hardware: tighten mq-deadline read_expire for jellyfin coexistence

modules/hardware.nix

@@ -12,7 +12,7 @@ let
     parent=''${1%%[0-9]*}
     dev="/sys/block/$parent"
     [ -d "$dev/queue/iosched" ] || exit 0
-    echo 15000 > "$dev/queue/iosched/read_expire"
+    echo 500 > "$dev/queue/iosched/read_expire"
     echo 15000 > "$dev/queue/iosched/write_expire"
     echo 128 > "$dev/queue/iosched/fifo_batch"
     echo 16 > "$dev/queue/iosched/writes_starved"
@@ -36,11 +36,17 @@ in
   hardware.cpu.amd.updateMicrocode = true;
   hardware.enableRedistributableFirmware = true;
 
-  # HDD I/O tuning for torrent seeding workload (high-concurrency random reads).
+  # HDD I/O tuning for torrent seeding workload (high-concurrency random reads)
+  # sharing the pool with latency-sensitive sequential reads (Jellyfin playback).
   #
   # mq-deadline sorts requests into elevator sweeps, reducing seek distance.
-  # Aggressive deadlines (15s) let the scheduler accumulate more ops before dispatching,
-  # maximizing coalescence — latency is irrelevant since torrent peers tolerate 30-60s.
+  # read_expire=500ms keeps reads bounded so a Jellyfin segment can't queue for
+  # seconds behind a torrent burst; write_expire=15s lets the scheduler batch
+  # writes for coalescence (torrent writes are async and tolerate delay).
+  # The bulk of read coalescence already happens above the scheduler via ZFS
+  # aggregation (zfs_vdev_aggregation_limit=4M, read_gap_limit=128K,
+  # async_read_max=32), so the scheduler deadline only needs to be large enough
+  # to keep the elevator sweep coherent -- 500ms is plenty on rotational disks.
   # fifo_batch=128 keeps sweeps long; writes_starved=16 heavily favors reads.
   # 4 MiB readahead matches libtorrent piece extent affinity for sequential prefetch.
   #