rusty_xml-parser 0.5.0

//! DTD subset parser (internal + caller-supplied external). No network.

use rusty_xml_tree::{AttrDecl, AttrDefault, ElementDecl, XmlDtd};
use crate::error::XmlError;

/// `xmlParseDTD` — parse a DTD from memory (caller already loaded the bytes).
#[doc(alias = "xmlParseDTD")]
pub fn xml_parse_dtd(
    buffer: &[u8],
    public_id: Option<&str>,
    system_id: Option<&str>,
) -> Result<XmlDtd, XmlError> {
    let text = String::from_utf8_lossy(buffer);
    let mut dtd = parse_external_subset(&text, false)?;
    dtd.public_id = public_id.map(str::to_string);
    dtd.system_id = system_id.map(str::to_string);
    Ok(dtd)
}

/// Parse a DTD internal/external subset into declarations.
pub fn parse_dtd_subset(src: &str, old10: bool) -> Result<XmlDtd, XmlError> {
    parse_subset(src, true, old10)
}

/// Parse an external subset, where conditional sections are legal and a
/// parameter entity may supply part of a declaration.
pub fn parse_external_subset(src: &str, old10: bool) -> Result<XmlDtd, XmlError> {
    parse_subset(src, false, old10)
}

fn parse_subset(src: &str, internal: bool, old10: bool) -> Result<XmlDtd, XmlError> {
    let expanded = expand_pe(src, internal)?;
    let mut dtd = XmlDtd::default();
    dtd.int_subset = Some(src.to_string());
    let mut p = DtdParser {
        src: expanded.as_str(),
        pos: 0,
        dtd: &mut dtd,
        internal,
        old10,
    };
    p.parse_markup()?;
    check_entity_graph(&dtd)?;
    Ok(dtd)
}

fn expand_pe(src: &str, internal: bool) -> Result<String, XmlError> {
    // Multi-pass PE expansion so `%percent;` can invent new PE names.
    let mut cur = src.to_string();
    for _ in 0..16 {
        let mut pes: std::collections::HashMap<String, String> = std::collections::HashMap::new();
        harvest_pe(&cur, &mut pes);
        let next = subst_pe(&cur, &pes, internal)?;
        if next == cur {
            return Ok(cur);
        }
        cur = next;
    }
    Ok(cur)
}

fn harvest_pe(src: &str, pes: &mut std::collections::HashMap<String, String>) {
    let bytes = src.as_bytes();
    let mut i = 0;
    while i + 8 < bytes.len() {
        if bytes[i] == b'<' && bytes.get(i..i + 9) == Some(b"<!ENTITY ") {
            i += 9;
            while i < bytes.len() && bytes[i].is_ascii_whitespace() {
                i += 1;
            }
            if i < bytes.len() && bytes[i] == b'%' {
                i += 1;
                while i < bytes.len() && bytes[i].is_ascii_whitespace() {
                    i += 1;
                }
                let start = i;
                while i < bytes.len() && !bytes[i].is_ascii_whitespace() && bytes[i] != b'"' && bytes[i] != b'\'' {
                    i += 1;
                }
                let name = src[start..i].to_string();
                while i < bytes.len() && bytes[i].is_ascii_whitespace() {
                    i += 1;
                }
                if i < bytes.len() && (bytes[i] == b'"' || bytes[i] == b'\'') {
                    let q = bytes[i];
                    i += 1;
                    let vs = i;
                    while i < bytes.len() && bytes[i] != q {
                        i += 1;
                    }
                    // The harvest pass is a pre-scan; a malformed value is
                    // reported later, by the declaration parser proper.
                    let Ok(val) = decode_charrefs(&src[vs..i]) else {
                        continue;
                    };
                    pes.insert(name, val);
                }
            }
        } else {
            i += 1;
        }
    }
}

fn subst_pe(
    src: &str,
    pes: &std::collections::HashMap<String, String>,
    internal: bool,
) -> Result<String, XmlError> {
    let mut out = String::new();
    let mut chars = src.chars().peekable();
    let mut in_comment = false;
    // In the internal subset a parameter entity reference may only occur where
    // a markup declaration can occur -- never inside one. `<!ELEMENT %pe;` and
    // `<!ENTITY foo "%e;">` were both expanded happily.
    let mut in_decl = false;
    // A PE reference is recognized inside an EntityValue but NOT inside an
    // attribute default, where `%` is an ordinary character. `<!ATTLIST d a
    // CDATA "%e;">` is a valid document and must stay one.
    let mut decl_is_entity = false;
    let mut in_quote: Option<char> = None;
    while let Some(c) = chars.next() {
        if in_comment {
            out.push(c);
            if c == '-' && chars.peek() == Some(&'-') {
                out.push(chars.next().unwrap());
                if chars.peek() == Some(&'>') {
                    out.push(chars.next().unwrap());
                    in_comment = false;
                }
            }
            continue;
        }
        if c == '>' && in_decl {
            in_decl = false;
            out.push(c);
            continue;
        }
        if in_decl {
            match in_quote {
                Some(q) if c == q => in_quote = None,
                None if c == '"' || c == '\'' => in_quote = Some(c),
                _ => {}
            }
        }
        if c == '<' && chars.peek() == Some(&'!') {
            in_decl = true;
            decl_is_entity = false;
            in_quote = None;
            out.push(c);
            out.push(chars.next().unwrap());
            if chars.peek() == Some(&'-') {
                out.push(chars.next().unwrap());
                if chars.peek() == Some(&'-') {
                    out.push(chars.next().unwrap());
                    in_comment = true;
                    in_decl = false;
                }
            } else {
                // Which declaration this is decides whether a `%` inside its
                // literals is a reference at all.
                let rest: String = chars.clone().take(6).collect();
                decl_is_entity = rest.starts_with("ENTITY");
            }
            continue;
        }
        if c == '%' {
            // `%` followed by anything that cannot start a Name is the PE
            // marker of an `<!ENTITY % name ...>` declaration, not a reference.
            let is_ref = chars
                .peek()
                .is_some_and(|n| crate::chvalid::xml_is_name_start_char(*n as u32, false));
            if !is_ref {
                // PEReference ::= '%' Name ';' -- `%;` has no name at all.
                if chars.peek() == Some(&';') {
                    return Err(XmlError::new(
                        crate::error::XML_ERR_ENTITYREF_NO_NAME,
                        "PEReference: no name",
                        0,
                        0,
                    ));
                }
                out.push('%');
                continue;
            }
            // Inside an attribute default the `%` is literal; leave it alone.
            if in_decl && in_quote.is_some() && !decl_is_entity {
                out.push('%');
                continue;
            }
            if internal && in_decl {
                return Err(XmlError::new(
                    crate::error::XML_ERR_ENTITYREF_NO_NAME,
                    "PEReferences forbidden in internal subset",
                    0,
                    0,
                ));
            }
            let mut name = String::new();
            let mut terminated = false;
            while let Some(&n) = chars.peek() {
                if n == ';' {
                    chars.next();
                    terminated = true;
                    break;
                }
                if !crate::chvalid::xml_is_name_char(n as u32, false) {
                    break;
                }
                name.push(n);
                chars.next();
            }
            // `%paaa` and `%paaa ;` were both accepted. The semicolon is not
            // optional, and no whitespace may come before it.
            if !terminated {
                return Err(XmlError::new(
                    crate::error::XML_ERR_ENTITYREF_SEMICOL_MISSING,
                    "PEReference: expecting ';'",
                    0,
                    0,
                ));
            }
            if let Some(v) = pes.get(&name) {
                out.push_str(v);
            } else {
                out.push('%');
                out.push_str(&name);
                out.push(';');
            }
            continue;
        }
        out.push(c);
    }
    Ok(out)
}

/// Decode character references in an entity value or attribute default.
///
/// EntityValue forbids a bare `&`: it must begin a character or entity
/// reference. Nothing checked that, so `&49;` was kept as literal text and
/// `&#002f;` -- digits followed by a non-digit -- was silently left alone
/// instead of being reported as an invalid decimal value.
///
/// General entity references are kept verbatim; they are expanded at the point
/// of use, not here.
fn decode_charrefs(s: &str) -> Result<String, &'static str> {
    let mut out = String::new();
    let mut it = s.chars().peekable();
    while let Some(c) = it.next() {
        if c != '&' {
            out.push(c);
            continue;
        }
        if it.peek() == Some(&'#') {
            it.next();
            let hex = matches!(it.peek(), Some('x') | Some('X'));
            let upper_x = it.peek() == Some(&'X');
            if hex {
                it.next();
            }
            let mut digits = String::new();
            while let Some(&d) = it.peek() {
                if hex && d.is_ascii_hexdigit() || !hex && d.is_ascii_digit() {
                    digits.push(d);
                    it.next();
                } else {
                    break;
                }
            }
            // `&#X41;` -- the production spells the marker lowercase only.
            if upper_x || digits.is_empty() || it.next() != Some(';') {
                return Err(if hex {
                    "CharRef: invalid hexadecimal value"
                } else {
                    "CharRef: invalid decimal value"
                });
            }
            let radix = if hex { 16 } else { 10 };
            let v = u32::from_str_radix(&digits, radix)
                .map_err(|_| "CharRef: value out of range")?;
            match char::from_u32(v).filter(|ch| crate::chvalid::xml_is_char(*ch as u32)) {
                Some(ch) => out.push(ch),
                None => return Err("CharRef: invalid XML character"),
            }
            continue;
        }
        // A general entity reference: keep it, but it must be well formed.
        let mut name = String::new();
        while let Some(&d) = it.peek() {
            if crate::chvalid::xml_is_name_char(d as u32, false) {
                name.push(d);
                it.next();
            } else {
                break;
            }
        }
        if name.is_empty() || it.next() != Some(';') {
            return Err("EntityValue: '&' forbidden except for entities references");
        }
        out.push('&');
        out.push_str(&name);
        out.push(';');
    }
    Ok(out)
}

struct DtdParser<'a> {
    src: &'a str,
    pos: usize,
    dtd: &'a mut XmlDtd,
    /// The internal subset carries rules the external one does not: no
    /// conditional sections, and a parameter entity may not supply part of a
    /// declaration.
    internal: bool,
    /// XML 1.0 before the 5th edition: the narrower name character classes.
    /// The DTD parser had no idea this option existed, so a name illegal
    /// under the old rules sailed through in a declaration -- and a PI
    /// target inside the subset was never checked at all.
    old10: bool,
}

impl<'a> DtdParser<'a> {
    fn rest(&self) -> &'a str {
        &self.src[self.pos..]
    }
    /// Whitespace only. Where the grammar says S it means S, not "whatever
    /// happens to be in the way" -- skip_ws_and_comments swallows the SGML
    /// `-- comment --` form and PIs, which is exactly how a malformed
    /// declaration slipped past.
    fn skip_ws(&mut self) {
        let r = self.rest();
        let trimmed = r.trim_start_matches([' ', '\t', '\r', '\n']);
        self.pos += r.len() - trimmed.len();
    }

    fn skip_ws_and_comments(&mut self) -> Result<(), XmlError> {
        loop {
            let r = self.rest();
            let trimmed = r.trim_start();
            let n = r.len() - trimmed.len();
            self.pos += n;
            if self.rest().starts_with("<!--") {
                if let Some(e) = self.rest().find("-->") {
                    self.pos += e + 3;
                    continue;
                }
            }
            if self.rest().starts_with("<?") {
                // An XML declaration is only legal at the very start of the
                // document. Inside the internal subset it is a PI whose target
                // is reserved, and this loop skipped every PI without looking.
                let after = &self.rest()[2..];
                // .get(..3), not [..3]: a byte index that lands inside a
                // multi-byte character panics, and a PI target is arbitrary
                // text. The suite hit this on the first run.
                let is_xml_decl = after
                    .get(..3)
                    .is_some_and(|k| k.eq_ignore_ascii_case("xml"))
                    && after[3..]
                        .chars()
                        .next()
                        .is_none_or(|c| c.is_whitespace() || c == '?');
                if is_xml_decl {
                    return Err(self.err(
                        "XML declaration allowed only at the start of the document",
                    ));
                }
                // A PI inside the subset was skipped without a glance at its
                // target. That is where the suite puts its illegal-name cases
                // -- roughly three hundred of them.
                let after = &self.rest()[2..];
                let mut target_len = 0usize;
                for (i, ch) in after.char_indices() {
                    let ok = if i == 0 {
                        crate::chvalid::xml_is_name_start_char(ch as u32, self.old10)
                    } else {
                        crate::chvalid::xml_is_name_char(ch as u32, self.old10)
                    };
                    if !ok {
                        break;
                    }
                    target_len = i + ch.len_utf8();
                }
                if target_len == 0 {
                    return Err(self.err("xmlParsePI : no target name"));
                }
                // The target has to END there too. `<?_` followed by a
                // character that is not a name character is not a PI with the
                // target `_`; it is a PI with an illegal character in its
                // target, which is what the suite is testing.
                let tail = &after[target_len..];
                let ends_cleanly = tail.is_empty()
                    || tail.starts_with("?>")
                    || tail.chars().next().is_some_and(char::is_whitespace);
                if !ends_cleanly {
                    return Err(self.err("xmlParsePI : invalid character in target name"));
                }
                if let Some(e) = self.rest().find("?>") {
                    self.pos += e + 2;
                    continue;
                }
            }
            break;
        }
        Ok(())
    }
    fn parse_markup(&mut self) -> Result<(), XmlError> {
        loop {
            self.skip_ws_and_comments()?;
            if self.pos >= self.src.len() {
                break;
            }
            if self.rest().starts_with("<!ELEMENT") {
                self.parse_element()?;
            } else if self.rest().starts_with("<!ATTLIST") {
                self.parse_attlist()?;
            } else if self.rest().starts_with("<!ENTITY") {
                self.parse_entity()?;
            } else if self.rest().starts_with("<!NOTATION") {
                self.parse_notation()?;
            } else if self.rest().starts_with("<![") {
                // INCLUDE and IGNORE sections are external-subset only.
                if self.internal {
                    return Err(self.err("Content error in the internal subset"));
                }
                self.skip_cond()?;
            } else if self.rest().starts_with('<') {
                // Anything else beginning with '<' is not a markup
                // declaration, and skipping it accepted `<ELEMENT ...>` with
                // the bang missing, `<!Attlist ...>` and `<!notation ...>`
                // with the keyword miscased, and every other near-miss.
                return Err(self.err("Content error in the internal subset"));
            } else {
                self.pos += self.rest().chars().next().unwrap().len_utf8();
            }
        }
        Ok(())
    }
    fn skip_decl(&mut self) -> Result<(), XmlError> {
        if let Some(i) = self.rest().find('>') {
            self.pos += i + 1;
            Ok(())
        } else {
            self.pos = self.src.len();
            Ok(())
        }
    }
    fn skip_cond(&mut self) -> Result<(), XmlError> {
        let mut depth = 0i32;
        let bytes = self.rest().as_bytes();
        let mut i = 0;
        while i < bytes.len() {
            if bytes[i] == b'<' && bytes.get(i..i + 3) == Some(b"<![") {
                depth += 1;
                i += 3;
                continue;
            }
            if bytes[i] == b']' && bytes.get(i..i + 3) == Some(b"]]>") {
                depth -= 1;
                i += 3;
                if depth == 0 {
                    self.pos += i;
                    return Ok(());
                }
                continue;
            }
            i += 1;
        }
        self.pos = self.src.len();
        Ok(())
    }
    fn bump(&mut self, n: usize) {
        self.pos += n;
    }
    fn parse_name(&mut self) -> String {
        // A misplaced XML declaration is reported by the markup loop; here we
        // only need the position advanced.
        let _ = self.skip_ws_and_comments();
        let r = self.rest();
        let mut n = 0;
        // Names here were ASCII-only, so `<!ELEMENT เจมส์ (#PCDATA)>` -- a
        // perfectly valid declaration -- came back empty and the document was
        // rejected. The document body accepted the same name happily; only the
        // DTD disagreed.
        for (i, c) in r.char_indices() {
            let ok = if i == 0 {
                crate::chvalid::xml_is_name_start_char(c as u32, self.old10)
            } else {
                crate::chvalid::xml_is_name_char(c as u32, self.old10)
            };
            if !ok {
                break;
            }
            n = i + c.len_utf8();
        }
        let s = r[..n].to_string();
        self.bump(n);
        s
    }
    /// Read a quoted literal from the internal subset.
    ///
    /// This returned a bare String and so could not report anything. An
    /// ATTLIST default or entity value holding a C0 control byte was
    /// therefore accepted, copied into every element that took the default,
    /// and written back out as U+FFFD -- a value the document never
    /// contained. C stops at the declaration with "invalid character in
    /// entity value". Found by the fuzz round-trip check, which saw the
    /// first save escape the character and the second not.
    fn parse_quoted(&mut self) -> Result<String, XmlError> {
        self.skip_ws_and_comments()?;
        let r = self.rest();
        if r.starts_with('"') || r.starts_with('\'') {
            let q = r.as_bytes()[0] as char;
            self.bump(1);
            if let Some(e) = self.rest().find(q) {
                let s = decode_charrefs(&self.rest()[..e]).map_err(|m| self.err(m))?;
                self.bump(e + 1);
                if let Some(bad) =
                    s.chars().find(|c| !crate::chvalid::xml_is_char(*c as u32))
                {
                    return Err(XmlError::new(
                        crate::error::XML_ERR_INVALID_CHAR,
                        format!("invalid character 0x{:X} in entity value", bad as u32),
                        0,
                        0,
                    ));
                }
                return Ok(s);
            }
        }
        Ok(String::new())
    }
    fn parse_element(&mut self) -> Result<(), XmlError> {
        self.bump("<!ELEMENT".len());
        if !self.require_ws() {
            return Err(self.err("Space required after '<!ELEMENT'"));
        }
        let name = self.parse_name();
        if name.is_empty() {
            return Err(self.err("Element name expected"));
        }
        if !self.require_ws() {
            return Err(self.err("Space required after the element name"));
        }
        let decl = if self.rest().starts_with("EMPTY") {
            self.bump(5);
            ElementDecl::Empty
        } else if self.rest().starts_with("ANY") {
            self.bump(3);
            ElementDecl::Any
        } else if self.rest().starts_with('(') {
            let mut spec = self.take_until_gt_paren();
            // take_until_gt_paren stops at the closing paren, so a trailing
            // occurrence indicator is still in the stream. It is part of the
            // content spec and Mixed content is not valid without it.
            if let Some(q @ ('?' | '*' | '+')) = self.rest().chars().next() {
                self.bump(1);
                spec.push(q);
            }
            // The content model was never checked, only scanned for '#PCDATA'
            // and split on '|'. Everything else was accepted: `(a & b)`,
            // `(a b)`, `(a|b,c)` mixing connectors, `(doc*?)`, `()`. That is
            // 73 conformance cases, and the unchecked loop behind it was the
            // 32 GB allocation.
            let mixed = validate_contentspec(&spec).map_err(|e| self.err(e))?;
            if mixed {
                let mut names = Vec::new();
                for part in spec.split('|') {
                    let t = part.trim().trim_matches(|c: char| c == '(' || c == ')' || c == '*');
                    if t != "#PCDATA" && !t.is_empty() {
                        names.push(t.to_string());
                    }
                }
                ElementDecl::Mixed(names)
            } else {
                ElementDecl::Children(spec)
            }
        } else {
            return Err(self.err("xmlParseElementDecl: 'EMPTY', 'ANY' or '(' expected"));
        };
        if self.dtd.elements.contains_key(&name) {
            self.dtd.duplicate_elements.push(name.clone());
        }
        self.dtd.elements.insert(name, decl);
        self.expect_decl_end("Element")
    }
    fn take_until_gt_paren(&mut self) -> String {
        let r = self.rest();
        let mut depth = 0i32;
        let mut i = 0;
        for (off, c) in r.char_indices() {
            match c {
                '(' => depth += 1,
                ')' => {
                    depth -= 1;
                    if depth == 0 {
                        i = off + 1;
                        break;
                    }
                }
                '>' if depth == 0 => {
                    i = off;
                    break;
                }
                _ => {}
            }
            i = off + c.len_utf8();
        }
        let s = r[..i].to_string();
        self.bump(i);
        s
    }
    fn parse_attlist(&mut self) -> Result<(), XmlError> {
        self.bump("<!ATTLIST".len());
        if !self.require_ws() {
            return Err(self.err("Space required after '<!ATTLIST'"));
        }
        let elem = self.parse_name();
        if elem.is_empty() {
            return Err(self.err("Element name expected in ATTLIST"));
        }
        loop {
            // AttDef ::= S Name S AttType S DefaultDecl -- every one of those
            // S is required, and none of them was checked.
            let had_ws = self.require_ws();
            self.skip_ws_and_comments()?;
            if self.rest().starts_with('>') {
                self.bump(1);
                break;
            }
            if self.pos >= self.src.len() {
                return Err(self.err("xmlParseAttributeListDecl: not terminated"));
            }
            if !had_ws {
                return Err(self.err("Space required after the attribute name"));
            }
            let aname = self.parse_name();
            if aname.is_empty() {
                return Err(self.err("Attribute name expected"));
            }
            if !self.require_ws() {
                return Err(self.err("Space required after the attribute name"));
            }
            self.skip_ws_and_comments()?;
            let mut enumerated = Vec::new();
            let att_type = if self.rest().starts_with('(') {
                let spec = self.take_until_gt_paren();
                // Enumeration ::= '(' S? Nmtoken (S? '|' S? Nmtoken)* S? ')'
                // Only '|' separates. `(foo,bar)` used to be accepted because
                // this split on '|' and shrugged at whatever else was inside.
                let body = spec.trim();
                if !body.starts_with('(') || !body.ends_with(')') {
                    return Err(self.err("')' required to finish ATTLIST enumeration"));
                }
                for part in body[1..body.len() - 1].split('|') {
                    let t = part.trim();
                    if t.is_empty() || !t.chars().all(|c| crate::chvalid::xml_is_name_char(c as u32, false)) {
                        return Err(self.err("')' required to finish ATTLIST enumeration"));
                    }
                    enumerated.push(t.to_string());
                }
                "ENUMERATION".into()
            } else {
                let t = self.parse_name();
                // AttType is a closed set. `NAME` is not in it, and was taken
                // as a perfectly good type.
                const TYPES: &[&str] = &[
                    "CDATA", "ID", "IDREF", "IDREFS", "ENTITY", "ENTITIES", "NMTOKEN",
                    "NMTOKENS", "NOTATION",
                ];
                if !TYPES.contains(&t.as_str()) {
                    return Err(self.err("'(' required to start ATTLIST enumeration"));
                }
                if t == "NOTATION" {
                    if !self.require_ws() {
                        return Err(self.err("Space required after 'NOTATION'"));
                    }
                    if !self.rest().starts_with('(') {
                        return Err(self.err("'(' required to start ATTLIST enumeration"));
                    }
                    let spec = self.take_until_gt_paren();
                    for part in spec.trim().trim_matches(['(', ')']).split('|') {
                        let n = part.trim();
                        // NotationType ::= 'NOTATION' S '(' S? Name (S? '|' S?
                        // Name)* S? ')' -- every entry is a Name, and an empty
                        // or malformed one was quietly dropped.
                        let ok = !n.is_empty()
                            && n.chars().enumerate().all(|(i, c)| {
                                if i == 0 {
                                    crate::chvalid::xml_is_name_start_char(c as u32, self.old10)
                                } else {
                                    crate::chvalid::xml_is_name_char(c as u32, self.old10)
                                }
                            });
                        if !ok {
                            return Err(self.err("Name expected in NOTATION declaration"));
                        }
                        enumerated.push(n.to_string());
                    }
                }
                t
            };
            if !self.require_ws() {
                return Err(self.err("Space required after the attribute type"));
            }
            self.skip_ws_and_comments()?;
            let (default, default_value) = if self.rest().starts_with("#REQUIRED") {
                self.bump(9);
                (AttrDefault::Required, None)
            } else if self.rest().starts_with("#IMPLIED") {
                self.bump(8);
                (AttrDefault::Implied, None)
            } else if self.rest().starts_with("#FIXED") {
                self.bump(6);
                if !self.require_ws() {
                    return Err(self.err("Space required after '#FIXED'"));
                }
                if !self.at_quote() {
                    return Err(self.err("AttValue: \" or ' expected"));
                }
                (AttrDefault::Fixed, Some(self.parse_quoted()?))
            } else {
                // A default value is an AttValue, which is quoted. `v1` bare
                // was accepted and silently became an empty string.
                if !self.at_quote() {
                    return Err(self.err("AttValue: \" or ' expected"));
                }
                (AttrDefault::Value, Some(self.parse_quoted()?))
            };
            // "When more than one definition is provided for the same
            // attribute of a given element type, the FIRST declaration is
            // binding and later declarations are ignored." We were inserting
            // into a map, so the last one won and a later #REQUIRED overrode
            // an earlier default.
            self.dtd.attributes.entry((elem.clone(), aname)).or_insert(
                AttrDecl {
                    att_type,
                    default,
                    default_value,
                    enumerated,
                },
            );
        }
        Ok(())
    }
    /// `NotationDecl ::= '<!NOTATION' S Name S (ExternalID | PublicID) S? '>'`
    ///
    /// This went to skip_decl, which took everything up to the next '>' and
    /// asked no questions: a missing space, a missing name, a public
    /// identifier holding characters the production forbids, all accepted.
    fn parse_notation(&mut self) -> Result<(), XmlError> {
        self.bump("<!NOTATION".len());
        if !self.require_ws() {
            return Err(self.err("Space required after '<!NOTATION'"));
        }
        let name = self.parse_name();
        if name.is_empty() {
            return Err(self.err("Notation name expected"));
        }
        if !self.require_ws() {
            return Err(self.err("Space required after the notation name"));
        }
        let public = if self.rest().starts_with("PUBLIC") {
            true
        } else if self.rest().starts_with("SYSTEM") {
            false
        } else {
            return Err(self.err("'PUBLIC' or 'SYSTEM' expected in NOTATION"));
        };
        self.bump(6);
        if !self.require_ws() {
            return Err(self.err("Space required after the external ID keyword"));
        }
        if !self.at_quote() {
            return Err(self.err("Unfinished System or Public ID \" or ' expected"));
        }
        let first = self.parse_quoted()?;
        if public {
            if let Some(bad) = first.chars().find(|c| !is_pubid_char(*c)) {
                return Err(self.err(&format!(
                    "Invalid character 0x{:X} in public identifier",
                    bad as u32
                )));
            }
            // PublicID (notation only) may stop after the public identifier;
            // ExternalID continues with a system literal.
            let before = self.pos;
            self.skip_ws();
            if self.at_quote() {
                if self.pos == before {
                    return Err(self.err("Space required after the Public Identifier"));
                }
                self.parse_quoted()?;
            }
        }
        self.expect_decl_end("Notation")
    }

    fn parse_entity(&mut self) -> Result<(), XmlError> {
        self.bump("<!ENTITY".len());
        if !self.require_ws() {
            return Err(self.err("Space required after '<!ENTITY'"));
        }
        let pe = self.rest().starts_with('%');
        if pe {
            self.bump(1);
            if !self.require_ws() {
                return Err(self.err("Space required after '%'"));
            }
        }
        let name = self.parse_name();
        if name.is_empty() {
            return Err(self.err("Entity name expected"));
        }
        // EntityDecl requires S between the name and the definition. Without
        // this, `<!ENTITY foo"some text">` was accepted.
        if !self.require_ws() {
            return Err(self.err("Space required after the entity name"));
        }
        if self.rest().starts_with("SYSTEM") || self.rest().starts_with("PUBLIC") {
            let public = self.rest().starts_with("PUBLIC");
            self.bump(6);
            if !self.require_ws() {
                return Err(self.err("Space required after the external ID keyword"));
            }
            if public {
                // ExternalID ::= 'PUBLIC' S PubidLiteral S SystemLiteral --
                // two literals, with space between them. One was accepted, and
                // so was `"whatever""e.ent"` with no space.
                let pid = self.parse_quoted()?;
                if let Some(bad) = pid.chars().find(|c| !is_pubid_char(*c)) {
                    return Err(self.err(&format!(
                        "Invalid character 0x{:X} in public identifier",
                        bad as u32
                    )));
                }
                if !self.require_ws() {
                    return Err(self.err("Space required after the Public Identifier"));
                }
                if !self.at_quote() {
                    return Err(self.err("SystemLiteral expected"));
                }
            }
            self.parse_quoted()?;
            // NDataDecl is the only thing allowed to follow, and it needs the
            // space before it. Measure BEFORE skipping, or the skip eats the
            // very thing being checked for.
            let ws_before_ndata = {
                let before = self.pos;
                self.skip_ws();
                self.pos > before
            };
            self.skip_ws_and_comments()?;
            if self.rest().starts_with("NDATA") {
                // A parameter entity is always parsed; NDATA is for unparsed
                // general entities only.
                if pe {
                    return Err(self.err("xmlParseEntityDecl: entity not terminated"));
                }
                if !ws_before_ndata {
                    return Err(self.err("Space required before 'NDATA'"));
                }
                self.bump(5);
                if !self.require_ws() {
                    return Err(self.err("Space required after 'NDATA'"));
                }
                if self.parse_name().is_empty() {
                    return Err(self.err("Notation name expected after 'NDATA'"));
                }
                // An NDATA entity is unparsed, and only an unparsed entity may
                // be the value of an ENTITY attribute.
                self.dtd.unparsed_entities.insert(name.clone());
                self.skip_ws();
            }
            return self.expect_decl_end("entity");
        }
        if !self.at_quote() {
            return Err(self.err("Entity value expected"));
        }
        let val = self.parse_quoted()?;
        if pe {
            self.dtd.parameter_entities.insert(name, val);
        } else {
            self.dtd.entities.insert(name, val);
        }
        self.skip_ws();
        self.expect_decl_end("entity")
    }

    /// Position of the parser as a line and column, so an error points at the
    /// declaration rather than at 0:0.
    fn line_col(&self) -> (u32, u32) {
        let mut line = 1u32;
        let mut col = 1u32;
        for c in self.src[..self.pos.min(self.src.len())].chars() {
            if c == '\n' {
                line += 1;
                col = 1;
            } else {
                col += 1;
            }
        }
        (line, col)
    }

    fn err(&self, msg: &str) -> XmlError {
        let (line, col) = self.line_col();
        XmlError::new(crate::error::XML_ERR_SPACE_REQUIRED, msg, line, col)
    }

    /// Consume required whitespace, reporting whether any was there.
    fn require_ws(&mut self) -> bool {
        let before = self.pos;
        self.skip_ws();
        self.pos > before || self.pos >= self.src.len()
    }

    fn at_quote(&self) -> bool {
        self.rest().starts_with('"') || self.rest().starts_with('\'')
    }

    /// A declaration ends at '>' and nothing else. It used to fall through to
    /// skip_decl(), which swallowed whatever was in the way -- including the
    /// SGML `-- comment --` form that XML does not have.
    fn expect_decl_end(&mut self, what: &str) -> Result<(), XmlError> {
        self.skip_ws();
        if self.rest().starts_with('>') {
            self.bump(1);
            Ok(())
        } else {
            Err(self.err(&format!("xmlParse{what}Decl: not terminated")))
        }
    }
}

/// Merge `src` into `dst` (external subset onto internal).
pub fn merge_dtd(dst: &mut XmlDtd, src: XmlDtd) {
    dst.entities.extend(src.entities);
    dst.unparsed_entities.extend(src.unparsed_entities);
    dst.duplicate_elements.extend(src.duplicate_elements);
    dst.parameter_entities.extend(src.parameter_entities);
    dst.elements.extend(src.elements);
    dst.attributes.extend(src.attributes);
    if dst.public_id.is_none() {
        dst.public_id = src.public_id;
    }
    if dst.system_id.is_none() {
        dst.system_id = src.system_id;
    }
}

/// Validate a content specification against XML 1.0 productions 46-51.
///
/// Returns `Ok(true)` for Mixed content, `Ok(false)` for a children model.
///
/// ```text
/// Mixed    ::= '(' S? '#PCDATA' (S? '|' S? Name)* S? ')*'
///            | '(' S? '#PCDATA' S? ')'
/// children ::= (choice | seq) ('?' | '*' | '+')?
/// cp       ::= (Name | choice | seq) ('?' | '*' | '+')?
/// choice   ::= '(' S? cp ( S? '|' S? cp )+ S? ')'
/// seq      ::= '(' S? cp ( S? ',' S? cp )* S? ')'
/// ```
///
/// The two rules that catch most malformed models: a group may not mix `,` and
/// `|` at the same level, and Mixed content that names elements must close
/// with `)*`.
fn validate_contentspec(spec: &str) -> Result<bool, &'static str> {
    let mut p = SpecParser {
        b: spec.as_bytes(),
        i: 0,
        depth: 0,
    };
    p.ws();
    if !p.eat(b'(') {
        return Err("ContentDecl : '(' expected");
    }
    p.ws();
    if p.b[p.i..].starts_with(b"#PCDATA") {
        p.i += 7;
        let mut named = false;
        loop {
            p.ws();
            if p.eat(b')') {
                break;
            }
            if !p.eat(b'|') {
                return Err("ContentDecl : ',' '|' or ')' expected");
            }
            p.ws();
            p.name()?;
            named = true;
        }
        // `(#PCDATA|a)` without the star is not a legal Mixed model.
        let star = p.eat(b'*');
        if named && !star {
            return Err("Element content model is not finished with ')*'");
        }
        p.ws();
        return if p.i == p.b.len() {
            Ok(true)
        } else {
            Err("trailing content after the content model")
        };
    }
    // A children model: rewind to the '(' and read it as a group.
    p.i = 0;
    p.ws();
    p.group()?;
    p.quant();
    p.ws();
    if p.i != p.b.len() {
        return Err("ContentDecl : garbage after the content model");
    }
    Ok(false)
}

struct SpecParser<'a> {
    b: &'a [u8],
    i: usize,
    depth: u32,
}

impl SpecParser<'_> {
    fn ws(&mut self) {
        while matches!(self.b.get(self.i), Some(b' ' | b'\t' | b'\r' | b'\n')) {
            self.i += 1;
        }
    }
    fn peek(&self) -> Option<u8> {
        self.b.get(self.i).copied()
    }
    fn eat(&mut self, c: u8) -> bool {
        if self.peek() == Some(c) {
            self.i += 1;
            true
        } else {
            false
        }
    }
    fn quant(&mut self) {
        if matches!(self.peek(), Some(b'?' | b'*' | b'+')) {
            self.i += 1;
        }
    }
    fn name(&mut self) -> Result<(), &'static str> {
        let start = self.i;
        // Names here are ASCII in practice, but a UTF-8 name must not be cut
        // mid-character, so decode properly.
        let rest = match std::str::from_utf8(&self.b[self.i..]) {
            Ok(r) => r,
            Err(_) => return Err("invalid UTF-8 in the content model"),
        };
        let mut chars = rest.char_indices();
        match chars.next() {
            Some((_, c)) if crate::chvalid::xml_is_name_start_char(c as u32, false) => {
                self.i += c.len_utf8();
            }
            _ => return Err("Name expected in the content model"),
        }
        for (off, c) in chars {
            if !crate::chvalid::xml_is_name_char(c as u32, false) {
                self.i = start + off;
                return Ok(());
            }
            self.i = start + off + c.len_utf8();
        }
        Ok(())
    }
    /// choice or seq. Which one is decided by the first separator, and the
    /// group must then use only that one.
    fn group(&mut self) -> Result<(), &'static str> {
        // `((((((...` must not recurse the stack away.
        self.depth += 1;
        if self.depth > 256 {
            return Err("content model nested too deeply");
        }
        if !self.eat(b'(') {
            return Err("ContentDecl : '(' expected");
        }
        self.ws();
        self.cp()?;
        self.ws();
        let sep = match self.peek() {
            Some(b')') => {
                self.i += 1;
                self.depth -= 1;
                return Ok(());
            }
            Some(c @ (b'|' | b',')) => c,
            _ => return Err("ContentDecl : ',' '|' or ')' expected"),
        };
        loop {
            if !self.eat(sep) {
                // A different connector at the same level: `(a|b,c)`.
                return Err("ContentDecl : ',' '|' or ')' expected");
            }
            self.ws();
            self.cp()?;
            self.ws();
            match self.peek() {
                Some(b')') => {
                    self.i += 1;
                    self.depth -= 1;
                    return Ok(());
                }
                Some(c) if c == sep => continue,
                _ => return Err("ContentDecl : ',' '|' or ')' expected"),
            }
        }
    }
    fn cp(&mut self) -> Result<(), &'static str> {
        if self.peek() == Some(b'(') {
            self.group()?;
        } else {
            self.name()?;
        }
        self.quant();
        Ok(())
    }
}

/// PubidChar ::= #x20 | #xD | #xA | [a-zA-Z0-9] | [-'()+,./:=?;!*#@$_%]
///
/// A public identifier is a restricted character set, not free text. Nothing
/// checked it, so `<!NOTATION n PUBLIC "a^b">` was accepted.
pub fn is_pubid_char(c: char) -> bool {
    matches!(c, ' ' | '\r' | '\n')
        || c.is_ascii_alphanumeric()
        || matches!(
            c,
            '-' | '\'' | '(' | ')' | '+' | ',' | '.' | '/' | ':'
                | '=' | '?' | ';' | '!' | '*' | '#' | '@' | '$' | '_' | '%'
        )
}

/// Well-formedness constraints on the entity graph, checked once the whole
/// subset is parsed.
///
/// `decode_charrefs` keeps general entity references verbatim, because they are
/// expanded at the point of use. Nothing then looked at them, so an entity
/// value or an ATTLIST default could reference an entity that was never
/// declared, or one declared NDATA (which may not be referenced at all), or
/// itself by way of a cycle. All three were accepted silently, and the cycle
/// only surfaced later as a depth-limit error pointing at the wrong entity.
fn check_entity_graph(dtd: &XmlDtd) -> Result<(), XmlError> {
    const PREDEFINED: &[&str] = &["lt", "gt", "amp", "apos", "quot"];
    let err = |m: String| XmlError::new(crate::error::XML_ERR_UNDECLARED_ENTITY, m, 0, 0);

    // Every reference in a literal must name a declared, parsed entity.
    let mut refs: std::collections::HashMap<&str, Vec<String>> = Default::default();
    let literals = dtd
        .entities
        .iter()
        .map(|(k, v)| (k.as_str(), v.as_str()))
        .chain(
            dtd.attributes
                .iter()
                .filter_map(|((_, a), d)| d.default_value.as_deref().map(|v| (a.as_str(), v))),
        );
    for (owner, text) in literals {
        for name in entity_refs_in(text) {
            if PREDEFINED.contains(&name.as_str()) {
                continue;
            }
            if dtd.unparsed_entities.contains(&name) {
                return Err(err(format!("Entity reference to unparsed entity {name}")));
            }
            if !dtd.entities.contains_key(&name) {
                return Err(err(format!("Entity '{name}' not defined")));
            }
            refs.entry(owner).or_default().push(name);
        }
    }

    // A cycle in the reference graph is a well-formedness error, not something
    // to discover by running out of depth.
    for start in dtd.entities.keys() {
        let mut seen = std::collections::HashSet::new();
        let mut stack = vec![start.as_str()];
        while let Some(cur) = stack.pop() {
            if !seen.insert(cur) {
                continue;
            }
            for next in refs.get(cur).into_iter().flatten() {
                if next == start {
                    return Err(err("Detected an entity reference loop".into()));
                }
                stack.push(next.as_str());
            }
        }
    }
    Ok(())
}

/// The general entity references in a literal, as `&name;` occurrences.
fn entity_refs_in(text: &str) -> Vec<String> {
    let mut out = Vec::new();
    let mut it = text.chars().peekable();
    while let Some(c) = it.next() {
        if c != '&' || it.peek() == Some(&'#') {
            continue;
        }
        let mut name = String::new();
        while let Some(&d) = it.peek() {
            if crate::chvalid::xml_is_name_char(d as u32, false) {
                name.push(d);
                it.next();
            } else {
                break;
            }
        }
        if !name.is_empty() && it.peek() == Some(&';') {
            it.next();
            out.push(name);
        }
    }
    out
}