libexpr: detect cycles when converting values to JSON/XML by dyegoaurelio · Pull Request #15888 · NixOS/nix

dyegoaurelio · 2026-05-20T04:35:54Z

Motivation

Serializing a self-referential value such as let a = { b = a; }; in a was
broken in two ways depending on the target:

builtins.toJSON and nix eval --json failed with an opaque
stack overflow; max-call-depth exceeded error.
nix-instantiate --xml streamed thousands of nested <attrs>/<list>
elements to stdout

Context

Fixes #12289.

Add 👍 to pull requests you find important.

The Nix maintainer team uses a GitHub project board to schedule and track reviews.

roberth · 2026-05-20T16:58:01Z

        }
+        const void * key = v.attrs();
+        if (!seen.insert(key).second)
+            state.error<InfiniteRecursionError>("infinite recursion encountered while converting value to JSON")


Could add a bit more detail referring to the cyclic nature.
(and then do the same for nList)

Suggested change

state.error<InfiniteRecursionError>("infinite recursion encountered while converting value to JSON")

state.error<InfiniteRecursionError>("infinite recursion encountered while converting Nix value to JSON, because only cycle-free data can be represented")

roberth · 2026-05-20T20:27:41Z

I don't know if we want to change anything about the XML dump.
I am unaware of its use cases, so the change looks potentially breaking to me.
@xokdvium @edolstra wdyt?

I'd say omit the XML change for now, maybe put it on a branch, unless you use it yourself.

I'm ok removing it from this PR and I don't use this feature myself.

But IMO the current behavior is much worse on the XML dump than on the json side.

nix-instantiate --xml --eval --strict --expr 'let a = { b = a; }; in a' just starts an infinite loop

Failing eagerly would make more sense. Changing a diverging computation -> early error is perfectly safe.

diverging computation -> early error is perfectly safe.

I think I agree actually. The old behavior was not usable, so compatibility is not much of a concern.

xokdvium · 2026-05-21T18:08:43Z

+        const void * key = v.attrs();
+        auto [it, fresh] = seen.try_emplace(key, currentPath ? *currentPath : ValuePath{});
+        if (!fresh)
+            cycleError(key, it->second);
+        Finally cleanup([&] { seen.erase(key); });
+        if (auto i = v.attrs()->get(state.s.outPath)) {
+            if (currentPath)
+                currentPath->emplace_back(state.s.outPath);
+            Finally popOutPath([&] {
+                if (currentPath)
+                    currentPath->pop_back();
+            });
+            return valueToJSON(state, strict, *i->value, i->pos, context, copyToStore, seen, currentPath);
+        } else {
            out = json::object();
            for (auto & a : v.attrs()->lexicographicOrder(state.symbols)) {
+                if (currentPath)
+                    currentPath->emplace_back(a->name);
+                Finally popSegment([&] {
+                    if (currentPath)
+                        currentPath->pop_back();
+                });
                try {
                    out.emplace(
                        state.symbols[a->name],
-                        printValueAsJSON(state, strict, *a->value, a->pos, context, copyToStore));
+                        valueToJSON(state, strict, *a->value, a->pos, context, copyToStore, seen, currentPath));


This does quite a bit more work in the happy (non-error) path. Does this measurably affect performance?

I've just ran some benchs and it did add a 2-4% slowdown on ""realistic"" inputs and ~30% slowdown on deeply nested arrays/attrsets ([ [ [ [ [ 0 ] ] ] ] ])

Bench results

➜ builddir (915772aa5) ✗ ./src/libexpr-tests/nix-expr-benchmarks --benchmark_filter='BM_PrintValueAsJSON_.*' --benchmark_repetitions=10 --benchmark_report_aggregates_only=true 2026-05-23T14:15:19-03:00 Running ./src/libexpr-tests/nix-expr-benchmarks Run on (12 X 5407.65 MHz CPU s) CPU Caches: L1 Data 48 KiB (x6) L1 Instruction 32 KiB (x6) L2 Unified 1024 KiB (x6) L3 Unified 32768 KiB (x1) Load Average: 1.24, 2.25, 2.94 -------------------------------------------------------------------------------------------------------- Benchmark Time CPU Iterations UserCounters... -------------------------------------------------------------------------------------------------------- BM_PrintValueAsJSON_DeepAttrs/100_mean 10355 ns 10346 ns 10 items_per_second=9.66602M/s BM_PrintValueAsJSON_DeepAttrs/100_median 10358 ns 10348 ns 10 items_per_second=9.66331M/s BM_PrintValueAsJSON_DeepAttrs/100_stddev 19.2 ns 19.5 ns 10 items_per_second=18.228k/s BM_PrintValueAsJSON_DeepAttrs/100_cv 0.19 % 0.19 % 10 items_per_second=0.19% BM_PrintValueAsJSON_DeepAttrs/1000_mean 107731 ns 107596 ns 10 items_per_second=9.2942M/s BM_PrintValueAsJSON_DeepAttrs/1000_median 107626 ns 107499 ns 10 items_per_second=9.30239M/s BM_PrintValueAsJSON_DeepAttrs/1000_stddev 433 ns 425 ns 10 items_per_second=36.6456k/s BM_PrintValueAsJSON_DeepAttrs/1000_cv 0.40 % 0.39 % 10 items_per_second=0.39% BM_PrintValueAsJSON_DeepAttrs/5000_mean 565658 ns 564697 ns 10 items_per_second=8.85452M/s BM_PrintValueAsJSON_DeepAttrs/5000_median 565417 ns 564423 ns 10 items_per_second=8.8586M/s BM_PrintValueAsJSON_DeepAttrs/5000_stddev 2890 ns 2969 ns 10 items_per_second=46.4683k/s BM_PrintValueAsJSON_DeepAttrs/5000_cv 0.51 % 0.53 % 10 items_per_second=0.52% BM_PrintValueAsJSON_DeepLists/100_mean 7726 ns 7717 ns 10 items_per_second=12.9599M/s BM_PrintValueAsJSON_DeepLists/100_median 7711 ns 7700 ns 10 items_per_second=12.9864M/s BM_PrintValueAsJSON_DeepLists/100_stddev 63.1 ns 63.3 ns 10 items_per_second=105.394k/s BM_PrintValueAsJSON_DeepLists/100_cv 0.82 % 0.82 % 10 items_per_second=0.81% BM_PrintValueAsJSON_DeepLists/1000_mean 81499 ns 81415 ns 10 items_per_second=12.2829M/s BM_PrintValueAsJSON_DeepLists/1000_median 81429 ns 81350 ns 10 items_per_second=12.2926M/s BM_PrintValueAsJSON_DeepLists/1000_stddev 154 ns 149 ns 10 items_per_second=22.3901k/s BM_PrintValueAsJSON_DeepLists/1000_cv 0.19 % 0.18 % 10 items_per_second=0.18% BM_PrintValueAsJSON_DeepLists/5000_mean 468968 ns 468349 ns 10 items_per_second=10.6762M/s BM_PrintValueAsJSON_DeepLists/5000_median 468764 ns 468179 ns 10 items_per_second=10.6797M/s BM_PrintValueAsJSON_DeepLists/5000_stddev 2934 ns 2935 ns 10 items_per_second=66.8169k/s BM_PrintValueAsJSON_DeepLists/5000_cv 0.63 % 0.63 % 10 items_per_second=0.63% BM_PrintValueAsJSON_BalancedAttrs/4/6_mean 505421 ns 504739 ns 10 items_per_second=10.8198M/s BM_PrintValueAsJSON_BalancedAttrs/4/6_median 503841 ns 503196 ns 10 items_per_second=10.8526M/s BM_PrintValueAsJSON_BalancedAttrs/4/6_stddev 3245 ns 3200 ns 10 items_per_second=68.3479k/s BM_PrintValueAsJSON_BalancedAttrs/4/6_cv 0.64 % 0.63 % 10 items_per_second=0.63% BM_PrintValueAsJSON_BalancedAttrs/8/4_mean 463897 ns 463382 ns 10 items_per_second=10.102M/s BM_PrintValueAsJSON_BalancedAttrs/8/4_median 463920 ns 463355 ns 10 items_per_second=10.1024M/s BM_PrintValueAsJSON_BalancedAttrs/8/4_stddev 2314 ns 2287 ns 10 items_per_second=49.4949k/s BM_PrintValueAsJSON_BalancedAttrs/8/4_cv 0.50 % 0.49 % 10 items_per_second=0.49% BM_PrintValueAsJSON_BalancedAttrs/16/3_mean 473908 ns 473439 ns 10 items_per_second=9.22856M/s BM_PrintValueAsJSON_BalancedAttrs/16/3_median 473676 ns 473227 ns 10 items_per_second=9.23237M/s BM_PrintValueAsJSON_BalancedAttrs/16/3_stddev 3089 ns 3051 ns 10 items_per_second=59.2301k/s BM_PrintValueAsJSON_BalancedAttrs/16/3_cv 0.65 % 0.64 % 10 items_per_second=0.64% BM_PrintValueAsJSON_PackageSet/1000_mean 260759 ns 260438 ns 10 items_per_second=3.83973M/s BM_PrintValueAsJSON_PackageSet/1000_median 260360 ns 260038 ns 10 items_per_second=3.8456M/s BM_PrintValueAsJSON_PackageSet/1000_stddev 851 ns 838 ns 10 items_per_second=12.3212k/s BM_PrintValueAsJSON_PackageSet/1000_cv 0.33 % 0.32 % 10 items_per_second=0.32% BM_PrintValueAsJSON_PackageSet/5000_mean 1552765 ns 1550252 ns 10 items_per_second=3.22592M/s BM_PrintValueAsJSON_PackageSet/5000_median 1547737 ns 1545242 ns 10 items_per_second=3.23574M/s BM_PrintValueAsJSON_PackageSet/5000_stddev 23391 ns 23080 ns 10 items_per_second=47.3982k/s BM_PrintValueAsJSON_PackageSet/5000_cv 1.51 % 1.49 % 10 items_per_second=1.47% BM_PrintValueAsJSON_PackageSet/10000_mean 3297638 ns 3292583 ns 10 items_per_second=3.03736M/s BM_PrintValueAsJSON_PackageSet/10000_median 3282901 ns 3277862 ns 10 items_per_second=3.05077M/s BM_PrintValueAsJSON_PackageSet/10000_stddev 30951 ns 30691 ns 10 items_per_second=28.1163k/s BM_PrintValueAsJSON_PackageSet/10000_cv 0.94 % 0.93 % 10 items_per_second=0.93% ➜ builddir (infinite-recursion-json-decoding) ✗ ./src/libexpr-tests/nix-expr-benchmarks --benchmark_filter='BM_PrintValueAsJSON_.*' --benchmark_repetitions=10 --benchmark_report_aggregates_only=true 2026-05-23T14:17:28-03:00 Running ./src/libexpr-tests/nix-expr-benchmarks Run on (12 X 5345.19 MHz CPU s) CPU Caches: L1 Data 48 KiB (x6) L1 Instruction 32 KiB (x6) L2 Unified 1024 KiB (x6) L3 Unified 32768 KiB (x1) Load Average: 1.59, 2.07, 2.78 -------------------------------------------------------------------------------------------------------- Benchmark Time CPU Iterations UserCounters... -------------------------------------------------------------------------------------------------------- BM_PrintValueAsJSON_DeepAttrs/100_mean 13692 ns 13669 ns 10 items_per_second=7.31601M/s BM_PrintValueAsJSON_DeepAttrs/100_median 13730 ns 13707 ns 10 items_per_second=7.29581M/s BM_PrintValueAsJSON_DeepAttrs/100_stddev 92.7 ns 92.4 ns 10 items_per_second=49.8798k/s BM_PrintValueAsJSON_DeepAttrs/100_cv 0.68 % 0.68 % 10 items_per_second=0.68% BM_PrintValueAsJSON_DeepAttrs/1000_mean 145453 ns 145191 ns 10 items_per_second=6.88756M/s BM_PrintValueAsJSON_DeepAttrs/1000_median 145331 ns 145073 ns 10 items_per_second=6.89307M/s BM_PrintValueAsJSON_DeepAttrs/1000_stddev 516 ns 503 ns 10 items_per_second=23.7781k/s BM_PrintValueAsJSON_DeepAttrs/1000_cv 0.35 % 0.35 % 10 items_per_second=0.35% BM_PrintValueAsJSON_DeepAttrs/5000_mean 751088 ns 749646 ns 10 items_per_second=6.67011M/s BM_PrintValueAsJSON_DeepAttrs/5000_median 750557 ns 749179 ns 10 items_per_second=6.67398M/s BM_PrintValueAsJSON_DeepAttrs/5000_stddev 5454 ns 5300 ns 10 items_per_second=46.6692k/s BM_PrintValueAsJSON_DeepAttrs/5000_cv 0.73 % 0.71 % 10 items_per_second=0.70% BM_PrintValueAsJSON_DeepLists/100_mean 9832 ns 9819 ns 10 items_per_second=10.185M/s BM_PrintValueAsJSON_DeepLists/100_median 9775 ns 9763 ns 10 items_per_second=10.2431M/s BM_PrintValueAsJSON_DeepLists/100_stddev 98.4 ns 97.0 ns 10 items_per_second=100.158k/s BM_PrintValueAsJSON_DeepLists/100_cv 1.00 % 0.99 % 10 items_per_second=0.98% BM_PrintValueAsJSON_DeepLists/1000_mean 108991 ns 108830 ns 10 items_per_second=9.18864M/s BM_PrintValueAsJSON_DeepLists/1000_median 109030 ns 108864 ns 10 items_per_second=9.18579M/s BM_PrintValueAsJSON_DeepLists/1000_stddev 194 ns 205 ns 10 items_per_second=17.2999k/s BM_PrintValueAsJSON_DeepLists/1000_cv 0.18 % 0.19 % 10 items_per_second=0.19% BM_PrintValueAsJSON_DeepLists/5000_mean 545378 ns 544444 ns 10 items_per_second=9.18371M/s BM_PrintValueAsJSON_DeepLists/5000_median 545477 ns 544448 ns 10 items_per_second=9.18361M/s BM_PrintValueAsJSON_DeepLists/5000_stddev 860 ns 866 ns 10 items_per_second=14.6199k/s BM_PrintValueAsJSON_DeepLists/5000_cv 0.16 % 0.16 % 10 items_per_second=0.16% BM_PrintValueAsJSON_BalancedAttrs/4/6_mean 520316 ns 519477 ns 10 items_per_second=10.5127M/s BM_PrintValueAsJSON_BalancedAttrs/4/6_median 520534 ns 519572 ns 10 items_per_second=10.5106M/s BM_PrintValueAsJSON_BalancedAttrs/4/6_stddev 2453 ns 2364 ns 10 items_per_second=47.774k/s BM_PrintValueAsJSON_BalancedAttrs/4/6_cv 0.47 % 0.46 % 10 items_per_second=0.45% BM_PrintValueAsJSON_BalancedAttrs/8/4_mean 484706 ns 483517 ns 10 items_per_second=9.68868M/s BM_PrintValueAsJSON_BalancedAttrs/8/4_median 478137 ns 477374 ns 10 items_per_second=9.80589M/s BM_PrintValueAsJSON_BalancedAttrs/8/4_stddev 14991 ns 14346 ns 10 items_per_second=282.474k/s BM_PrintValueAsJSON_BalancedAttrs/8/4_cv 3.09 % 2.97 % 10 items_per_second=2.92% BM_PrintValueAsJSON_BalancedAttrs/16/3_mean 486369 ns 485096 ns 10 items_per_second=9.00762M/s BM_PrintValueAsJSON_BalancedAttrs/16/3_median 485174 ns 483827 ns 10 items_per_second=9.03011M/s BM_PrintValueAsJSON_BalancedAttrs/16/3_stddev 5781 ns 5795 ns 10 items_per_second=107.483k/s BM_PrintValueAsJSON_BalancedAttrs/16/3_cv 1.19 % 1.19 % 10 items_per_second=1.19% BM_PrintValueAsJSON_PackageSet/1000_mean 271135 ns 270743 ns 10 items_per_second=3.69361M/s BM_PrintValueAsJSON_PackageSet/1000_median 270766 ns 270355 ns 10 items_per_second=3.69884M/s BM_PrintValueAsJSON_PackageSet/1000_stddev 1343 ns 1299 ns 10 items_per_second=17.6638k/s BM_PrintValueAsJSON_PackageSet/1000_cv 0.50 % 0.48 % 10 items_per_second=0.48% BM_PrintValueAsJSON_PackageSet/5000_mean 1589454 ns 1586677 ns 10 items_per_second=3.15127M/s BM_PrintValueAsJSON_PackageSet/5000_median 1589694 ns 1586950 ns 10 items_per_second=3.1507M/s BM_PrintValueAsJSON_PackageSet/5000_stddev 5485 ns 5455 ns 10 items_per_second=10.8287k/s BM_PrintValueAsJSON_PackageSet/5000_cv 0.35 % 0.34 % 10 items_per_second=0.34% BM_PrintValueAsJSON_PackageSet/10000_mean 3365166 ns 3359508 ns 10 items_per_second=2.97668M/s BM_PrintValueAsJSON_PackageSet/10000_median 3369174 ns 3362968 ns 10 items_per_second=2.97356M/s BM_PrintValueAsJSON_PackageSet/10000_stddev 15649 ns 15460 ns 10 items_per_second=13.7507k/s BM_PrintValueAsJSON_PackageSet/10000_cv 0.47 % 0.46 % 10 items_per_second=0.46%

Bench code

commit 25643e9e4741cc0fb073fa398aaf8e9499833914 Author: Dyego Aurélio <dyegoaurelio@gmail.com> Date: Sat May 23 13:31:07 2026 -0300 libexpr-tests: add value-to-json benchmark Covers four value shapes: DeepAttrs — linear chain of single-attr attrsets at depth 100/1k/5k DeepLists — linear chain of single-element lists at depth 100/1k/5k BalancedAttrs — balanced attrset tree at (breadth, depth) = (4,6), (8,4), (16,3); each yields ~5k total nodes PackageSet — wide top-level attrset of derivation-like entries with outPath shortcut, at 1k/5k/10k entries; mimics the shape of an evaluated hackage-packages.nix diff --git a/src/libexpr-tests/meson.build b/src/libexpr-tests/meson.build index 50d158209..55ed20527 100644 --- a/src/libexpr-tests/meson.build +++ b/src/libexpr-tests/meson.build @@ -100,6 +100,7 @@ if get_option('benchmarks') 'dynamic-attrs-bench.cc', 'get-drvs-bench.cc', 'regex-cache-bench.cc', + 'value-to-json-bench.cc', ) benchmark_exe = executable( diff --git a/src/libexpr-tests/value-to-json-bench.cc b/src/libexpr-tests/value-to-json-bench.cc new file mode 100644 index 000000000..acb064ec8 --- /dev/null +++ b/src/libexpr-tests/value-to-json-bench.cc @@ -0,0 +1,196 @@ +#include <benchmark/benchmark.h> +#include <nlohmann/json.hpp> + +#include "nix/expr/value-to-json.hh" +#include "nix/expr/eval.hh" +#include "nix/expr/eval-settings.hh" +#include "nix/fetchers/fetch-settings.hh" +#include "nix/store/store-open.hh" +#include "nix/util/fmt.hh" + +namespace nix { + +static std::shared_ptr<EvalState> makeEvalState() +{ + static auto store = openStore("dummy://"); + static fetchers::Settings fetchSettings; + static bool readOnlyMode = true; + static EvalSettings evalSettings = [] { + EvalSettings s{readOnlyMode}; + s.nixPath = {}; + return s; + }(); + return std::make_shared<EvalState>(LookupPath{}, store, fetchSettings, evalSettings, nullptr); +} + +// Build a linear chain of attrs: { a = { a = { ... { a = 0; } ... } } } of depth `depth`. +static void buildDeepAttrs(EvalState & state, Value & root, size_t depth) +{ + Value * cur = &root; + auto sym = state.symbols.create("a"); + for (size_t i = 0; i < depth; ++i) { + auto b = state.buildBindings(1); + auto & child = b.alloc(sym); + cur->mkAttrs(b.finish()); + cur = &child; + } + cur->mkInt(0); +} + +// Build a linear chain of lists: [[[ ... [ 0 ] ... ]]] of depth `depth`. +static void buildDeepLists(EvalState & state, Value & root, size_t depth) +{ + Value * cur = &root; + for (size_t i = 0; i < depth; ++i) { + auto lb = state.buildList(1); + auto * child = state.allocValue(); + lb[0] = child; + cur->mkList(lb); + cur = child; + } + cur->mkInt(0); +} + +// Build a balanced attrset tree: each node has `names.size()` attrs, recursing `depth` levels. +// Total node count = breadth^depth. +static void buildBalancedAttrs(EvalState & state, Value & v, const std::vector<std::string> & names, size_t depth) +{ + if (depth == 0) { + v.mkInt(0); + return; + } + auto b = state.buildBindings(names.size()); + for (auto & n : names) { + auto & child = b.alloc(state.symbols.create(n)); + buildBalancedAttrs(state, child, names, depth - 1); + } + v.mkAttrs(b.finish()); +} + +static void BM_PrintValueAsJSON_DeepAttrs(benchmark::State & bstate) +{ + auto statePtr = makeEvalState(); + auto & state = *statePtr; + const auto depth = static_cast<size_t>(bstate.range(0)); + Value root; + buildDeepAttrs(state, root, depth); + + NixStringContext context; + for (auto _ : bstate) { + auto j = printValueAsJSON(state, /*strict=*/true, root, noPos, context, /*copyToStore=*/false); + benchmark::DoNotOptimize(j); + } + bstate.SetItemsProcessed(bstate.iterations() * depth); +} + +BENCHMARK(BM_PrintValueAsJSON_DeepAttrs)->Arg(100)->Arg(1'000)->Arg(5'000); + +static void BM_PrintValueAsJSON_DeepLists(benchmark::State & bstate) +{ + auto statePtr = makeEvalState(); + auto & state = *statePtr; + const auto depth = static_cast<size_t>(bstate.range(0)); + Value root; + buildDeepLists(state, root, depth); + + NixStringContext context; + for (auto _ : bstate) { + auto j = printValueAsJSON(state, /*strict=*/true, root, noPos, context, /*copyToStore=*/false); + benchmark::DoNotOptimize(j); + } + bstate.SetItemsProcessed(bstate.iterations() * depth); +} + +BENCHMARK(BM_PrintValueAsJSON_DeepLists)->Arg(100)->Arg(1'000)->Arg(5'000); + +static void BM_PrintValueAsJSON_BalancedAttrs(benchmark::State & bstate) +{ + auto statePtr = makeEvalState(); + auto & state = *statePtr; + const auto breadth = static_cast<size_t>(bstate.range(0)); + const auto depth = static_cast<size_t>(bstate.range(1)); + + std::vector<std::string> names; + names.reserve(breadth); + for (size_t i = 0; i < breadth; ++i) + names.push_back(fmt("k%|1$04d|", i)); + + Value root; + buildBalancedAttrs(state, root, names, depth); + + size_t totalNodes = 1; + size_t p = 1; + for (size_t i = 0; i < depth; ++i) { + p *= breadth; + totalNodes += p; + } + + NixStringContext context; + for (auto _ : bstate) { + auto j = printValueAsJSON(state, /*strict=*/true, root, noPos, context, /*copyToStore=*/false); + benchmark::DoNotOptimize(j); + } + bstate.SetItemsProcessed(bstate.iterations() * totalNodes); +} + +BENCHMARK(BM_PrintValueAsJSON_BalancedAttrs) + ->ArgPair(4, 6) // ~5k nodes + ->ArgPair(8, 4) // ~4.7k nodes + ->ArgPair(16, 3); // ~4.4k nodes + +// Build a "package set" shape: a wide top-level attrset where every entry is +// a small derivation-like attrset (pname, version, description, license, +// broken, outPath). The outPath attr triggers the JSON serializer's shortcut +// to emit only the store path. Mimics the shape of nixpkgs/.../hackage-packages.nix +// after evaluation. +static void buildPackageSet(EvalState & state, Value & root, size_t numPackages) +{ + auto pnameSym = state.symbols.create("pname"); + auto versionSym = state.symbols.create("version"); + auto descSym = state.symbols.create("description"); + auto licenseSym = state.symbols.create("license"); + auto brokenSym = state.symbols.create("broken"); + auto outPathSym = state.s.outPath; + + auto topB = state.buildBindings(numPackages); + for (size_t i = 0; i < numPackages; ++i) { + auto pkgB = state.buildBindings(6); + + auto & op = pkgB.alloc(outPathSym); + op.mkString(fmt("/nix/store/000000000000000000000000000000000-pkg-%|1$06d|-1.2.3", i), state.mem); + auto & pn = pkgB.alloc(pnameSym); + pn.mkString(fmt("pkg-%|1$06d|", i), state.mem); + auto & ver = pkgB.alloc(versionSym); + ver.mkString("1.2.3", state.mem); + auto & d = pkgB.alloc(descSym); + d.mkString("A benchmark package", state.mem); + auto & l = pkgB.alloc(licenseSym); + l.mkString("BSD3", state.mem); + auto & b = pkgB.alloc(brokenSym); + b.mkBool(false); + + auto & pkg = topB.alloc(state.symbols.create(fmt("pkg-%|1$06d|", i))); + pkg.mkAttrs(pkgB.finish()); + } + root.mkAttrs(topB.finish()); +} + +static void BM_PrintValueAsJSON_PackageSet(benchmark::State & bstate) +{ + auto statePtr = makeEvalState(); + auto & state = *statePtr; + const auto numPackages = static_cast<size_t>(bstate.range(0)); + Value root; + buildPackageSet(state, root, numPackages); + + NixStringContext context; + for (auto _ : bstate) { + auto j = printValueAsJSON(state, /*strict=*/true, root, noPos, context, /*copyToStore=*/false); + benchmark::DoNotOptimize(j); + } + bstate.SetItemsProcessed(bstate.iterations() * numPackages); +} + +BENCHMARK(BM_PrintValueAsJSON_PackageSet)->Arg(1'000)->Arg(5'000)->Arg(10'000); + +} // namespace nix

Previously, `builtins.toJSON` on a self-referential value (e.g. `let a = { b = a; }; in a`) crashed the evaluator with a stack overflow because `printValueAsJSON` recursed unboundedly. Track visited Bindings (for attrsets) and Value pointers (for lists) in a scoped seen-set; on a collision, throw `InfiniteRecursionError`. To name both ends of the cycle in the error (e.g. `the value at .branch.twig is the same as the value at .branch`), the walk runs twice: recording a path for every node is overhead that's wasted on cycle-free values, so the first pass uses a minimal seen-set and throws an internal sentinel on collision, and only the re-run records paths to build the rich message.

Mirrors the cycle detection added for `printValueAsJSON`. The XML walker already handled cycles for derivations (via `drvsSeen` keyed on `drvPath`) but not for arbitrary attrsets or lists, which could loop indefinitely on self-referential values. Throws `InfiniteRecursionError` on collision, naming both ends of the cycle (e.g. `the value at .branch.twig is the same as the value at .branch`) via the same two-pass walk used by `printValueAsJSON`: the fast pass streams XML to the user's `ostream` and throws an internal sentinel on collision; the wrapper then re-runs against a discarded `XMLWriter` with path tracking to build the rich message.

dyegoaurelio requested a review from edolstra as a code owner May 20, 2026 04:35

github-actions Bot added documentation with-tests Issues related to testing. PRs with tests have some priority labels May 20, 2026

roberth reviewed May 20, 2026

View reviewed changes

dyegoaurelio force-pushed the infinite-recursion-json-decoding branch 2 times, most recently from 120c51d to f4b2fbc Compare May 21, 2026 17:24

dyegoaurelio requested review from roberth and xokdvium May 21, 2026 17:54

xokdvium reviewed May 21, 2026

View reviewed changes

dyegoaurelio force-pushed the infinite-recursion-json-decoding branch from f4b2fbc to e74168f Compare May 23, 2026 17:22

dyegoaurelio added 3 commits May 23, 2026 14:30

doc: add release note for JSON/XML cycle detection

058b490

dyegoaurelio force-pushed the infinite-recursion-json-decoding branch from e74168f to 058b490 Compare May 23, 2026 17:32

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

libexpr: detect cycles when converting values to JSON/XML#15888

libexpr: detect cycles when converting values to JSON/XML#15888
dyegoaurelio wants to merge 3 commits into
NixOS:masterfrom
dyegoaurelio:infinite-recursion-json-decoding

dyegoaurelio commented May 20, 2026

Uh oh!

roberth May 20, 2026

Uh oh!

Uh oh!

roberth May 20, 2026

Uh oh!

dyegoaurelio May 20, 2026

Uh oh!

xokdvium May 20, 2026

Uh oh!

roberth May 20, 2026

Uh oh!

xokdvium May 21, 2026

Uh oh!

dyegoaurelio May 23, 2026 •

edited

Loading

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

	state.error<InfiniteRecursionError>("infinite recursion encountered while converting value to JSON")
	state.error<InfiniteRecursionError>("infinite recursion encountered while converting Nix value to JSON, because only cycle-free data can be represented")

Uh oh!

Conversation

dyegoaurelio commented May 20, 2026

Motivation

Context

Uh oh!

roberth May 20, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

roberth May 20, 2026

Choose a reason for hiding this comment

Uh oh!

dyegoaurelio May 20, 2026

Choose a reason for hiding this comment

Uh oh!

xokdvium May 20, 2026

Choose a reason for hiding this comment

Uh oh!

roberth May 20, 2026

Choose a reason for hiding this comment

Uh oh!

xokdvium May 21, 2026

Choose a reason for hiding this comment

Uh oh!

dyegoaurelio May 23, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

dyegoaurelio May 23, 2026 •

edited

Loading