Tribblix: manual page: cpc

CPC_BIND_EVENT(3CPC) CPU Performance Counters Library Functions

NAME

cpc_bind_event, cpc_take_sample, cpc_rele - use CPU performance
counters on lwps

SYNOPSIS

cc [ flag... ] file... -lcpc [ library... ]
#include <libcpc.h>

int cpc_bind_event(cpc_event_t *event, int flags);

int cpc_take_sample(cpc_event_t *event);

int cpc_rele(void);

DESCRIPTION

Once the events to be sampled have been selected using, for example,
cpc_strtoevent(3CPC), the event selections can be bound to the
calling LWP using cpc_bind_event(). If cpc_bind_event() returns
successfully, the system has associated performance counter context
with the calling LWP. The context allows the system to virtualize the
hardware counters to that specific LWP, and the counters are enabled.

Two flags are defined that can be passed into the routine to allow
the behavior of the interface to be modified, as described below.

Counter values can be sampled at any time by calling
cpc_take_sample(), and dereferencing the fields of the ce_pic[] array
returned. The ce_hrt field contains the timestamp at which the kernel
last sampled the counters.

To immediately remove the performance counter context on an LWP, the
cpc_rele() interface should be used. Otherwise, the context will be
destroyed after the LWP or process exits.

The caller should take steps to ensure that the counters are sampled
often enough to avoid the 32-bit counters wrapping. The events most
prone to wrap are those that count processor clock cycles. If such an
event is of interest, sampling should occur frequently so that less
than 4 billion clock cycles can occur between samples. Practically
speaking, this is only likely to be a problem for otherwise idle
systems, or when processes are bound to processors, since normal
context switching behavior will otherwise hide this problem.

RETURN VALUES

Upon successful completion, cpc_bind_event() and cpc_take_sample()
return 0. Otherwise, these functions return -1, and set errno to
indicate the error.

ERRORS

The cpc_bind_event() and cpc_take_sample() functions will fail if:

EACCES
For cpc_bind_event(), access to the requested hypervisor
event was denied.

EAGAIN
Another process may be sampling system-wide CPU
statistics. For cpc_bind_event(), this implies that no new
contexts can be created. For cpc_take_sample(), this
implies that the performance counter context has been
invalidated and must be released with cpc_rele(). Robust
programs should be coded to expect this behavior and
recover from it by releasing the now invalid context by
calling cpc_rele() sleeping for a while, then attempting
to bind and sample the event once more.

EINVAL
The cpc_take_sample() function has been invoked before the
context is bound.

ENOTSUP
The caller has attempted an operation that is illegal or
not supported on the current platform, such as attempting
to specify signal delivery on counter overflow on a CPU
that doesn't generate an interrupt on counter overflow.

USAGE

Prior to calling cpc_bind_event(), applications should call
cpc_access(3CPC) to determine if the counters are accessible on the
system.

EXAMPLES

Example 1: Use hardware performance counters to measure events in a

process.

The example below shows how a standalone program can be instrumented
with the libcpc routines to use hardware performance counters to
measure events in a process. The program performs 20 iterations of a
computation, measuring the counter values for each iteration. By
default, the example makes the counters measure external cache
references and external cache hits; these options are only
appropriate for UltraSPARC processors. By setting the PERFEVENTS
environment variable to other strings (a list of which can be gleaned
from the -h flag of the cpustat or cputrack utilities), other events
can be counted. The error() routine below is assumed to be a user-
provided routine analogous to the familiar printf(3C) routine from
the C library but which also performs an exit(2) after printing the
message.

#include <inttypes.h>
#include <stdlib.h>
#include <stdio.h>
#include <unistd.h>
#include <libcpc.h>
int
main(int argc, char *argv[])
{
int cpuver, iter;
char *setting = NULL;
cpc_event_t event;

if (cpc_version(CPC_VER_CURRENT) != CPC_VER_CURRENT)
error("application:library cpc version mismatch!");

if ((cpuver = cpc_getcpuver()) == -1)
error("no performance counter hardware!");

if ((setting = getenv("PERFEVENTS")) == NULL)
setting = "pic0=EC_ref,pic1=EC_hit";

if (cpc_strtoevent(cpuver, setting, &event) != 0)
error("can't measure '%s' on this processor", setting);
setting = cpc_eventtostr(&event);

if (cpc_access() == -1)
error("can't access perf counters: %s", strerror(errno));

if (cpc_bind_event(&event, 0) == -1)
error("can't bind lwp%d: %s", _lwp_self(), strerror(errno));

for (iter = 1; iter <= 20; iter++) {
cpc_event_t before, after;

if (cpc_take_sample(&before) == -1)
break;

/* ==> Computation to be measured goes here <== */

if (cpc_take_sample(&after) == -1)
break;
(void) printf("%3d: %" PRId64 " %" PRId64 "\n", iter,
after.ce_pic[0] - before.ce_pic[0],
after.ce_pic[1] - before.ce_pic[1]);
}

if (iter != 20)
error("can't sample '%s': %s", setting, strerror(errno));

free(setting);
return (0);
}

Example 2: Write a signal handler to catch overflow signals.

This example builds on Example 1, but demonstrates how to write the
signal handler to catch overflow signals. The counters are preset so
that counter zero is 1000 counts short of overflowing, while counter
one is set to zero. After 1000 counts on counter zero, the signal
handler will be invoked.

First the signal handler:

#define PRESET0 (UINT64_MAX - UINT64_C(999))
#define PRESET1 0

void
emt_handler(int sig, siginfo_t *sip, void *arg)
{
ucontext_t *uap = arg;
cpc_event_t sample;

if (sig != SIGEMT || sip->si_code != EMT_CPCOVF) {
psignal(sig, "example");
psiginfo(sip, "example");
return;
}

(void) printf("lwp%d - si_addr %p ucontext: %%pc %p %%sp %p\n",
_lwp_self(), (void *)sip->si_addr,
(void *)uap->uc_mcontext.gregs[PC],
(void *)uap->uc_mcontext.gregs[USP]);

if (cpc_take_sample(&sample) == -1)
error("can't sample: %s", strerror(errno));

(void) printf("0x%" PRIx64 " 0x%" PRIx64 "\n",
sample.ce_pic[0], sample.ce_pic[1]);
(void) fflush(stdout);

sample.ce_pic[0] = PRESET0;
sample.ce_pic[1] = PRESET1;
if (cpc_bind_event(&sample, CPC_BIND_EMT_OVF) == -1)
error("cannot bind lwp%d: %s", _lwp_self(), strerror(errno));
}

and second the setup code (this can be placed after the code that
selects the event to be measured):

struct sigaction act;
cpc_event_t event;
...
act.sa_sigaction = emt_handler;
bzero(&act.sa_mask, sizeof (act.sa_mask));
act.sa_flags = SA_RESTART|SA_SIGINFO;
if (sigaction(SIGEMT, &act, NULL) == -1)
error("sigaction: %s", strerror(errno));
event.ce_pic[0] = PRESET0;
event.ce_pic[1] = PRESET1;
if (cpc_bind_event(&event, CPC_BIND_EMT_OVF) == -1)
error("cannot bind lwp%d: %s", _lwp_self(), strerror(errno));

for (iter = 1; iter <= 20; iter++) {
/* ==> Computation to be measured goes here <== */
}

cpc_bind_event(NULL, 0); /* done */

Note that a more general version of the signal handler would use
write(2) directly instead of depending on the signal-unsafe semantics
of stderr and stdout. Most real signal handlers will probably do more
with the samples than just print them out.

ATTRIBUTES

NOTES

The cpc_bind_event(), cpc_take_sample(), and cpc_rele() functions
exist for binary compatibility only. Source containing these
functions will not compile. These functions are obsolete and might be
removed in a future release. Applications should use
cpc_bind_curlwp(3CPC), cpc_set_sample(3CPC), and cpc_unbind(3CPC)
instead.

Sometimes, even the overhead of performing a system call will be too
disruptive to the events being measured. Once a call to
cpc_bind_event() has been issued, it is possible to directly access
the performance hardware registers from within the application. If
the performance counter context is active, then the counters will
count on behalf of the current LWP.

SPARC

rd %pic, %rN ! All UltraSPARC
wr %rN, %pic ! (ditto, but see text)

x86
rdpmc ! Pentium II only

If the counter context is not active or has been invalidated, the
%pic register (SPARC), and the rdpmc instruction (Pentium) will
become unavailable.

Note that the two 32-bit UltraSPARC performance counters are kept in
the single 64-bit %pic register so a couple of additional
instructions are required to separate the values. Also note that when
the %pcr register bit has been set that configures the %pic register
as readable by an application, it is also writable. Any values
written will be preserved by the context switching mechanism.

Pentium II processors support the non-privileged rdpmc instruction
which requires [5] that the counter of interest be specified in %ecx,
and returns a 40-bit value in the %edx:%eax register pair. There is
no non-privileged access mechanism for Pentium I processors.

Handling counter overflow

As described above, when counting events, some processors allow their
counter registers to silently overflow. More recent CPUs such as
UltraSPARC III and Pentium II, however, are capable of generating an
interrupt when the hardware counter overflows. Some processors offer
more control over when interrupts will actually be generated. For
example, they might allow the interrupt to be programmed to occur
when only one of the counters overflows. See cpc_strtoevent(3CPC) for
the syntax.

The most obvious use for this facility is to ensure that the full
64-bit counter values are maintained without repeated sampling.
However, current hardware does not record which counter overflowed. A
more subtle use for this facility is to preset the counter to a value
to a little less than the maximum value, then use the resulting
interrupt to catch the counter overflow associated with that event.
The overflow can then be used as an indication of the frequency of
the occurrence of that event.

Note that the interrupt generated by the processor may not be
particularly precise. That is, the particular instruction that
caused the counter overflow may be earlier in the instruction stream
than is indicated by the program counter value in the ucontext.

When cpc_bind_event() is called with the CPC_BIND_EMT_OVF flag set,
then as before, the control registers and counters are preset from
the 64-bit values contained in event. However, when the flag is set,
the kernel arranges to send the calling process a SIGEMT signal when
the overflow occurs, with the si_code field of the corresponding
siginfo structure set to EMT_CPCOVF, and the si_addr field is the
program counter value at the time the overflow interrupt was
delivered. Counting is disabled until the next call to
cpc_bind_event(). Even in a multithreaded process, during execution
of the signal handler, the thread behaves as if it is temporarily
bound to the running LWP.

Different processors have different counter ranges available, though
all processors supported by Solaris allow at least 31 bits to be
specified as a counter preset value; thus portable preset values lie
in the range UINT64_MAX to UINT64_MAX-INT32_MAX.

The appropriate preset value will often need to be determined
experimentally. Typically, it will depend on the event being
measured, as well as the desire to minimize the impact of the act of
measurement on the event being measured; less frequent interrupts and
samples lead to less perturbation of the system.

If the processor cannot detect counter overflow, this call will fail
(ENOTSUP). Specifying a null event unbinds the context from the
underlying LWP and disables signal delivery. Currently, only user
events can be measured using this technique. See Example 2, above.

Inheriting events onto multiple LWPs
By default, the library binds the performance counter context to the
current LWP only. If the CPC_BIND_LWP_INHERIT flag is set, then any
subsequent LWPs created by that LWP will automatically inherit the
same performance counter context. The counters will be initialized
to 0 as if a cpc_bind_event() had just been issued. This automatic
inheritance behavior can be useful when dealing with multithreaded
programs to determine aggregate statistics for the program as a
whole.

If the CPC_BIND_EMT_OVF flag is also set, the process will
immediately dispatch a SIGEMT signal to the freshly created LWP so
that it can preset its counters appropriately on the new LWP. This
initialization condition can be detected using cpc_take_sample() to
check that both ce_pic[] values are set to UINT64_MAX.

September 10, 2013 CPC_BIND_EVENT(3CPC)