See chapters 18,19 of the Intel Architecture Software Developer's Manual.
Several ioctls are provided to control performance data collection.
ssize_t ioctl_cpuperf_get_properties(int fd, cpuperf_properties_t* props);
Return various aspects of PMU properties in |*props|.
Returns 0 on success or a negative error code.
ssize_t ioctl_cpuperf_alloc_trace(int fd, const ioctl_cpuperf_alloc_t* alloc);
Allocate various resources needed to perform a trace. This must be called before staging a configuration and starting a trace. This must be called while tracing is stopped.
Returns 0 on success or a negative error code.
ssize_t ioctl_cpuperf_free_trace(int fd);
Free all resources allocated by a preceding all to ioctl_cpuperf_alloc_trace(). This must be called while tracing is stopped.
Returns 0 on success or a negative error code.
ssize_t ioctl_cpuperf_get_alloc(int fd, cpuperf_alloc_t* alloc);
Fetch the trace configuration passed in to a preceding call to ioctl_cpuperf_alloc_trace().
Returns 0 on success or a negative error code.
ssize_t ioctl_cpuperf_stage_config(int fd, const cpuperf_config_t* config);
Configure data collection. |*config| specifies the events to collect and the rate at which to collect them. This must be called while tracing is stopped.
Returns 0 on success or a negative error code.
ssize_t ioctl_cpuperf_get_config(int fd, cpuperf_config_t* config);
Fetch the configuration passed in to a preceding call to ioctl_cpuperf_stage_config().
Returns 0 on success or a negative error code.
ssize_t ioctl_cpuperf_get_buffer_handle(
int fd, const ioctl_cpuperf_buffer_handle_req_t* rqst,
zx_handle_t* handle);
Fetch the handle of the VMO for the given descriptor. Each CPU is given a separate VMO and the descriptor is the cpu's number. This must be called while tracing is stopped.
Returns 0 on success or a negative error code.
ssize_t ioctl_cpuperf_start(int fd);
Start data collection. This must be called while tracing is stopped.
Returns 0 on success or a negative error code.
ssize_t ioctl_cpuperf_stop(int fd);
Stop tracing and collect any remaining data from each cpu. This may be called even if tracing is already stopped.
Returns 0 on success or a negative error code.
Here's a sketch of typical usage: