xref: /xnu-11417.121.6/doc/observability/cpu_counters.md (revision a1e26a70f38d1d7daa7b49b258e2f8538ad81650)
1*a1e26a70SApple OSS Distributions# CPU Counters
2*a1e26a70SApple OSS Distributions
3*a1e26a70SApple OSS DistributionsThe xnu subsystems that manage CPU performance counters.
4*a1e26a70SApple OSS Distributions
5*a1e26a70SApple OSS Distributions## Overview
6*a1e26a70SApple OSS Distributions
7*a1e26a70SApple OSS DistributionsCPU performance counters are hardware registers that count events of interest to efficient CPU execution.
8*a1e26a70SApple OSS DistributionsCounters that measure events closely correlated with each CPU's execution pipeline are managed by the Core Performance Monitoring Unit (CPMU).
9*a1e26a70SApple OSS DistributionsThe CPMU contains both fixed instructions and cycles counters, as well as configurable counters that can be programmed to count any of several hundred possible events.
10*a1e26a70SApple OSS DistributionsIn addition to the CPMU, the Last Level Cache (LLC) hosts the Uncore Performance Monitoring Unit (UPMU), which measures effects that aren't necessarily correlated to a single CPU.
11*a1e26a70SApple OSS DistributionsAll counters in the UPMU are configurable.
12*a1e26a70SApple OSS Distributions
13*a1e26a70SApple OSS DistributionsCounters are typically used in one of two ways:
14*a1e26a70SApple OSS Distributions
15*a1e26a70SApple OSS Distributions1. In "counting" mode, their counts are periodically queried and tallied up for a duration of interest.
16*a1e26a70SApple OSS Distributions2. In "sampling" mode, the counters are programmed to generate a Performance Monitor Interrupt (PMI) periodically, during which the currently running code can be sampled, like a time profiler.
17*a1e26a70SApple OSS Distributions
18*a1e26a70SApple OSS Distributions## Subsystems
19*a1e26a70SApple OSS Distributions
20*a1e26a70SApple OSS DistributionsThere are several subsystems that provide access to CPU counter hardware:
21*a1e26a70SApple OSS Distributions
22*a1e26a70SApple OSS Distributions- kpc: The Kernel Performance Counter system is the oldest subsystem and still manages the configurable CPMU counters.
23*a1e26a70SApple OSS DistributionsIt can use PMIs from these counters to trigger kperf samples and counter values can be recorded in kperf samples.
24*a1e26a70SApple OSS Distributions
25*a1e26a70SApple OSS Distributions- Monotonic: The Monotonic system provides access to the fixed CPMU counters with limited support for PMIs.
26*a1e26a70SApple OSS DistributionsAdditionally, the UPMU is entirely provided by a Monotonic dev node interface.
27*a1e26a70SApple OSS Distributions
28*a1e26a70SApple OSS Distributions- cpc: The CPU Performance Counter subsystem provides a policy layer on top of kpc and Monotonic to prevent malicious use of the hardware.
29*a1e26a70SApple OSS Distributions
30*a1e26a70SApple OSS DistributionsEventually, cpc will subsume kpc's and Monotonic's roles in the system.
31*a1e26a70SApple OSS Distributions
32*a1e26a70SApple OSS Distributions## Integrations
33*a1e26a70SApple OSS Distributions
34*a1e26a70SApple OSS Distributions- The Recount subsystem makes extensive use of the fixed CPMU counters to attribute CPU resources back to threads and processes.
35*a1e26a70SApple OSS Distributions
36*a1e26a70SApple OSS Distributions- Microstackshot telemetry is sampled periodically using the CPMU's cycle PMI trigger.
37*a1e26a70SApple OSS Distributions
38*a1e26a70SApple OSS Distributions- Stackshot includes cycles and instructions for each thread container in its kcdata.
39*a1e26a70SApple OSS Distributions
40*a1e26a70SApple OSS Distributions- The kperf profiling system can trigger samples of thread states and call stacks using CPMU PMIs, allowing it to sample thread states and call stacks.
41*a1e26a70SApple OSS DistributionsAnd CPU counter values can be sampled by kperf on other triggers, like timers or kdebug events.
42*a1e26a70SApple OSS Distributions
43*a1e26a70SApple OSS Distributions## See Also
44*a1e26a70SApple OSS Distributions
45*a1e26a70SApple OSS Distributions- <doc:recount>
46