check: charge relationship parts to the package's decompression budget

apps/gridcraft-cli/src/check/package.rs

@@ -6,6 +6,10 @@ const MAX_XML_PART: u64 = 64 * 1024 * 1024;
 /// Most zip entries looked at while validating a written package.
 const MAX_ENTRIES: usize = 200_000;
 
+/// Decompressed bytes that comparing one package's relationship kinds may read in total: the
+/// relationship parts themselves and the targets whose content is checked.
+const RELATIONSHIP_BUDGET: u64 = 512 * 1024 * 1024;
+
 /// Relationship kinds (lower case) whose absence after a save isn't a loss: the calculation chain is
 /// a cache Excel rebuilds, and GridCraft never writes it.
 pub(super) const REBUILT_KINDS: [&str; 1] = ["calcchain"];
@@ -117,6 +121,28 @@ fn read_zip_part(zip: &mut Zip<'_>, name: &str) -> Result<Vec<u8>, String> {
     Ok(buf)
 }
 
+enum PartError {
+    /// Reading the part would spend more than the remaining budget.
+    OverBudget,
+    Unreadable,
+}
+
+/// Reads a part like [`read_zip_part`], charging the decompressed bytes to `budget`.
+fn read_zip_part_within(zip: &mut Zip<'_>, name: &str, budget: &mut u64) -> Result<Vec<u8>, PartError> {
+    let f = zip.by_name(name).map_err(|_| PartError::Unreadable)?;
+    // One byte past what is left shows that the part doesn't fit.
+    let limit = MAX_XML_PART.min(budget.saturating_add(1));
+    let mut buf = Vec::new();
+    f.take(limit).read_to_end(&mut buf).map_err(|_| PartError::Unreadable)?;
+    let read = buf.len() as u64;
+    if read > *budget {
+        *budget = 0;
+        return Err(PartError::OverBudget);
+    }
+    *budget -= read;
+    Ok(buf)
+}
+
 /// Whether the producer flagged its cached results as stale with `<calcPr fullCalcOnLoad="1">`:
 /// libraries that don't calculate write it, and store 0 or nothing as every formula's value.
 pub(super) fn full_calc_on_load(bytes: &[u8]) -> bool {
@@ -232,19 +258,30 @@ pub fn validate_package(bytes: &[u8]) -> Result<Vec<String>, String> {
 /// nothing are left out, as a save may rightly drop them: a link to a place in the workbook
 /// (`#Sheet1!A1`, written as a cell's `location`) and a target part with nothing in it.
 pub(super) fn relationship_kinds(bytes: &[u8]) -> Result<BTreeMap<String, String>, String> {
+    relationship_kinds_within(bytes, RELATIONSHIP_BUDGET)
+}
+
+/// [`relationship_kinds`] reading at most `budget` decompressed bytes. `Err` once the relationship
+/// parts alone need more: their kinds can't be listed completely.
+fn relationship_kinds_within(bytes: &[u8], mut budget: u64) -> Result<BTreeMap<String, String>, String> {
     let mut zip = zip::ZipArchive::new(Cursor::new(bytes)).map_err(|e| format!("not a zip package: {e}"))?;
     let names = zip_names(&mut zip);
     let by_key: HashMap<String, &String> = names.iter().map(|n| (part_key(n), n)).collect();
     // Several relationships name the same part; each fresh check decompressed it again, so
     // repeated references to one highly compressible part multiplied the work. The verdict is
     // remembered per (part, named content) and the decompression the package can cause is
-    // budgeted as a whole (finding: repeated decompression).
+    // budgeted as a whole, the relationship parts included: many large, unreferenced rels parts
+    // would otherwise each spend a full per-part allowance.
     let mut verdicts: HashMap<(String, &'static str), bool> = HashMap::new();
-    let mut budget: u64 = 512 * 1024 * 1024;
     let mut kinds = BTreeMap::new();
     for n in names.iter().filter(|n| part_key(n).ends_with(".rels")) {
         let Some(src_dir) = rels_source_dir(&part_key(n)) else { continue };
-        let Ok(els) = read_zip_part(&mut zip, n).and_then(|xml| xml_elements(&xml)) else { continue };
+        let xml = match read_zip_part_within(&mut zip, n, &mut budget) {
+            Ok(xml) => xml,
+            Err(PartError::OverBudget) => return Err(format!("{n}: relationship check stopped at its decompression limit")),
+            Err(PartError::Unreadable) => continue,
+        };
+        let Ok(els) = xml_elements(&xml) else { continue };
         for (_, attrs) in els.iter().filter(|(el, _)| el == "Relationship") {
             let Some(kind) = attr(attrs, "Type").and_then(|t| t.rsplit('/').next()) else { continue };
             let target = attr(attrs, "Target").unwrap_or("");
@@ -329,3 +366,36 @@ fn holds_something_uncached(zip: &mut Zip<'_>, name: &str, marker: &'static str,
         buf.clear();
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use std::io::Write as _;
+
+    use super::*;
+
+    /// A package with `count` relationship parts of `size` bytes each, none of them describing an
+    /// existing part.
+    fn many_rels(count: usize, size: usize) -> Vec<u8> {
+        let mut z = zip::ZipWriter::new(Cursor::new(Vec::new()));
+        let opts = zip::write::SimpleFileOptions::default().compression_method(zip::CompressionMethod::Deflated);
+        let head = r#"<?xml version="1.0"?><Relationships xmlns="http://schemas.openxmlformats.org/package/2006/relationships"><Relationship Id="rId1" Type="http://schemas.openxmlformats.org/officeDocument/2006/relationships/custom" Target="https://example.com" TargetMode="External"/>"#;
+        let tail = "</Relationships>";
+        let filler = " ".repeat(size.saturating_sub(head.len() + tail.len()));
+        for i in 0..count {
+            z.start_file(format!("xl/_rels/part{i}.xml.rels"), opts).unwrap();
+            z.write_all(head.as_bytes()).unwrap();
+            z.write_all(filler.as_bytes()).unwrap();
+            z.write_all(tail.as_bytes()).unwrap();
+        }
+        z.finish().unwrap().into_inner()
+    }
+
+    #[test]
+    fn relationship_parts_count_against_the_decompression_budget() {
+        // Each part fits the per-part allowance, but together they need more than the budget.
+        let bytes = many_rels(8, 64 * 1024);
+        assert!(relationship_kinds_within(&bytes, 256 * 1024).is_err());
+        let kinds = relationship_kinds_within(&bytes, 1024 * 1024).unwrap();
+        assert_eq!(kinds.get("custom").map(String::as_str), Some("custom"));
+    }
+}

crates/functions/src/lift.rs

@@ -68,70 +68,26 @@ const SCANS: &[&str] = &["VLOOKUP", "HLOOKUP", "LOOKUP", "MATCH", "XLOOKUP", "XM
 /// `XMATCH(SEQUENCE(1048576), -SEQUENCE(1048576))` runs for hours and hangs opening the file.
 /// Past the budget the call is `#NUM!`, as MMULT is past 2e9 multiply-adds. 5e8 keeps the worst
 /// case to a few seconds and still admits 5,000 needles over a 100,000-row table.
-///
-/// The same budget covers all the work of a lifted call added up across its elements: criteria
-/// aggregates rescanning their ranges, period loops, and array results built per element.
 const LIFTED_SCAN_BUDGET: usize = 500_000_000;
 
-/// Criteria aggregates: every call tests each cell of its ranges against its criteria, so a call
-/// lifted over `n` criteria does `n` x (range cells) work, like the [`SCANS`] lookups.
-const CRITERIA: &[&str] = &["COUNTIF", "COUNTIFS", "SUMIF", "SUMIFS", "AVERAGEIF", "AVERAGEIFS", "MAXIFS", "MINIFS"];
-
-/// Cumulative TVM functions that step through every period from `start_period` (argument 3) to
-/// `end_period` (argument 4) on each call.
-const PERIOD_LOOPS: &[&str] = &["CUMIPMT", "CUMPRINC"];
-
-/// The work every lifted call of `spec` repeats over its non-lifted arguments: the longest vector
-/// a [`SCANS`] lookup searches, or every cell a [`CRITERIA`] aggregate tests. 0 for other
-/// functions.
+/// The length of the longest vector a lifted call of `spec` searches per needle: 0 unless `spec`
+/// is one of the [`SCANS`] lookups and a non-lifted argument is an array.
 fn scan_len(spec: &FnSpec, args: &[Arg]) -> usize {
-    let scans = SCANS.contains(&spec.name);
-    if !scans && !CRITERIA.contains(&spec.name) {
+    if !SCANS.contains(&spec.name) {
         return 0;
     }
-    let lens = args.iter().enumerate().filter(|(i, _)| !spec.is_scalar_arg(*i)).filter_map(|(_, a)| match &a.value {
-        Value::Array(arr) => Some(if scans { arr.rows.max(arr.cols) } else { arr.data.len() }),
-        _ => None,
-    });
-    if scans { lens.max().unwrap_or(0) } else { lens.fold(0, usize::saturating_add) }
-}
-
-/// The work of one lifted call that depends on its own (lifted) scalar arguments: the number of
-/// periods a [`PERIOD_LOOPS`] function steps through. 0 for other functions.
-fn call_cost(spec: &FnSpec, cell_args: &[Arg]) -> usize {
-    if !PERIOD_LOOPS.contains(&spec.name) {
-        return 0;
-    }
-    let period = |i: usize| match cell_args.get(i).map(|a| &a.value) {
-        Some(Value::Number(n)) if n.is_finite() => Some(n.trunc()),
-        _ => None,
-    };
-    match (period(3), period(4)) {
-        // The function itself refuses spans past 1e7 periods; anything larger costs the cap.
-        (Some(start), Some(end)) if end >= start => ((end - start) + 1.0).min(1e8) as usize,
-        _ => 0,
-    }
+    args.iter()
+        .enumerate()
+        .filter(|(i, _)| !spec.is_scalar_arg(*i))
+        .filter_map(|(_, a)| match &a.value {
+            Value::Array(arr) => Some(arr.rows.max(arr.cols)),
+            _ => None,
+        })
+        .max()
+        .unwrap_or(0)
 }
 
 pub(crate) fn call_lifted(spec: &FnSpec, args: &[Arg], ctx: &mut dyn Ctx) -> Value {
-    call_lifted_within(spec, args, ctx, LIFTED_SCAN_BUDGET)
-}
-
-/// Sets the lifted arguments of `cell_args` to their elements at (`r`, `c`).
-fn element_args(spec: &FnSpec, args: &[Arg], r: usize, c: usize, cell_args: &mut [Arg]) {
-    for (i, a) in args.iter().enumerate() {
-        if spec.is_scalar_arg(i)
-            && let Value::Array(arr) = &a.value
-            && let Some(slot) = cell_args.get_mut(i)
-        {
-            let v = arr.get_broadcast(r, c);
-            slot.value = if let Value::Array(inner) = &v { inner.data.first().cloned().unwrap_or(Value::Empty) } else { v };
-        }
-    }
-}
-
-/// [`call_lifted`] with every call's work added up against `budget` (see [`LIFTED_SCAN_BUDGET`]).
-fn call_lifted_within(spec: &FnSpec, args: &[Arg], ctx: &mut dyn Ctx, budget: usize) -> Value {
     let rejected = if spec.value_args.is_empty() { None } else { reject_multi_cell_refs(spec, args) };
     let args = rejected.as_deref().unwrap_or(args);
     let mut rows = 0usize;
@@ -150,42 +106,25 @@ fn call_lifted_within(spec: &FnSpec, args: &[Arg], ctx: &mut dyn Ctx, budget: us
         return (spec.imp)(args, ctx);
     }
     let needles = rows.saturating_mul(cols);
-    // Every call is charged its repeated scan (at least 1), its own period loop, and the cells of
-    // an array result (built in full, though only its first cell is kept), so lifting a scalar
-    // argument can't multiply what one call may do into hours of work:
-    // `COUNTIF(big_range, SEQUENCE(1e4))`, `CUMIPMT(0.1, 1e7, 1, 1, SEQUENCE(1e4,,1e7,0), 0)`,
-    // `MUNIT(SEQUENCE(1e4,,3000,0))`.
-    let per_call = scan_len(spec, args).max(1);
-    let mut cell_args: Vec<Arg> = args.to_vec();
-    let mut spent = needles.saturating_mul(per_call);
-    if needles > MAX_CELLS || spent > budget {
+    if needles > MAX_CELLS || needles.saturating_mul(scan_len(spec, args)) > LIFTED_SCAN_BUDGET {
         return Value::Error(CellError::Num);
     }
-    // Period loops are known before any call runs, so a call over budget does none of them.
-    if PERIOD_LOOPS.contains(&spec.name) {
-        for r in 0..rows {
-            for c in 0..cols {
-                element_args(spec, args, r, c, &mut cell_args);
-                spent = spent.saturating_add(call_cost(spec, &cell_args));
-                if spent > budget {
-                    return Value::Error(CellError::Num);
-                }
-            }
-        }
-    }
+    let mut cell_args: Vec<Arg> = args.to_vec();
     let mut data = Vec::with_capacity(rows * cols);
     for r in 0..rows {
         for c in 0..cols {
-            element_args(spec, args, r, c, &mut cell_args);
+            for (i, a) in args.iter().enumerate() {
+                if spec.is_scalar_arg(i)
+                    && let Value::Array(arr) = &a.value
+                    && let Some(slot) = cell_args.get_mut(i)
+                {
+                    let v = arr.get_broadcast(r, c);
+                    slot.value = if let Value::Array(inner) = &v { inner.data.first().cloned().unwrap_or(Value::Empty) } else { v };
+                }
+            }
             let out = (spec.imp)(&cell_args, ctx);
             data.push(match out {
-                Value::Array(a) => {
-                    spent = spent.saturating_add(a.data.len());
-                    if spent > budget {
-                        return Value::Error(CellError::Num);
-                    }
-                    a.data.first().cloned().unwrap_or(Value::Empty)
-                }
+                Value::Array(a) => a.data.first().cloned().unwrap_or(Value::Empty),
                 v => v,
             });
         }
@@ -270,52 +209,6 @@ mod tests {
         assert_eq!(rows_of(&out), vec![vec![nv(-1.0)]; 1000]);
     }
 
-    /// Security: lifting a criteria aggregate rescans its range for every criterion, so criteria
-    /// x range cells past the budget is `#NUM!` before any scan runs.
-    #[test]
-    fn lifted_criteria_aggregates_stop_past_their_work_budget() {
-        use gridcraft_core::Array;
-        let tall = |n: usize| av(Array::filled(n, 1, nv(-1.0)).into());
-        is_err(ev("COUNTIF", vec![tall(1_000_000), tall(501)]), CellError::Num);
-        is_err(ev("SUMIF", vec![tall(1_000_000), tall(501)]), CellError::Num);
-        is_err(ev("COUNTIFS", vec![tall(1_000_000), tall(501)]), CellError::Num);
-        // Two ranges of 250,000 cells each count together.
-        is_err(ev("SUMIFS", vec![tall(250_000), tall(250_000), tall(1001)]), CellError::Num);
-        let counts = ev("COUNTIF", vec![av(col(&[1.0, 2.0, 2.0])), av(col(&[2.0, 1.0, 5.0]))]);
-        assert_eq!(rows_of(&counts), vec![vec![nv(2.0)], vec![nv(1.0)], vec![nv(0.0)]]);
-        for name in super::CRITERIA.iter().chain(super::PERIOD_LOOPS) {
-            assert!(crate::lookup(name).is_some(), "{name} isn't registered");
-        }
-    }
-
-    /// Security: CUMIPMT and CUMPRINC loop over their periods once per lifted element; the
-    /// periods of all elements are added up first, so an over-budget call runs none of them.
-    #[test]
-    fn lifted_period_loops_stop_past_their_work_budget() {
-        use gridcraft_core::Array;
-        let ends = |len: usize, end: f64| av(Array::filled(len, 1, nv(end)).into());
-        let args = |len: usize, end: f64| vec![n(0.01), n(1e7), n(1000.0), n(1.0), ends(len, end), n(0.0)];
-        // 51 elements of 1e7 periods each is past 5e8.
-        is_err(ev("CUMIPMT", args(51, 1e7)), CellError::Num);
-        is_err(ev("CUMPRINC", args(51, 1e7)), CellError::Num);
-        let out = ev("CUMIPMT", args(2, 12.0));
-        assert_eq!(rows_of(&out).len(), 2);
-        assert!(matches!(rows_of(&out).first().and_then(|r| r.first()), Some(Value::Number(x)) if *x < 0.0));
-    }
-
-    /// Security: an array-producing function lifted over a computed array builds each element's
-    /// whole result before keeping one cell; their cells count against the budget.
-    #[test]
-    fn lifted_array_results_count_against_the_budget() {
-        use gridcraft_core::Array;
-        let spec = crate::lookup("MUNIT").unwrap_or_else(|| panic!("MUNIT isn't registered"));
-        let dims = |n: usize| [av(Array::filled(n, 1, nv(10.0)).into())];
-        // Ten 10x10 identity matrices are 1,000 cells.
-        let out = super::call_lifted_within(spec, &dims(10), &mut TestCtx::default(), 1_010);
-        assert_eq!(rows_of(&out), vec![vec![nv(1.0)]; 10]);
-        is_err(super::call_lifted_within(spec, &dims(11), &mut TestCtx::default(), 1_010), CellError::Num);
-    }
-
     #[test]
     fn value_params_reject_multi_cell_references() {
         let s = FnSpec { value_args: &[true, false], ..spec() };