pdfcraft/crates/cos/src/security.rs

//! Bridging the COS object model and `pdfcraft-crypt`: reading `/Encrypt` dictionaries and
//! decrypting / encrypting the strings and streams of an object (§7.6.2).
//!
//! Not encrypted, per the spec: the `/Encrypt` dictionary itself, cross-reference streams, the
//! trailer `/ID`, objects inside object streams (the object stream is encrypted as a whole),
//! the `/Contents` of signature dictionaries, and metadata streams when `/EncryptMetadata` is
//! false. Streams naming a crypt filter (`/Crypt` in `/Filter`) use that filter; embedded files
//! use `/EFF`.

use std::cell::Cell;

use pdfcraft_crypt::{EncryptDict, Method, SecurityHandler, StreamKind};
use pdfcraft_filters::MAX_CHAIN;

use crate::document::Held;
use crate::{Dict, Document, Object, PdfString, Stream};

/// How many references one read or write of an object may follow, with all that it loads on the
/// way. How an object is treated can be given by reference, and what that loads can do the same: a
/// file could make one read a tree of them, each object read again once the cache has let it go.
/// An object takes a few of these, and two dozen if every part of it is given by reference; past
/// this they read as absent. The parameters of an object stream, read together, count as one.
const MAX_FOLLOWED: u32 = 32;

std::thread_local! {
    /// How many [`FollowScope`]s this thread is inside.
    static SCOPES: Cell<u32> = const { Cell::new(0) };
    /// The references followed in the outermost of them.
    static FOLLOWED: Cell<u32> = const { Cell::new(0) };
    /// The references refused in the outermost of them, past [`MAX_FOLLOWED`].
    static REFUSED: Cell<u32> = const { Cell::new(0) };
}

/// One read or write of an object, with everything it loads on the way: what [`MAX_FOLLOWED`] counts.
pub(crate) struct FollowScope;

impl FollowScope {
    pub(crate) fn enter() -> Self {
        if SCOPES.replace(SCOPES.get().saturating_add(1)) == 0 {
            FOLLOWED.set(0);
            REFUSED.set(0);
        }
        FollowScope
    }
}

impl Drop for FollowScope {
    fn drop(&mut self) {
        SCOPES.set(SCOPES.get().saturating_sub(1));
    }
}

/// What `f` gives, counted against [`MAX_FOLLOWED`], or `spent` past it. Called inside a
/// [`FollowScope`], which is what starts the count again.
pub(crate) fn following<T>(spent: T, f: impl FnOnce() -> T) -> T {
    if FOLLOWED.get() >= MAX_FOLLOWED {
        REFUSED.set(REFUSED.get().saturating_add(1));
        return spent;
    }
    FOLLOWED.set(FOLLOWED.get() + 1);
    f()
}

/// `o`, or the object it refers to, for what decides how a stream or dictionary is treated. Past
/// [`MAX_FOLLOWED`] a reference reads as absent.
fn followed<'o>(doc: &Document, o: &'o Object) -> Held<'o> {
    match o {
        Object::Ref(_) => following(Held::Direct(&Object::Null), || doc.held(o)),
        o => Held::Direct(o),
    }
}

fn bytes_of(doc: &Document, d: &Dict, key: &[u8]) -> Vec<u8> {
    d.get(key).map(|v| doc.resolve(v)).and_then(|v| v.as_string().map(|s| s.bytes.clone())).unwrap_or_default()
}

/// Read the handler parameters from an `/Encrypt` dictionary.
pub(crate) fn encrypt_dict(doc: &Document, d: &Dict) -> EncryptDict {
    let int = |k: &[u8], default: i64| d.get(k).map(|v| doc.resolve(v)).and_then(|v| v.as_int()).unwrap_or(default);
    let name = |d: &Dict, k: &[u8]| d.get(k).and_then(|v| doc.resolve(v).as_name().map(<[u8]>::to_vec));
    let mut crypt_filters = Vec::new();
    if let Some(cf) = d.get(b"CF").map(|c| doc.resolve(c)).and_then(|c| c.as_dict().cloned()) {
        for (k, v) in cf.iter() {
            let Some(f) = doc.resolve(v).as_dict().cloned() else { continue };
            let method = match name(&f, b"CFM").as_deref() {
                Some(b"V2") => Method::Rc4,
                Some(b"AESV2") => Method::Aes128,
                Some(b"AESV3") => Method::Aes256,
                _ => Method::Identity,
            };
            crypt_filters.push((k.clone(), method));
        }
    }
    let v = int(b"V", 0);
    // V4: the key length comes from the standard crypt filter's /Length (bytes, or bits in
    // some files); /Length on the dictionary is often a stale 40. A length that isn't
    // positive means nothing and reads as if absent.
    let cf_len = d
        .get(b"CF")
        .map(|c| doc.resolve(c))
        .and_then(|c| c.as_dict().and_then(|c| c.get(b"StdCF").map(|f| doc.resolve(f))))
        .and_then(|f| f.as_dict().and_then(|f| f.get(b"Length").map(|n| doc.resolve(n))))
        .and_then(|n| n.as_int())
        .filter(|n| *n > 0);
    let length_bits = match (v, cf_len) {
        (5.., _) => 256,
        (4, Some(n)) if n <= 32 => n * 8,
        (4, Some(n)) => n,
        (4, None) => 128,
        _ => int(b"Length", 40),
    };
    EncryptDict {
        filter: name(d, b"Filter").unwrap_or_default(),
        v,
        r: int(b"R", 0),
        length_bits,
        o: bytes_of(doc, d, b"O"),
        u: bytes_of(doc, d, b"U"),
        oe: bytes_of(doc, d, b"OE"),
        ue: bytes_of(doc, d, b"UE"),
        perms: bytes_of(doc, d, b"Perms"),
        p: int(b"P", -1) as i32,
        encrypt_metadata: !matches!(d.get(b"EncryptMetadata").map(|v| doc.resolve(v)).as_deref(), Some(Object::Bool(false))),
        crypt_filters,
        stm_f: name(d, b"StmF").unwrap_or_else(|| b"Identity".to_vec()),
        str_f: name(d, b"StrF").unwrap_or_else(|| b"Identity".to_vec()),
        ef_f: name(d, b"EFF").unwrap_or_default(),
    }
}

/// Whether `d`, an `/Encrypt` dictionary, has crypt filters that encrypt (`/CFM` `V2`, `AESV2` or
/// `AESV3`), each authorised only when an embedded file is opened (`/AuthEvent /EFOpen`, ISO
/// 32000-2 §7.6.6, Table 25): what a document asks for its embedded files alone, which it opens
/// without (§7.6.6: "EFOpen Authorization shall be required when accessing embedded files"). The
/// files use them through `/EFF` or through a `/Crypt` filter of their own. Its strings and streams
/// have to be in the clear too, which [`SecurityHandler::without_key`] checks.
pub(crate) fn efopen_only(doc: &Document, d: &Dict) -> bool {
    let name = |d: &Dict, k: &[u8]| d.get(k).and_then(|v| doc.resolve(v).as_name().map(<[u8]>::to_vec));
    let Some(cf) = d.get(b"CF").map(|c| doc.resolve(c)).and_then(|c| c.as_dict().cloned()) else { return false };
    let mut encrypting = Vec::new();
    for (k, v) in cf.iter() {
        let Some(f) = doc.resolve(v).as_dict().cloned() else { continue };
        if !matches!(name(&f, b"CFM").as_deref(), Some(b"V2" | b"AESV2" | b"AESV3")) {
            continue;
        }
        if name(&f, b"AuthEvent").as_deref() != Some(b"EFOpen") {
            return false;
        }
        encrypting.push(k.clone());
    }
    !encrypting.is_empty()
}

/// Whether `s`, read from a document whose handler has no file key, holds data that handler would
/// decrypt, which therefore reads as stored: still encrypted.
pub(crate) fn stream_reads_as_stored(doc: &Document, h: &SecurityHandler, s: &Stream) -> bool {
    if h.has_key() {
        return false;
    }
    let _scope = FollowScope::enter();
    let named = named_crypt_filter(doc, &s.dict);
    let ty = s.dict.get(b"Type").map(|t| followed(doc, t));
    let kind = match (&named, ty.as_deref().and_then(Object::as_name)) {
        (Some(n), _) => StreamKind::Named(n),
        (None, Some(b"EmbeddedFile")) => StreamKind::EmbeddedFile,
        _ => StreamKind::Normal,
    };
    h.decrypts(kind)
}

/// Whether `f`, an entry of a `/Filter` chain or its only filter, is `/Crypt`.
fn is_crypt(doc: &Document, f: &Object) -> bool {
    followed(doc, f).as_name() == Some(b"Crypt")
}

/// The crypt filter a stream names in its `/Filter` chain (`/Crypt` + `/DecodeParms /Name`), each
/// given directly or by reference, in an array or not. A chain longer than any decoder takes names
/// none: nothing can read that stream anyway, and every stream of a file can share one long array.
/// A `/Name` that reads as absent means `/Identity`, unless a reference on the way to it was refused
/// (see [`MAX_FOLLOWED`]): that says nothing about the stream, so it names none, and the document's
/// own filter is used.
fn named_crypt_filter(doc: &Document, dict: &Dict) -> Option<Vec<u8>> {
    let refused = REFUSED.get();
    let filter = followed(doc, dict.get(b"Filter")?);
    let i = match &*filter {
        Object::Array(a) if a.len() <= MAX_CHAIN + 1 => a.iter().position(|f| is_crypt(doc, f))?,
        f if is_crypt(doc, f) => 0,
        _ => return None,
    };
    let parms = dict.get(b"DecodeParms").map(|p| followed(doc, p));
    let entry = match parms.as_deref() {
        Some(Object::Array(a)) => a.get(i).map(|p| followed(doc, p)),
        Some(p) if i == 0 => Some(Held::Direct(p)),
        _ => None,
    };
    let name = entry.as_deref().and_then(Object::as_dict).and_then(|p| p.get(b"Name")).map(|n| followed(doc, n));
    match name.as_deref().and_then(Object::as_name) {
        Some(name) => Some(name.to_vec()),
        // /Identity here would read ciphertext as data, and write the stream in the clear.
        None if REFUSED.get() > refused => None,
        None => Some(b"Identity".to_vec()),
    }
}

/// A signature dictionary (field signature, document timestamp or usage-rights signature),
/// whose `/Contents` is never encrypted (ISO 32000-2 §7.6.2, §12.8.1): the signer writes a
/// placeholder, hashes the file around it and patches the value in place, so encrypting it
/// would make signing impossible. Identified as other readers do: by `/Type`, or by a string
/// `/Contents` alongside an array `/ByteRange` (`/Type` is optional). A signature dictionary
/// has no `/Subtype`; a dictionary with one (an annotation, whose `/Contents` is the comment
/// text) or with another `/Type` is never exempt, whatever else it carries.
/// With `doc` its entries are read through references; a dictionary nested in another object
/// has none (see [`transform`]).
fn is_signature_dict(doc: Option<&Document>, d: &Dict) -> bool {
    // All this decides is what happens to such a /Contents.
    if !matches!(d.get(b"Contents"), Some(Object::String(_))) {
        return false;
    }
    // A null entry, or one that refers to nothing, is an absent one (ISO 32000-2 §7.3.7, §7.3.10).
    let entry = |key: &[u8]| {
        let value = d.get(key)?;
        let held = doc.map_or(Held::Direct(value), |doc| followed(doc, value));
        (!matches!(*held, Object::Null)).then_some(held)
    };
    if entry(b"Subtype").is_some() {
        return false;
    }
    match entry(b"Type") {
        Some(t) => matches!(t.as_name(), Some(b"Sig" | b"DocTimeStamp")),
        None => matches!(entry(b"ByteRange").as_deref(), Some(Object::Array(_))),
    }
}

/// Longer than any filter's name. A name can be of any length, and a chain every stream shares can
/// hold a long one.
const MAX_FILTER_NAME: usize = 127;

/// The dictionary of `source` without the `/Crypt` entries of its filter chain and their decode
/// parameters. The two come out direct when given by reference, which every stream of a file can
/// share, so of what was shared only small values are copied: filter names, and the parameters
/// decoders read ([`Document::with_direct_parms`]). The chain is short, as [`named_crypt_filter`]
/// found a filter in it.
fn drop_crypt_filter(doc: &Document, source: &Stream) -> Dict {
    let entry = |f: &Object| match f {
        f if is_crypt(doc, f) => Object::name("Crypt"),
        Object::Name(n) if n.len() <= MAX_FILTER_NAME => f.clone(),
        Object::Ref(_) => f.clone(),
        _ => Object::Null,
    };
    let filter = source.dict.get(b"Filter").map(|f| followed(doc, f));
    let chain: Vec<Object> = match filter.as_deref() {
        Some(Object::Array(a)) => a.iter().map(entry).collect(),
        Some(f) => vec![entry(f)],
        None => Vec::new(),
    };
    let crypt: Vec<bool> = chain.iter().map(|f| f.as_name() == Some(b"Crypt")).collect();
    let mut chained = Stream::from_raw(source.dict.clone(), Vec::new());
    chained.dict.set(b"Filter".to_vec(), Object::Array(chain.clone()));
    // A lone parameter dictionary belongs to the first filter (as `named_crypt_filter` reads it).
    let parms = match doc.with_direct_parms(&chained).dict.get(b"DecodeParms") {
        Some(Object::Array(p)) => Some(Object::Array(p.iter().zip(&crypt).filter(|(_, c)| !**c).map(|(p, _)| p.clone()).collect())),
        Some(Object::Dict(_)) if crypt.first() == Some(&true) => None,
        _ => source.dict.get(b"DecodeParms").cloned(),
    };
    let mut dict = source.dict.clone();
    match parms {
        Some(p) => dict.set(b"DecodeParms".to_vec(), p),
        None => {
            dict.remove(b"DecodeParms");
        }
    }
    let filters: Vec<Object> = chain.into_iter().zip(&crypt).filter(|(_, c)| !**c).map(|(f, _)| f).collect();
    if filters.is_empty() {
        dict.remove(b"Filter");
        if matches!(dict.get(b"DecodeParms"), Some(Object::Array(p)) if p.is_empty()) {
            dict.remove(b"DecodeParms");
        }
    } else {
        dict.set(b"Filter".to_vec(), Object::Array(filters));
    }
    dict
}

/// What [`transform`] does to an object.
#[derive(Clone, Copy, PartialEq, Eq)]
pub(crate) enum Pass {
    /// Decrypt, with the handler the document was opened with.
    Decrypt,
    /// Encrypt with the handler the document was opened with: streams keep their crypt filters.
    Encrypt,
    /// Encrypt with protection applied by `set_encryption`. A stream's crypt filter came from
    /// the source's `/Encrypt`; unless the new one defines that name it is dropped, so the
    /// stream follows the new policy instead of reading as `/Identity` and staying in clear.
    EncryptReplacing,
}

/// Decrypt or encrypt every string and stream in an indirect object, except a signature
/// dictionary's `/Contents` (see [`is_signature_dict`]).
///
/// `o` is as `doc` holds it: what a stream's dictionary gives by reference is read from `doc`, so
/// its references have to be the document's, not a save's renumbered ones.
pub(crate) fn transform(doc: &Document, h: &SecurityHandler, o: &Object, num: u32, generation: u16, pass: Pass) -> Object {
    let _scope = FollowScope::enter();
    let decrypt = pass == Pass::Decrypt;
    let strings = |s: &PdfString| {
        let bytes = if decrypt { h.decrypt_string(num, generation, &s.bytes) } else { h.encrypt_string(num, generation, &s.bytes) };
        PdfString { bytes, hex: s.hex || !decrypt }
    };
    // References are followed here only, for the object itself, never for what is nested in it.
    // `walk` goes deep, and a load from down there adds that depth again for each object loaded,
    // which overflowed the stack on a 10 KB file. A stream can't be nested, and a dictionary that
    // is reads as it says directly.
    match o {
        Object::Stream(s) => stream(doc, s, &strings, h, num, generation, pass),
        Object::Dict(d) if is_signature_dict(Some(doc), d) => signature_dict(d, &strings),
        o => walk(o, &strings),
    }
}

/// `d`, a signature dictionary, with the strings of everything but its `/Contents` transformed.
fn signature_dict(d: &Dict, strings: &dyn Fn(&PdfString) -> PdfString) -> Object {
    Object::Dict(
        d.iter()
            .map(|(k, v)| match (k.as_slice(), v) {
                (b"Contents", Object::String(_)) => (k.clone(), v.clone()),
                _ => (k.clone(), walk(v, strings)),
            })
            .collect(),
    )
}

fn walk(o: &Object, strings: &dyn Fn(&PdfString) -> PdfString) -> Object {
    match o {
        Object::String(s) => Object::String(strings(s)),
        Object::Array(a) => Object::Array(a.iter().map(|x| walk(x, strings)).collect()),
        Object::Dict(d) if is_signature_dict(None, d) => signature_dict(d, strings),
        Object::Dict(d) => Object::Dict(d.iter().map(|(k, v)| (k.clone(), walk(v, strings))).collect()),
        other => other.clone(),
    }
}

fn stream(
    doc: &Document,
    s: &Stream,
    strings: &dyn Fn(&PdfString) -> PdfString,
    h: &SecurityHandler,
    num: u32,
    generation: u16,
    pass: Pass,
) -> Object {
    let decrypt = pass == Pass::Decrypt;
    let walked = |d: &Dict| -> Dict { d.iter().map(|(k, v)| (k.clone(), walk(v, strings))).collect() };
    let ty = s.dict.get(b"Type").map(|t| followed(doc, t));
    let ty = ty.as_deref().and_then(Object::as_name);
    // Cross-reference streams are never encrypted (§7.5.8), but only reading may trust the
    // label: a stream a producer calls /XRef stays plaintext if writing believes it, and the
    // writer builds its own cross-reference streams outside this transform.
    if (decrypt && ty == Some(b"XRef")) || (ty == Some(b"Metadata") && !h.encrypts_metadata()) {
        return Object::Stream(Stream { dict: walked(&s.dict), raw: s.raw.clone() });
    }
    let mut named = named_crypt_filter(doc, &s.dict);
    // Before the walk, which then encrypts what is in the dictionary this leaves.
    let foreign = pass == Pass::EncryptReplacing && named.as_deref().is_some_and(|n| !h.dict().crypt_filters.iter().any(|(f, _)| f == n));
    let dropped = foreign.then(|| drop_crypt_filter(doc, s));
    if dropped.is_some() {
        named = None;
    }
    let dict = walked(dropped.as_ref().unwrap_or(&s.dict));
    let kind = match (&named, ty) {
        (Some(n), _) => StreamKind::Named(n),
        (None, Some(b"EmbeddedFile")) => StreamKind::EmbeddedFile,
        _ => StreamKind::Normal,
    };
    let raw = if decrypt { h.decrypt_stream(num, generation, &s.raw, kind) } else { h.encrypt_stream(num, generation, &s.raw, kind) };
    Object::Stream(Stream { dict, raw: raw.into() })
}