This prevents failures at the database layer, given that aliases in the
ON clause are not quoted.
Systematically quoting aliases even in FilteredRelation is tracked in
https://code.djangoproject.com/ticket/36795.
Backport of 005d60d97c from main.
Before, `order_by()` treated a period in a field name as a sign that it
was requested via `.extra(order_by=...)` and thus should be passed
through as raw table and column names, even if `extra()` was not used.
Since periods are permitted in aliases, this meant user-controlled
aliases could force the `order_by()` clause to resolve to a raw table
and column pair instead of the actual target field for the alias.
In practice, only `FilteredRelation` was affected, as the other
expressions we tested, e.g. `F`, aggressively optimize away the ordering
expressions into ordinal positions, e.g. ORDER BY 2, instead of ORDER BY
"table".column.
Thanks Solomon Kebede for the report, and Simon Charette and Jake Howard
for reviews.
Backport of 69065ca869 from main.
Control characters in FilteredRelation column aliases could be used for
SQL injection attacks. This affected QuerySet.annotate(), aggregate(),
extra(), values(), values_list(), and alias() when using dictionary
expansion with **kwargs.
Thanks Solomon Kebede for the report, and Simon Charette, Jacob Walls,
and Natalia Bidart for reviews.
Backport of e891a84c7e from main.
The `TruncateHTMLParser` used `deque.remove()` to remove tags from the
stack when processing end tags. With crafted input containing many
unmatched end tags, this caused repeated full scans of the tag stack,
leading to quadratic time complexity.
The fix uses LIFO semantics, only removing a tag from the stack when it
matches the most recently opened tag. This avoids linear scans for
unmatched end tags and reduces complexity to linear time.
Refs #30686 and 6ee37ada32.
Thanks Seokchan Yoon for the report, and Jake Howard and Jacob Walls for
reviews.
Backport of a33540b3e2 from main.
Regression in a16eedcf9c.
The UNNEST strategy is affected by the same problem bulk_update has wrt/
to silent data truncation due to its usage of db_type which always returns
a parametrized subtype.
Backport of d6ae2ed868 from main.
This variable should only be set once. Under forkserver, this module
was getting executed multiple times, causing nested temporary dirs
that didn't clean up properly, raising FileNotFoundError.
This similar to #27890 although a slightly different cause.
Backport of cd6278c4c0 from main.
Previously, `getInnerText()` recursively used `list.extend()` on strings,
which added each character from child nodes as a separate list element.
On deeply nested XML content, this caused the overall deserialization
work to grow quadratically with input size, potentially allowing
disproportionate CPU consumption for crafted XML.
The fix separates collection of inner texts from joining them, so that
each subtree is joined only once, reducing the complexity to linear in
the size of the input. These changes also include a mitigation for a
xml.dom.minidom performance issue.
Thanks Seokchan Yoon (https://ch4n3.kr/) for report.
Co-authored-by: Jacob Walls <jacobtylerwalls@gmail.com>
Co-authored-by: Natalia <124304+nessita@users.noreply.github.com>
Backport of 50efb718b3 from main.
Follow-up to CVE-2025-57833.
Thanks Stackered for the report, and Simon Charette and Mariusz Felisiak
for the reviews.
Backport of 5b90ca1e75 from main.
Ideally, this will be reverted when an upstream solution is available for
https://github.com/python/cpython/issues/141560.
Thanks Patrick Rauscher for the report and Augusto Pontes for the
first iteration and test.
Backport of 34186e731c from main.
The fix landed in a8cf8c292c will be
backported to 5.1 and 4.2 since the 2048 limit was rolled out as part of
the security release for CVE-2025-64458.
Backport of 18b13cf6c4 from main.
Refs CVE-2025-64458.
The previous limit of 2048 characters reused the URLValidator constant
and proved too restrictive for legitimate redirects to some third-party
services. This change introduces a separate `MAX_URL_REDIRECT_LENGTH`
constant (defaulting to 16384) and uses it in HttpResponseRedirectBase.
Thanks Jacob Walls for report and review.
Backport of a8cf8c292c from main.
Added warning in DateField documentation about type differences when using
timedelta on PostgreSQL and MySQL. Mentioned Cast() and integer arithmetic
solutions.
Backport of 55af4749b9 from main.
GitHub Actions defaults to a 360-minute (6-hour) timeout. We've had jobs
hang due to issues in the parallel test runner, causing them to run for
the full 6 hours. This wastes resources and negatively impacts CI
availability, so explicit timeouts have been added to prevent
long-running hangs.
Backport of e48527f91d from main.
Regression in b8e5a8a9a2.
Refs #36404.
The replace_expressions method was innapropriately dealing with falsey
but not None source expressions causing them to also be potentially
evaluated when __bool__ was invoked (e.g. QuerySet.__bool__ evaluates
the queryset).
The changes introduced in b8e5a8a9a2, which were to deal with a similar
issue, surfaced the problem as aggregation over an annotated queryset
requires an inlining (or pushdown) of aggregate references which is
achieved through replace_expressions.
In cases where an empty Q object was provided as an aggregate filter,
such as when the admin facetting feature was used as reported, it would
wrongly be turned into None, instead of an empty WhereNode, causing a
crash at aggregate filter compilation.
Note that the crash signature differed depending on whether or not the
backend natively supports aggregate filtering
(supports_aggregate_filter_clause) as the fallback, which makes use
Case / When expressions, would result in a TypeError instead of a
NoneType AttributeError.
Thanks Rafael Urben for the report, Antoliny and Youngkwang Yang for
the triage.
Backport of 2a6e0bd72d from main
