Update from the latency front

Efforts to track down and eliminate sources of latency in the 2.6 kernel continue. It seems, however, that most of the low-hanging fruit has been found; with the current iteration of the voluntary preemption patch, the remaining problems are rare and relatively hard to track down. So Ingo Molnar built himself a new tool to help with those harder cases.

Ingo's problem with the previous preempt timing patch was that, while it showed where a lengthy latency took place, it yielded little information about what was happening during the high-latency event. So he adapted the profiling mechanism to bring a little light to the situation. With the latency tracing option turned on, a little tracing function gets called as part of every kernel function call. This trace code stores the time of the call into a large (4000 entries), per-CPU array. If the kernel avoids scheduling for too long, that array of function call information gets copied into a static array which is made available via /proc/latency.

Ingo included some example output with his patch:

  preemption latency trace v1.0
  -----------------------------
   latency: 121 us, entries: 1032 (1032)
   process: default.hotplug/1470, uid: 0
   nice: -10, policy: 0, rt_priority: 0
  =======>
   0.000ms (+0.000ms): page_address (kmap_high)
   0.000ms (+0.000ms): page_slot (page_address)
   0.000ms (+0.000ms): flush_all_zero_pkmaps (kmap_high)
   0.000ms (+0.000ms): set_page_address (flush_all_zero_pkmaps)
  [...]
   0.118ms (+0.000ms): page_slot (set_page_address)
   0.118ms (+0.000ms): check_preempt_timing (sub_preempt_count)

The output shows the function call, and, in parentheses, the caller of each function. In this case, the output identifies flush_all_zero_pkmaps() as the real villain.

Other changes to this patch include making hardware and software interrupts (which have been redirected into kernel threads) preemptible by default ("I reviewed a number of softirq users and it appears to be safe"), and, of course, the breaking up of more code which holds locks for too long.

---

(Log in to post comments)