SYNOPSIS

#include <perfmon/pfmlib.h>

---

int pfm_dispatch_events(pfmlib_input_param_t *p, void *mod_in, pfmlib_output_param_t *q,void *mod_out);

---

DESCRIPTION

This function is the central piece of the library. It is important to understand that the library does not effectively program the PMU, i.e., it does not make the operating system calls. The PMU is never actually accessed by the library. Instead, the library helps applications prepare the arguments to pass to the kernel. In particular, it sets up the values to program into the PMU configuration registers (PMC). The list of used data registers (PMD) is also returned.

The input argument are divided into two categories: the generic arguments in p and the optional PMU model specific arguments in mod_in. The same applies for the output arguments: q contains the generic output arguments and mod_out the optional PMU model specific arguments.

An application describes what it wants to measure in the in and if it uses some model specific features, such as opcode matching on Itanium 2 processors, it must pass a pointer to the relevant model-specific input parameters in mod_in. The generic output parameters contains the register index and values for the PMC and PMD registers needed to make the measurement. The index mapping is guaranteed to match the mapping used by the Linux perfmon2 interface. In case the library is not used on this system, the hardware register addresses or indexes can also be retrieved from the output structure.

The pfmlib_input_param_t structure is defined as follows:

typedef struct 
        int           event;
        unsigned int  plm;
        unsigned long flags;
        unsigned int  unit_masks[PFMLIB_MAX_MASKS_PER_EVENT];
        unsigned int  num_masks;
} pfmlib_event_t;
typedef struct {
        unsigned int     pfp_event_count;
        unsigned int     pfp_dfl_plm;
        unsigned int     pfp_flags;
        pfmlib_event_t   pfp_events[PFMLIB_MAX_PMCS];
        pfmlib_regmask_t pfp_unavail_pmcs;
} pfmlib_input_param_t;

The structure mostly contains one table, called pfp_events which describes the events to be measured. The number of submitted events is indicated by pfp_event_count.

Each event is described in the pfp_events table by an opaque descriptor stored in the event field. This descriptor is obtained with the pfm_find_full_event() or derivative functions. For some events, it may be necessary to specify at least one unit mask in the unit_masks table. A unit mask is yet another opaque descriptor obtained via pfm_find_event_mask() or pfm_find_full_event. Typically, if an event supports multiple unit masks, they can be combined in which case more than one entry in unit_masks must be specified. The actual number of unit mask descriptors passed must be indicated in num_masks. When no unit mask is used, this field must be set to 0.

A privilege level mask for the event can be provided in plm. This is a bitmask where each bit indicates a privilege level at which to monitor, more than one bit can be set. The library supports up to four levels, but depending on the PMU model, some levels may not be available. The levels are as follows:

PFM_PLM0

monitor at the privilege level 0. For many architectures, this means kernel level

PFM_PLM1

monitor at privilege level 1

PFM_PLM2

monitor at privilege level 2

PFM_PLM3

monitor at the privilege level 3. For many architectures, this means user level

Events with a plm value of 0 will use the default privilege level mask as indicated by pfp_dfl_plm which must be set to any combinations of values described above. It is illegal to have a value of 0 for this field.

The pfp_flags field contains a set of flags that affect the whole set of events to be monitored. The currently defined flags are:

PFMLIB_PFP_SYSTEMWIDE

indicates that the monitors are to be used in a system-wide monitoring session. This could influence the way the library sets up some register values.

The pfp_unavail_pmcs bitmask can be used by applications to communicate to the library the list of PMC registers which are not available on the system. Some kernels may allocate certain PMC registers (and associated data registers) for other purposes. Those registers must not be used by the library otherwise the assignement of events to PMC registers may be rejected by the kernel. Applications must figure out which registers are available using a kernel interface at their disposal, the library does not provide this service. The library expect the restrictions to be expressed using the Linux perfmon2 PMC register mapping.

Refer to the PMU specific manual for a description of the model-specific input parameters to be passed in mod_in.

The generic output parameters are contained in the fBpfmlib_output_param_t structure which is defined as:

typedef struct {
    unsigned long long reg_value;
    unsigned int       reg_num;
    unsigned long      reg_addr;
} pfmlib_reg_t;
typedef struct {
    unsigned int pfp_pmc_count;
    unsigned int pfp_pmd_count;
    pfmlib_reg_t pfp_pmcs[PFMLIB_MAX_PMCS];
    pfmlib_reg_t pfp_pmds[PFMLIB_MAX_PMDS];
} pfmlib_output_param_t;

The number of valid entries in the pfp_pmcs table is indicated by pfp_pmc_count. The number of valid entries in the pfp_pmds table is indicated by pfp_pmd_count. Each entry in both tables is of type pfmlib_reg_t.

In the pfp_pmcs table, the reg_num contains the PMC register index (perfmon2 mapping), and the reg_value contains a 64-bit value to be used to program the PMC register. The reg_addr indicates the hardware address or index for the PMC register.

In the pfp_pmds table, the reg_num contains the PMD register index (perfmon2 mapping). the reg_value is ignored. The reg_addr indicates the hardware address or index for the PMC register.

Refer to the PMU specific manual for a description of the model-specific output parameters to be returned in mod_out.

The current implementation of the pfm_dispatch_events completely overwrites the pfmlib_output_param structure. In other words, results do not accumulate into the pfp_pmcs table across multiple calls. Unused fields are guaranteed to be zeroed upon successful return.

Depending on the PMU model, there may not always be a one to one mapping between a PMC register and a data register. Register dependencies may be more intricate. However the pfm_dispatch_events guarantees certain ordering between the pfp_pmcs and pfp_pmds tables. In particular, it guarantees that the pfp_pmds table always starts with the counters corresponding, in the same order, to the events as provided in the pfp_event table on input. There is always one counter per event. Additional PMD registers, if any, come after.

EXAMPLE

Here is a typical sequence using the perfmon2 interface:

        #include <perfmon/pfmlib.h>
        ...
        pfmlib_input_param_t inp;
        pfmlib_output_param_t outp;
        pfarg_ctx_t ctx;
        pfarg_pmd_t pd[1];
        pfarg_pmc_t pc[1];
        pfarg_load_t load_arg;
        int fd, i;
        int ret;
        if (pfm_initialize() != PFMLIB_SUCCESS) {
                fprintf(stderr, "can't initialize library\n");
                exit(1);
        }
        memset(&ctx,0, sizeof(ctx));
        memset(&inp,0, sizeof(inp));
        memset(&outp,0, sizeof(outp));
        memset(pd, 0, sizeof(pd));
        memset(pc, 0, sizeof(pc));
        memset(&load_arg, 0, sizeof(load_arg));
        ret = pfm_get_cycle_event(&inp.pfp_events[0]);
        if (ret != PFMLIB_SUCCESS) {
                fprintf(stderr, "cannot find cycle event\n");
                exit(1);
        }
        inp.pfp_dfl_plm     = PFM_PLM3; 
        inp.pfp_event_count = 1;
        ret = pfm_dispatch_events(&inp, NULL, &outp, NULL);
        if (ret != PFMLIB_SUCCESS) {
                fprintf(stderr, "cannot dispatch events: %s\n", pfm_strerror(ret));
                exit(1);
        }
        /* propagate pmc value to perfmon2 structures */
        for(i=0; i < outp.pfp_pmc_count; i++) {
                pc[i].reg_num   = outp.pfp_pmcs[i].reg_num;
                pc[i].reg_value = outp.pfp_pmcs[i].reg_value;
        }
        for(i=0; i < outp.pfp_pmd_count; i++) {
                pd[i].reg_num   = outp.pfp_pmds[i].reg_num;
                pd[i].reg_value = 0;
        }
        ...
        if (pfm_create_context(&ctx, NULL, 0) == -1 ) {
                ...
        }
        fd = ctx.ctx_fd;
        if (pfm_write_pmcs(fd, pc, outp.pfp_pmc_count) == -1) {
                ...
        }
        if (pfm_write_pmds(fd, pd, outp.pfp_pmd_count) == -1) {
                ...
        }
        load_arg.load_pid = getpid();
        if (pfm_load_context(fd, &load_arg) == -1) {
                ...
        }
        pfm_start(fd, NULL);
        /* code to monitor */
        pfm_stop(fd);
        if (pfm_read_pmds(fd, pd, evt.pfp_event_count) == -1) {
                ...
        }
        printf("results: %llu, pd[0].reg_value);
        ...
        close(fd);
        ...

RETURN

The function returns whether or not the call was successful. A return value of PFMLIB_SUCCESS indicates sucess, otherwise the value is the error code.

AUTHOR

Stephane Eranian <[email protected]>