_box_
/w/d/capu.php?z=$script,third-party
||liveonlinetv247.com/images/muvixx-150x50-watch-now-in-hd-play-btn.gif
Examples of filters that could NOT be moved to a bidi-trie:
-adap.$domain=~l-adap.org
/tsc.php?*&ses=
||ibsrv.net/*forumsponsor$domain=[...]
@@||imgspice.com/jquery.cookie.js|$script
||view.atdmt.com^*/iview/$third-party
||postimg.cc/image/$csp=[...]
Ideally the filters above should be able to be moved to a
bidi-trie since they are basically plain patterns, or at
least partially moved to a bidi-trie when there is only a
single wildcard (i.e. made of two plain patterns).
Also, there were two distinct bidi-tries in which
plain-pattern filters can be moved to: one for patterns
without hostname anchoring and another one for patterns
with hostname-anchoring. This was required because the
hostname-anchored patterns have an extra condition which
is outside the bidi-trie knowledge.
This commit expands the number of filters which can be
stored in the bidi-trie, and also remove the need to
use two distinct bidi-tries.
- Added ability to associate a pattern with an integer
in the bidi-trie [1].
- The bidi-trie match code passes this externally
provided integer when calling an externally
provided method used for testing extra conditions
that may be present for a plain pattern found to
be matching in the bidi-trie.
- Decomposed existing filters into smaller logical units:
- FilterPlainLeftAnchored =>
FilterPatternPlain +
FilterAnchorLeft
- FilterPlainRightAnchored =>
FilterPatternPlain +
FilterAnchorRight
- FilterExactMatch =>
FilterPatternPlain +
FilterAnchorLeft +
FilterAnchorRight
- FilterPlainHnAnchored =>
FilterPatternPlain +
FilterAnchorHn
- FilterWildcard1 =>
FilterPatternPlain + [
FilterPatternLeft or
FilterPatternRight
]
- FilterWildcard1HnAnchored =>
FilterPatternPlain + [
FilterPatternLeft or
FilterPatternRight
] +
FilterAnchorHn
- FilterGenericHnAnchored =>
FilterPatternGeneric +
FilterAnchorHn
- FilterGenericHnAndRightAnchored =>
FilterPatternGeneric +
FilterAnchorRight +
FilterAnchorHn
- FilterOriginMixedSet =>
FilterOriginMissSet +
FilterOriginHitSet
- Instances of FilterOrigin[...], FilterDataHolder
can also be added to a composite filter to
represent `domain=` and `csp=` options.
- Added a new filter class, FilterComposite, for
filters which are a combination of two or more
logical units. A FilterComposite instance is a
match when *all* filters composing it are a
match.
Since filters are now encoded into combination of
smaller units, it becomes possible to extract the
FilterPatternPlain component and store it in the
bidi-trie, and use the integer as a handle for the
remaining extra conditions, if any.
Since a single pattern in the bidi-trie may be a
component for different filters, the associated
integer points to a sequence of extra conditions,
and a match occurs as soon as one of the extra
conditions (which may itself be a sequence of
conditions) is fulfilled.
Decomposing filters which are currently single
instance into sequences of smaller logical filters
means increasing the storage and CPU overhead when
evaluating such filters. The CPU overhead is
compensated by the fact that more filters can now
moved into the bidi-trie, where the first match is
efficiently evaluated. The extra conditions have to
be evaluated if and only if there is a match in the
bidi-trie.
The storage overhead is compensated by the
bidi-trie's intrinsic nature of merging similar
patterns.
Furthermore, the storage overhead is reduced by no
longer using JavaScript array to store collection
of filters (which is what FilterComposite is):
the same technique used in [2] is imported to store
sequences of filters.
A sequence of filters is a sequence of integer pairs
where the first integer is an index to an actual
filter instance stored in a global array of filters
(`filterUnits`), while the second integer is an index
to the next pair in the sequence -- which means all
sequences of filters are encoded in one single array
of integers (`filterSequences` => Uint32Array). As
a result, a sequence of filters can be represented by
one single integer -- an index to the first pair --
regardless of the number of filters in the sequence.
This representation is further leveraged to replace
the use of JavaScript array in FilterBucket [3],
which used a JavaScript array to store collection
of filters. Doing so means there is no more need for
FilterPair [4], which purpose was to be a lightweight
representation when there was only two filters in a
collection.
As a result of the above changes, the map of `token`
(integer) => filter instance (object) used to
associate tokens to filters or collections of filters
is replaced with a more efficient map of `token`
(integer) to filter unit index (integer) to lookup a
filter object from the global `filterUnits` array.
Another consequence of using one single global
array to store all filter instances means we can reuse
existing instances when a logical filter instance is
parameter-less, which is the case for FilterAnchorLeft,
FilterAnchorRight, FilterAnchorHn, the index to these
single instances is reused where needed.
`urlTokenizer` now stores the character codes of the
scanned URL into a bidi-trie buffer, for reuse when
string matching methods are called.
New method: `tokenHistogram()`, used to generate
histograms of occurrences of token extracted from URLs
in built-in benchmark. The top results of the "miss"
histogram are used as "bad tokens", i.e. tokens to
avoid if possible when compiling filter lists.
All plain pattern strings are now stored in the
bidi-trie memory buffer, regardless of whether they
will be used in the trie proper or not.
Three methods have been added to the bidi-trie to test
stored string against the URL which is also stored in
then bidi-trie.
FilterParser is now instanciated on demand and
released when no longer used.
***
[1] 135a45a878/src/js/strie.js (L120)
[2] e94024d350
[3] 135a45a878/src/js/static-net-filtering.js (L1630)
[4] 135a45a878/src/js/static-net-filtering.js (L1566)
Related commit:
- fb4e94f92c
A bidi-trie can't store strings longer than 255 characters
because the string segment lengths are encoded into a single
byte. This commit ensures only strings smaller than
256 characters are stored in the bidi-tries.
Plain filters can be any length, while the bidi-trie was
assuming max length of 256. The origin of this error
is from when the bidi-trie code was originally imported
from the hntrie code as start to develop bidi-trie.
The erroneous code could cause issue when the following
conditions were met:
- Plain filter longer than 256 characters
- Free space in bidi-trie's character buffer was less
than 256 bytes
Likely a rare occurrence, but some filter lists do contains
long plain filters, for example:
https://gitlab.com/curben/urlhaus-filter/raw/master/urlhaus-filter.txt
Related feedback:
- https://www.reddit.com/r/uBlockOrigin/comments/cs1y26/
The bidirectional trie allows storing the right
and left parts of a string into a trie given a
pivot position.
Releated issue:
- https://github.com/uBlockOrigin/uBlock-issues/issues/528
Additionally, the mandatory token-at-index-0 rule
for FilterPlainHnAnchored has been lifted, thus
allowing the engine to pick a potentially better token
at any position in the filter string.
***
TODO: Eventually rename `strie.js` to `biditrie.js`.
TODO: Fix dump() method, it currently only show the
right-hand side of a filter string.
To be used at the console, as an investigation tool for
development purpose.
Using it to verify the content of the largest
FilterHostnameDict instance, I spotted an all-uppercase
hostname in the HNTrieRef instance:
µBlock.staticNetFilteringEngine.categories.get(0).get(0x10000000).dict.dump();
Thus the changes to static-net-filtering.js are to fix
the erroneous insertion of filters with uppercase
characters. The single instance found was a hostname entry
in Malware Domain List (TRIANGLESERVICESLTD dot COM).
The purpose of using a custom base128 encoder is to
convert array buffers into strings, to allow a direct
string-to-array buffer conversion at load time:
string => array buffer
Whereas a JSON array would require an extra step:
JSON array as string => JS array => array buffer
Turns out that the current use of a custom base128 encoding
results in a significantly larger selfie storage usage when
converting array buffers into strings.
Speculation: possibly the browser convert the strings to
save into JSON strings internally. Since the custom base128
encoder is likely to cause the resulting string to contain
a lot of unprintable ASCII characters, these will need to
be escaped when converted to JSON -- escaped characters
occupy more space than non-escaped ones.
Using a sequence of base 64 numbers means only printable
will be present in the output string, hence no escaping
necessary. I have observed significant reduction in
storage usage for selfie purpose.
Related issue:
- https://github.com/uBlockOrigin/uBlock-issues/issues/528#issuecomment-484408622
Following STrie-related work in above issue, I noticed that a large
number of filters in EasyList were filters which only had to match
against the document origin. For instance, among just the top 10
most populous buckets, there were four such buckets with over
hundreds of entries each:
- bits: 72, token: "http", 146 entries
- bits: 72, token: "https", 139 entries
- bits: 88, token: "http", 122 entries
- bits: 88, token: "https", 118 entries
These filters in these buckets have to be matched against all
the network requests.
In order to leverage HNTrie for these filters[1], they are now handled
in a special way so as to ensure they all end up in a single HNTrie
(per bucket), which means that instead of scanning hundreds of entries
per URL, there is now a single scan per bucket per URL for these
apply-everywhere filters.
Now, any filter which fulfill ALL the following condition will be
processed in a special manner internally:
- Is of the form `|https://` or `|http://` or `*`; and
- Does have a `domain=` option; and
- Does not have a negated domain in its `domain=` option; and
- Does not have `csp=` option; and
- Does not have a `redirect=` option
If a filter does not fulfill ALL the conditions above, no change
in behavior.
A filter which matches ALL of the above will be processed in a special
manner:
- The `domain=` option will be decomposed so as to create as many
distinct filter as there is distinct value in the `domain=` option
- This also apply to the `badfilter` version of the filter, which
means it now become possible to `badfilter` only one of the
distinct filter without having to `badfilter` all of them.
- The logger will always report these special filters with only a
single hostname in the `domain=` option.
***
[1] HNTrie is currently WASM-ed on Firefox.
Implement a plain string trie container class: STrieContainer.
Make use of STrieContainer where beneficial
Some filter buckets can grow quite large, and in such case
coalescing "trieable" filter classes into a single trie reduces
lookup performance and memory usage.
For instance, at time of commit, the filter bucket for the
`ad` keyword contains 919 entries[1].
Coalescing trieable filters of the same class into a single plain
string trie reduced the size of the bucket into 50 entries + two
tries which are scanned only once each whenever the bucket is
visited.
[1] Enter the following code at uBO's dev console:
µBlock.staticNetFilteringEngine.categories.get(0).get(µBlock.urlTokenizer.tokenHashFromString('ad'))
Refactor static network filtering engine code to make use of
ES6's syntactic sugar `class`.
Change first auto-update run from 7 to 5 minutes.
