DOKK / manpages / debian 12 / libpfm4-dev / pfm_get_os_event_encoding.3.en
LIBPFM(3) Linux Programmer's Manual LIBPFM(3)

pfm_get_os_event_encoding - get event encoding for a specific operating system

#include <perfmon/pfmlib.h>

int pfm_get_os_event_encoding(const char *str, int dfl_plm, pfm_os_t os,  void *arg);

This is the key function to retrieve the encoding of an event for a specific operating system interface. The event string passed in str is parsed and encoded for the operating system specified by os. Only one event per call can be encoded. As such, str can contain only one symbolic event name. The event is encoded to monitor at the privilege levels specified by the dfl_plm mask, if supported, otherwise this parameter is ignored. The operating system specific input and output arguments are passed in arg.

The event string, str, may contains sub-event masks (umask) and any other supported modifiers. Only one event is parsed from the string. For convenience, it is possible to pass a comma-separated list of events in str but only the first event is encoded.

The following values are supported for os:

This value causes the event to be encoded purely as specified by the PMU hardware. The arg argument must be a pointer to a pfm_pmu_encode_arg_t structure which is defined as follows: 

typedef struct {


    uint64_t    *codes;


    char        **fstr;


    size_t      size;


    int         count;


    int         idx;
} pfm_pmu_encode_arg_t;

The fields are defined as follows:

A pointer to an array of 64-bit values. On input, if codes is NULL, then the library allocates whatever is necessary to store the encoding of the event. If codes is not NULL on input, then count must reflect its actual number of elements. If count is big enough, the library stores the encoding at the address provided. Otherwise, an error is returned.
On input, the field contains the maximum number of elements in the array codes. Upon return, it contains the number of actual entries in codes. If codes is NULL, then count must be zero.
If the caller is interested in retrieving the fully qualified event string where all used unit masks and all modifiers are spelled out, this field must be set to a non-null address of a pointer to a string (char **). Upon return, if fstr was not NULL, then the string pointer passed on entry points to the event string. The string is dynamically allocated and must eventually be freed by the caller. If fstr was NULL on entry, then nothing is returned in this field. The typical calling sequence looks as follows: 


   char *fstr = NULL


   pfm_pmu_encode_arg_t arg;


   arg.fstr = &fstr;


   ret = pfm_get_os_event_encoding("event",


                                   PFM_PLM0|PFM_PLM3,


                                   PFM_OS_NONE,


                                   &e);


   if (ret == PFM_SUCCESS) {


      printf("fstr=%s\n", fstr);


      free(fstr);


   }
This field contains the size of the struct passed. This field is used to provide for extensibility of the struct without compromising backward compatibility. The value should be set to sizeof(pfm_pmu_encode_arg_t). If instead, a value of 0 is specified, the library assumes the struct passed is identical to the first ABI version which size is PFM_RAW_ENCODE_ABI0. Thus, if fields were added after the first ABI, they will not be set by the library. The library does check that bytes beyond what is implemented are zeroes.
Upon return, this field contains the opaque unique identifier for the event described in str. This index can be used to retrieve information about the event using pfm_get_event_info(), for instance.
This value causes the event to be encoded for the perf_event Linux kernel interface (available since 2.6.31). The arg must be a pointer to a pfm_perf_encode_arg_t structure. The PFM_OS_PERF_EVENT layer provides the modifiers exported by the underlying PMU hardware, some of which may actually be overridden by the perf_event interface, such as the monitoring privilege levels. The PFM_OS_PERF_EVENT_EXT extends PFM_OS_PERF_EVENT to add modifiers controlled only by the perf_event interface, such as sampling period (period), frequency (freq) and exclusive resource access (excl). 

typedef struct {


    struct perf_event_attr *attr;


    char **fstr;


    size_t size;


    int idx;


    int cpu;


    int flags;
} pfm_perf_encode_arg_t;
The fields are defined as follows:
A pointer to a struct perf_event_attr as defined in perf_event.h. This field cannot be NULL on entry. The struct is not completely overwritten by the call. The library only modifies the fields it knows about, thereby allowing perf_event ABI mismatch between caller and library.
Same behavior as is described for PFM_OS_NONE above.
This field contains the size of the struct passed. This field is used to provide for extensibility of the struct without compromising backward compatibility. The value should be set to sizeof(pfm_perf_encode_arg_t). If instead, a value of 0 is specified, the library assumes the struct passed is identical to the first ABI version which size is PFM_PERF_ENCODE_ABI0. Thus, if fields were added after the first ABI, they will not be set by the library. The library does check that bytes beyond what is implemented are zeroes.
Upon return, this field contains the opaque unique identifier for the event described in str. This index can be used to retrieve information about the event using pfm_get_event_info(), for instance.
Not used yet.
Not used yet.

Here is a example of how this function could be used with PFM_OS_NONE:

#include <inttypes.h>
#include <err.h>
#include <perfmon/pfmlib.h>
int main(int argc, char **argv)
{


   pfm_pmu_encode_t arg;


   int ret;


   ret = pfm_initialize();


   if (ret != PFMLIB_SUCCESS)


      errx(1, "cannot initialize library %s", pfm_strerror(ret));


   memset(&arg, 0, sizeof(arg));


   ret = pfm_get_os_event_encoding("RETIRED_INSTRUCTIONS", PFM_PLM3, PFM_OS_NONE, &arg);


   if (ret != PFM_SUCCESS)


      err(1", cannot get encoding %s", pfm_strerror(ret));


   for(i=0; i < arg.count; i++)


      printf("count[%d]=0x%"PRIx64"\n", i, arg.codes[i]);


   free(arg.codes);


   return 0;
}

The function returns in arg the encoding of the event for the os passed in os. The content of arg depends on the os argument. Upon success, PFM_SUCCESS is returned otherwise a specific error code is returned.

The code argument is too small for the encoding.
The code or count argument is NULL or the str contains more than one symbolic event.
Not enough memory.
Event not found.
Invalid event attribute (unit mask or modifier)
Invalid modifier value.
attribute already set, cannot be changed.
Missing unit mask.
Unit masks or features cannot be combined into a single event.

Stephane Eranian <eranian@gmail.com>

January, 2011