1LIBPFM(3)                  Linux Programmer's Manual                 LIBPFM(3)
2
3
4

NAME

6       pfm_dispatch_events  -  determine  PMC  registers  values  for a set of
7       events to measure
8

SYNOPSIS

10       #include <perfmon/pfmlib.h>
11
12       int pfm_dispatch_events(pfmlib_input_param_t *p, void *mod_in, pfmlib_output_param_t *q,void *mod_out);
13
14

DESCRIPTION

16       This function is the central piece of the library. It is  important  to
17       understand that the library does not effectively program the PMU, i.e.,
18       it does not make the operating system calls. The PMU is never  actually
19       accessed  by  the library. Instead, the library helps applications pre‐
20       pare the arguments to pass to the kernel. In particular, it sets up the
21       values  to program into the PMU configuration registers (PMC). The list
22       of used data registers (PMD) is also returned.
23
24       The input argument are divided into two categories: the  generic  argu‐
25       ments  in  p  and  the optional PMU model specific arguments in mod_in.
26       The same applies for the output arguments: q contains the generic  out‐
27       put arguments and mod_out the optional PMU model specific arguments.
28
29       An  application  describes what it wants to measure in the in and if it
30       uses some model specific features, such as opcode matching on Itanium 2
31       processors, it must pass a pointer to the relevant model-specific input
32       parameters in mod_in. The generic output parameters contains the regis‐
33       ter  index  and values for the PMC and PMD registers needed to make the
34       measurement. The index mapping is guaranteed to match the mapping  used
35       by  the  Linux  perfmon2  interface. In case the library is not used on
36       this system, the hardware register addresses or  indexes  can  also  be
37       retrieved from the output structure.
38
39       The pfmlib_input_param_t structure is defined as follows:
40
41       typedef struct
42            int           event;
43            unsigned int  plm;
44            unsigned long flags;
45            unsigned int  unit_masks[PFMLIB_MAX_MASKS_PER_EVENT];
46            unsigned int  num_masks;
47       } pfmlib_event_t;
48
49       typedef struct {
50            unsigned int     pfp_event_count;
51            unsigned int     pfp_dfl_plm;
52            unsigned int     pfp_flags;
53            pfmlib_event_t   pfp_events[PFMLIB_MAX_PMCS];
54            pfmlib_regmask_t pfp_unavail_pmcs;
55       } pfmlib_input_param_t;
56
57       The  structure  mostly  contains  one  table,  called  pfp_events which
58       describes the events to be measured. The number of submitted events  is
59       indicated by pfp_event_count.
60
61       Each event is described in the pfp_events table by an opaque descriptor
62       stored in the  event  field.  This  descriptor  is  obtained  with  the
63       pfm_find_full_event() or derivative functions.  For some events, it may
64       be necessary to specify at least one unit mask in the unit_masks table.
65       A   unit   mask   is   yet   another  opaque  descriptor  obtained  via
66       pfm_find_event_mask() or pfm_find_full_event. Typically,  if  an  event
67       supports  multiple  unit masks, they can be combined in which case more
68       than one entry in unit_masks must be specified. The  actual  number  of
69       unit  mask  descriptors  passed must be indicated in num_masks. When no
70       unit mask is used, this field must be set to 0.
71
72       A privilege level mask for the event can be provided in plm.  This is a
73       bitmask where each bit indicates a privilege level at which to monitor,
74       more than one bit can be set.  The library supports up to four  levels,
75       but  depending  on the PMU model, some levels may not be available. The
76       levels are as follows:
77
78       PFM_PLM0
79              monitor at the privilege level 0. For many  architectures,  this
80              means kernel level
81
82       PFM_PLM1
83              monitor at privilege level 1
84
85       PFM_PLM2
86              monitor at privilege level 2
87
88       PFM_PLM3
89              monitor  at  the privilege level 3. For many architectures, this
90              means user level
91
92       Events with a plm value of 0 will use the default privilege level  mask
93       as  indicated  by  pfp_dfl_plm which must be set to any combinations of
94       values described above. It is illegal to have a value  of  0  for  this
95       field.
96
97       The  pfp_flags  field contains a set of flags that affect the whole set
98       of events to be monitored. The currently defined flags are:
99
100       PFMLIB_PFP_SYSTEMWIDE
101              indicates that the monitors are to be used in a system-wide mon‐
102              itoring  session.  This could influence the way the library sets
103              up some register values.
104
105
106       The pfp_unavail_pmcs bitmask can be used by applications to communicate
107       to the library the list of PMC registers which are not available on the
108       system.  Some kernels may allocate certain PMC registers  (and  associ‐
109       ated  data  registers)  for other purposes. Those registers must not be
110       used by the library otherwise the assignement of events to  PMC  regis‐
111       ters  may be rejected by the kernel. Applications must figure out which
112       registers are available using a kernel interface at their disposal, the
113       library  does not provide this service. The library expect the restric‐
114       tions to be expressed using the Linux perfmon2 PMC register mapping.
115
116       Refer to the PMU specific manual for a description  of  the  model-spe‐
117       cific input parameters to be passed in mod_in.
118
119       The  generic  output  parameters  are  contained  in  the fBpfmlib_out‐
120       put_param_t structure which is defined as:
121
122       typedef struct {
123           unsigned long long reg_value;
124           unsigned int       reg_num;
125           unsigned long      reg_addr;
126       } pfmlib_reg_t;
127
128       typedef struct {
129           unsigned int pfp_pmc_count;
130           unsigned int pfp_pmd_count;
131           pfmlib_reg_t pfp_pmcs[PFMLIB_MAX_PMCS];
132           pfmlib_reg_t pfp_pmds[PFMLIB_MAX_PMDS];
133       } pfmlib_output_param_t;
134
135       The number of valid entries  in the  pfp_pmcs  table  is  indicated  by
136       pfp_pmc_count.   The  number of valid entries  in the pfp_pmds table is
137       indicated by pfp_pmd_count.  Each entry in both tables is of type  pfm‐
138       lib_reg_t.
139
140       In  the  pfp_pmcs  table,  the  reg_num contains the PMC register index
141       (perfmon2 mapping), and the reg_value contains a  64-bit  value  to  be
142       used  to program the PMC register.  The reg_addr indicates the hardware
143       address or index for the PMC register.
144
145       In the pfp_pmds table, the reg_num  contains  the  PMD  register  index
146       (perfmon2  mapping).  the  reg_value is ignored. The reg_addr indicates
147       the hardware address or index for the PMC register.
148
149       Refer to the PMU specific manual for a description  of  the  model-spe‐
150       cific output parameters to be returned in mod_out.
151
152       The  current implementation of the pfm_dispatch_events completely over‐
153       writes the pfmlib_output_param structure. In other  words,  results  do
154       not  accumulate  into  the pfp_pmcs table across multiple calls. Unused
155       fields are guaranteed to be zeroed upon successful return.
156
157       Depending on the PMU model, there may not always be a one to  one  map‐
158       ping  between a PMC register and a data register. Register dependencies
159       may be more intricate.  However the pfm_dispatch_events guarantees cer‐
160       tain  ordering between the pfp_pmcs and pfp_pmds tables. In particular,
161       it guarantees that the pfp_pmds table always starts with  the  counters
162       corresponding,  in  the  same  order,  to the events as provided in the
163       pfp_event table on input.  There is always one counter per event. Addi‐
164       tional PMD registers, if any, come after.
165

EXAMPLE

167       Here is a typical sequence using the perfmon2 interface:
168            #include <perfmon/pfmlib.h>
169            ...
170            pfmlib_input_param_t inp;
171            pfmlib_output_param_t outp;
172            pfarg_ctx_t ctx;
173            pfarg_pmd_t pd[1];
174            pfarg_pmc_t pc[1];
175            pfarg_load_t load_arg;
176            int fd, i;
177            int ret;
178
179            if (pfm_initialize() != PFMLIB_SUCCESS) {
180                 fprintf(stderr, "can't initialize library\n");
181                 exit(1);
182            }
183            memset(&ctx,0, sizeof(ctx));
184            memset(&inp,0, sizeof(inp));
185            memset(&outp,0, sizeof(outp));
186            memset(pd, 0, sizeof(pd));
187            memset(pc, 0, sizeof(pc));
188            memset(&load_arg, 0, sizeof(load_arg));
189
190            ret = pfm_get_cycle_event(&inp.pfp_events[0]);
191            if (ret != PFMLIB_SUCCESS) {
192                 fprintf(stderr, "cannot find cycle event\n");
193                 exit(1);
194            }
195            inp.pfp_dfl_plm     = PFM_PLM3;
196            inp.pfp_event_count = 1;
197
198            ret = pfm_dispatch_events(&inp, NULL, &outp, NULL);
199            if (ret != PFMLIB_SUCCESS) {
200                 fprintf(stderr, "cannot dispatch events: %s\n", pfm_strerror(ret));
201                 exit(1);
202            }
203            /* propagate pmc value to perfmon2 structures */
204            for(i=0; i < outp.pfp_pmc_count; i++) {
205                 pc[i].reg_num   = outp.pfp_pmcs[i].reg_num;
206                 pc[i].reg_value = outp.pfp_pmcs[i].reg_value;
207            }
208            for(i=0; i < outp.pfp_pmd_count; i++) {
209                 pd[i].reg_num   = outp.pfp_pmds[i].reg_num;
210                 pd[i].reg_value = 0;
211            }
212            ...
213            if (pfm_create_context(&ctx, NULL, 0) == -1 ) {
214                 ...
215            }
216            fd = ctx.ctx_fd;
217
218            if (pfm_write_pmcs(fd, pc, outp.pfp_pmc_count) == -1) {
219                 ...
220            }
221            if (pfm_write_pmds(fd, pd, outp.pfp_pmd_count) == -1) {
222                 ...
223            }
224
225            load_arg.load_pid = getpid();
226            if (pfm_load_context(fd, &load_arg) == -1) {
227                 ...
228            }
229
230            pfm_start(fd, NULL);
231            /* code to monitor */
232            pfm_stop(fd);
233
234            if (pfm_read_pmds(fd, pd, evt.pfp_event_count) == -1) {
235                 ...
236            }
237            printf("results: %llu0, pd[0].reg_value);
238            ...
239            close(fd);
240            ...
241
242

RETURN

244       The  function returns whether or not the call was successful.  A return
245       value of PFMLIB_SUCCESS indicates sucess, otherwise the  value  is  the
246       error code.
247

ERRORS

249       PFMLIB_ERR_NOINIT The library has not been initialized properly.
250
251       PFMLIB_ERR_INVAL
252              Some arguments were invalid. For instance the value of *count is
253              zero.  This can also be due to he content of the  pfmlib_param_t
254              structure.
255
256       PFMLIB_ERR_NOTFOUND
257              No matching event was found.
258
259       PFMLIB_ERR_TOOMANY
260              The number of events to monitor exceed the number of implemented
261              counters.
262
263       PFMLIB_ERR_NOASSIGN
264              The events cannot be dispatched to the PMC because  events  have
265              conflicting constraints.
266
267       PFMLIB_ERR_MAGIC
268              The  model specific extension does not have the right magic num‐
269              ber.
270
271       PFMLIB_ERR_FEATCOMB
272              The set of events and features cannot be combined.
273
274       PFMLIB_ERR_EVTMANY
275              An event has  been  supplied  more  than  once  and  is  causing
276              resource (PMC) conflicts.
277
278       PFMLIB_ERR_IRRINVAL
279              Invalid code range restriction (Itanium, Itanium 2).
280
281       PFMLIB_ERR_IRRALIGN
282              Code range has invalid alignment (Itanium, Itanium 2).
283
284       PFMLIB_ERR_IRRTOOMANY
285              Cannot satisfy all the code ranges (Itanium, Itanium 2).
286
287       PFMLIB_ERR_DRRTOOMANY
288              Cannot satisfy all the data ranges (Itanium, Itanium 2).
289
290       PFMLIB_ERR_DRRINVAL
291              Invalid data range restriction (Itanium, Itanium 2).
292
293       PFMLIB_ERR_EVTSET
294              Some events belong to incompatible sets (Itanium 2).
295
296       PFMLIB_ERR_EVTINCOMP
297              Some events cannot be measured at the same time (Itanium 2).
298
299       PFMLIB_ERR_IRRTOOBIG
300              Code range is too big (Itanium 2).
301
302       PFMLIB_ERR_UMASK
303              Invalid or missing unit mask.
304

SEE ALSO

306       libpfm_itanium(3),   libpfm_itanium2(3),  pfm_regmask_set(3),  pfm_reg‐
307       mask_clr(3), pfm_find_event_code_mask(3)
308

AUTHOR

310       Stephane Eranian <eranian@hpl.hp.com>
311
312                                  July , 2003                        LIBPFM(3)
Impressum