iopsystems/rezolus v5.25.0

on GitHub

The per-cgroup BPF cost reductions, the cgroup_attribution option for five
more samplers and the hindsight stop fix from 6.0, backported in #1406.

Changed

• cpu_usage, cpu_migrations, cpu_tlb_flush, scheduler_runqueue and
  syscall_counts honour cgroup_attribution, on by default as for
  cpu_perf, so nothing changes unless a config sets it. With
  cgroup_attribution = false the per-cgroup path (the task-group read, the
  new-cgroup check and the per-cgroup adds) is folded out of each sampler's
  hook and the cgroup_* series are absent; host-level series are
  unchanged. A config that already sets [defaults] cgroup_attribution = false now reaches these five samplers as well. The cgroups view finds
  cgroup names from any per-cgroup series when cgroup_cpu_usage is absent.
• syscall_counts attaches to sys_enter as a raw tracepoint (tp_btf,
  or raw_tp without kernel BTF) instead of the raw_syscalls/sys_enter
  tracepoint, and its per-cgroup path reads the task group once, as direct
  loads from a BTF task pointer on kernels from 5.11. On bare metal (EPYC
  4564P, 6.12, measured on 6.0) the program costs 33–39 ns per syscall with
  cgroup_attribution on, down from 140–151 ns.
• Every sampler with a per-event cgroup path reads the task group once, as
  direct loads from a BTF task pointer where the program is tp_btf,
  fentry or fexit, instead of through up to seven
  bpf_probe_read_kernel() calls. cpu_tlb_flush attaches to tlb_flush
  as tp_btf where the kernel has BTF.
• scheduler_runqueue, cpu_migrations and cpu_usage read the task's
  pid, state and CPU times as direct loads from BTF pointers.
  syscall_latency attaches to sys_enter and sys_exit as raw
  tracepoints instead of the raw_syscalls tracepoints. With bpftrace
  keeping classic programs on those tracepoints, the two syscall samplers
  together cost about 121 ns more per syscall than on classic tracepoints;
  with no other syscall tracer they cost about 82 ns less (KVM guest,
  measured on 6.0).
• The series of every sampler above are unchanged. Kernels older than 5.11
  keep the bpf_probe_read_kernel() paths.

Fixed

• Hindsight exits on SIGTERM and SIGINT. Hindsight treated SIGTERM and
  SIGINT like SIGHUP: it captured the buffer and kept recording. systemctl stop and restart therefore waited out TimeoutStopSec (120 s) and
  SIGKILLed the daemon, leaving its buffer directory behind. SIGTERM and
  SIGINT now capture the buffer and exit with status 0, or 1 if that capture
  failed. A stop during a SIGHUP capture exits when that capture completes;
  a second stop exits at once with status 2 and removes the buffer
  directory. SIGHUP captures and keeps recording, and a SIGHUP during a
  capture is ignored.