Skip to content
zit Artifact [58162493d5]
Login

Artifact [58162493d5]

Check-ins Using Download Hex

Artifact 58162493d5b8a5e9c546b4bdd42402368c802a52740a255c13ae0df518ea2ae2:

//! `zit verify` — this repository's own fsck.
//!
//! NOT AN RFC-MANDATED VERB. repo-state §7.3 blesses a full rebuild from the
//! artifacts as a recovery fallback, but it specifies nothing about a
//! verification command; the shape of this one is zit's own design decision and
//! no spec citation is claimed for it.
//!
//! It exists because the open-time scan is no longer where damage is noticed.
//! The scan used to stream every payload byte through the crc on every open,
//! which made corruption detection a side effect of paying Θ(repository bytes)
//! to open. That cost is being deleted; this verb is the on-demand replacement,
//! the same split git draws between ordinary operation and `git fsck`.
//!
//! TWO PASSES, DELIBERATELY LAYERED, because they fail independently:
//!
//!  1. The whole-log frame pass (`logstore.FrameVerifier`) reads every frame's
//!     bytes and checks the whole-frame checksum. It is built directly on the
//!     store FILE, so it works on a store no `open` would accept — which is the
//!     entire reason an fsck exists. It sees config records and shun rows,
//!     which the second pass structurally cannot.
//!  2. The per-artifact pass reconstructs every stored artifact through its
//!     delta chain and re-hashes it against its own name. It catches a frame
//!     whose checksum is perfectly valid over the wrong bytes — a fault the
//!     first pass is blind to by construction.
//!
//! Scope stops at content artifacts: manifests are not re-parsed. A manifest
//! whose bytes hash to its name is byte-identical to what `zit commit` wrote and
//! what sync validated on ingest, so re-parsing it would catch a zit bug rather
//! than storage rot. `verify_help` says so.

const std = @import("std");
const cli = @import("../cli.zig");
const logstore = @import("zit").logstore;
const storage = @import("zit").storage;
const artifact = @import("zit").artifact;

// Shared CLI helpers that live in main.zig, aliased so this file reads like its
// siblings in src/cmd/.
const main = @import("../main.zig");
const repo_filename = main.repo_filename;
const reasonOf = main.reasonOf;
const stateOom = main.stateOom;
const not_a_repo_msg = main.not_a_repo_msg;
const repoReadFail = main.repoReadFail;
const digestToWireName = main.digestToWireName;
const verify_help = main.verify_help;

// The two named-length couplings this file's output depends on, tied at
// compile time rather than by coincidence: a frame's hash is at most
// `max_hash_len` bytes and `digestToWireName` asserts its input fits
// `digest_len`, so `name_buf` below is only the right size while the two agree.
comptime {
    std.debug.assert(main.digest_len == logstore.max_hash_len);
}

/// Running tallies across both passes, so the verdict can state the byte
/// accounting as a separate fact from the fault count — and can state WHICH
/// pass found what. The split is not cosmetic: one physical bitflip in a
/// content frame is found by both passes (a bad checksum, then an artifact that
/// will not reconstruct), and a single aggregate "2 faults" would read as two
/// damaged frames.
const Tally = struct {
    frame_faults: u64 = 0,
    artifact_faults: u64 = 0,
    /// Artifacts the per-artifact pass reconstructed, or null when that pass
    /// did not run. Null rather than zero on purpose: "0 artifacts" would read
    /// as "this store holds none", which is a different claim from "nobody
    /// looked".
    artifacts: ?u64 = null,
    /// The per-artifact pass could not open the store at all. Its own kind of
    /// fault, and one that has to raise the exit code: a repository every other
    /// verb would refuse to open is not a clean repository, and `verify_help`'s
    /// exit table promises 1 when "the store could not be read".
    unreadable: bool = false,

    fn faults(self: Tally) u64 {
        return self.frame_faults + self.artifact_faults + @intFromBool(self.unreadable);
    }
};

/// Checks this repository's store for damage. Read-only in the strongest sense:
/// it takes no lock, opens nothing for write, and never truncates — so it is
/// safe to run against a store a `zit serve` has open.
///
/// Findings stream to stderr in walk order as they are found; the clean summary
/// is the only thing that ever reaches stdout, so a faulty run leaves stdout
/// empty and nothing downstream mistakes damage for a report.
pub fn runVerify(
    gpa: std.mem.Allocator,
    io: std.Io,
    arena: std.mem.Allocator,
    verb_args: []const []const u8,
    streams: cli.Streams,
    amb: cli.Ambient,
) cli.Exit {
    const help = switch (cli.parseNoPositionalArgs(verb_args)) {
        .ok => |a| a.help,
        .unknown_option => |tok| return cli.usageError(streams, "verify", .unknown_option, tok),
        .extra_argument => |tok| return cli.usageError(streams, "verify", .extra_argument, tok),
    };
    if (help) {
        streams.out.writeAll(verify_help) catch {};
        return .ok;
    }

    // `-R` is a global option parsed before dispatch, so it never reaches the
    // verb's own parser; `resolveRoot` is where the two routes converge.
    const root = (cli.resolveRoot(arena, io, amb.repo_override) catch return stateOom(streams, "verify")) orelse {
        streams.err.writeAll(not_a_repo_msg) catch {};
        return .failure;
    };
    var root_dir = root.dir;
    defer root_dir.close(io);

    var file = root_dir.openFile(io, repo_filename, .{}) catch return repoReadFail(streams, "verify");
    defer file.close(io);
    const st = file.stat(io) catch return repoReadFail(streams, "verify");
    var fs = storage.FileStorage.init(io, file, st.size);

    var tally: Tally = .{};

    // --- pass 1: every frame's whole-frame checksum ------------------------
    var v = logstore.FrameVerifier.init(fs.storageFile()) catch |e| {
        // The FILE header itself. Nothing past it can be read as frames at all,
        // so there is no inventory to take and no second pass to attempt.
        streams.err.print("zit verify: the store's file header is unreadable: {s}\n", .{reasonOf(e)}) catch {};
        streams.err.writeAll(advisory) catch {};
        return .failure;
    };
    var fatal = false;
    // STATED LIMIT ON OUTPUT: one stderr line per damaged frame, bounded by the
    // store's frame count and streamed as it is found. Deliberately uncapped —
    // the inventory IS the product here, and an operator who wants less can
    // pipe it; a cap would make the verb unable to answer the question it
    // exists to answer.
    while (v.next()) |fault| {
        tally.frame_faults += 1;
        renderFrameFault(streams, fault);
        if (fault.fatal()) {
            fatal = true;
            break;
        }
    }
    const totals = v.totals();
    if (!totals.complete and !fatal) {
        // DEFENCE AGAINST THE WORST OUTPUT THIS VERB CAN PRODUCE. Today the
        // only way the walk ends short is a fatal fault, which is already
        // reported and already raises the exit code — so this branch is
        // unreachable. It exists because the coupling is invisible from here:
        // if a future change ever lets the walk stop early without saying why,
        // a short walk must be a fault rather than a silent "0 faults" over an
        // unexamined store.
        tally.frame_faults += 1;
        streams.err.print(
            "zit verify: the frame walk ended at offset {d} without reporting why\n",
            .{totals.bytes_checked},
        ) catch {};
    }
    if (totals.torn_tail_at) |at| {
        // A NOTE, NOT A FAULT, and it must not move the exit code: `open`
        // recovers a torn tail silently because it means an interrupted write.
        // Sending an operator to re-clone over an ordinary crash would be a
        // worse failure than the one being reported.
        streams.err.print(
            "zit verify: incomplete final frame at offset {d} from an interrupted write; the next write truncates it, this is not corruption\n",
            .{at},
        ) catch {};
    }

    // --- pass 2: every artifact reconstructed and re-hashed -----------------
    if (fatal) {
        streams.err.writeAll("zit verify: per-artifact verification skipped: the frame walk could not continue\n") catch {};
    } else if (totals.torn_tail_at != null) {
        // `openForVerify` never truncates, so a torn tail is `error.Corrupt`
        // to it. Say which of the two passes did not run rather than reporting
        // a clean second pass that never happened.
        streams.err.writeAll("zit verify: per-artifact verification skipped: the store has an incomplete final frame\n") catch {};
    } else {
        verifyArtifacts(gpa, streams, fs.storageFile(), &tally);
    }

    // --- verdict ------------------------------------------------------------
    if (tally.faults() == 0) {
        if (tally.artifacts) |n| {
            streams.out.print(
                "verified: {d} {s}, {d} {s}, {d} bytes, 0 faults\n",
                .{ totals.frames, plural(totals.frames, "frame", "frames"), n, plural(n, "artifact", "artifacts"), totals.bytes_checked },
            ) catch {};
        } else {
            // The per-artifact pass did not run; stderr already said why. The
            // summary must not imply it did.
            streams.out.print(
                "verified: {d} {s}, {d} bytes, 0 faults; per-artifact verification did not run\n",
                .{ totals.frames, plural(totals.frames, "frame", "frames"), totals.bytes_checked },
            ) catch {};
        }
        return .ok;
    }
    // Each pass's count is stated on its own, because one physical bitflip in a
    // content frame produces a finding from BOTH — and "2 faults" would read as
    // two damaged frames. "N of M bytes checked" is mandatory and is a SEPARATE
    // FACT again: a verb that reports one fault after stopping at 8% of the
    // store has lied about the other 92%.
    if (tally.artifacts != null) {
        streams.err.print(
            "zit verify: {d} frame {s}, {d} artifact {s}, {d} of {d} bytes checked\n",
            .{
                tally.frame_faults,    plural(tally.frame_faults, "fault", "faults"),
                tally.artifact_faults, plural(tally.artifact_faults, "fault", "faults"),
                totals.bytes_checked,  totals.total,
            },
        ) catch {};
    } else {
        streams.err.print(
            "zit verify: {d} frame {s}, {d} of {d} bytes checked; per-artifact verification did not run\n",
            .{ tally.frame_faults, plural(tally.frame_faults, "fault", "faults"), totals.bytes_checked, totals.total },
        ) catch {};
    }
    streams.err.writeAll(advisory) catch {};
    return .failure;
}

/// One unconditional closing line, never branched on shun state: an operator
/// reading a damage report needs the repair route stated the same way every
/// time, and a conditional hint is one more thing to be wrong about.
const advisory =
    "zit verify: cloning from a healthy peer is the reliable repair; 'zit pack' drops artifacts already shunned\n";

fn plural(n: u64, one: []const u8, many: []const u8) []const u8 {
    return if (n == 1) one else many;
}

/// Renders one frame-pass finding. THE FULL LOWERCASE HEX NAME, never
/// abbreviated: the line's job is to be pasted into `zit artifact` or
/// `zit shun`, and a truncated name cannot be.
///
/// The name is available on a payload fault at all because `readFrameHeader`
/// checksum-verifies the prefix and header windows SEPARATELY from the payload,
/// and the hash lives in the header window — so a payload bitflip leaves the
/// name intact. That is what makes damage nameable from the frame pass alone.
fn renderFrameFault(streams: cli.Streams, fault: logstore.FrameFault) void {
    const kind = fault.kind orelse {
        const why = switch (fault.cause) {
            error.Corrupt => "header checksum failed",
            error.Storage => "input/output error",
        };
        streams.err.print(
            "zit verify: frame at offset {d}: {s}; cannot verify past this point\n",
            .{ fault.offset, why },
        ) catch {};
        return;
    };
    var name_buf: [2 * logstore.max_hash_len]u8 = undefined;
    const name = digestToWireName(fault.name(), &name_buf);
    // A post-header fault is one of two things, and they are not the same
    // advice: damaged bytes on a working disk, or a disk that would not hand
    // the bytes over at all. The walk continues in both cases.
    const what = switch (fault.cause) {
        error.Corrupt => "payload checksum mismatch",
        error.Storage => "payload bytes could not be read",
    };
    switch (kind) {
        .config => streams.err.print(
            "zit verify: frame at offset {d} (config record): {s}\n",
            .{ fault.offset, what },
        ) catch {},
        .shun => streams.err.print(
            "zit verify: frame at offset {d} (shun row for {s}): {s}\n",
            .{ fault.offset, name, what },
        ) catch {},
        else => streams.err.print(
            "zit verify: frame at offset {d} ({s}, {s}): {s}\n",
            .{ fault.offset, name, kindLabel(kind), what },
        ) catch {},
    }
}

/// A frame's payload encoding in operator words rather than field names.
fn kindLabel(kind: logstore.FrameKind) []const u8 {
    return switch (kind) {
        .raw => "full content",
        .zlib => "compressed content",
        .delta => "delta",
        .config => "config record",
        .shun => "shun row",
        _ => "unknown frame kind",
    };
}

/// Pass 2. Opens the store through `openForVerify` — NOT `openReadOnly`, for
/// two reasons that both matter here: `openReadOnly` scans `.verify_payloads`,
/// so on exactly the store being fsck-ed it would abort and this pass would
/// never run; and it sets `verified_on_open`, which makes every frame read below
/// skip its own checksum.
fn verifyArtifacts(
    gpa: std.mem.Allocator,
    streams: cli.Streams,
    file: storage.StorageFile,
    tally: *Tally,
) void {
    // A FAULT, NOT A SKIP. The frame pass walks raw bytes and can therefore
    // come up clean on a store `scan` still rejects — a checksum-valid but
    // malformed shun row is exactly that shape. Reporting "0 faults" and
    // exiting 0 on a repository every other verb refuses to open is the one
    // answer this verb must never give, and `verify_help`'s exit table already
    // promises 1 when "the store could not be read".
    var store = logstore.LogStore.openForVerify(gpa, file) catch |e| {
        tally.unreadable = true;
        streams.err.print("zit verify: the store could not be opened for per-artifact verification: {s}\n", .{reasonOf(e)}) catch {};
        return;
    };
    defer store.deinit(gpa);

    const rids = store.ridsInPersistedOrder(gpa) catch {
        tally.unreadable = true;
        streams.err.writeAll("zit verify: the store could not be inventoried for per-artifact verification: out of memory\n") catch {};
        return;
    };
    defer gpa.free(rids);
    // From here the pass HAS run, so the count is a real zero rather than
    // "nobody looked" even if every rid turns out to be a phantom.
    tally.artifacts = 0;

    var cache = logstore.PlaintextCache.init();
    defer cache.deinit(gpa);

    for (rids) |rid| {
        var name_buf: [2 * logstore.max_hash_len]u8 = undefined;
        const name = digestToWireName(store.hashOf(rid), &name_buf);
        // A shunned artifact whose bytes are still on disk IS verified and IS
        // reported. `zit pack` filters before verifying so corruption cannot
        // block its own removal; verify has no removal to block, and must not
        // claim clean about bytes it declined to look at.
        // `isShunned` asserts a COMPLETE digest width — the shun table is keyed
        // by complete artifact name and cannot answer about a prefix. A store
        // holding a frame of some other width is not something to trap on
        // inside an fsck, so the width is checked rather than assumed.
        const shunned = logstore.isCompleteDigest(store.hashOf(rid).len) and store.isShunned(store.hashOf(rid));
        // `contentUnmasked`, not `content`: the masked reader answers null for a
        // shunned artifact, so a masked loop would report "0 faults" about bytes
        // it never read.
        const content = store.contentUnmasked(gpa, rid, &cache) catch |e| {
            // A read error is a FINDING, not an abort — the next artifact may
            // be fine, and an fsck that stopped here would under-report.
            tally.artifact_faults += 1;
            streams.err.print(
                "zit verify: artifact {s}{s}: cannot reconstruct: {s}\n",
                .{ name, if (shunned) " (shunned, not yet reclaimed)" else "", reasonOf(e) },
            ) catch {};
            continue;
        } orelse continue; // unreachable: `ridsInPersistedOrder` already drops phantoms
        defer gpa.free(content);

        tally.artifacts = tally.artifacts.? + 1;
        // `artifact.verify` is total — a name of any length it does not
        // recognise is `false`, never a trap — so unlike `isShunned` above it
        // needs no width guard. It does mean a store carrying some hash width
        // other than 20 or 32 bytes would be reported as a name mismatch;
        // nothing zit writes produces one, and a store that holds one is
        // damaged by any reading.
        if (!artifact.verify(name, content)) {
            tally.artifact_faults += 1;
            streams.err.print(
                "zit verify: artifact {s}{s}: content does not hash to its name\n",
                .{ name, if (shunned) " (shunned, not yet reclaimed)" else "" },
            ) catch {};
        }
    }
}