1 files changed, 50 insertions, 5 deletions

diff --git a/src/ninep/events.zig b/src/ninep/events.zig
index 98799774..d7079c5e 100644
--- a/src/ninep/events.zig
+++ b/src/ninep/events.zig
@@ -292,11 +292,13 @@ pub fn dropMessage(p: *Pardes, text: []const u8) void {
 /// counts that one instead: `<record> (x2)`, the count being every time it
 /// was said. One a follower has already read is not rewritten: the repeat
 /// is a new line carrying the running count, `(x3)`.
-fn pushCounting(p: *Pardes, record: []u8) void {
+fn pushCounting(p: *Pardes, raw: []u8) void {
     // A client retrying a write that fails the same way would fill the ring
     // with one line, so a repeat of the newest record is that record counted,
     // `(x3)`, as +Messages counts its repeats; unless a follower has read it
     // already and so waits on the repeat as a line of its own.
+    var clean: [4 * 4096 + 256]u8 = undefined;
+    const record = sanitize(raw, &clean);
     const last = newest(p) orelse return pushLog(p, record);
     var said = last.text;
     var times: u32 = 1;
@@ -340,10 +342,47 @@ fn followerRead(p: *Pardes, seq: u64) bool {
 
 /// The log is one ring that records whether or not anyone reads it. A record
 /// is one line: a newline in a message or a name would read as two records.
-fn pushLog(p: *Pardes, record: []u8) void {
-    for (record[0 .. record.len - 1]) |*c| if (c.* < ' ') {
-        c.* = ' ';
-    };
+/// A record as one line of text: a control character, DEL or a C1 control
+/// (U+0080-U+009F) is a space each, and a byte that is not UTF-8 is
+/// written `\xNN`, so a reader splitting on newlines and decoding UTF-8
+/// never trips. The record's own newline, last, is kept.
+fn sanitize(record: []const u8, out: []u8) []u8 {
+    var w: usize = 0;
+    var i: usize = 0;
+    const body = record[0 .. record.len - 1];
+    while (i < body.len and w + 4 < out.len) {
+        const c = body[i];
+        if (c < ' ' or c == 0x7f) {
+            out[w] = ' ';
+            w += 1;
+            i += 1;
+            continue;
+        }
+        const n = std.unicode.utf8ByteSequenceLength(c) catch 0;
+        if (n == 0 or i + n > body.len or !std.unicode.utf8ValidateSlice(body[i .. i + n])) {
+            _ = std.fmt.bufPrint(out[w..], "\\x{x:0>2}", .{c}) catch break;
+            w += 4;
+            i += 1;
+            continue;
+        }
+        if (n == 2 and c == 0xC2 and body[i + 1] <= 0x9F) {
+            out[w] = ' ';
+            w += 1;
+            i += 2;
+            continue;
+        }
+        if (w + n + 1 > out.len) break;
+        @memcpy(out[w..][0..n], body[i..][0..n]);
+        w += n;
+        i += n;
+    }
+    out[w] = '\n';
+    return out[0 .. w + 1];
+}
+
+fn pushLog(p: *Pardes, raw: []u8) void {
+    var clean: [4 * 4096 + 256]u8 = undefined;
+    const record = sanitize(raw, &clean);
     // One record larger than the ring would push every other out and then
     // not fit itself; cut it to what fits instead, on a character boundary.
     var end = @min(record.len, p.fs.log.cap - 4) - 1;
@@ -1244,6 +1283,12 @@ test "a long msg record is cut between words at its cap, with an ellipsis" {
     _ = call(p, .{ .tag = 3, .op = .release, .node = log, .handle = h });
 }
 
+test "a record is one line of UTF-8: DEL and C1 are spaces, bytes not UTF-8 are escaped" {
+    var out: [64]u8 = undefined;
+    try testing.expectEqualStrings("msg - a b c \\xff d\n", sanitize("msg - a\x7fb\xc2\x85c \xff d\n", &out));
+    try testing.expectEqualStrings("msg - caf\xc3\xa9\n", sanitize("msg - caf\xc3\xa9\n", &out));
+}
+
 test "a long err record is cut between words, with an ellipsis" {
     const p = try withFile(testing.allocator, "x\n");
     defer p.deinit();