Auto merge of #90461 - pietroalbini:bidi-beta, r=nikomatsakis

[beta] Fix CVE-2021-42574

This PR implements new lints to mitigate the impact of [CVE-2021-42574], caused by the presence of bidirectional-override Unicode codepoints in the compiled source code. [See the advisory][advisory] for more information about the vulnerability.

The changes in this PR will be released in tomorrow's beta release.

[CVE-2021-42574]: https://cve.mitre.org/cgi-bin/cvename.cgi?name=CVE-2021-42574
[advisory]: https://blog.rust-lang.org/2021/11/01/cve-2021-42574.html

Cargo.lock

@@ -4223,6 +4223,7 @@ dependencies = [
  "rustc_span",
  "tracing",
  "unicode-normalization",
+ "unicode-width",
 ]
 
 [[package]]

RELEASES.md

@@ -1,3 +1,11 @@
+Version 1.56.1 (2021-11-01)
+===========================
+
+- New lints to detect the presence of bidirectional-override Unicode
+  codepoints in the compiled source code ([CVE-2021-42574])
+
+[CVE-2021-42574]: https://cve.mitre.org/cgi-bin/cvename.cgi?name=CVE-2021-42574
+
 Version 1.56.0 (2021-10-21)
 ========================
 

compiler/rustc_errors/src/emitter.rs

@@ -2054,8 +2054,26 @@ fn num_decimal_digits(num: usize) -> usize {
     MAX_DIGITS
 }
 
+// We replace some characters so the CLI output is always consistent and underlines aligned.
+const OUTPUT_REPLACEMENTS: &[(char, &str)] = &[
+    ('\t', "    "),   // We do our own tab replacement
+    ('\u{202A}', ""), // The following unicode text flow control characters are inconsistently
+    ('\u{202B}', ""), // supported accross CLIs and can cause confusion due to the bytes on disk
+    ('\u{202D}', ""), // not corresponding to the visible source code, so we replace them always.
+    ('\u{202E}', ""),
+    ('\u{2066}', ""),
+    ('\u{2067}', ""),
+    ('\u{2068}', ""),
+    ('\u{202C}', ""),
+    ('\u{2069}', ""),
+];
+
 fn replace_tabs(str: &str) -> String {
-    str.replace('\t', "    ")
+    let mut s = str.to_string();
+    for (c, replacement) in OUTPUT_REPLACEMENTS {
+        s = s.replace(*c, replacement);
+    }
+    s
 }
 
 fn draw_col_separator(buffer: &mut StyledBuffer, line: usize, col: usize) {

compiler/rustc_lint/src/context.rs

@@ -16,6 +16,7 @@
 
 use self::TargetLint::*;
 
+use crate::hidden_unicode_codepoints::UNICODE_TEXT_FLOW_CHARS;
 use crate::levels::{is_known_lint_tool, LintLevelsBuilder};
 use crate::passes::{EarlyLintPassObject, LateLintPassObject};
 use rustc_ast as ast;
@@ -39,7 +40,7 @@ use rustc_session::lint::{BuiltinLintDiagnostics, ExternDepSpec};
 use rustc_session::lint::{FutureIncompatibleInfo, Level, Lint, LintBuffer, LintId};
 use rustc_session::Session;
 use rustc_span::lev_distance::find_best_match_for_name;
-use rustc_span::{symbol::Symbol, MultiSpan, Span, DUMMY_SP};
+use rustc_span::{symbol::Symbol, BytePos, MultiSpan, Span, DUMMY_SP};
 use rustc_target::abi;
 use tracing::debug;
 
@@ -597,6 +598,42 @@ pub trait LintContext: Sized {
             // Now, set up surrounding context.
             let sess = self.sess();
             match diagnostic {
+                BuiltinLintDiagnostics::UnicodeTextFlow(span, content) => {
+                    let spans: Vec<_> = content
+                        .char_indices()
+                        .filter_map(|(i, c)| {
+                            UNICODE_TEXT_FLOW_CHARS.contains(&c).then(|| {
+                                let lo = span.lo() + BytePos(2 + i as u32);
+                                (c, span.with_lo(lo).with_hi(lo + BytePos(c.len_utf8() as u32)))
+                            })
+                        })
+                        .collect();
+                    let (an, s) = match spans.len() {
+                        1 => ("an ", ""),
+                        _ => ("", "s"),
+                    };
+                    db.span_label(span, &format!(
+                        "this comment contains {}invisible unicode text flow control codepoint{}",
+                        an,
+                        s,
+                    ));
+                    for (c, span) in &spans {
+                        db.span_label(*span, format!("{:?}", c));
+                    }
+                    db.note(
+                        "these kind of unicode codepoints change the way text flows on \
+                         applications that support them, but can cause confusion because they \
+                         change the order of characters on the screen",
+                    );
+                    if !spans.is_empty() {
+                        db.multipart_suggestion_with_style(
+                            "if their presence wasn't intentional, you can remove them",
+                            spans.into_iter().map(|(_, span)| (span, "".to_string())).collect(),
+                            Applicability::MachineApplicable,
+                            SuggestionStyle::HideCodeAlways,
+                        );
+                    }
+                },
                 BuiltinLintDiagnostics::Normal => (),
                 BuiltinLintDiagnostics::BareTraitObject(span, is_global) => {
                     let (sugg, app) = match sess.source_map().span_to_snippet(span) {

compiler/rustc_lint/src/hidden_unicode_codepoints.rs

@@ -0,0 +1,161 @@
+use crate::{EarlyContext, EarlyLintPass, LintContext};
+use rustc_ast as ast;
+use rustc_errors::{Applicability, SuggestionStyle};
+use rustc_span::{BytePos, Span, Symbol};
+
+declare_lint! {
+    /// The `text_direction_codepoint_in_literal` lint detects Unicode codepoints that change the
+    /// visual representation of text on screen in a way that does not correspond to their on
+    /// memory representation.
+    ///
+    /// ### Explanation
+    ///
+    /// The unicode characters `\u{202A}`, `\u{202B}`, `\u{202D}`, `\u{202E}`, `\u{2066}`,
+    /// `\u{2067}`, `\u{2068}`, `\u{202C}` and `\u{2069}` make the flow of text on screen change
+    /// its direction on software that supports these codepoints. This makes the text "abc" display
+    /// as "cba" on screen. By leveraging software that supports these, people can write specially
+    /// crafted literals that make the surrounding code seem like it's performing one action, when
+    /// in reality it is performing another. Because of this, we proactively lint against their
+    /// presence to avoid surprises.
+    ///
+    /// ### Example
+    ///
+    /// ```rust,compile_fail
+    /// #![deny(text_direction_codepoint_in_literal)]
+    /// fn main() {
+    ///     println!("{:?}", '‮');
+    /// }
+    /// ```
+    ///
+    /// {{produces}}
+    ///
+    pub TEXT_DIRECTION_CODEPOINT_IN_LITERAL,
+    Deny,
+    "detect special Unicode codepoints that affect the visual representation of text on screen, \
+     changing the direction in which text flows",
+}
+
+declare_lint_pass!(HiddenUnicodeCodepoints => [TEXT_DIRECTION_CODEPOINT_IN_LITERAL]);
+
+crate const UNICODE_TEXT_FLOW_CHARS: &[char] = &[
+    '\u{202A}', '\u{202B}', '\u{202D}', '\u{202E}', '\u{2066}', '\u{2067}', '\u{2068}', '\u{202C}',
+    '\u{2069}',
+];
+
+impl HiddenUnicodeCodepoints {
+    fn lint_text_direction_codepoint(
+        &self,
+        cx: &EarlyContext<'_>,
+        text: Symbol,
+        span: Span,
+        padding: u32,
+        point_at_inner_spans: bool,
+        label: &str,
+    ) {
+        // Obtain the `Span`s for each of the forbidden chars.
+        let spans: Vec<_> = text
+            .as_str()
+            .char_indices()
+            .filter_map(|(i, c)| {
+                UNICODE_TEXT_FLOW_CHARS.contains(&c).then(|| {
+                    let lo = span.lo() + BytePos(i as u32 + padding);
+                    (c, span.with_lo(lo).with_hi(lo + BytePos(c.len_utf8() as u32)))
+                })
+            })
+            .collect();
+
+        cx.struct_span_lint(TEXT_DIRECTION_CODEPOINT_IN_LITERAL, span, |lint| {
+            let mut err = lint.build(&format!(
+                "unicode codepoint changing visible direction of text present in {}",
+                label
+            ));
+            let (an, s) = match spans.len() {
+                1 => ("an ", ""),
+                _ => ("", "s"),
+            };
+            err.span_label(
+                span,
+                &format!(
+                    "this {} contains {}invisible unicode text flow control codepoint{}",
+                    label, an, s,
+                ),
+            );
+            if point_at_inner_spans {
+                for (c, span) in &spans {
+                    err.span_label(*span, format!("{:?}", c));
+                }
+            }
+            err.note(
+                "these kind of unicode codepoints change the way text flows on applications that \
+                 support them, but can cause confusion because they change the order of \
+                 characters on the screen",
+            );
+            if point_at_inner_spans && !spans.is_empty() {
+                err.multipart_suggestion_with_style(
+                    "if their presence wasn't intentional, you can remove them",
+                    spans.iter().map(|(_, span)| (*span, "".to_string())).collect(),
+                    Applicability::MachineApplicable,
+                    SuggestionStyle::HideCodeAlways,
+                );
+                err.multipart_suggestion(
+                    "if you want to keep them but make them visible in your source code, you can \
+                    escape them",
+                    spans
+                        .into_iter()
+                        .map(|(c, span)| {
+                            let c = format!("{:?}", c);
+                            (span, c[1..c.len() - 1].to_string())
+                        })
+                        .collect(),
+                    Applicability::MachineApplicable,
+                );
+            } else {
+                // FIXME: in other suggestions we've reversed the inner spans of doc comments. We
+                // should do the same here to provide the same good suggestions as we do for
+                // literals above.
+                err.note("if their presence wasn't intentional, you can remove them");
+                err.note(&format!(
+                    "if you want to keep them but make them visible in your source code, you can \
+                     escape them: {}",
+                    spans
+                        .into_iter()
+                        .map(|(c, _)| { format!("{:?}", c) })
+                        .collect::<Vec<String>>()
+                        .join(", "),
+                ));
+            }
+            err.emit();
+        });
+    }
+}
+impl EarlyLintPass for HiddenUnicodeCodepoints {
+    fn check_attribute(&mut self, cx: &EarlyContext<'_>, attr: &ast::Attribute) {
+        if let ast::AttrKind::DocComment(_, comment) = attr.kind {
+            if comment.as_str().contains(UNICODE_TEXT_FLOW_CHARS) {
+                self.lint_text_direction_codepoint(cx, comment, attr.span, 0, false, "doc comment");
+            }
+        }
+    }
+
+    fn check_expr(&mut self, cx: &EarlyContext<'_>, expr: &ast::Expr) {
+        // byte strings are already handled well enough by `EscapeError::NonAsciiCharInByteString`
+        let (text, span, padding) = match &expr.kind {
+            ast::ExprKind::Lit(ast::Lit { token, kind, span }) => {
+                let text = token.symbol;
+                if !text.as_str().contains(UNICODE_TEXT_FLOW_CHARS) {
+                    return;
+                }
+                let padding = match kind {
+                    // account for `"` or `'`
+                    ast::LitKind::Str(_, ast::StrStyle::Cooked) | ast::LitKind::Char(_) => 1,
+                    // account for `r###"`
+                    ast::LitKind::Str(_, ast::StrStyle::Raw(val)) => *val as u32 + 2,
+                    _ => return,
+                };
+                (text, span, padding)
+            }
+            _ => return,
+        };
+        self.lint_text_direction_codepoint(cx, text, *span, padding, true, "literal");
+    }
+}

compiler/rustc_lint/src/lib.rs

@@ -48,6 +48,7 @@ pub mod builtin;
 mod context;
 mod early;
 mod enum_intrinsics_non_enums;
+pub mod hidden_unicode_codepoints;
 mod internal;
 mod late;
 mod levels;
@@ -78,6 +79,7 @@ use rustc_span::Span;
 use array_into_iter::ArrayIntoIter;
 use builtin::*;
 use enum_intrinsics_non_enums::EnumIntrinsicsNonEnums;
+use hidden_unicode_codepoints::*;
 use internal::*;
 use methods::*;
 use non_ascii_idents::*;
@@ -129,6 +131,7 @@ macro_rules! early_lint_passes {
                 DeprecatedAttr: DeprecatedAttr::new(),
                 WhileTrue: WhileTrue,
                 NonAsciiIdents: NonAsciiIdents,
+                HiddenUnicodeCodepoints: HiddenUnicodeCodepoints,
                 IncompleteFeatures: IncompleteFeatures,
                 RedundantSemicolons: RedundantSemicolons,
                 UnusedDocComment: UnusedDocComment,

compiler/rustc_lint_defs/src/builtin.rs

@@ -3518,6 +3518,34 @@ declare_lint! {
     @feature_gate = sym::non_exhaustive_omitted_patterns_lint;
 }
 
+declare_lint! {
+    /// The `text_direction_codepoint_in_comment` lint detects Unicode codepoints in comments that
+    /// change the visual representation of text on screen in a way that does not correspond to
+    /// their on memory representation.
+    ///
+    /// ### Example
+    ///
+    /// ```rust,compile_fail
+    /// #![deny(text_direction_codepoint_in_comment)]
+    /// fn main() {
+    ///     println!("{:?}"); // '‮');
+    /// }
+    /// ```
+    ///
+    /// {{produces}}
+    ///
+    /// ### Explanation
+    ///
+    /// Unicode allows changing the visual flow of text on screen in order to support scripts that
+    /// are written right-to-left, but a specially crafted comment can make code that will be
+    /// compiled appear to be part of a comment, depending on the software used to read the code.
+    /// To avoid potential problems or confusion, such as in CVE-2021-42574, by default we deny
+    /// their use.
+    pub TEXT_DIRECTION_CODEPOINT_IN_COMMENT,
+    Deny,
+    "invisible directionality-changing codepoints in comment"
+}
+
 declare_lint! {
     /// The `deref_into_dyn_supertrait` lint is output whenever there is a use of the
     /// `Deref` implementation with a `dyn SuperTrait` type as `Output`.

compiler/rustc_lint_defs/src/lib.rs

@@ -306,6 +306,7 @@ pub enum BuiltinLintDiagnostics {
     TrailingMacro(bool, Ident),
     BreakWithLabelAndLoop(Span),
     NamedAsmLabel(String),
+    UnicodeTextFlow(Span, String),
 }
 
 /// Lints that are buffered up early on in the `Session` before the

compiler/rustc_parse/Cargo.toml

@@ -18,3 +18,4 @@ rustc_session = { path = "../rustc_session" }
 rustc_span = { path = "../rustc_span" }
 rustc_ast = { path = "../rustc_ast" }
 unicode-normalization = "0.1.11"
+unicode-width = "0.1.4"