commit 3729b1f0d54a88f119facacfacfc506b7fd1523e
author Ryan Savitski <rsavitski@google.com> Thu Feb 22 16:30:06 2024 +0000
committer Ryan Savitski <rsavitski@google.com> Thu Feb 22 16:30:06 2024 +0000
tree ce6c07d62c979f2016232de239484c2509de7da4
parent de1ec22020b84fb2864e7fb59a9459c0aefb9479

ftrace: add drain_buffer_percent option for poll-based ftrace ring buffer reads

Firstly, explaining net changes to existing configs that utilise only
periodic reads:
* removed per-tick-period quotas of number of pages that the reader
  extracts from the kernel buffers before it starts shedding load. There
  are situations where this is beneficial, but we're changing our
  approach here to instead scale with the underlying ftrace bandwidth,
  even if it increases the cpu% utilisation. Either way, I'm unconvinced
  these quotas were ever being hit in typical tracing configurations.
* small: for read tick period calculations, we now consider only the
  active data sources (as opposed to all setup ones).
* niche: for configs using non-boot ftrace clocks, moved the clock
  snapshotting closer to the beginning of a read pass.

The poll-based reads are enabled in addition to the periodic ReadTick if
the config requests them. If all concurrent configs are using poll and
none set an explicit drain_period_ms, we'll relax the ReadTick period to
1 second. The tick is still beneficial as it spreads out work more
evenly, helps procfs scrape freshness, and is a general fallback in case
polling doesn't work as expected. The polling is scoped to individual
ftrace instances (highly niche feature for now), but the tick is still
global.

Code-wise:
* config_muxer manages the tracefs/buffer_percent file, picking the
  smallest value out of active configs.
* ftrace_controller manages the polling and callbacks on the per-cpu
  ring buffer fds.

For now the feature is nop'd out on kernels below v6.1 (prior to which
the polls returned as soon as a buffer was non-empty), but I'm looking
into relaxing it on Android GKI builds that are known to have the fix
backported.

(Most of the testing has been manual, the mock-based unittests are too
brittle for me to want to pile on more at the moment.)

Change-Id: Icba56768ded03bb4d725fbb596e965c86195c33c