behind is never cleared on backlog dropSeverity: MEDIUM
Status: ✅ FIXED
Files: src/server/server.zig:2830-2837, src/server/server.zig:2766-2772, src/server/server.zig:2890
Verification status: the logic was read from live source and the state machine is unambiguous; the reachability (a client accumulating >16 MiB behind) was not demonstrated. Treat as a lead until reproduced.
When a display client falls far enough behind, the server abandons its backlog:
// src/server/server.zig:2830-2837
if (dc.out_buf.items.len + total_len > MAX_OUT_BUF) {
_ = self.flushDisplayClient(dc);
if (dc.out_buf.items.len + total_len > MAX_OUT_BUF) {
dc.out_buf.clearRetainingCapacity();
dc.last_cells.clearRetainingCapacity();
return false;
}
}
The drop path clears out_buf and last_cells but leaves dc.behind == true
and behind_count set. That combination deadlocks the client:
out_buf.items.len == 0, so the top-of-loop
flushDisplayClient is skipped.writeFrameToClient short-circuits on if (dc.behind) return true;
(:2890) and writes nothing.if (dc.behind or !self.flushDisplayClient(dc)) all_flushed = false;
(:3352) short-circuits on the first term, so flushDisplayClient is never
called and behind can never clear.The result is that all_flushed stays false permanently, self.dirty is never
cleared, and that client never receives another frame — even after it drains.
appendClientOutUrgent (:2764) has the same omission.
When dropping the backlog, clear the flag in the same step:
if (dc.behind) {
dc.behind = false;
self.behind_count -|= 1;
}
and, if no client remains behind, call resumePaneReads(). Factor the cap check
into one helper so both appendClientOut variants share it (see bug #520,
site 1).
All backlog-clearing and client-disconnect sites (appendClientOutUrgent,
appendClientOutFrame, removeClient, writeFrameToClient, and
flushDisplayClient) now clear the flag through one helper:
// src/server/server.zig
fn clearBehind(self: *Server, dc: *DisplayClient) void {
if (dc.behind) {
dc.behind = false;
self.behind_count -|= 1;
if (!self.anyDisplayClientBehind()) {
self.resumePaneReads();
}
}
}
The saturating decrement matches every other behind_count site, so a stale
flag can never underflow the counter and make anyDisplayClientBehind() lie.
When behind_count reaches zero, resumePaneReads() is called immediately to
re-arm POLL.IN on every pane master pty (which was disarmed during throttling
at server.zig:811). Without this, a force-drop left pane reads disarmed forever
because out_buf was emptied and the flush catch-up path could never fire.
The same fix applies to client disconnects (removeClient) and broken-pipe
errors (writeFrameToClient / flushDisplayClient).
clearBehind resets the flag and the counter — bug #522 and
clearBehind resumes pane reads when behind_count reaches zero — bug #522 follow-up
(src/server/server.zig) verify flag clearing, underflow safety, and that
POLL.IN on pane master ptys is re-armed if and only if no client remains behind.
The drop sites themselves need a >16 MiB backlog (MAX_OUT_BUF = 16 << 20) to
reach, which is too heavy for this suite’s shared DebugAllocator budget, so
they are covered by inspection rather than by a test that fills the buffer.
Verified with the full suite on Zig 0.16.0 and 0.17.0: 1053/1053 passing on both.
Bugs #297, #298, #376 and #403 all concern display backpressure; #376 in particular is about responses being dropped while a client is marked behind.