drivers/perf/hisilicon/hisi_uncore_pmu.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * HiSilicon SoC Hardware event counters support
   4  *
   5  * Copyright (C) 2017 Hisilicon Limited
   6  * Author: Anurup M <anurup.m@huawei.com>
   7  *         Shaokun Zhang <zhangshaokun@hisilicon.com>
   8  *
   9  * This code is based on the uncore PMUs like arm-cci and arm-ccn.
  10  */
  11 #include <linux/bitmap.h>
  12 #include <linux/bitops.h>
  13 #include <linux/bug.h>
  14 #include <linux/err.h>
  15 #include <linux/errno.h>
  16 #include <linux/interrupt.h>
  17
  18 #include <asm/local64.h>
  19
  20 #include "hisi_uncore_pmu.h"
  21
  22 #define HISI_GET_EVENTID(ev) (ev->hw.config_base & 0xff)
  23 #define HISI_MAX_PERIOD(nr) (BIT_ULL(nr) - 1)
  24
  25 /*
  26  * PMU format attributes
  27  */
  28 ssize_t hisi_format_sysfs_show(struct device *dev,
  29                                struct device_attribute *attr, char *buf)
  30 {
  31         struct dev_ext_attribute *eattr;
  32
  33         eattr = container_of(attr, struct dev_ext_attribute, attr);
  34
  35         return sprintf(buf, "%s\n", (char *)eattr->var);
  36 }
  37
  38 /*
  39  * PMU event attributes
  40  */
  41 ssize_t hisi_event_sysfs_show(struct device *dev,
  42                               struct device_attribute *attr, char *page)
  43 {
  44         struct dev_ext_attribute *eattr;
  45
  46         eattr = container_of(attr, struct dev_ext_attribute, attr);
  47
  48         return sprintf(page, "config=0x%lx\n", (unsigned long)eattr->var);
  49 }
  50
  51 /*
  52  * sysfs cpumask attributes. For uncore PMU, we only have a single CPU to show
  53  */
  54 ssize_t hisi_cpumask_sysfs_show(struct device *dev,
  55                                 struct device_attribute *attr, char *buf)
  56 {
  57         struct hisi_pmu *hisi_pmu = to_hisi_pmu(dev_get_drvdata(dev));
  58
  59         return sprintf(buf, "%d\n", hisi_pmu->on_cpu);
  60 }
  61
  62 static bool hisi_validate_event_group(struct perf_event *event)
  63 {
  64         struct perf_event *sibling, *leader = event->group_leader;
  65         struct hisi_pmu *hisi_pmu = to_hisi_pmu(event->pmu);
  66         /* Include count for the event */
  67         int counters = 1;
  68
  69         if (!is_software_event(leader)) {
  70                 /*
  71                  * We must NOT create groups containing mixed PMUs, although
  72                  * software events are acceptable
  73                  */
  74                 if (leader->pmu != event->pmu)
  75                         return false;
  76
  77                 /* Increment counter for the leader */
  78                 if (leader != event)
  79                         counters++;
  80         }
  81
  82         for_each_sibling_event(sibling, event->group_leader) {
  83                 if (is_software_event(sibling))
  84                         continue;
  85                 if (sibling->pmu != event->pmu)
  86                         return false;
  87                 /* Increment counter for each sibling */
  88                 counters++;
  89         }
  90
  91         /* The group can not count events more than the counters in the HW */
  92         return counters <= hisi_pmu->num_counters;
  93 }
  94
  95 int hisi_uncore_pmu_counter_valid(struct hisi_pmu *hisi_pmu, int idx)
  96 {
  97         return idx >= 0 && idx < hisi_pmu->num_counters;
  98 }
  99
 100 int hisi_uncore_pmu_get_event_idx(struct perf_event *event)
 101 {
 102         struct hisi_pmu *hisi_pmu = to_hisi_pmu(event->pmu);
 103         unsigned long *used_mask = hisi_pmu->pmu_events.used_mask;
 104         u32 num_counters = hisi_pmu->num_counters;
 105         int idx;
 106
 107         idx = find_first_zero_bit(used_mask, num_counters);
 108         if (idx == num_counters)
 109                 return -EAGAIN;
 110
 111         set_bit(idx, used_mask);
 112
 113         return idx;
 114 }
 115
 116 static void hisi_uncore_pmu_clear_event_idx(struct hisi_pmu *hisi_pmu, int idx)
 117 {
 118         if (!hisi_uncore_pmu_counter_valid(hisi_pmu, idx)) {
 119                 dev_err(hisi_pmu->dev, "Unsupported event index:%d!\n", idx);
 120                 return;
 121         }
 122
 123         clear_bit(idx, hisi_pmu->pmu_events.used_mask);
 124 }
 125
 126 int hisi_uncore_pmu_event_init(struct perf_event *event)
 127 {
 128         struct hw_perf_event *hwc = &event->hw;
 129         struct hisi_pmu *hisi_pmu;
 130
 131         if (event->attr.type != event->pmu->type)
 132                 return -ENOENT;
 133
 134         /*
 135          * We do not support sampling as the counters are all
 136          * shared by all CPU cores in a CPU die(SCCL). Also we
 137          * do not support attach to a task(per-process mode)
 138          */
 139         if (is_sampling_event(event) || event->attach_state & PERF_ATTACH_TASK)
 140                 return -EOPNOTSUPP;
 141
 142         /*
 143          *  The uncore counters not specific to any CPU, so cannot
 144          *  support per-task
 145          */
 146         if (event->cpu < 0)
 147                 return -EINVAL;
 148
 149         /*
 150          * Validate if the events in group does not exceed the
 151          * available counters in hardware.
 152          */
 153         if (!hisi_validate_event_group(event))
 154                 return -EINVAL;
 155
 156         hisi_pmu = to_hisi_pmu(event->pmu);
 157         if (event->attr.config > hisi_pmu->check_event)
 158                 return -EINVAL;
 159
 160         if (hisi_pmu->on_cpu == -1)
 161                 return -EINVAL;
 162         /*
 163          * We don't assign an index until we actually place the event onto
 164          * hardware. Use -1 to signify that we haven't decided where to put it
 165          * yet.
 166          */
 167         hwc->idx                = -1;
 168         hwc->config_base        = event->attr.config;
 169
 170         /* Enforce to use the same CPU for all events in this PMU */
 171         event->cpu = hisi_pmu->on_cpu;
 172
 173         return 0;
 174 }
 175
 176 /*
 177  * Set the counter to count the event that we're interested in,
 178  * and enable interrupt and counter.
 179  */
 180 static void hisi_uncore_pmu_enable_event(struct perf_event *event)
 181 {
 182         struct hisi_pmu *hisi_pmu = to_hisi_pmu(event->pmu);
 183         struct hw_perf_event *hwc = &event->hw;
 184
 185         hisi_pmu->ops->write_evtype(hisi_pmu, hwc->idx,
 186                                     HISI_GET_EVENTID(event));
 187
 188         hisi_pmu->ops->enable_counter_int(hisi_pmu, hwc);
 189         hisi_pmu->ops->enable_counter(hisi_pmu, hwc);
 190 }
 191
 192 /*
 193  * Disable counter and interrupt.
 194  */
 195 static void hisi_uncore_pmu_disable_event(struct perf_event *event)
 196 {
 197         struct hisi_pmu *hisi_pmu = to_hisi_pmu(event->pmu);
 198         struct hw_perf_event *hwc = &event->hw;
 199
 200         hisi_pmu->ops->disable_counter(hisi_pmu, hwc);
 201         hisi_pmu->ops->disable_counter_int(hisi_pmu, hwc);
 202 }
 203
 204 void hisi_uncore_pmu_set_event_period(struct perf_event *event)
 205 {
 206         struct hisi_pmu *hisi_pmu = to_hisi_pmu(event->pmu);
 207         struct hw_perf_event *hwc = &event->hw;
 208
 209         /*
 210          * The HiSilicon PMU counters support 32 bits or 48 bits, depending on
 211          * the PMU. We reduce it to 2^(counter_bits - 1) to account for the
 212          * extreme interrupt latency. So we could hopefully handle the overflow
 213          * interrupt before another 2^(counter_bits - 1) events occur and the
 214          * counter overtakes its previous value.
 215          */
 216         u64 val = BIT_ULL(hisi_pmu->counter_bits - 1);
 217
 218         local64_set(&hwc->prev_count, val);
 219         /* Write start value to the hardware event counter */
 220         hisi_pmu->ops->write_counter(hisi_pmu, hwc, val);
 221 }
 222
 223 void hisi_uncore_pmu_event_update(struct perf_event *event)
 224 {
 225         struct hisi_pmu *hisi_pmu = to_hisi_pmu(event->pmu);
 226         struct hw_perf_event *hwc = &event->hw;
 227         u64 delta, prev_raw_count, new_raw_count;
 228
 229         do {
 230                 /* Read the count from the counter register */
 231                 new_raw_count = hisi_pmu->ops->read_counter(hisi_pmu, hwc);
 232                 prev_raw_count = local64_read(&hwc->prev_count);
 233         } while (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
 234                                  new_raw_count) != prev_raw_count);
 235         /*
 236          * compute the delta
 237          */
 238         delta = (new_raw_count - prev_raw_count) &
 239                 HISI_MAX_PERIOD(hisi_pmu->counter_bits);
 240         local64_add(delta, &event->count);
 241 }
 242
 243 void hisi_uncore_pmu_start(struct perf_event *event, int flags)
 244 {
 245         struct hisi_pmu *hisi_pmu = to_hisi_pmu(event->pmu);
 246         struct hw_perf_event *hwc = &event->hw;
 247
 248         if (WARN_ON_ONCE(!(hwc->state & PERF_HES_STOPPED)))
 249                 return;
 250
 251         WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE));
 252         hwc->state = 0;
 253         hisi_uncore_pmu_set_event_period(event);
 254
 255         if (flags & PERF_EF_RELOAD) {
 256                 u64 prev_raw_count =  local64_read(&hwc->prev_count);
 257
 258                 hisi_pmu->ops->write_counter(hisi_pmu, hwc, prev_raw_count);
 259         }
 260
 261         hisi_uncore_pmu_enable_event(event);
 262         perf_event_update_userpage(event);
 263 }
 264
 265 void hisi_uncore_pmu_stop(struct perf_event *event, int flags)
 266 {
 267         struct hw_perf_event *hwc = &event->hw;
 268
 269         hisi_uncore_pmu_disable_event(event);
 270         WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
 271         hwc->state |= PERF_HES_STOPPED;
 272
 273         if (hwc->state & PERF_HES_UPTODATE)
 274                 return;
 275
 276         /* Read hardware counter and update the perf counter statistics */
 277         hisi_uncore_pmu_event_update(event);
 278         hwc->state |= PERF_HES_UPTODATE;
 279 }
 280
 281 int hisi_uncore_pmu_add(struct perf_event *event, int flags)
 282 {
 283         struct hisi_pmu *hisi_pmu = to_hisi_pmu(event->pmu);
 284         struct hw_perf_event *hwc = &event->hw;
 285         int idx;
 286
 287         hwc->state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
 288
 289         /* Get an available counter index for counting */
 290         idx = hisi_pmu->ops->get_event_idx(event);
 291         if (idx < 0)
 292                 return idx;
 293
 294         event->hw.idx = idx;
 295         hisi_pmu->pmu_events.hw_events[idx] = event;
 296
 297         if (flags & PERF_EF_START)
 298                 hisi_uncore_pmu_start(event, PERF_EF_RELOAD);
 299
 300         return 0;
 301 }
 302
 303 void hisi_uncore_pmu_del(struct perf_event *event, int flags)
 304 {
 305         struct hisi_pmu *hisi_pmu = to_hisi_pmu(event->pmu);
 306         struct hw_perf_event *hwc = &event->hw;
 307
 308         hisi_uncore_pmu_stop(event, PERF_EF_UPDATE);
 309         hisi_uncore_pmu_clear_event_idx(hisi_pmu, hwc->idx);
 310         perf_event_update_userpage(event);
 311         hisi_pmu->pmu_events.hw_events[hwc->idx] = NULL;
 312 }
 313
 314 void hisi_uncore_pmu_read(struct perf_event *event)
 315 {
 316         /* Read hardware counter and update the perf counter statistics */
 317         hisi_uncore_pmu_event_update(event);
 318 }
 319
 320 void hisi_uncore_pmu_enable(struct pmu *pmu)
 321 {
 322         struct hisi_pmu *hisi_pmu = to_hisi_pmu(pmu);
 323         int enabled = bitmap_weight(hisi_pmu->pmu_events.used_mask,
 324                                     hisi_pmu->num_counters);
 325
 326         if (!enabled)
 327                 return;
 328
 329         hisi_pmu->ops->start_counters(hisi_pmu);
 330 }
 331
 332 void hisi_uncore_pmu_disable(struct pmu *pmu)
 333 {
 334         struct hisi_pmu *hisi_pmu = to_hisi_pmu(pmu);
 335
 336         hisi_pmu->ops->stop_counters(hisi_pmu);
 337 }
 338
 339 /*
 340  * Read Super CPU cluster and CPU cluster ID from MPIDR_EL1.
 341  * If multi-threading is supported, CCL_ID is the low 3-bits in MPIDR[Aff2]
 342  * and SCCL_ID is the upper 5-bits of Aff2 field; if not, SCCL_ID
 343  * is in MPIDR[Aff2] and CCL_ID is in MPIDR[Aff1].
 344  */
 345 static void hisi_read_sccl_and_ccl_id(int *sccl_id, int *ccl_id)
 346 {
 347         u64 mpidr = read_cpuid_mpidr();
 348
 349         if (mpidr & MPIDR_MT_BITMASK) {
 350                 int aff2 = MPIDR_AFFINITY_LEVEL(mpidr, 2);
 351
 352                 if (sccl_id)
 353                         *sccl_id = aff2 >> 3;
 354                 if (ccl_id)
 355                         *ccl_id = aff2 & 0x7;
 356         } else {
 357                 if (sccl_id)
 358                         *sccl_id = MPIDR_AFFINITY_LEVEL(mpidr, 2);
 359                 if (ccl_id)
 360                         *ccl_id = MPIDR_AFFINITY_LEVEL(mpidr, 1);
 361         }
 362 }
 363
 364 /*
 365  * Check whether the CPU is associated with this uncore PMU
 366  */
 367 static bool hisi_pmu_cpu_is_associated_pmu(struct hisi_pmu *hisi_pmu)
 368 {
 369         int sccl_id, ccl_id;
 370
 371         if (hisi_pmu->ccl_id == -1) {
 372                 /* If CCL_ID is -1, the PMU only shares the same SCCL */
 373                 hisi_read_sccl_and_ccl_id(&sccl_id, NULL);
 374
 375                 return sccl_id == hisi_pmu->sccl_id;
 376         }
 377
 378         hisi_read_sccl_and_ccl_id(&sccl_id, &ccl_id);
 379
 380         return sccl_id == hisi_pmu->sccl_id && ccl_id == hisi_pmu->ccl_id;
 381 }
 382
 383 int hisi_uncore_pmu_online_cpu(unsigned int cpu, struct hlist_node *node)
 384 {
 385         struct hisi_pmu *hisi_pmu = hlist_entry_safe(node, struct hisi_pmu,
 386                                                      node);
 387
 388         if (!hisi_pmu_cpu_is_associated_pmu(hisi_pmu))
 389                 return 0;
 390
 391         cpumask_set_cpu(cpu, &hisi_pmu->associated_cpus);
 392
 393         /* If another CPU is already managing this PMU, simply return. */
 394         if (hisi_pmu->on_cpu != -1)
 395                 return 0;
 396
 397         /* Use this CPU in cpumask for event counting */
 398         hisi_pmu->on_cpu = cpu;
 399
 400         /* Overflow interrupt also should use the same CPU */
 401         WARN_ON(irq_set_affinity(hisi_pmu->irq, cpumask_of(cpu)));
 402
 403         return 0;
 404 }
 405
 406 int hisi_uncore_pmu_offline_cpu(unsigned int cpu, struct hlist_node *node)
 407 {
 408         struct hisi_pmu *hisi_pmu = hlist_entry_safe(node, struct hisi_pmu,
 409                                                      node);
 410         cpumask_t pmu_online_cpus;
 411         unsigned int target;
 412
 413         if (!cpumask_test_and_clear_cpu(cpu, &hisi_pmu->associated_cpus))
 414                 return 0;
 415
 416         /* Nothing to do if this CPU doesn't own the PMU */
 417         if (hisi_pmu->on_cpu != cpu)
 418                 return 0;
 419
 420         /* Give up ownership of the PMU */
 421         hisi_pmu->on_cpu = -1;
 422
 423         /* Choose a new CPU to migrate ownership of the PMU to */
 424         cpumask_and(&pmu_online_cpus, &hisi_pmu->associated_cpus,
 425                     cpu_online_mask);
 426         target = cpumask_any_but(&pmu_online_cpus, cpu);
 427         if (target >= nr_cpu_ids)
 428                 return 0;
 429
 430         perf_pmu_migrate_context(&hisi_pmu->pmu, cpu, target);
 431         /* Use this CPU for event counting */
 432         hisi_pmu->on_cpu = target;
 433         WARN_ON(irq_set_affinity(hisi_pmu->irq, cpumask_of(target)));
 434
 435         return 0;
 436 }