Reported internally by @gwarser.
In rare occasion, a timing issue could cause uBO to redirect
to a web accessible resource meant to be used for another
network request. This is a regression introduced with the
following commit:
- 2e5d32e967
Additionally, I identified another issue which would cause
cached redirection to fail when a cache entry with redirection
to a web accessible resource was being reused, an issue which
could especially affect pages which are generated dynamically
(i.e. without full page reload).
filterUnits is now treated as a buffer which is
pre-allocated and which will grow in chunks so as
to minimize memory allocations. Entries are never
released, just null-ed.
Additionally, move urlTokenizer into the static
network filtering engine, since it's not used
anywhere else.
Notably, defer the post-load optimization operations
to a few seconds after the filters have been all
loaded in memory -- this is not a critical step for
the filtering engine to work properly, hence this
can be delayed in order to ensure readiness as soon
as possible.
Most notably, the `denyallow=` option now requires
the presence of a valid `domain=` option to not be
rejected.
Using `denyallow=` without narrowing down using the
`domain=` option leads to catastrophic blocking
behvior, hence the requirement for a valid `domain=`
option.
Related commit:
- b265f2644d
The optimization in the commit above was meant to
improve the performance of lookup operations of
modifier filters, but I forgot to enable the
optimisation for that class of filters.
This means this commit brings another significant
performance gain on top of the previous commit, as
shown by the built-in benchmark.
Additionally a few minor code rearrangements.
Performance-related work.
There is a fair number of filters which can't be tokenized
in uBO's own filter lists. Majority of those filters also
declare a `domain=` option, examples:
*$script,redirect-rule=noopjs,domain=...
*$script,3p,domain=...,denyallow=...
*$frame,3p,domain=...
Such filters can be found in uBO's asset viewer using the
following search expression:
/^\*?\$[^\n]*?domain=/
Some filter buckets will contain many of those filters, for
instance one of the bucket holding untokenizable `redirect=`
filters has over 170 entries, which must be all visited when
collating all matching `redirect=` filters.
When a bucket contains many such filters, I found that it's
worth to extract all the non-negated hostname values from
`domain=` options into a single hntrie and perform a pre-test
at match() time to find out whether the current origin of a
network request matches any one of the collected hostnames,
so as to avoid iterating through all the filters.
Since there is rarely a match() for vast majority of network
requests with `domain=` option, this pre-test saves a good
amount of work, and this is measurable with the built-in
benchmark.
This commit moves the parsing, compiling and enforcement
of the `redirect=` and `redirect-rule=` network filter
options into the static network filtering engine as
modifier options -- just like `csp=` and `queryprune=`.
This solves the two following issues:
- https://github.com/gorhill/uBlock/issues/3590
- https://github.com/uBlockOrigin/uBlock-issues/issues/1008#issuecomment-716164214
Additionally, `redirect=` option is not longer afflicted
by static network filtering syntax quirks, `redirect=`
filters can be used with any other static filtering
modifier options, can be excepted using `@@` and can be
badfilter-ed.
Since more than one `redirect=` directives could be found
to apply to a single network request, the concept of
redirect priority is introduced.
By default, `redirect=` directives have an implicit
priority of 0. Filter authors can declare an explicit
priority by appending `:[integer]` to the token of the
`redirect=` option, for example:
||example.com/*.js$1p,script,redirect=noopjs:100
The priority dictates which redirect token out of many
will be ultimately used. Cases of multiple `redirect=`
directives applying to a single blocked network request
are expected to be rather unlikely.
Explicit redirect priority should be used if and only if
there is a case of redirect ambiguity to solve.
