perf: Support PERF_SAMPLE_READ with inherit
This change allows events to use PERF_SAMPLE_READ with inherit
so long as PERF_SAMPLE_TID is also set. This enables sample based
profiling of a group of counters over a hierarchy of processes or
threads. This is useful, for example, for collecting per-thread
counters/metrics, event based sampling of multiple counters as a unit,
access to the enabled and running time when using multiplexing and so
on.
Prior to this, users were restricted to either collecting aggregate
statistics for a multi-threaded/-process application (e.g. with
"perf stat"), or to sample individual threads, or to profile the entire
system (which requires root or CAP_PERFMON, and may produce much more
data than is required). Theoretically a tool could poll for or otherwise
monitor thread/process creation and construct whatever events the user
is interested in using perf_event_open, for each new thread or process,
but this is racy, can lead to file-descriptor exhaustion, and ultimately
just replicates the behaviour of inherit, but in userspace.
This configuration differs from inherit without PERF_SAMPLE_READ in that
the accumulated event count, and consequently any sample (such as if
triggered by overflow of sample_period) will be on a per-thread rather
than on an aggregate basis.
The meaning of read_format::value field of both PERF_RECORD_READ and
PERF_RECORD_SAMPLE is changed such that if the sampled event uses this
new configuration then the values reported will be per-thread rather
than the global aggregate value. This is a change from the existing
semantics of read_format (where PERF_SAMPLE_READ is used without
inherit), but it is necessary to expose the per-thread counter values,
and it avoids reinventing a separate "read_format_thread" field that
otherwise replicates the same behaviour. This change should not break
existing tools, since this configuration was not previously valid and
was rejected by the kernel. Tools that opt into this new mode will need
to account for this when calculating the counter delta for a given
sample. Tools that wish to have both the per-thread and aggregate value
can perform the global aggregation themselves from the per-thread
values.
The change to read_format::value does not affect existing valid
perf_event_attr configurations, nor does it change the behaviour of
calls to "read" on an event descriptor. Both continue to report the
aggregate value for the entire thread/process hierarchy. The difference
between the results reported by "read" and PERF_RECORD_SAMPLE in this
new configuration is justified on the basis that it is not (easily)
possible for "read" to target a specific thread (the caller only has
the fd for the original parent event).
Signed-off-by: Ben Gainey <ben.gainey@arm.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Link: https://lkml.kernel.org/r/20240730084417.7693-3-ben.gainey@arm.com
2 files changed
