virperf.c 11.1 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40
/*
 * virperf.c: methods for managing perf events
 *
 * This library is free software; you can redistribute it and/or
 * modify it under the terms of the GNU Lesser General Public
 * License as published by the Free Software Foundation; either
 * version 2.1 of the License, or (at your option) any later version.
 *
 * This library is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 * Lesser General Public License for more details.
 *
 * You should have received a copy of the GNU Lesser General Public
 * License along with this library.  If not, see
 * <http://www.gnu.org/licenses/>.
 *
 * Authors:
 *  Ren Qiaowei <qiaowei.ren@intel.com>
 */
#include <config.h>

#include <sys/ioctl.h>
#if defined HAVE_SYS_SYSCALL_H
# include <sys/syscall.h>
#endif

#include "virperf.h"
#include "viralloc.h"
#include "virerror.h"
#include "virlog.h"
#include "virfile.h"
#include "virstring.h"
#include "virtypedparam.h"

VIR_LOG_INIT("util.perf");

#define VIR_FROM_THIS VIR_FROM_PERF

VIR_ENUM_IMPL(virPerfEvent, VIR_PERF_EVENT_LAST,
Q
Qiaowei Ren 已提交
41 42
              "cmt", "mbmt", "mbml",
              "cpu_cycles", "instructions",
43
              "cache_references", "cache_misses",
44
              "branch_instructions", "branch_misses",
45
              "bus_cycles", "stalled_cycles_frontend",
46
              "stalled_cycles_backend", "ref_cpu_cycles",
47
              "cpu_clock", "task_clock", "page_faults",
48
              "context_switches", "cpu_migrations",
49
              "page_faults_min", "page_faults_maj");
50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71

struct virPerfEvent {
    int type;
    int fd;
    bool enabled;
    union {
        /* cmt */
        struct {
            int scale;
        } cmt;
    } efields;
};
typedef struct virPerfEvent *virPerfEventPtr;

struct virPerf {
    struct virPerfEvent events[VIR_PERF_EVENT_LAST];
};

#if defined(__linux__) && defined(HAVE_SYS_SYSCALL_H)

# include <linux/perf_event.h>

72 73 74 75 76 77 78 79 80 81
struct virPerfEventAttr {
    int type;
    unsigned int attrType;
    unsigned long long attrConfig;
};

static struct virPerfEventAttr attrs[] = {
    {.type = VIR_PERF_EVENT_CMT, .attrType = 0, .attrConfig = 1},
    {.type = VIR_PERF_EVENT_MBMT, .attrType = 0, .attrConfig = 2},
    {.type = VIR_PERF_EVENT_MBML, .attrType = 0, .attrConfig = 3},
Q
Qiaowei Ren 已提交
82 83 84 85 86 87 88 89 90 91 92 93
    {.type = VIR_PERF_EVENT_CPU_CYCLES,
     .attrType = PERF_TYPE_HARDWARE,
     .attrConfig = PERF_COUNT_HW_CPU_CYCLES},
    {.type = VIR_PERF_EVENT_INSTRUCTIONS,
     .attrType = PERF_TYPE_HARDWARE,
     .attrConfig = PERF_COUNT_HW_INSTRUCTIONS},
    {.type = VIR_PERF_EVENT_CACHE_REFERENCES,
     .attrType = PERF_TYPE_HARDWARE,
     .attrConfig = PERF_COUNT_HW_CACHE_REFERENCES},
    {.type = VIR_PERF_EVENT_CACHE_MISSES,
     .attrType = PERF_TYPE_HARDWARE,
     .attrConfig = PERF_COUNT_HW_CACHE_MISSES},
94 95 96
    {.type = VIR_PERF_EVENT_BRANCH_INSTRUCTIONS,
     .attrType = PERF_TYPE_HARDWARE,
     .attrConfig = PERF_COUNT_HW_BRANCH_INSTRUCTIONS},
97 98 99
    {.type = VIR_PERF_EVENT_BRANCH_MISSES,
     .attrType = PERF_TYPE_HARDWARE,
     .attrConfig = PERF_COUNT_HW_BRANCH_MISSES},
100 101 102
    {.type = VIR_PERF_EVENT_BUS_CYCLES,
     .attrType = PERF_TYPE_HARDWARE,
     .attrConfig = PERF_COUNT_HW_BUS_CYCLES},
103 104 105
    {.type = VIR_PERF_EVENT_STALLED_CYCLES_FRONTEND,
     .attrType = PERF_TYPE_HARDWARE,
     .attrConfig = PERF_COUNT_HW_STALLED_CYCLES_FRONTEND},
106 107 108
    {.type = VIR_PERF_EVENT_STALLED_CYCLES_BACKEND,
     .attrType = PERF_TYPE_HARDWARE,
     .attrConfig = PERF_COUNT_HW_STALLED_CYCLES_BACKEND},
109
    {.type = VIR_PERF_EVENT_REF_CPU_CYCLES,
110
# ifdef PERF_COUNT_HW_REF_CPU_CYCLES
111
     .attrType = PERF_TYPE_HARDWARE,
112 113 114 115 116 117
     .attrConfig = PERF_COUNT_HW_REF_CPU_CYCLES
# else
     .attrType = 0,
     .attrConfig = 0,
# endif
    },
118 119 120
    {.type = VIR_PERF_EVENT_CPU_CLOCK,
     .attrType = PERF_TYPE_SOFTWARE,
     .attrConfig = PERF_COUNT_SW_CPU_CLOCK},
121 122 123
    {.type = VIR_PERF_EVENT_TASK_CLOCK,
     .attrType = PERF_TYPE_SOFTWARE,
     .attrConfig = PERF_COUNT_SW_TASK_CLOCK},
124 125 126
    {.type = VIR_PERF_EVENT_PAGE_FAULTS,
     .attrType = PERF_TYPE_SOFTWARE,
     .attrConfig = PERF_COUNT_SW_PAGE_FAULTS},
127 128 129
    {.type = VIR_PERF_EVENT_CONTEXT_SWITCHES,
     .attrType = PERF_TYPE_SOFTWARE,
     .attrConfig = PERF_COUNT_SW_CONTEXT_SWITCHES},
130 131 132
    {.type = VIR_PERF_EVENT_CPU_MIGRATIONS,
     .attrType = PERF_TYPE_SOFTWARE,
     .attrConfig = PERF_COUNT_SW_CPU_MIGRATIONS},
133 134 135
    {.type = VIR_PERF_EVENT_PAGE_FAULTS_MIN,
     .attrType = PERF_TYPE_SOFTWARE,
     .attrConfig = PERF_COUNT_SW_PAGE_FAULTS_MIN},
136 137 138
    {.type = VIR_PERF_EVENT_PAGE_FAULTS_MAJ,
     .attrType = PERF_TYPE_SOFTWARE,
     .attrConfig = PERF_COUNT_SW_PAGE_FAULTS_MAJ},
139 140 141
};
typedef struct virPerfEventAttr *virPerfEventAttrPtr;

142

143 144 145 146
static virPerfEventAttrPtr
virPerfGetEventAttr(virPerfEventType type)
{
    size_t i;
147 148 149 150 151 152 153
    if (type >= VIR_PERF_EVENT_LAST) {
        virReportError(VIR_ERR_INTERNAL_ERROR,
                       _("Event '%d' is not supported"),
                       type);
        return NULL;
    }

154 155 156 157 158 159
    for (i = 0; i < VIR_PERF_EVENT_LAST; i++) {
        if (i == type)
            return attrs + i;
    }

    return NULL;
160 161
}

162

163
static int
164
virPerfRdtAttrInit(void)
165 166 167
{
    char *buf = NULL;
    char *tmp = NULL;
168
    unsigned int attr_type = 0;
169

170
    if (virFileReadAllQuiet("/sys/devices/intel_cqm/type", 10, &buf) < 0)
171
        goto error;
172 173 174 175

    if ((tmp = strchr(buf, '\n')))
        *tmp = '\0';

176
    if (virStrToLong_ui(buf, NULL, 10, &attr_type) < 0) {
177
        virReportSystemError(errno, "%s",
178
                             _("failed to get rdt event type"));
179
        goto error;
180 181 182
    }
    VIR_FREE(buf);

183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223
    attrs[VIR_PERF_EVENT_CMT].attrType = attr_type;
    attrs[VIR_PERF_EVENT_MBMT].attrType = attr_type;
    attrs[VIR_PERF_EVENT_MBML].attrType = attr_type;

    return 0;

 error:
    VIR_FREE(buf);
    return -1;
}


static virPerfEventPtr
virPerfGetEvent(virPerfPtr perf,
                virPerfEventType type)
{
    if (!perf)
        return NULL;

    if (type >= VIR_PERF_EVENT_LAST) {
        virReportError(VIR_ERR_INTERNAL_ERROR,
                       _("Event '%d' is not supported"),
                       type);
        return NULL;
    }

    return perf->events + type;
}

int
virPerfEventEnable(virPerfPtr perf,
                   virPerfEventType type,
                   pid_t pid)
{
    char *buf = NULL;
    struct perf_event_attr attr;
    virPerfEventPtr event = virPerfGetEvent(perf, type);
    virPerfEventAttrPtr event_attr = virPerfGetEventAttr(type);

    if (!event || !event_attr)
        return -1;
224

225 226 227
    if (event->enabled)
        return 0;

228 229 230
    if (event_attr->attrType == 0 && (type == VIR_PERF_EVENT_CMT ||
                                       type == VIR_PERF_EVENT_MBMT ||
                                       type == VIR_PERF_EVENT_MBML)) {
231
        virReportError(VIR_ERR_ARGUMENT_UNSUPPORTED,
232
                       _("unable to enable host cpu perf event for %s"),
233
                       virPerfEventTypeToString(event->type));
234 235 236 237
        return -1;
    }

    if (type == VIR_PERF_EVENT_CMT) {
238 239 240 241
        if (virFileReadAll("/sys/devices/intel_cqm/events/llc_occupancy.scale",
                           10, &buf) < 0)
            goto error;

242
        if (virStrToLong_i(buf, NULL, 10, &event->efields.cmt.scale) < 0) {
243 244 245 246
            virReportSystemError(errno, "%s",
                                 _("failed to get cmt scaling factor"));
            goto error;
        }
247 248

        VIR_FREE(buf);
249 250
    }

251 252 253 254 255 256 257 258 259
    memset(&attr, 0, sizeof(attr));
    attr.size = sizeof(attr);
    attr.inherit = 1;
    attr.disabled = 1;
    attr.enable_on_exec = 0;
    attr.type = event_attr->attrType;
    attr.config = event_attr->attrConfig;

    event->fd = syscall(__NR_perf_event_open, &attr, pid, -1, -1, 0);
260 261
    if (event->fd < 0) {
        virReportSystemError(errno,
262
                             _("unable to open host cpu perf event for %s"),
263
                             virPerfEventTypeToString(event->type));
264
        goto error;
265 266 267
    }

    if (ioctl(event->fd, PERF_EVENT_IOC_ENABLE) < 0) {
268
        virReportSystemError(errno,
269
                             _("unable to enable host cpu perf event for %s"),
270
                             virPerfEventTypeToString(event->type));
271
        goto error;
272 273 274 275 276
    }

    event->enabled = true;
    return 0;

277
 error:
278 279 280 281 282 283 284 285 286 287 288 289 290
    VIR_FORCE_CLOSE(event->fd);
    VIR_FREE(buf);
    return -1;
}

int
virPerfEventDisable(virPerfPtr perf,
                    virPerfEventType type)
{
    virPerfEventPtr event = virPerfGetEvent(perf, type);
    if (event == NULL)
        return -1;

291 292 293
    if (!event->enabled)
        return 0;

294 295
    if (ioctl(event->fd, PERF_EVENT_IOC_DISABLE) < 0) {
        virReportSystemError(errno,
296 297
                             _("unable to disable host cpu perf event for %s"),
                             virPerfEventTypeToString(event->type));
298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324
        return -1;
    }

    event->enabled = false;
    VIR_FORCE_CLOSE(event->fd);
    return 0;
}

bool virPerfEventIsEnabled(virPerfPtr perf,
                           virPerfEventType type)
{
    virPerfEventPtr event = virPerfGetEvent(perf, type);
    if (event == NULL)
        return false;

    return event->enabled;
}

int
virPerfReadEvent(virPerfPtr perf,
                 virPerfEventType type,
                 uint64_t *value)
{
    virPerfEventPtr event = virPerfGetEvent(perf, type);
    if (event == NULL || !event->enabled)
        return -1;

325
    if (saferead(event->fd, value, sizeof(uint64_t)) < 0) {
326 327 328 329 330 331 332 333 334 335 336 337
        virReportSystemError(errno, "%s",
                             _("Unable to read cache data"));
        return -1;
    }

    if (type == VIR_PERF_EVENT_CMT)
        *value *= event->efields.cmt.scale;

    return 0;
}

#else
338 339 340 341 342 343 344
static int
virPerfRdtAttrInit(void)
{
    return 0;
}


345 346
int
virPerfEventEnable(virPerfPtr perf ATTRIBUTE_UNUSED,
R
Roman Bogorodskiy 已提交
347
                   virPerfEventType type ATTRIBUTE_UNUSED,
348 349 350 351 352 353 354 355 356
                   pid_t pid ATTRIBUTE_UNUSED)
{
    virReportSystemError(ENXIO, "%s",
                         _("Perf not supported on this platform"));
    return -1;
}

int
virPerfEventDisable(virPerfPtr perf ATTRIBUTE_UNUSED,
R
Roman Bogorodskiy 已提交
357
                    virPerfEventType type ATTRIBUTE_UNUSED)
358 359 360 361 362 363 364
{
    virReportSystemError(ENXIO, "%s",
                         _("Perf not supported on this platform"));
    return -1;
}

bool
R
Roman Bogorodskiy 已提交
365 366
virPerfEventIsEnabled(virPerfPtr perf ATTRIBUTE_UNUSED,
                      virPerfEventType type ATTRIBUTE_UNUSED)
367 368 369 370 371
{
    return false;
}

int
R
Roman Bogorodskiy 已提交
372 373 374
virPerfReadEvent(virPerfPtr perf ATTRIBUTE_UNUSED,
                 virPerfEventType type ATTRIBUTE_UNUSED,
                 uint64_t *value ATTRIBUTE_UNUSED)
375 376 377 378 379 380 381
{
    virReportSystemError(ENXIO, "%s",
                         _("Perf not supported on this platform"));
    return -1;
}

#endif
382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397

virPerfPtr
virPerfNew(void)
{
    size_t i;
    virPerfPtr perf;

    if (VIR_ALLOC(perf) < 0)
        return NULL;

    for (i = 0; i < VIR_PERF_EVENT_LAST; i++) {
        perf->events[i].type = i;
        perf->events[i].fd = -1;
        perf->events[i].enabled = false;
    }

398 399 400
    if (virPerfRdtAttrInit() < 0)
        virResetLastError();

401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418
    return perf;
}

void
virPerfFree(virPerfPtr perf)
{
    size_t i;

    if (perf == NULL)
        return;

    for (i = 0; i < VIR_PERF_EVENT_LAST; i++) {
        if (perf->events[i].enabled)
            virPerfEventDisable(perf, i);
    }

    VIR_FREE(perf);
}