perf_event_uncore.c File Reference

Include dependency graph for perf_event_uncore.c:

Go to the source code of this file.

Macros
#define	PERF_EVENTS_OPENED   0x01
#define	PERF_EVENTS_RUNNING   0x02
#define	HANDLE_STRING_ERROR   {fprintf(stderr,"%s:%i unexpected string function error.\n",__FILE__,__LINE__); exit(-1);}
#define	READ_BUFFER_SIZE   (3 + (2 * PERF_EVENT_MAX_MPX_COUNTERS))

Functions
static int	_peu_set_domain (hwd_control_state_t *ctl, int domain)
static int	_peu_shutdown_component (void)
static unsigned int	get_read_format (unsigned int multiplex, unsigned int inherit, int format_group)
static long	sys_perf_event_open (struct perf_event_attr *hw_event, pid_t pid, int cpu, int group_fd, unsigned long flags)
static int	map_perf_event_errors_to_papi (int perf_event_error)
static int	check_scheduability (pe_context_t *ctx, pe_control_t *ctl)
static int	open_pe_events (pe_context_t *ctx, pe_control_t *ctl)
static int	close_pe_events (pe_context_t *ctx, pe_control_t *ctl)
static int	_peu_init_thread (hwd_context_t *hwd_ctx)
static int	_peu_init_control_state (hwd_control_state_t *ctl)
static int	_peu_init_component (int cidx)
int	_peu_update_control_state (hwd_control_state_t *ctl, NativeInfo_t *native, int count, hwd_context_t *ctx)
static int	_peu_shutdown_thread (hwd_context_t *ctx)
static int	_peu_reset (hwd_context_t *ctx, hwd_control_state_t *ctl)
static int	_peu_write (hwd_context_t *ctx, hwd_control_state_t *ctl, long long *from)
static int	_peu_read (hwd_context_t *ctx, hwd_control_state_t *ctl, long long **events, int flags)
static int	_peu_start (hwd_context_t *ctx, hwd_control_state_t *ctl)
static int	_peu_stop (hwd_context_t *ctx, hwd_control_state_t *ctl)
static int	_peu_ctl (hwd_context_t *ctx, int code, _papi_int_option_t *option)
static int	_peu_ntv_enum_events (unsigned int *PapiEventCode, int modifier)
static int	_peu_ntv_name_to_code (const char *name, unsigned int *event_code)
static int	_peu_ntv_code_to_name (unsigned int EventCode, char *ntv_name, int len)
static int	_peu_ntv_code_to_descr (unsigned int EventCode, char *ntv_descr, int len)
static int	_peu_ntv_code_to_info (unsigned int EventCode, PAPI_event_info_t *info)

Variables
papi_vector_t	_perf_event_uncore_vector
struct native_event_table_t	uncore_native_event_table
static int	our_cidx

Macro Definition Documentation

HANDLE_STRING_ERROR

#define HANDLE_STRING_ERROR   {fprintf(stderr,"%s:%i unexpected string function error.\n",__FILE__,__LINE__); exit(-1);}

Definition at line 66 of file perf_event_uncore.c.

PERF_EVENTS_OPENED

#define PERF_EVENTS_OPENED   0x01

Definition at line 60 of file perf_event_uncore.c.

PERF_EVENTS_RUNNING

#define PERF_EVENTS_RUNNING   0x02

Definition at line 61 of file perf_event_uncore.c.

READ_BUFFER_SIZE

#define READ_BUFFER_SIZE   (3 + (2 * PERF_EVENT_MAX_MPX_COUNTERS))

Definition at line 222 of file perf_event_uncore.c.

Function Documentation

_peu_ctl()

static int _peu_ctl ( hwd_context_t *  ctx, int  code, _papi_int_option_t *  option  )

static

Definition at line 1147 of file perf_event_uncore.c.

{
   int ret;
   pe_context_t *pe_ctx = ( pe_context_t *) ctx;
   pe_control_t *pe_ctl = NULL;
 
   switch ( code ) {
      case PAPI_MULTIPLEX:
       pe_ctl = ( pe_control_t * ) ( option->multiplex.ESI->ctl_state );
 
       pe_ctl->multiplexed = 1;
       ret = _peu_update_control_state( pe_ctl, NULL,
                        pe_ctl->num_events, pe_ctx );
       if (ret != PAPI_OK) {
          pe_ctl->multiplexed = 0;
       }
       return ret;
 
      case PAPI_ATTACH:
       pe_ctl = ( pe_control_t * ) ( option->attach.ESI->ctl_state );
 
       pe_ctl->tid = option->attach.tid;
 
       /* If events have been already been added, something may */
       /* have been done to the kernel, so update */
       ret =_peu_update_control_state( pe_ctl, NULL,
                        pe_ctl->num_events, pe_ctx);
 
       return ret;
 
      case PAPI_DETACH:
       pe_ctl = ( pe_control_t *) ( option->attach.ESI->ctl_state );
 
       pe_ctl->tid = 0;
       return PAPI_OK;
 
      case PAPI_CPU_ATTACH:
       pe_ctl = ( pe_control_t *) ( option->cpu.ESI->ctl_state );
 
       /* this tells the kernel not to count for a thread   */
       /* should we warn if we try to set both?  perf_event */
       /* will reject it.                                   */
       pe_ctl->tid = -1;
 
       pe_ctl->cpu = option->cpu.cpu_num;
 
       return PAPI_OK;
 
      case PAPI_DOMAIN:
       pe_ctl = ( pe_control_t *) ( option->domain.ESI->ctl_state );
 
       /* looks like we are allowed, so set event set level counting domains */
       pe_ctl->domain = option->domain.domain;
       return PAPI_OK;
 
      case PAPI_GRANUL:
       pe_ctl = (pe_control_t *) ( option->granularity.ESI->ctl_state );
 
       /* FIXME: we really don't support this yet */
 
           switch ( option->granularity.granularity  ) {
              case PAPI_GRN_PROCG:
              case PAPI_GRN_SYS_CPU:
              case PAPI_GRN_PROC:
           return PAPI_ECMP;
 
          /* Currently we only support thread and CPU granularity */
              case PAPI_GRN_SYS:
           pe_ctl->granularity=PAPI_GRN_SYS;
           break;
 
              case PAPI_GRN_THR:
           pe_ctl->granularity=PAPI_GRN_THR;
           break;
 
 
              default:
           return PAPI_EINVAL;
       }
           return PAPI_OK;
 
      case PAPI_INHERIT:
       pe_ctl = (pe_control_t *) ( option->inherit.ESI->ctl_state );
 
       if (option->inherit.inherit) {
          /* children will inherit counters */
          pe_ctl->inherit = 1;
       } else {
          /* children won't inherit counters */
          pe_ctl->inherit = 0;
       }
       return PAPI_OK;
 
      case PAPI_DATA_ADDRESS:
       return PAPI_ENOSUPP;
 
      case PAPI_INSTR_ADDRESS:
       return PAPI_ENOSUPP;
 
      case PAPI_DEF_ITIMER:
       return PAPI_ENOSUPP;
 
      case PAPI_DEF_MPX_NS:
       return PAPI_ENOSUPP;
 
      case PAPI_DEF_ITIMER_NS:
       return PAPI_ENOSUPP;
 
      default:
       return PAPI_ENOSUPP;
   }
}

Here is the call graph for this function:

_peu_init_component()

static int _peu_init_component ( int  cidx )

static

Definition at line 593 of file perf_event_uncore.c.

{
 
   int retval;
   int paranoid_level;
 
   FILE *fff;
   char *strCpy;
 
   our_cidx=cidx;
 
   /* The is the official way to detect if perf_event support exists */
   /* The file is called perf_counter_paranoid on 2.6.31             */
   /* currently we are lazy and do not support 2.6.31 kernels        */
 
   fff=fopen("/proc/sys/kernel/perf_event_paranoid","r");
   if (fff==NULL) {
     strCpy=strncpy(_papi_hwd[cidx]->cmp_info.disabled_reason,
        "perf_event support not detected",PAPI_MAX_STR_LEN);
     if (strCpy == NULL) HANDLE_STRING_ERROR;
     retval = PAPI_ECMP;
     goto fn_fail;
   }
   retval=fscanf(fff,"%d",&paranoid_level);
   if (retval!=1) fprintf(stderr,"Error reading paranoid level\n");
   fclose(fff);
 
 
   /* Run the libpfm4-specific setup */
 
   retval = _papi_libpfm4_init(_papi_hwd[cidx]);
   if (retval) {
     strCpy=strncpy(_papi_hwd[cidx]->cmp_info.disabled_reason,
         "Error initializing libpfm4",PAPI_MAX_STR_LEN);
     _peu_shutdown_component( );
     if (strCpy == NULL) HANDLE_STRING_ERROR;
     retval = PAPI_ECMP;
     goto fn_fail;
   }
 
 
   /* Run the uncore specific libpfm4 setup */
 
   retval = _peu_libpfm4_init(_papi_hwd[cidx], our_cidx,
                   &uncore_native_event_table,
                               PMU_TYPE_UNCORE);
   if (retval) {
     strCpy=strncpy(_papi_hwd[cidx]->cmp_info.disabled_reason,
         "Error setting up libpfm4",PAPI_MAX_STR_LEN);
     _peu_shutdown_component( );
     if (strCpy == NULL) HANDLE_STRING_ERROR;
     retval = PAPI_ECMP;
     goto fn_fail;
   }
 
   /* Check if no uncore events found */
 
   if (_papi_hwd[cidx]->cmp_info.num_native_events==0) {
     strCpy=strncpy(_papi_hwd[cidx]->cmp_info.disabled_reason,
         "No uncore PMUs or events found",PAPI_MAX_STR_LEN);
     _peu_shutdown_component( );
     if (strCpy == NULL) HANDLE_STRING_ERROR;
     retval = PAPI_ECMP;
     goto fn_fail;
   }
 
   /* Check if we have enough permissions for uncore */
 
   /* 2 means no kernel measurements allowed   */
   /* 1 means normal counter access            */
   /* 0 means you can access CPU-specific data */
   /* -1 means no restrictions                 */
 
   if ((paranoid_level>0) && (getuid()!=0)) {
      strCpy=strncpy(_papi_hwd[cidx]->cmp_info.disabled_reason,
        "Insufficient permissions for uncore access.  Set /proc/sys/kernel/perf_event_paranoid to 0 or run as root.",
        PAPI_MAX_STR_LEN);
      _peu_shutdown_component( );
     if (strCpy == NULL) HANDLE_STRING_ERROR;
     retval = PAPI_ECMP;
     goto fn_fail;
   }
 
  fn_exit:
   _papi_hwd[cidx]->cmp_info.disabled = retval;
   return retval;
  fn_fail:
   goto fn_exit;
 
}

Here is the call graph for this function:

_peu_init_control_state()

static int _peu_init_control_state ( hwd_control_state_t *  ctl )

static

Definition at line 568 of file perf_event_uncore.c.

{
  pe_control_t *pe_ctl = ( pe_control_t *) ctl;
 
  /* clear the contents */
  memset( pe_ctl, 0, sizeof ( pe_control_t ) );
 
  /* Set the default domain */
  _peu_set_domain( ctl, _perf_event_uncore_vector.cmp_info.default_domain );
 
  /* Set the default granularity */
  pe_ctl->granularity=_perf_event_uncore_vector.cmp_info.default_granularity;
 
  pe_ctl->cidx=our_cidx;
 
  /* Set cpu number in the control block to show events */
  /* are not tied to specific cpu                       */
  pe_ctl->cpu = -1;
  return PAPI_OK;
}

Here is the call graph for this function:

_peu_init_thread()

static int _peu_init_thread ( hwd_context_t *  hwd_ctx )

static

Definition at line 551 of file perf_event_uncore.c.

{
 
  pe_context_t *pe_ctx = ( pe_context_t *) hwd_ctx;
 
  /* clear the context structure and mark as initialized */
  memset( pe_ctx, 0, sizeof ( pe_context_t ) );
  pe_ctx->initialized=1;
 
  pe_ctx->event_table=&uncore_native_event_table;
  pe_ctx->cidx=our_cidx;
 
  return PAPI_OK;
}

_peu_ntv_code_to_descr()

static int _peu_ntv_code_to_descr ( unsigned int  EventCode, char *  ntv_descr, int  len  )

static

Definition at line 1293 of file perf_event_uncore.c.

                                                      {
 
   if (_perf_event_uncore_vector.cmp_info.disabled) return PAPI_ENOEVNT;
 
   return _pe_libpfm4_ntv_code_to_descr(EventCode,ntv_descr,len,
                                          &uncore_native_event_table);
}

_peu_ntv_code_to_info()

static int _peu_ntv_code_to_info ( unsigned int  EventCode, PAPI_event_info_t *  info  )

static

Definition at line 1303 of file perf_event_uncore.c.

                                                   {
 
  if (_perf_event_uncore_vector.cmp_info.disabled) return PAPI_ENOEVNT;
 
  return _pe_libpfm4_ntv_code_to_info(EventCode, info,
                                        &uncore_native_event_table);
}

Here is the call graph for this function:

_peu_ntv_code_to_name()

static int _peu_ntv_code_to_name ( unsigned int  EventCode, char *  ntv_name, int  len  )

static

Definition at line 1282 of file perf_event_uncore.c.

                                                   {
 
   if (_perf_event_uncore_vector.cmp_info.disabled) return PAPI_ENOEVNT;
 
   return _pe_libpfm4_ntv_code_to_name(EventCode,
                                         ntv_name, len, 
                     &uncore_native_event_table);
}

_peu_ntv_enum_events()

static int _peu_ntv_enum_events ( unsigned int *  PapiEventCode, int  modifier  )

static

Definition at line 1262 of file perf_event_uncore.c.

{
 
  if (_perf_event_uncore_vector.cmp_info.disabled) return PAPI_ENOEVNT;
 
 
  return _pe_libpfm4_ntv_enum_events(PapiEventCode, modifier, our_cidx,
                                       &uncore_native_event_table);
}

_peu_ntv_name_to_code()

static int _peu_ntv_name_to_code ( const char *  name, unsigned int *  event_code  )

static

Definition at line 1273 of file perf_event_uncore.c.

                                                                   {
 
  if (_perf_event_uncore_vector.cmp_info.disabled) return PAPI_ENOEVNT;
 
  return _pe_libpfm4_ntv_name_to_code(name,event_code, our_cidx,
                                        &uncore_native_event_table);
}

_peu_read()

static int _peu_read ( hwd_context_t *  ctx, hwd_control_state_t *  ctl, long long **  events, int  flags  )

static

Definition at line 911 of file perf_event_uncore.c.

{
    SUBDBG("ENTER: ctx: %p, ctl: %p, events: %p, flags: %#x\n", ctx, ctl, events, flags);
 
   ( void ) flags;           /*unused */
   int i, ret = -1;
   /* pe_context_t *pe_ctx = ( pe_context_t *) ctx; */ 
   (void) ctx; /*unused*/
   pe_control_t *pe_ctl = ( pe_control_t *) ctl;
   long long papi_pe_buffer[READ_BUFFER_SIZE];
   long long tot_time_running, tot_time_enabled, scale;
 
   /* Handle case where we are multiplexing */
   if (pe_ctl->multiplexed) {
 
      /* currently we handle multiplexing by having individual events */
      /* so we read from each in turn.                                */
 
      for ( i = 0; i < pe_ctl->num_events; i++ ) {
 
         ret = read( pe_ctl->events[i].event_fd, papi_pe_buffer,
            sizeof ( papi_pe_buffer ) );
         if ( ret == -1 ) {
        PAPIERROR("read returned an error: %s", strerror( errno ));
       SUBDBG("EXIT: PAPI_ESYS\n");
        return PAPI_ESYS;
     }
 
     /* We should read 3 64-bit values from the counter */
     if (ret<(signed)(3*sizeof(long long))) {
        PAPIERROR("Error!  short read!\n");
       SUBDBG("EXIT: PAPI_ESYS\n");
        return PAPI_ESYS;
     }
 
         SUBDBG("read: fd: %2d, tid: %ld, cpu: %d, ret: %d\n",
            pe_ctl->events[i].event_fd,
        (long)pe_ctl->tid, pe_ctl->events[i].cpu, ret);
         SUBDBG("read: %lld %lld %lld\n",papi_pe_buffer[0],
            papi_pe_buffer[1],papi_pe_buffer[2]);
 
         tot_time_enabled = papi_pe_buffer[1];
         tot_time_running = papi_pe_buffer[2];
 
         SUBDBG("count[%d] = (papi_pe_buffer[%d] %lld * "
        "tot_time_enabled %lld) / tot_time_running %lld\n",
        i, 0,papi_pe_buffer[0],
        tot_time_enabled,tot_time_running);
 
         if (tot_time_running == tot_time_enabled) {
        /* No scaling needed */
        pe_ctl->counts[i] = papi_pe_buffer[0];
         } else if (tot_time_running && tot_time_enabled) {
        /* Scale factor of 100 to avoid overflows when computing */
        /*enabled/running */
 
        scale = (tot_time_enabled * 100LL) / tot_time_running;
        scale = scale * papi_pe_buffer[0];
        scale = scale / 100LL;
        pe_ctl->counts[i] = scale;
     } else {
       /* This should not happen, but Phil reports it sometime does. */
        SUBDBG("perf_event kernel bug(?) count, enabled, "
           "running: %lld, %lld, %lld\n",
           papi_pe_buffer[0],tot_time_enabled,
           tot_time_running);
 
        pe_ctl->counts[i] = papi_pe_buffer[0];
     }
      }
   }
 
   /* Handle cases where we cannot use FORMAT GROUP */
   else if (pe_ctl->inherit) {
 
      /* we must read each counter individually */
      for ( i = 0; i < pe_ctl->num_events; i++ ) {
 
         ret = read( pe_ctl->events[i].event_fd, papi_pe_buffer, 
            sizeof ( papi_pe_buffer ) );
         if ( ret == -1 ) {
        PAPIERROR("read returned an error: %s", strerror( errno ));
       SUBDBG("EXIT: PAPI_ESYS\n");
        return PAPI_ESYS;
     }
 
     /* we should read one 64-bit value from each counter */
     if (ret!=sizeof(long long)) {
        PAPIERROR("Error!  short read!\n");
        PAPIERROR("read: fd: %2d, tid: %ld, cpu: %d, ret: %d\n",
           pe_ctl->events[i].event_fd,
           (long)pe_ctl->tid, pe_ctl->events[i].cpu, ret);
       SUBDBG("EXIT: PAPI_ESYS\n");
        return PAPI_ESYS;
     }
 
         SUBDBG("read: fd: %2d, tid: %ld, cpu: %d, ret: %d\n",
            pe_ctl->events[i].event_fd, (long)pe_ctl->tid,
        pe_ctl->events[i].cpu, ret);
         SUBDBG("read: %lld\n",papi_pe_buffer[0]);
 
     pe_ctl->counts[i] = papi_pe_buffer[0];
      }
   }
 
 
   /* Handle cases where we are using FORMAT_GROUP   */
   /* We assume only one group leader, in position 0 */
 
   else {
      if (pe_ctl->events[0].group_leader_fd!=-1) {
     PAPIERROR("Was expecting group leader!\n");
      }
 
      ret = read( pe_ctl->events[0].event_fd, papi_pe_buffer,
          sizeof ( papi_pe_buffer ) );
 
      if ( ret == -1 ) {
     PAPIERROR("read returned an error: %s", strerror( errno ));
       SUBDBG("EXIT: PAPI_ESYS\n");
     return PAPI_ESYS;
      }
 
      /* we read 1 64-bit value (number of events) then     */
      /* num_events more 64-bit values that hold the counts */
      if (ret<(signed)((1+pe_ctl->num_events)*sizeof(long long))) {
     PAPIERROR("Error! short read!\n");
       SUBDBG("EXIT: PAPI_ESYS\n");
     return PAPI_ESYS;
      }
 
      SUBDBG("read: fd: %2d, tid: %ld, cpu: %d, ret: %d\n",
         pe_ctl->events[0].event_fd,
         (long)pe_ctl->tid, pe_ctl->events[0].cpu, ret);
      {
     int j;
     for(j=0;j<ret/8;j++) {
            SUBDBG("read %d: %lld\n",j,papi_pe_buffer[j]);
     }
      }
 
      /* Make sure the kernel agrees with how many events we have */
      if (papi_pe_buffer[0]!=pe_ctl->num_events) {
     PAPIERROR("Error!  Wrong number of events!\n");
       SUBDBG("EXIT: PAPI_ESYS\n");
     return PAPI_ESYS;
      }
 
      /* put the count values in their proper location */
      for(i=0;i<pe_ctl->num_events;i++) {
         pe_ctl->counts[i] = papi_pe_buffer[1+i];
      }
   }
 
   /* point PAPI to the values we read */
   *events = pe_ctl->counts;
 
   SUBDBG("EXIT: PAPI_OK\n");
   return PAPI_OK;
}

Here is the call graph for this function:

_peu_reset()

static int _peu_reset ( hwd_context_t *  ctx, hwd_control_state_t *  ctl  )

static

Definition at line 856 of file perf_event_uncore.c.

{
   int i, ret;
   pe_control_t *pe_ctl = ( pe_control_t *) ctl;
 
   ( void ) ctx;             /*unused */
 
   /* We need to reset all of the events, not just the group leaders */
   for( i = 0; i < pe_ctl->num_events; i++ ) {
      ret = ioctl( pe_ctl->events[i].event_fd, PERF_EVENT_IOC_RESET, NULL );
      if ( ret == -1 ) {
     PAPIERROR("ioctl(%d, PERF_EVENT_IOC_RESET, NULL) "
           "returned error, Linux says: %s",
           pe_ctl->events[i].event_fd, strerror( errno ) );
     return PAPI_ESYS;
      }
   }
 
   return PAPI_OK;
}

Here is the call graph for this function:

Here is the caller graph for this function:

_peu_set_domain()

static int _peu_set_domain ( hwd_control_state_t *  ctl, int  domain  )

static

Definition at line 829 of file perf_event_uncore.c.

{
   pe_control_t *pe_ctl = ( pe_control_t *) ctl;
 
   SUBDBG("old control domain %d, new domain %d\n",
      pe_ctl->domain,domain);
 
   pe_ctl->domain = domain;
   return PAPI_OK;
}

Here is the caller graph for this function:

_peu_shutdown_component()

static int _peu_shutdown_component ( void  )

static

Definition at line 686 of file perf_event_uncore.c.

                                {
 
    /* deallocate our event table */
    _pe_libpfm4_shutdown(&_perf_event_uncore_vector,
                &uncore_native_event_table);
 
    /* Shutdown libpfm4 */
    _papi_libpfm4_shutdown(&_perf_event_uncore_vector);
 
    return PAPI_OK;
}

Here is the caller graph for this function:

_peu_shutdown_thread()

static int _peu_shutdown_thread ( hwd_context_t *  ctx )

static

Definition at line 842 of file perf_event_uncore.c.

{
    pe_context_t *pe_ctx = ( pe_context_t *) ctx;
 
    pe_ctx->initialized=0;
 
    return PAPI_OK;
}

_peu_start()

static int _peu_start ( hwd_context_t *  ctx, hwd_control_state_t *  ctl  )

static

Definition at line 1075 of file perf_event_uncore.c.

{
   int ret;
   int i;
   int did_something = 0;
   pe_context_t *pe_ctx = ( pe_context_t *) ctx;
   pe_control_t *pe_ctl = ( pe_control_t *) ctl;
 
   /* Reset the counters first.  Is this necessary? */
   ret = _peu_reset( pe_ctx, pe_ctl );
   if ( ret ) {
      return ret;
   }
 
   /* Enable all of the group leaders                */
   /* All group leaders have a group_leader_fd of -1 */
   for( i = 0; i < pe_ctl->num_events; i++ ) {
      if (pe_ctl->events[i].group_leader_fd == -1) {
     SUBDBG("ioctl(enable): fd: %d\n", pe_ctl->events[i].event_fd);
     ret=ioctl( pe_ctl->events[i].event_fd, PERF_EVENT_IOC_ENABLE, NULL) ; 
 
     /* ioctls always return -1 on failure */
         if (ret == -1) {
            PAPIERROR("ioctl(PERF_EVENT_IOC_ENABLE) failed.\n");
            return PAPI_ESYS;
     }
 
     did_something++;
      } 
   }
 
   if (!did_something) {
      PAPIERROR("Did not enable any counters.\n");
      return PAPI_EBUG;
   }
 
   pe_ctx->state |= PERF_EVENTS_RUNNING;
 
   return PAPI_OK;
 
}

Here is the call graph for this function:

_peu_stop()

static int _peu_stop ( hwd_context_t *  ctx, hwd_control_state_t *  ctl  )

static

Definition at line 1119 of file perf_event_uncore.c.

{
 
   int ret;
   int i;
   pe_context_t *pe_ctx = ( pe_context_t *) ctx;
   pe_control_t *pe_ctl = ( pe_control_t *) ctl;
 
   /* Just disable the group leaders */
   for ( i = 0; i < pe_ctl->num_events; i++ ) {
      if ( pe_ctl->events[i].group_leader_fd == -1 ) {
     ret=ioctl( pe_ctl->events[i].event_fd, PERF_EVENT_IOC_DISABLE, NULL);
     if ( ret == -1 ) {
        PAPIERROR( "ioctl(%d, PERF_EVENT_IOC_DISABLE, NULL) "
               "returned error, Linux says: %s",
               pe_ctl->events[i].event_fd, strerror( errno ) );
        return PAPI_EBUG;
     }
      }
   }
 
   pe_ctx->state &= ~PERF_EVENTS_RUNNING;
 
   return PAPI_OK;
}

Here is the call graph for this function:

_peu_update_control_state()

int _peu_update_control_state	(	hwd_control_state_t *	ctl,
		NativeInfo_t *	native,
		int	count,
		hwd_context_t *	ctx
	)

Definition at line 703 of file perf_event_uncore.c.

{
    int i;
    int j;
    int ret;
    int skipped_events=0;
    struct native_event_t *ntv_evt;
   pe_context_t *pe_ctx = ( pe_context_t *) ctx;
   pe_control_t *pe_ctl = ( pe_control_t *) ctl;
 
   /* close all of the existing fds and start over again */
   /* In theory we could have finer-grained control and know if             */
   /* things were changed, but it's easier to tear things down and rebuild. */
   close_pe_events( pe_ctx, pe_ctl );
 
   /* Calling with count==0 should be OK, it's how things are deallocated */
   /* when an eventset is destroyed.                                      */
   if ( count == 0 ) {
      SUBDBG( "Called with count == 0\n" );
      return PAPI_OK;
   }
 
   /* set up all the events */
   for( i = 0; i < count; i++ ) {
      if ( native ) {
            // get the native event pointer used for this papi event
            int ntv_idx = _papi_hwi_get_ntv_idx((unsigned)(native[i].ni_papi_code));
            if (ntv_idx < -1) {
                SUBDBG("papi_event_code: %#x known by papi but not by the component\n", native[i].ni_papi_code);
                continue;
            }
            // if native index is -1, then we have an event without a mask and need to find the right native index to use
            if (ntv_idx == -1) {
                // find the native event index we want by matching for the right papi event code
                for (j=0 ; j<pe_ctx->event_table->num_native_events ; j++) {
                    if (pe_ctx->event_table->native_events[j].papi_event_code == native[i].ni_papi_code) {
                        ntv_idx = j;
                    }
                }
            }
 
            // if native index is still negative, we did not find event we wanted so just return error
            if (ntv_idx < 0) {
                SUBDBG("papi_event_code: %#x not found in native event tables\n", native[i].ni_papi_code);
                continue;
            }
 
            // this native index is positive so there was a mask with the event, the ntv_idx identifies which native event to use
            ntv_evt = (struct native_event_t *)(&(pe_ctx->event_table->native_events[ntv_idx]));
 
            SUBDBG("ntv_evt: %p\n", ntv_evt);
 
            SUBDBG("i: %d, pe_ctx->event_table->num_native_events: %d\n", i, pe_ctx->event_table->num_native_events);
 
            // Move this events hardware config values and other attributes to the perf_events attribute structure
            memcpy (&pe_ctl->events[i].attr, &ntv_evt->attr, sizeof(perf_event_attr_t));
 
            // may need to update the attribute structure with information from event set level domain settings (values set by PAPI_set_domain)
            // only done if the event mask which controls each counting domain was not provided
 
            // get pointer to allocated name, will be NULL when adding preset events to event set
            char *aName = ntv_evt->allocated_name;
            if ((aName == NULL)  ||  (strstr(aName, ":u=") == NULL)) {
                SUBDBG("set exclude_user attribute from eventset level domain flags, encode: %d, eventset: %d\n", pe_ctl->events[i].attr.exclude_user, !(pe_ctl->domain & PAPI_DOM_USER));
                pe_ctl->events[i].attr.exclude_user = !(pe_ctl->domain & PAPI_DOM_USER);
            }
            if ((aName == NULL)  ||  (strstr(aName, ":k=") == NULL)) {
                SUBDBG("set exclude_kernel attribute from eventset level domain flags, encode: %d, eventset: %d\n", pe_ctl->events[i].attr.exclude_kernel, !(pe_ctl->domain & PAPI_DOM_KERNEL));
                pe_ctl->events[i].attr.exclude_kernel = !(pe_ctl->domain & PAPI_DOM_KERNEL);
            }
 
            // set the cpu number provided with an event mask if there was one (will be -1 if mask not provided)
            pe_ctl->events[i].cpu = ntv_evt->cpu;
            // if cpu event mask not provided, then set the cpu to use to what may have been set on call to PAPI_set_opt (will still be -1 if not called)
            if (pe_ctl->events[i].cpu == -1) {
                pe_ctl->events[i].cpu = pe_ctl->cpu;
            }
      } else {
          // This case happens when called from _pe_set_overflow and _pe_ctl
          // Those callers put things directly into the pe_ctl structure so it is already set for the open call
      }
 
      // Copy the inherit flag into the attribute block that will be passed to the kernel
      pe_ctl->events[i].attr.inherit = pe_ctl->inherit;
 
      /* Set the position in the native structure */
      /* We just set up events linearly           */
      if ( native ) {
          native[i].ni_position = i;
          SUBDBG( "&native[%d]: %p, ni_papi_code: %#x, ni_event: %#x, ni_position: %d, ni_owners: %d\n",
            i, &(native[i]), native[i].ni_papi_code, native[i].ni_event, native[i].ni_position, native[i].ni_owners);
      }
   }
 
    if (count <= skipped_events) {
        SUBDBG("EXIT: No events to count, they all contained invalid umasks\n");
        return PAPI_ENOEVNT;
    }
 
  pe_ctl->num_events = count - skipped_events;
 
   /* actuall open the events */
   /* (why is this a separate function?) */
   ret = open_pe_events( pe_ctx, pe_ctl );
   if ( ret != PAPI_OK ) {
      SUBDBG("open_pe_events failed\n");
      /* Restore values ? */
      return ret;
   }
 
   SUBDBG( "EXIT: PAPI_OK\n" );
   return PAPI_OK;
}

Here is the call graph for this function:

Here is the caller graph for this function:

_peu_write()

static int _peu_write ( hwd_context_t *  ctx, hwd_control_state_t *  ctl, long long *  from  )

static

Definition at line 881 of file perf_event_uncore.c.

{
   ( void ) ctx;             /*unused */
   ( void ) ctl;             /*unused */
   ( void ) from;            /*unused */
   /*
    * Counters cannot be written.  Do we need to virtualize the
    * counters so that they can be written, or perhaps modify code so that
    * they can be written? FIXME ?
    */
 
    return PAPI_ENOSUPP;
}

check_scheduability()

static int check_scheduability ( pe_context_t *  ctx, pe_control_t *  ctl  )

static

Definition at line 232 of file perf_event_uncore.c.

{
   SUBDBG("ENTER: ctx: %p, ctl: %p\n", ctx, ctl);
   int retval = 0, cnt = -1;
   ( void ) ctx;         /*unused */
   long long papi_pe_buffer[READ_BUFFER_SIZE];
   int i;
 
   /* If the kernel isn't tracking scheduability right       */
   /* Then we need to start/stop/read to force the event     */
   /* to be scheduled and see if an error condition happens. */
 
   /* start all events */
   for( i = 0; i < ctl->num_events; i++) {
      retval = ioctl( ctl->events[i].event_fd, PERF_EVENT_IOC_ENABLE, NULL );
      if (retval == -1) {
     SUBDBG("EXIT: Enable failed event index: %d, num_events: %d, return PAPI_ESYS\n", i, ctl->num_events);
     return PAPI_ESYS;
      }
   }
 
   /* stop all events */
   for( i = 0; i < ctl->num_events; i++) {
      retval = ioctl(ctl->events[i].event_fd, PERF_EVENT_IOC_DISABLE, NULL );
      if (retval == -1) {
     SUBDBG("EXIT: Disable failed: event index: %d, num_events: %d, return PAPI_ESYS\n", i, ctl->num_events);
     return PAPI_ESYS;
      }
   }
 
   /* See if a read of each event returns results */
   for( i = 0; i < ctl->num_events; i++) {
      cnt = read( ctl->events[i].event_fd, papi_pe_buffer, sizeof(papi_pe_buffer));
      if ( cnt == -1 ) {
     SUBDBG( "EXIT: read failed: event index: %d, num_events: %d, return PAPI_ESYS.  Should never happen.\n", i, ctl->num_events);
     return PAPI_ESYS;
      }
 
      if ( cnt == 0 ) {
     /* We read 0 bytes if we could not schedule the event */
     /* The kernel should have detected this at open       */
     /* but various bugs (including NMI watchdog)          */
     /* result in this behavior                            */
 
     SUBDBG( "EXIT: read returned 0: event index: %d, num_events: %d, return PAPI_ECNFLCT.\n", i, ctl->num_events);
     return PAPI_ECNFLCT;
      }
   }
 
   /* Reset all of the counters (opened so far) back to zero      */
   /* from the above brief enable/disable call pair.              */
 
   /* We have to reset all events because reset of group leader      */
   /* does not reset all.                                            */
   /* we assume that the events are being added one by one and that  */
   /* we do not need to reset higher events (doing so may reset ones */
   /* that have not been initialized yet.                            */
 
   /* Note... PERF_EVENT_IOC_RESET does not reset time running       */
   /* info if multiplexing, so we should avoid coming here if        */
   /* we are multiplexing the event.                                 */
   for( i = 0; i < ctl->num_events; i++) {
      retval=ioctl( ctl->events[i].event_fd, PERF_EVENT_IOC_RESET, NULL );
      if (retval == -1) {
     SUBDBG("EXIT: Reset failed: event index: %d, num_events: %d, return PAPI_ESYS\n", i, ctl->num_events);
     return PAPI_ESYS;
      }
   }
   SUBDBG("EXIT: return PAPI_OK\n");
   return PAPI_OK;
}

Here is the call graph for this function:

Here is the caller graph for this function:

close_pe_events()

static int close_pe_events ( pe_context_t *  ctx, pe_control_t *  ctl  )

static

Definition at line 454 of file perf_event_uncore.c.

{
   int i;
   int num_closed=0;
   int events_not_opened=0;
 
   /* should this be a more serious error? */
   if ( ctx->state & PERF_EVENTS_RUNNING ) {
      SUBDBG("Closing without stopping first\n");
   }
 
   /* Close child events first */
   for( i=0; i<ctl->num_events; i++ ) {
 
      if (ctl->events[i].event_opened) {
 
         if (ctl->events[i].group_leader_fd!=-1) {
            if ( ctl->events[i].mmap_buf ) {
           if ( munmap ( ctl->events[i].mmap_buf,
                     ctl->events[i].nr_mmap_pages * getpagesize() ) ) {
              PAPIERROR( "munmap of fd = %d returned error: %s",
                 ctl->events[i].event_fd, strerror( errno ) );
              return PAPI_ESYS;
           }
        }
 
            if ( close( ctl->events[i].event_fd ) ) {
           PAPIERROR( "close of fd = %d returned error: %s",
               ctl->events[i].event_fd, strerror( errno ) );
           return PAPI_ESYS;
        } else {
           num_closed++;
        }
        ctl->events[i].event_opened=0;
     }
      }
      else {
    events_not_opened++;
      }
   }
 
   /* Close the group leaders last */
   for( i=0; i<ctl->num_events; i++ ) {
 
      if (ctl->events[i].event_opened) {
 
         if (ctl->events[i].group_leader_fd==-1) {
            if ( ctl->events[i].mmap_buf ) {
           if ( munmap ( ctl->events[i].mmap_buf,
                     ctl->events[i].nr_mmap_pages * getpagesize() ) ) {
              PAPIERROR( "munmap of fd = %d returned error: %s",
                 ctl->events[i].event_fd, strerror( errno ) );
              return PAPI_ESYS;
           }
        }
 
 
            if ( close( ctl->events[i].event_fd ) ) {
           PAPIERROR( "close of fd = %d returned error: %s",
               ctl->events[i].event_fd, strerror( errno ) );
           return PAPI_ESYS;
        } else {
           num_closed++;
        }
        ctl->events[i].event_opened=0;
     }
      }
   }
 
 
   if (ctl->num_events!=num_closed) {
      if (ctl->num_events!=(num_closed+events_not_opened)) {
         PAPIERROR("Didn't close all events: "
           "Closed %d Not Opened: %d Expected %d\n",
           num_closed,events_not_opened,ctl->num_events);
         return PAPI_EBUG;
      }
   }
 
   ctl->num_events=0;
 
   ctx->state &= ~PERF_EVENTS_OPENED;
 
   return PAPI_OK;
}

Here is the call graph for this function:

Here is the caller graph for this function:

get_read_format()

static unsigned int get_read_format ( unsigned int  multiplex, unsigned int  inherit, int  format_group  )

static

Definition at line 81 of file perf_event_uncore.c.

{
   unsigned int format = 0;
 
   /* if we need read format options for multiplexing, add them now */
   if (multiplex) {
      format |= PERF_FORMAT_TOTAL_TIME_ENABLED;
      format |= PERF_FORMAT_TOTAL_TIME_RUNNING;
   }
 
   /* If we are not using inherit, add the group read options     */
   if (!inherit) {
      if (format_group) {
     format |= PERF_FORMAT_GROUP;
      }
   }
 
   SUBDBG("multiplex: %d, inherit: %d, group_leader: %d, format: %#x\n",
      multiplex, inherit, format_group, format);
 
   return format;
}

Here is the call graph for this function:

Here is the caller graph for this function:

map_perf_event_errors_to_papi()

static int map_perf_event_errors_to_papi ( int  perf_event_error )

static

Definition at line 181 of file perf_event_uncore.c.

                                                               {
 
   int ret;
 
   /* These mappings are approximate.
      EINVAL in particular can mean lots of different things */
   switch(perf_event_error) {
      case EPERM:
      case EACCES:
           ret = PAPI_EPERM;
       break;
      case ENODEV:
      case EOPNOTSUPP:
       ret = PAPI_ENOSUPP;
           break;
      case ENOENT:
       ret = PAPI_ENOEVNT;
           break;
      case ENOSYS:
      case EAGAIN:
      case EBUSY:
      case E2BIG:
       ret = PAPI_ESYS;
       break;
      case ENOMEM:
       ret = PAPI_ENOMEM;
       break;
      case EINVAL:
      default:
       ret = PAPI_EINVAL;
           break;
   }
   return ret;
}

Here is the caller graph for this function:

open_pe_events()

static int open_pe_events ( pe_context_t *  ctx, pe_control_t *  ctl  )

static

Definition at line 307 of file perf_event_uncore.c.

{
 
   int i, ret = PAPI_OK;
   long pid;
 
   if (ctl->granularity==PAPI_GRN_SYS) {
      pid = -1;
   }
   else {
      pid = ctl->tid;
   }
 
   for( i = 0; i < ctl->num_events; i++ ) {
 
      ctl->events[i].event_opened=0;
 
      /* set up the attr structure.  We don't set up all fields here */
      /* as some have already been set up previously.                */
 
/*
 * The following code controls how the uncore component interfaces with the 
 * kernel for uncore events.  The code inside the ifdef will use grouping of 
 * uncore events which can make the cost of reading the results more efficient.
 * The problem with it is that the uncore component supports 20 different uncore 
 * PMU's.  The kernel requires that all events in a group must be for the same PMU.
 * This means that with grouping enabled papi applications can count events on only
 * one of the 20 PMU's during a run.
 * 
 * The code inside the else clause treats each event in the event set as 
 * independent.  When running in this mode the kernel allows the papi multiple 
 * uncore PMU's at the same time.
 * 
 * Example:
 *  An application wants to measure all the L3 cache write requests.
 *  The event to do this is part of a cbox pmu (there are 8 cbox pmu's).
 *  When built with the code in the ifdef, the application would have to be 
 *    run 8 times and count write requests from one pmu at a time.
 *  When built with the code in the else, the write requests in all 8 cbox 
 *    pmu's could be counted in the same run.
 * 
 */
// #define GROUPIT 1       // remove the comment on this line to force event grouping
#ifdef GROUPIT
      /* group leader (event 0) is special                */
      /* If we're multiplexed, everyone is a group leader */
      if (( i == 0 ) || (ctl->multiplexed)) {
         ctl->events[i].attr.pinned = !ctl->multiplexed;
     ctl->events[i].attr.disabled = 1;
     ctl->events[i].group_leader_fd=-1;
         ctl->events[i].attr.read_format = get_read_format(ctl->multiplexed,
                               ctl->inherit,
                               !ctl->multiplexed );
      } else {
     ctl->events[i].attr.pinned=0;
     ctl->events[i].attr.disabled = 0;
     ctl->events[i].group_leader_fd=ctl->events[0].event_fd,
         ctl->events[i].attr.read_format = get_read_format(ctl->multiplexed,
                               ctl->inherit,
                               0 );
      }
#else
             ctl->events[i].attr.pinned = !ctl->multiplexed;
             ctl->events[i].attr.disabled = 1;
             ctl->inherit = 1;
             ctl->events[i].group_leader_fd=-1;
             ctl->events[i].attr.read_format = get_read_format(ctl->multiplexed, ctl->inherit, 0 );
#endif
 
 
      /* try to open */
      ctl->events[i].event_fd = sys_perf_event_open( &ctl->events[i].attr,
                             pid,
                             ctl->events[i].cpu,
                   ctl->events[i].group_leader_fd,
                             0 /* flags */
                             );
 
      /* Try to match Linux errors to PAPI errors */
      if ( ctl->events[i].event_fd == -1 ) {
     SUBDBG("sys_perf_event_open returned error on event #%d."
        "  Error: %s\n",
        i, strerror( errno ) );
         ret=map_perf_event_errors_to_papi(errno);
 
    goto open_peu_cleanup;
      }
 
      SUBDBG ("sys_perf_event_open: tid: %ld, cpu_num: %d,"
              " group_leader/fd: %d, event_fd: %d,"
              " read_format: %"PRIu64"\n",
          pid, ctl->events[i].cpu, ctl->events[i].group_leader_fd,
          ctl->events[i].event_fd, ctl->events[i].attr.read_format);
 
      ctl->events[i].event_opened=1;
   }
 
 
   /* in many situations the kernel will indicate we opened fine */
   /* yet things will fail later.  So we need to double check    */
   /* we actually can use the events we've set up.               */
 
   /* This is not necessary if we are multiplexing, and in fact */
   /* we cannot do this properly if multiplexed because         */
   /* PERF_EVENT_IOC_RESET does not reset the time running info */
   if (!ctl->multiplexed) {
    ret = check_scheduability( ctx, ctl);
 
    if ( ret != PAPI_OK ) {
        /* the last event did open, so we need to bump the counter */
        /* before doing the cleanup                                */
        i++;
        goto open_peu_cleanup;
    }
   }
 
   /* Now that we've successfully opened all of the events, do whatever  */
   /* "tune-up" is needed to attach the mmap'd buffers, signal handlers, */
   /* and so on.                                                         */
   for ( i = 0; i < ctl->num_events; i++ ) {
 
      /* No sampling if uncore */
      ctl->events[i].mmap_buf = NULL;
   }
 
   /* Set num_evts only if completely successful */
   ctx->state |= PERF_EVENTS_OPENED;
 
   return PAPI_OK;
 
open_peu_cleanup:
   /* We encountered an error, close up the fds we successfully opened.  */
   /* We go backward in an attempt to close group leaders last, although */
   /* That's probably not strictly necessary.                            */
   while ( i > 0 ) {
      i--;
      if (ctl->events[i].event_fd>=0) {
     close( ctl->events[i].event_fd );
     ctl->events[i].event_opened=0;
      }
   }
 
   return ret;
}

Here is the call graph for this function:

Here is the caller graph for this function:

sys_perf_event_open()

static long sys_perf_event_open ( struct perf_event_attr *  hw_event, pid_t  pid, int  cpu, int  group_fd, unsigned long  flags  )

static

Definition at line 130 of file perf_event_uncore.c.

{
   int ret;
 
   SUBDBG("sys_perf_event_open(hw_event: %p, pid: %d, cpu: %d, group_fd: %d, flags: %lx\n",hw_event,pid,cpu,group_fd,flags);
   SUBDBG("   type: %d\n",hw_event->type);
   SUBDBG("   size: %d\n",hw_event->size);
   SUBDBG("   config: %#"PRIx64" (%"PRIu64")\n",hw_event->config,
      hw_event->config);
   SUBDBG("   sample_period: %"PRIu64"\n",hw_event->sample_period);
   SUBDBG("   sample_type: %"PRIu64"\n",hw_event->sample_type);
   SUBDBG("   read_format: %"PRIu64"\n",hw_event->read_format);
   SUBDBG("   disabled: %d\n",hw_event->disabled);
   SUBDBG("   inherit: %d\n",hw_event->inherit);
   SUBDBG("   pinned: %d\n",hw_event->pinned);
   SUBDBG("   exclusive: %d\n",hw_event->exclusive);
   SUBDBG("   exclude_user: %d\n",hw_event->exclude_user);
   SUBDBG("   exclude_kernel: %d\n",hw_event->exclude_kernel);
   SUBDBG("   exclude_hv: %d\n",hw_event->exclude_hv);
   SUBDBG("   exclude_idle: %d\n",hw_event->exclude_idle);
   SUBDBG("   mmap: %d\n",hw_event->mmap);
   SUBDBG("   comm: %d\n",hw_event->comm);
   SUBDBG("   freq: %d\n",hw_event->freq);
   SUBDBG("   inherit_stat: %d\n",hw_event->inherit_stat);
   SUBDBG("   enable_on_exec: %d\n",hw_event->enable_on_exec);
   SUBDBG("   task: %d\n",hw_event->task);
   SUBDBG("   watermark: %d\n",hw_event->watermark);
   SUBDBG("   precise_ip: %d\n",hw_event->precise_ip);
   SUBDBG("   mmap_data: %d\n",hw_event->mmap_data);
   SUBDBG("   sample_id_all: %d\n",hw_event->sample_id_all);
   SUBDBG("   exclude_host: %d\n",hw_event->exclude_host);
   SUBDBG("   exclude_guest: %d\n",hw_event->exclude_guest);
   SUBDBG("   exclude_callchain_kernel: %d\n",hw_event->exclude_callchain_kernel);
   SUBDBG("   exclude_callchain_user: %d\n",hw_event->exclude_callchain_user);
   SUBDBG("   wakeup_watermark: %d\n",hw_event->wakeup_watermark);
   SUBDBG("   bp_type: %d\n",hw_event->bp_type);
   SUBDBG("   config1: %#lx (%lu)\n",hw_event->config1,hw_event->config1);
   SUBDBG("   config2: %#lx (%lu)\n",hw_event->config2,hw_event->config2);
   SUBDBG("   branch_sample_type: %lu\n",hw_event->branch_sample_type);
   SUBDBG("   sample_regs_user: %lu\n",hw_event->sample_regs_user);
   SUBDBG("   sample_stack_user: %d\n",hw_event->sample_stack_user);
 
    ret = syscall( __NR_perf_event_open, hw_event, pid, cpu, group_fd, flags );
    SUBDBG("Returned %d %d %s\n",ret,
           ret<0?errno:0,
           ret<0?strerror(errno):" ");
    return ret;
}

Here is the caller graph for this function:

Variable Documentation

_perf_event_uncore_vector

papi_vector_t _perf_event_uncore_vector

Definition at line 49 of file perf_event_uncore.c.

our_cidx

int our_cidx

static

Definition at line 53 of file perf_event_uncore.c.

uncore_native_event_table

struct native_event_table_t uncore_native_event_table

Definition at line 52 of file perf_event_uncore.c.