PMC_CONTROL(2)						      BSD System Calls Manual						    PMC_CONTROL(2)

NAME

     pmc_control, pmc_get_info -- Hardware Performance Monitoring Interface

LIBRARY

     Standard C Library (libc, -lc)

SYNOPSIS

     #include <sys/pmc.h>

     int
     pmc_control(int ctr, int op, void *argp);

     int
     pmc_get_info(int ctr, int op, void *argp);

DESCRIPTION

     pmc_get_info() returns the number of counters in the system or information on a specified counter ctr.  The possible values for op are:

     PMC_INFO_NCOUNTERS
	    When querying the number of counters in the system, ctr is ignored and argp is of type int *.  Upon return, the integer pointed to by
	    argp will contain the number of counters that are available in the system.

     PMC_INFO_CPUCTR_TYPE
	    When querying the type of a counter in the system, ctr refers to the counter being queried, and argp is of type int *.  Upon return,
	    the integer pointed to by argp will contain the implementation-dependent type of the specified counter.

	    If ctr is -1, the integer pointed to by argp will contain the machine-dependent type describing the CPU or counter configuration.

     PMC_INFO_COUNTER_VALUE
	    When querying the value of a counter in the system, ctr refers to the counter being queried, and argp is of type uint64_t *.  Upon
	    return, the 64-bit integer pointed to by argp will contain the value of the specified counter.

     PMC_INFO_ACCUMULATED_COUNTER_VALUE
	    When querying the value of a counter in the system, ctr refers to the counter being queried, and argp is of type uint64_t *.  Upon
	    return, the 64-bit integer pointed to by argp will contain the sum of the accumulated values of specified counter in all exited sub-
	    processes of the current process.

     pmc_control() manipulates the specified counter ctr in one of several fashions.  The op parameter determines the action taken by the kernel
     and also the interpretation of the argp parameter.  The possible values for op are:

     PMC_OP_START
	    Starts the specified ctr running.  It must be preceded by a call with PMC_OP_CONFIGURE.  argp is ignored in this case and may be NULL.

     PMC_OP_STOP
	    Stops the specified ctr from running.  argp is ignored in this case and may be NULL.

     PMC_OP_CONFIGURE
	    Configures the specified ctr prior to running.  argp is a pointer to a struct pmc_counter_cfg.

		    struct pmc_counter_cfg {
			    pmc_evid_t	    event_id;
			    pmc_ctr_t	    reset_value;
			    uint32_t	    flags;
		    };

	    event_id
		   is the event ID to be counted.

	    reset_value
		   is a value to which the counter should be reset on overflow (if supported by the implementation).  This is most useful when
		   profiling (see PMC_OP_PROFSTART, below).  This value is defined to be the number of counter ticks before the next overflow.
		   So, to get a profiling tick on every hundredth data cache miss, set the event_id to the proper value for ``dcache-miss'' and
		   set reset_value to 100.

	    flags  Currently unused.

     PMC_OP_PROFSTART
	    Configures the specified ctr for use in profiling.	argp is a pointer to a struct pmc_counter_cfg as in PMC_OP_CONFIGURE, above.  This
	    request allocates a kernel counter, which will fail if any process is using the requested counter.	Not all implementations or coun-
	    ters may support this option.

     PMC_OP_PROFSTOP
	    Stops the specified ctr from being used for profiling.  argp is ignored in this case and may be NULL.

RETURN VALUES

     A return value of 0 indicates that the call succeeded.  Otherwise, -1 is returned and the global variable errno is set to indicate the error.

ERRORS

     Among the possible error codes from pmc_control() and pmc_get_info() are

     [EFAULT]		The address specified for the argp is invalid.

     [ENXIO]		Specified counter is not yet configured.

     [EINPROGRESS]	PMC_OP_START was passed for a counter that is already running.

     [EINVAL]		Specified counter was invalid.

     [EBUSY]		If the requested counter is already in use--either by the current process or by the kernel.

     [ENODEV]		If and only if the specified counter event is not valid for the specified counter when configuring a counter or starting
			profiling.

     [ENOMEM]		If the kernel is unable to allocate memory.

SEE ALSO

     pmc(1), pmc(9)

HISTORY

     The pmc_control() and pmc_get_info() system calls appeared in NetBSD 2.0.

BSD
								 October 27, 2005							       BSD