zig/lib/std/Build/Fuzz.zig
Kendall Condon e66b269333 greatly improve capabilities of the fuzzer
This PR significantly improves the capabilities of the fuzzer.

The changes made to the fuzzer to accomplish this feat mostly include
tracking memory reads from .rodata to determine fresh inputs, new
mutations (especially the ones that insert const values from .rodata
reads and __sanitizer_conv_const_cmp), and minimizing found inputs.
Additionally, the runs per second has greatly been increased due to
generating smaller inputs and avoiding clearing the 8-bit pc counters.

An additional feature added is that the length of the input file is now
stored and the old input file is rerun upon start.

Other changes made to the fuzzer include more logical initialization,
using one shared file `in` for inputs, creating corpus files with
proper sizes, and using hexadecimal-numbered corpus files for
simplicity.

Furthermore, I added several new fuzz tests to gauge the fuzzer's
efficiency. I also tried to add a test for zstandard decompression,
which it crashed within 60,000 runs (less than a second.)

Bug fixes include:
* Fixed a race conditions when multiple fuzzer processes needed to use
the same coverage file.
* Web interface stats now update even when unique runs is not changing.
* Fixed tokenizer.testPropertiesUpheld to allow stray carriage returns
since they are valid whitespace.
2025-09-18 18:56:10 -04:00

449 lines
17 KiB
Zig

const std = @import("../std.zig");
const Build = std.Build;
const Cache = Build.Cache;
const Step = std.Build.Step;
const assert = std.debug.assert;
const fatal = std.process.fatal;
const Allocator = std.mem.Allocator;
const log = std.log;
const Coverage = std.debug.Coverage;
const abi = Build.abi.fuzz;
const Fuzz = @This();
const build_runner = @import("root");
ws: *Build.WebServer,
/// Allocated into `ws.gpa`.
run_steps: []const *Step.Run,
wait_group: std.Thread.WaitGroup,
prog_node: std.Progress.Node,
/// Protects `coverage_files`.
coverage_mutex: std.Thread.Mutex,
coverage_files: std.AutoArrayHashMapUnmanaged(u64, CoverageMap),
queue_mutex: std.Thread.Mutex,
queue_cond: std.Thread.Condition,
msg_queue: std.ArrayListUnmanaged(Msg),
const Msg = union(enum) {
coverage: struct {
id: u64,
run: *Step.Run,
},
entry_point: struct {
coverage_id: u64,
addr: u64,
},
};
const CoverageMap = struct {
mapped_memory: []align(std.heap.page_size_min) const u8,
coverage: Coverage,
source_locations: []Coverage.SourceLocation,
/// Elements are indexes into `source_locations` pointing to the unit tests that are being fuzz tested.
entry_points: std.ArrayListUnmanaged(u32),
start_timestamp: i64,
fn deinit(cm: *CoverageMap, gpa: Allocator) void {
std.posix.munmap(cm.mapped_memory);
cm.coverage.deinit(gpa);
cm.* = undefined;
}
};
pub fn init(ws: *Build.WebServer) Allocator.Error!Fuzz {
const gpa = ws.gpa;
const run_steps: []const *Step.Run = steps: {
var steps: std.ArrayListUnmanaged(*Step.Run) = .empty;
defer steps.deinit(gpa);
const rebuild_node = ws.root_prog_node.start("Rebuilding Unit Tests", 0);
defer rebuild_node.end();
var rebuild_wg: std.Thread.WaitGroup = .{};
defer rebuild_wg.wait();
for (ws.all_steps) |step| {
const run = step.cast(Step.Run) orelse continue;
if (run.producer == null) continue;
if (run.fuzz_tests.items.len == 0) continue;
try steps.append(gpa, run);
ws.thread_pool.spawnWg(&rebuild_wg, rebuildTestsWorkerRun, .{ run, gpa, ws.ttyconf, rebuild_node });
}
if (steps.items.len == 0) fatal("no fuzz tests found", .{});
rebuild_node.setEstimatedTotalItems(steps.items.len);
break :steps try gpa.dupe(*Step.Run, steps.items);
};
errdefer gpa.free(run_steps);
for (run_steps) |run| {
assert(run.fuzz_tests.items.len > 0);
if (run.rebuilt_executable == null)
fatal("one or more unit tests failed to be rebuilt in fuzz mode", .{});
}
return .{
.ws = ws,
.run_steps = run_steps,
.wait_group = .{},
.prog_node = .none,
.coverage_files = .empty,
.coverage_mutex = .{},
.queue_mutex = .{},
.queue_cond = .{},
.msg_queue = .empty,
};
}
pub fn start(fuzz: *Fuzz) void {
const ws = fuzz.ws;
fuzz.prog_node = ws.root_prog_node.start("Fuzzing", fuzz.run_steps.len);
// For polling messages and sending updates to subscribers.
fuzz.wait_group.start();
_ = std.Thread.spawn(.{}, coverageRun, .{fuzz}) catch |err| {
fuzz.wait_group.finish();
fatal("unable to spawn coverage thread: {s}", .{@errorName(err)});
};
for (fuzz.run_steps) |run| {
for (run.fuzz_tests.items) |unit_test_index| {
assert(run.rebuilt_executable != null);
ws.thread_pool.spawnWg(&fuzz.wait_group, fuzzWorkerRun, .{
fuzz, run, unit_test_index,
});
}
}
}
pub fn deinit(fuzz: *Fuzz) void {
if (true) @panic("TODO: terminate the fuzzer processes");
fuzz.wait_group.wait();
fuzz.prog_node.end();
const gpa = fuzz.ws.gpa;
gpa.free(fuzz.run_steps);
}
fn rebuildTestsWorkerRun(run: *Step.Run, gpa: Allocator, ttyconf: std.Io.tty.Config, parent_prog_node: std.Progress.Node) void {
rebuildTestsWorkerRunFallible(run, gpa, ttyconf, parent_prog_node) catch |err| {
const compile = run.producer.?;
log.err("step '{s}': failed to rebuild in fuzz mode: {s}", .{
compile.step.name, @errorName(err),
});
};
}
fn rebuildTestsWorkerRunFallible(run: *Step.Run, gpa: Allocator, ttyconf: std.Io.tty.Config, parent_prog_node: std.Progress.Node) !void {
const compile = run.producer.?;
const prog_node = parent_prog_node.start(compile.step.name, 0);
defer prog_node.end();
const result = compile.rebuildInFuzzMode(gpa, prog_node);
const show_compile_errors = compile.step.result_error_bundle.errorMessageCount() > 0;
const show_error_msgs = compile.step.result_error_msgs.items.len > 0;
const show_stderr = compile.step.result_stderr.len > 0;
if (show_error_msgs or show_compile_errors or show_stderr) {
var buf: [256]u8 = undefined;
const w = std.debug.lockStderrWriter(&buf);
defer std.debug.unlockStderrWriter();
build_runner.printErrorMessages(gpa, &compile.step, .{ .ttyconf = ttyconf }, w, false) catch {};
}
const rebuilt_bin_path = result catch |err| switch (err) {
error.MakeFailed => return,
else => |other| return other,
};
run.rebuilt_executable = try rebuilt_bin_path.join(gpa, compile.out_filename);
}
fn fuzzWorkerRun(
fuzz: *Fuzz,
run: *Step.Run,
unit_test_index: u32,
) void {
const gpa = run.step.owner.allocator;
const test_name = run.cached_test_metadata.?.testName(unit_test_index);
const prog_node = fuzz.prog_node.start(test_name, 0);
defer prog_node.end();
run.rerunInFuzzMode(fuzz, unit_test_index, prog_node) catch |err| switch (err) {
error.MakeFailed => {
var buf: [256]u8 = undefined;
const w = std.debug.lockStderrWriter(&buf);
defer std.debug.unlockStderrWriter();
build_runner.printErrorMessages(gpa, &run.step, .{ .ttyconf = fuzz.ws.ttyconf }, w, false) catch {};
return;
},
else => {
log.err("step '{s}': failed to rerun '{s}' in fuzz mode: {s}", .{
run.step.name, test_name, @errorName(err),
});
return;
},
};
}
pub fn serveSourcesTar(fuzz: *Fuzz, req: *std.http.Server.Request) !void {
const gpa = fuzz.ws.gpa;
var arena_state: std.heap.ArenaAllocator = .init(gpa);
defer arena_state.deinit();
const arena = arena_state.allocator();
const DedupTable = std.ArrayHashMapUnmanaged(Build.Cache.Path, void, Build.Cache.Path.TableAdapter, false);
var dedup_table: DedupTable = .empty;
defer dedup_table.deinit(gpa);
for (fuzz.run_steps) |run_step| {
const compile_inputs = run_step.producer.?.step.inputs.table;
for (compile_inputs.keys(), compile_inputs.values()) |dir_path, *file_list| {
try dedup_table.ensureUnusedCapacity(gpa, file_list.items.len);
for (file_list.items) |sub_path| {
if (!std.mem.endsWith(u8, sub_path, ".zig")) continue;
const joined_path = try dir_path.join(arena, sub_path);
dedup_table.putAssumeCapacity(joined_path, {});
}
}
}
const deduped_paths = dedup_table.keys();
const SortContext = struct {
pub fn lessThan(this: @This(), lhs: Build.Cache.Path, rhs: Build.Cache.Path) bool {
_ = this;
return switch (std.mem.order(u8, lhs.root_dir.path orelse ".", rhs.root_dir.path orelse ".")) {
.lt => true,
.gt => false,
.eq => std.mem.lessThan(u8, lhs.sub_path, rhs.sub_path),
};
}
};
std.mem.sortUnstable(Build.Cache.Path, deduped_paths, SortContext{}, SortContext.lessThan);
return fuzz.ws.serveTarFile(req, deduped_paths);
}
pub const Previous = struct {
unique_runs: usize,
entry_points: usize,
pub const init: Previous = .{ .unique_runs = 0, .entry_points = 0 };
};
pub fn sendUpdate(
fuzz: *Fuzz,
socket: *std.http.Server.WebSocket,
prev: *Previous,
) !void {
fuzz.coverage_mutex.lock();
defer fuzz.coverage_mutex.unlock();
const coverage_maps = fuzz.coverage_files.values();
if (coverage_maps.len == 0) return;
// TODO: handle multiple fuzz steps in the WebSocket packets
const coverage_map = &coverage_maps[0];
const cov_header: *const abi.SeenPcsHeader = @ptrCast(coverage_map.mapped_memory[0..@sizeOf(abi.SeenPcsHeader)]);
// TODO: this isn't sound! We need to do volatile reads of these bits rather than handing the
// buffer off to the kernel, because we might race with the fuzzer process[es]. This brings the
// whole mmap strategy into question. Incidentally, I wonder if post-writergate we could pass
// this data straight to the socket with sendfile...
const seen_pcs = cov_header.seenBits();
const n_runs = @atomicLoad(usize, &cov_header.n_runs, .monotonic);
const unique_runs = @atomicLoad(usize, &cov_header.unique_runs, .monotonic);
{
if (unique_runs != 0 and prev.unique_runs == 0) {
// We need to send initial context.
const header: abi.SourceIndexHeader = .{
.directories_len = @intCast(coverage_map.coverage.directories.entries.len),
.files_len = @intCast(coverage_map.coverage.files.entries.len),
.source_locations_len = @intCast(coverage_map.source_locations.len),
.string_bytes_len = @intCast(coverage_map.coverage.string_bytes.items.len),
.start_timestamp = coverage_map.start_timestamp,
};
var iovecs: [5][]const u8 = .{
@ptrCast(&header),
@ptrCast(coverage_map.coverage.directories.keys()),
@ptrCast(coverage_map.coverage.files.keys()),
@ptrCast(coverage_map.source_locations),
coverage_map.coverage.string_bytes.items,
};
try socket.writeMessageVec(&iovecs, .binary);
}
const header: abi.CoverageUpdateHeader = .{
.n_runs = n_runs,
.unique_runs = unique_runs,
};
var iovecs: [2][]const u8 = .{
@ptrCast(&header),
@ptrCast(seen_pcs),
};
try socket.writeMessageVec(&iovecs, .binary);
prev.unique_runs = unique_runs;
}
if (prev.entry_points != coverage_map.entry_points.items.len) {
const header: abi.EntryPointHeader = .init(@intCast(coverage_map.entry_points.items.len));
var iovecs: [2][]const u8 = .{
@ptrCast(&header),
@ptrCast(coverage_map.entry_points.items),
};
try socket.writeMessageVec(&iovecs, .binary);
prev.entry_points = coverage_map.entry_points.items.len;
}
}
fn coverageRun(fuzz: *Fuzz) void {
defer fuzz.wait_group.finish();
fuzz.queue_mutex.lock();
defer fuzz.queue_mutex.unlock();
while (true) {
fuzz.queue_cond.wait(&fuzz.queue_mutex);
for (fuzz.msg_queue.items) |msg| switch (msg) {
.coverage => |coverage| prepareTables(fuzz, coverage.run, coverage.id) catch |err| switch (err) {
error.AlreadyReported => continue,
else => |e| log.err("failed to prepare code coverage tables: {s}", .{@errorName(e)}),
},
.entry_point => |entry_point| addEntryPoint(fuzz, entry_point.coverage_id, entry_point.addr) catch |err| switch (err) {
error.AlreadyReported => continue,
else => |e| log.err("failed to prepare code coverage tables: {s}", .{@errorName(e)}),
},
};
fuzz.msg_queue.clearRetainingCapacity();
}
}
fn prepareTables(fuzz: *Fuzz, run_step: *Step.Run, coverage_id: u64) error{ OutOfMemory, AlreadyReported }!void {
const ws = fuzz.ws;
const gpa = ws.gpa;
fuzz.coverage_mutex.lock();
defer fuzz.coverage_mutex.unlock();
const gop = try fuzz.coverage_files.getOrPut(gpa, coverage_id);
if (gop.found_existing) {
// We are fuzzing the same executable with multiple threads.
// Perhaps the same unit test; perhaps a different one. In any
// case, since the coverage file is the same, we only have to
// notice changes to that one file in order to learn coverage for
// this particular executable.
return;
}
errdefer _ = fuzz.coverage_files.pop();
gop.value_ptr.* = .{
.coverage = std.debug.Coverage.init,
.mapped_memory = undefined, // populated below
.source_locations = undefined, // populated below
.entry_points = .{},
.start_timestamp = ws.now(),
};
errdefer gop.value_ptr.coverage.deinit(gpa);
const rebuilt_exe_path = run_step.rebuilt_executable.?;
var debug_info = std.debug.Info.load(gpa, rebuilt_exe_path, &gop.value_ptr.coverage) catch |err| {
log.err("step '{s}': failed to load debug information for '{f}': {s}", .{
run_step.step.name, rebuilt_exe_path, @errorName(err),
});
return error.AlreadyReported;
};
defer debug_info.deinit(gpa);
const coverage_file_path: Build.Cache.Path = .{
.root_dir = run_step.step.owner.cache_root,
.sub_path = "v/" ++ std.fmt.hex(coverage_id),
};
var coverage_file = coverage_file_path.root_dir.handle.openFile(coverage_file_path.sub_path, .{}) catch |err| {
log.err("step '{s}': failed to load coverage file '{f}': {s}", .{
run_step.step.name, coverage_file_path, @errorName(err),
});
return error.AlreadyReported;
};
defer coverage_file.close();
const file_size = coverage_file.getEndPos() catch |err| {
log.err("unable to check len of coverage file '{f}': {s}", .{ coverage_file_path, @errorName(err) });
return error.AlreadyReported;
};
const mapped_memory = std.posix.mmap(
null,
file_size,
std.posix.PROT.READ,
.{ .TYPE = .SHARED },
coverage_file.handle,
0,
) catch |err| {
log.err("failed to map coverage file '{f}': {s}", .{ coverage_file_path, @errorName(err) });
return error.AlreadyReported;
};
gop.value_ptr.mapped_memory = mapped_memory;
const header: *const abi.SeenPcsHeader = @ptrCast(mapped_memory[0..@sizeOf(abi.SeenPcsHeader)]);
const pcs = header.pcAddrs();
const source_locations = try gpa.alloc(Coverage.SourceLocation, pcs.len);
errdefer gpa.free(source_locations);
// Unfortunately the PCs array that LLVM gives us from the 8-bit PC
// counters feature is not sorted.
var sorted_pcs: std.MultiArrayList(struct { pc: u64, index: u32, sl: Coverage.SourceLocation }) = .{};
defer sorted_pcs.deinit(gpa);
try sorted_pcs.resize(gpa, pcs.len);
@memcpy(sorted_pcs.items(.pc), pcs);
for (sorted_pcs.items(.index), 0..) |*v, i| v.* = @intCast(i);
sorted_pcs.sortUnstable(struct {
addrs: []const u64,
pub fn lessThan(ctx: @This(), a_index: usize, b_index: usize) bool {
return ctx.addrs[a_index] < ctx.addrs[b_index];
}
}{ .addrs = sorted_pcs.items(.pc) });
debug_info.resolveAddresses(gpa, sorted_pcs.items(.pc), sorted_pcs.items(.sl)) catch |err| {
log.err("failed to resolve addresses to source locations: {s}", .{@errorName(err)});
return error.AlreadyReported;
};
for (sorted_pcs.items(.index), sorted_pcs.items(.sl)) |i, sl| source_locations[i] = sl;
gop.value_ptr.source_locations = source_locations;
ws.notifyUpdate();
}
fn addEntryPoint(fuzz: *Fuzz, coverage_id: u64, addr: u64) error{ AlreadyReported, OutOfMemory }!void {
fuzz.coverage_mutex.lock();
defer fuzz.coverage_mutex.unlock();
const coverage_map = fuzz.coverage_files.getPtr(coverage_id).?;
const header: *const abi.SeenPcsHeader = @ptrCast(coverage_map.mapped_memory[0..@sizeOf(abi.SeenPcsHeader)]);
const pcs = header.pcAddrs();
// Since this pcs list is unsorted, we must linear scan for the best index.
const index = i: {
var best: usize = 0;
for (pcs[1..], 1..) |elem_addr, i| {
if (elem_addr == addr) break :i i;
if (elem_addr > addr) continue;
if (elem_addr > pcs[best]) best = i;
}
break :i best;
};
if (index >= pcs.len) {
log.err("unable to find unit test entry address 0x{x} in source locations (range: 0x{x} to 0x{x})", .{
addr, pcs[0], pcs[pcs.len - 1],
});
return error.AlreadyReported;
}
if (false) {
const sl = coverage_map.source_locations[index];
const file_name = coverage_map.coverage.stringAt(coverage_map.coverage.fileAt(sl.file).basename);
log.debug("server found entry point for 0x{x} at {s}:{d}:{d} - index {d} between {x} and {x}", .{
addr, file_name, sl.line, sl.column, index, pcs[index - 1], pcs[index + 1],
});
}
try coverage_map.entry_points.append(fuzz.ws.gpa, @intCast(index));
}