summaryrefslogtreecommitdiff
path: root/src/cli/probe.zig
diff options
context:
space:
mode:
authorGabriel Schneider <[email protected]>2026-11-10 11:49:22 -0300
committerGabriel Schneider <[email protected]>2026-11-10 12:19:06 -0300
commit297943b01e18bb3d1dd41a8d3e220811de25228b (patch)
tree4938e643abada811d3ab28b14c0e31dc94ebda62 /src/cli/probe.zig
parentdb923fa823354270217362db06356f2b1a41315e (diff)
downloadcomputeranim-297943b01e18bb3d1dd41a8d3e220811de25228b.tar.gz
computeranim-297943b01e18bb3d1dd41a8d3e220811de25228b.zip
ztrace probe: BPF on syscalls, scheduling, kprobes, uprobes and USDT
Programs built as instruction arrays for one process (no compiler, no libbpf), writing Events into a BPF ring buffer; a drop counter in a map; task_struct offsets from the kernel's BTF; a USDT probe's arguments read in the kernel at the probe. Without CAP_BPF and CAP_PERFMON it says so; --dry-run builds and lists everything. Tested offline by running the programs in a small BPF machine, and for real by test-kernel-vm, which boots this kernel in QEMU where we are root. ztrace perf: moments are no longer cut per event unless --cut; --max-size, and a full recording stops the target and says so. perf.Sorted merges two backends' events by time. Co-Authored-By: Claude Fable 5.2 <[email protected]>
Diffstat (limited to 'src/cli/probe.zig')
-rw-r--r--src/cli/probe.zig215
1 files changed, 215 insertions, 0 deletions
diff --git a/src/cli/probe.zig b/src/cli/probe.zig
new file mode 100644
index 0000000..f0e2d12
--- /dev/null
+++ b/src/cli/probe.zig
@@ -0,0 +1,215 @@
+//! ztrace probe: BPF programs in the kernel watching a program: its
+//! syscalls, its scheduling, any kernel or user function, static probes.
+//! For root; --dry-run shows what it would do without being it.
+const std = @import("std");
+const linux = std.os.linux;
+const zt = @import("ztrace");
+const util = @import("util.zig");
+const perf = @import("perf.zig");
+const fail = util.fail;
+const bpf = zt.bpf;
+
+pub const usage =
+ \\usage: ztrace probe [-o FILE] [OPTIONS] -- PROGRAM [ARGS]
+ \\
+ \\Runs PROGRAM at full speed with BPF programs attached for it in the
+ \\kernel, into a recording (default out.ztrace). Needs CAP_BPF and
+ \\CAP_PERFMON, which root has.
+ \\
+ \\ --syscalls every syscall, with the registers at its entry
+ \\ --sched every switch of a CPU to or from one of its threads
+ \\ --kprobe FUNCTION
+ \\ every call of a kernel function it makes: five arguments
+ \\ --uprobe [FILE:]SYMBOL
+ \\ every call of a function of its own, or of a library
+ \\ given by its file: five arguments, the registers
+ \\ --usdt [FILE:]PROVIDER:NAME
+ \\ a static probe (<sys/sdt.h>), with its own arguments
+ \\ --dry-run build the programs, find the places, print them; attach nothing
+ \\ --max-size BYTES stop when the recording is this large (default 1 GiB)
+ \\ --ring BYTES the ring the programs write to (a power of two, default 4 MiB)
+ \\
+;
+
+const Probe = struct { kind: enum { kprobe, uprobe, usdt }, what: [:0]const u8 };
+
+pub fn run(init: std.process.Init, args: []const [:0]const u8) !void {
+ const arena = init.arena.allocator();
+ var out: []const u8 = "out.ztrace";
+ var syscalls = false;
+ var sched = false;
+ var dry = false;
+ var max_bytes: ?u64 = null;
+ var ring_bytes: u32 = 1 << 22;
+ var probes: std.ArrayList(Probe) = .empty;
+ var i: usize = 0;
+ while (i < args.len) : (i += 1) {
+ const arg = args[i];
+ const value: ?[:0]const u8 = if (i + 1 < args.len) args[i + 1] else null;
+ if (std.mem.eql(u8, arg, "--")) {
+ i += 1;
+ break;
+ } else if (std.mem.eql(u8, arg, "--syscalls")) {
+ syscalls = true;
+ } else if (std.mem.eql(u8, arg, "--sched")) {
+ sched = true;
+ } else if (std.mem.eql(u8, arg, "--dry-run")) {
+ dry = true;
+ } else if (value == null) {
+ fail(usage, .{});
+ } else if (std.mem.eql(u8, arg, "-o")) {
+ out = value.?;
+ i += 1;
+ } else if (std.mem.eql(u8, arg, "--max-size")) {
+ max_bytes = std.fmt.parseInt(u64, value.?, 0) catch fail("--max-size needs bytes\n", .{});
+ i += 1;
+ } else if (std.mem.eql(u8, arg, "--ring")) {
+ ring_bytes = std.fmt.parseInt(u32, value.?, 0) catch 0;
+ if (!std.math.isPowerOfTwo(ring_bytes) or ring_bytes < 4096) fail("--ring needs a power of two, 4096 or more\n", .{});
+ i += 1;
+ } else if (std.meta.stringToEnum(@FieldType(Probe, "kind"), arg[@min(arg.len, 2)..])) |kind| {
+ try probes.append(arena, .{ .kind = kind, .what = value.? });
+ i += 1;
+ } else fail(usage, .{});
+ }
+ if (i >= args.len) fail(usage, .{});
+ const path = std.mem.span(try util.resolve(init, args[i]));
+
+ // Everything that can be wrong without the kernel's say is found first.
+ var places: std.ArrayList(Place) = .empty;
+ for (probes.items) |p| try places.append(arena, try place(arena, p, path));
+ var offsets: [2]u32 = .{ 0, 0 };
+ var btf: []const u8 = "";
+ if (sched) {
+ btf = std.Io.Dir.cwd().readFileAlloc(init.io, "/sys/kernel/btf/vmlinux", arena, .unlimited) catch fail("--sched needs the kernel's type information, and /sys/kernel/btf/vmlinux cannot be read\n", .{});
+ offsets = .{
+ bpf.fieldOffset(btf, "task_struct", "pid") orelse fail("no task_struct.pid in the kernel's BTF\n", .{}),
+ bpf.fieldOffset(btf, "task_struct", "tgid") orelse fail("no task_struct.tgid in the kernel's BTF\n", .{}),
+ };
+ }
+ var why: [512]u8 = undefined;
+ const missing = bpf.missing(&why);
+
+ if (dry) {
+ // The process id and the maps are not there yet: 0 stands for them.
+ const maps: bpf.Maps = .{ .ring = 0, .drops = 0 };
+ std.debug.print("for the process started from {s}:\n", .{path});
+ if (syscalls) {
+ list("raw tracepoint sys_enter: a syscall_enter and the registers", bpf.sysEnter(maps, 0));
+ list("raw tracepoint sys_exit: a syscall_exit", bpf.sysExit(maps, 0));
+ }
+ if (sched) {
+ std.debug.print("task_struct.pid is at {d}, .tgid at {d} (BTF)\n", .{ offsets[0], offsets[1] });
+ list("raw tracepoint sched_switch: a sched", bpf.schedSwitch(maps, 0, offsets[0], offsets[1]));
+ }
+ for (places.items) |p| {
+ std.debug.print("{t} {s}: {s} +{x}", .{ p.kind, p.name, p.target, p.offset });
+ if (p.usdt) |u| std.debug.print(", semaphore at +{x}, arguments {s}", .{ if (u.semaphore == 0) 0 else p.elf.?.offset(u.semaphore) orelse 0, u.args });
+ std.debug.print("\n", .{});
+ const id = zt.perf.stringId(p.name);
+ list("its program: a probe and the registers", if (p.usdt) |u| bpf.usdtProgram(maps, 0, id, u.args) catch fail("{s}: an argument of this probe ({s}) is in a form not understood yet\n", .{ p.name, u.args }) else bpf.probe(maps, 0, id, null));
+ }
+ std.debug.print("{s}\n", .{if (missing) |m| m else "nothing is missing to attach them"});
+ return;
+ }
+ if (missing) |m| fail("ztrace probe: {s}\n", .{m});
+
+ const argv = try util.targetArgv(arena, args[i..]);
+ var proc = try zt.ptrace.Process.spawn(path, argv, init.minimal.environ.block.slice.ptr);
+ errdefer proc.kill();
+ var rec = try zt.Recorder.create(init.gpa, init.io, out);
+ if (max_bytes) |bytes| rec.max_bytes = bytes;
+ // Two sources, the programs' ring and perf's: one sequence in the recording.
+ var sorted: zt.perf.Sorted = .{ .gpa = init.gpa, .to = rec.sink() };
+ defer sorted.deinit();
+ var tracer = bpf.Tracer.open(init.gpa, proc.pid, sorted.sink(), ring_bytes) catch |err| refused(err, null);
+ defer tracer.close();
+ if (syscalls) tracer.syscalls() catch |err| refused(err, &tracer);
+ if (sched) tracer.sched(btf) catch |err| refused(err, &tracer);
+ for (places.items) |p| {
+ const id = zt.perf.stringId(p.name);
+ try sorted.sink().emit(.{ .ts = zt.event.now(), .kind = .string, .a = .{ id, 0, 0, 0, 0, 0 } }, p.name);
+ (if (p.usdt) |u| tracer.usdtProbe(p.elf.?, p.target, u, id) else switch (p.kind) {
+ .kprobe => tracer.kprobe(p.target, id),
+ else => tracer.uprobe(p.target, p.offset, id),
+ }) catch |err| switch (err) {
+ error.NoSuchProbe => fail("{s}: the kernel has no such place to probe\n", .{p.name}),
+ error.BadUsdtArg => fail("{s}: an argument of this probe ({s}) is in a form not understood yet\n", .{ p.name, p.usdt.?.args }),
+ else => |e| refused(e, &tracer),
+ };
+ }
+ // The maps and the threads, which no program here reports.
+ var session = try zt.perf.Session.open(init.gpa, proc.pid, .{}, sorted.sink());
+ defer session.close();
+
+ zt.perf.release(proc);
+ const start = zt.event.now();
+ const exit_code: ?u64 = pump(&tracer, &session, &sorted) catch |err| switch (err) {
+ error.RecordingFull => null,
+ else => |e| return e,
+ };
+ if (exit_code == null) proc.kill();
+ try perf.finish(init, &rec, out, exit_code, tracer.lost() + session.lost, zt.event.now() - start, &.{}, &.{});
+}
+
+fn pump(tracer: *bpf.Tracer, session: *zt.perf.Session, sorted: *zt.perf.Sorted) !u64 {
+ // What both sources have reported past: the start of the round before.
+ var past: u64 = 0;
+ while (true) {
+ const now = zt.event.now();
+ const alive = try session.poll(0);
+ const records = try tracer.poll(if (alive) 20 else 0);
+ try sorted.flush(past);
+ past = now;
+ if (records == 0 and !alive) break;
+ }
+ const code = try perf.exited(tracer.pid, sorted.sink());
+ try sorted.flush(std.math.maxInt(u64));
+ return code;
+}
+
+fn refused(err: anyerror, tracer: ?*bpf.Tracer) noreturn {
+ const log = if (tracer) |t| std.mem.sliceTo(t.log, 0) else "";
+ switch (err) {
+ error.PermissionDenied => fail("ztrace probe: bpf() refused though the capabilities are there: inside a user namespace or a container they do not count, the kernel wants them in the initial namespace\n", .{}),
+ error.UnsafeProgram, error.InvalidProgram => fail("ztrace probe: the kernel's verifier refused a program ({t}):\n{s}\n", .{ err, log }),
+ else => fail("ztrace probe: {t}\n{s}", .{ err, log }),
+ }
+}
+
+/// Where a probe goes: a kernel function, or an offset into a file.
+const Place = struct {
+ kind: @FieldType(Probe, "kind"),
+ name: []const u8,
+ target: [:0]const u8,
+ offset: u64 = 0,
+ elf: ?zt.usdt.Elf = null,
+ usdt: ?zt.usdt.Probe = null,
+};
+
+fn place(arena: std.mem.Allocator, p: Probe, program: [:0]const u8) !Place {
+ if (p.kind == .kprobe) return .{ .kind = .kprobe, .name = p.what, .target = p.what };
+ // FILE: in front, or the program itself.
+ const colons = std.mem.count(u8, p.what, ":");
+ const named = colons == @as(usize, if (p.kind == .usdt) 2 else 1);
+ const split = if (named) std.mem.indexOfScalar(u8, p.what, ':').? else 0;
+ const file = if (named) try arena.dupeSentinel(u8, p.what[0..split], 0) else program;
+ const name = p.what[if (named) split + 1 else 0..];
+ const elf = zt.usdt.Elf.open(file) catch fail("{s}: not an ELF file that can be read\n", .{file});
+ if (p.kind == .uprobe) {
+ const sym = elf.symbol(name) orelse fail("{s}: no symbol {s}\n", .{ file, name });
+ return .{ .kind = .uprobe, .name = name, .target = file, .offset = elf.offset(sym.value) orelse fail("{s}: {s} is not in the file\n", .{ file, name }), .elf = elf };
+ }
+ for (try zt.usdt.probes(arena, elf, 0)) |found| {
+ const colon = std.mem.indexOfScalar(u8, name, ':') orelse fail("--usdt wants [FILE:]PROVIDER:NAME\n", .{});
+ if (!std.mem.eql(u8, found.provider, name[0..colon]) or !std.mem.eql(u8, found.name, name[colon + 1 ..])) continue;
+ return .{ .kind = .usdt, .name = name, .target = file, .offset = elf.offset(found.pc) orelse continue, .elf = elf, .usdt = found };
+ }
+ fail("{s}: no static probe {s} (readelf -n lists them)\n", .{ file, name });
+}
+
+/// A program as the verifier numbers it: opcode, registers, offset, immediate.
+fn list(what: []const u8, program: bpf.Program) void {
+ std.debug.print("{s}, {d} instructions\n", .{ what, program.len });
+ for (program.code(), 0..) |insn, at| std.debug.print(" {d:>3}: {x:0>2} r{d} r{d} {d:>5} {d}\n", .{ at, insn.code, insn.dst, insn.src, insn.off, insn.imm });
+}