drivers/perf/riscv_pmu.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * RISC-V performance counter support.
   4  *
   5  * Copyright (C) 2021 Western Digital Corporation or its affiliates.
   6  *
   7  * This implementation is based on old RISC-V perf and ARM perf event code
   8  * which are in turn based on sparc64 and x86 code.
   9  */
  10
  11 #include <linux/cpumask.h>
  12 #include <linux/irq.h>
  13 #include <linux/irqdesc.h>
  14 #include <linux/perf/riscv_pmu.h>
  15 #include <linux/printk.h>
  16 #include <linux/smp.h>
  17 #include <linux/sched_clock.h>
  18
  19 #include <asm/sbi.h>
  20
  21 static bool riscv_perf_user_access(struct perf_event *event)
  22 {
  23         return ((event->attr.type == PERF_TYPE_HARDWARE) ||
  24                 (event->attr.type == PERF_TYPE_HW_CACHE) ||
  25                 (event->attr.type == PERF_TYPE_RAW)) &&
  26                 !!(event->hw.flags & PERF_EVENT_FLAG_USER_READ_CNT) &&
  27                 (event->hw.idx != -1);
  28 }
  29
  30 void arch_perf_update_userpage(struct perf_event *event,
  31                                struct perf_event_mmap_page *userpg, u64 now)
  32 {
  33         struct clock_read_data *rd;
  34         unsigned int seq;
  35         u64 ns;
  36
  37         userpg->cap_user_time = 0;
  38         userpg->cap_user_time_zero = 0;
  39         userpg->cap_user_time_short = 0;
  40         userpg->cap_user_rdpmc = riscv_perf_user_access(event);
  41
  42 #ifdef CONFIG_RISCV_PMU
  43         /*
  44          * The counters are 64-bit but the priv spec doesn't mandate all the
  45          * bits to be implemented: that's why, counter width can vary based on
  46          * the cpu vendor.
  47          */
  48         if (userpg->cap_user_rdpmc)
  49                 userpg->pmc_width = to_riscv_pmu(event->pmu)->ctr_get_width(event->hw.idx) + 1;
  50 #endif
  51
  52         do {
  53                 rd = sched_clock_read_begin(&seq);
  54
  55                 userpg->time_mult = rd->mult;
  56                 userpg->time_shift = rd->shift;
  57                 userpg->time_zero = rd->epoch_ns;
  58                 userpg->time_cycles = rd->epoch_cyc;
  59                 userpg->time_mask = rd->sched_clock_mask;
  60
  61                 /*
  62                  * Subtract the cycle base, such that software that
  63                  * doesn't know about cap_user_time_short still 'works'
  64                  * assuming no wraps.
  65                  */
  66                 ns = mul_u64_u32_shr(rd->epoch_cyc, rd->mult, rd->shift);
  67                 userpg->time_zero -= ns;
  68
  69         } while (sched_clock_read_retry(seq));
  70
  71         userpg->time_offset = userpg->time_zero - now;
  72
  73         /*
  74          * time_shift is not expected to be greater than 31 due to
  75          * the original published conversion algorithm shifting a
  76          * 32-bit value (now specifies a 64-bit value) - refer
  77          * perf_event_mmap_page documentation in perf_event.h.
  78          */
  79         if (userpg->time_shift == 32) {
  80                 userpg->time_shift = 31;
  81                 userpg->time_mult >>= 1;
  82         }
  83
  84         /*
  85          * Internal timekeeping for enabled/running/stopped times
  86          * is always computed with the sched_clock.
  87          */
  88         userpg->cap_user_time = 1;
  89         userpg->cap_user_time_zero = 1;
  90         userpg->cap_user_time_short = 1;
  91 }
  92
  93 static unsigned long csr_read_num(int csr_num)
  94 {
  95 #define switchcase_csr_read(__csr_num, __val)           {\
  96         case __csr_num:                                 \
  97                 __val = csr_read(__csr_num);            \
  98                 break; }
  99 #define switchcase_csr_read_2(__csr_num, __val)         {\
 100         switchcase_csr_read(__csr_num + 0, __val)        \
 101         switchcase_csr_read(__csr_num + 1, __val)}
 102 #define switchcase_csr_read_4(__csr_num, __val)         {\
 103         switchcase_csr_read_2(__csr_num + 0, __val)      \
 104         switchcase_csr_read_2(__csr_num + 2, __val)}
 105 #define switchcase_csr_read_8(__csr_num, __val)         {\
 106         switchcase_csr_read_4(__csr_num + 0, __val)      \
 107         switchcase_csr_read_4(__csr_num + 4, __val)}
 108 #define switchcase_csr_read_16(__csr_num, __val)        {\
 109         switchcase_csr_read_8(__csr_num + 0, __val)      \
 110         switchcase_csr_read_8(__csr_num + 8, __val)}
 111 #define switchcase_csr_read_32(__csr_num, __val)        {\
 112         switchcase_csr_read_16(__csr_num + 0, __val)     \
 113         switchcase_csr_read_16(__csr_num + 16, __val)}
 114
 115         unsigned long ret = 0;
 116
 117         switch (csr_num) {
 118         switchcase_csr_read_32(CSR_CYCLE, ret)
 119         switchcase_csr_read_32(CSR_CYCLEH, ret)
 120         default :
 121                 break;
 122         }
 123
 124         return ret;
 125 #undef switchcase_csr_read_32
 126 #undef switchcase_csr_read_16
 127 #undef switchcase_csr_read_8
 128 #undef switchcase_csr_read_4
 129 #undef switchcase_csr_read_2
 130 #undef switchcase_csr_read
 131 }
 132
 133 /*
 134  * Read the CSR of a corresponding counter.
 135  */
 136 unsigned long riscv_pmu_ctr_read_csr(unsigned long csr)
 137 {
 138         if (csr < CSR_CYCLE || csr > CSR_HPMCOUNTER31H ||
 139            (csr > CSR_HPMCOUNTER31 && csr < CSR_CYCLEH)) {
 140                 pr_err("Invalid performance counter csr %lx\n", csr);
 141                 return -EINVAL;
 142         }
 143
 144         return csr_read_num(csr);
 145 }
 146
 147 u64 riscv_pmu_ctr_get_width_mask(struct perf_event *event)
 148 {
 149         int cwidth;
 150         struct riscv_pmu *rvpmu = to_riscv_pmu(event->pmu);
 151         struct hw_perf_event *hwc = &event->hw;
 152
 153         if (hwc->idx == -1)
 154                 /* Handle init case where idx is not initialized yet */
 155                 cwidth = rvpmu->ctr_get_width(0);
 156         else
 157                 cwidth = rvpmu->ctr_get_width(hwc->idx);
 158
 159         return GENMASK_ULL(cwidth, 0);
 160 }
 161
 162 u64 riscv_pmu_event_update(struct perf_event *event)
 163 {
 164         struct riscv_pmu *rvpmu = to_riscv_pmu(event->pmu);
 165         struct hw_perf_event *hwc = &event->hw;
 166         u64 prev_raw_count, new_raw_count;
 167         unsigned long cmask;
 168         u64 oldval, delta;
 169
 170         if (!rvpmu->ctr_read || (hwc->state & PERF_HES_UPTODATE))
 171                 return 0;
 172
 173         cmask = riscv_pmu_ctr_get_width_mask(event);
 174
 175         do {
 176                 prev_raw_count = local64_read(&hwc->prev_count);
 177                 new_raw_count = rvpmu->ctr_read(event);
 178                 oldval = local64_cmpxchg(&hwc->prev_count, prev_raw_count,
 179                                          new_raw_count);
 180         } while (oldval != prev_raw_count);
 181
 182         delta = (new_raw_count - prev_raw_count) & cmask;
 183         local64_add(delta, &event->count);
 184         local64_sub(delta, &hwc->period_left);
 185
 186         return delta;
 187 }
 188
 189 void riscv_pmu_stop(struct perf_event *event, int flags)
 190 {
 191         struct hw_perf_event *hwc = &event->hw;
 192         struct riscv_pmu *rvpmu = to_riscv_pmu(event->pmu);
 193
 194         if (!(hwc->state & PERF_HES_STOPPED)) {
 195                 if (rvpmu->ctr_stop) {
 196                         rvpmu->ctr_stop(event, 0);
 197                         hwc->state |= PERF_HES_STOPPED;
 198                 }
 199                 riscv_pmu_event_update(event);
 200                 hwc->state |= PERF_HES_UPTODATE;
 201         }
 202 }
 203
 204 int riscv_pmu_event_set_period(struct perf_event *event)
 205 {
 206         struct hw_perf_event *hwc = &event->hw;
 207         s64 left = local64_read(&hwc->period_left);
 208         s64 period = hwc->sample_period;
 209         int overflow = 0;
 210         uint64_t max_period = riscv_pmu_ctr_get_width_mask(event);
 211
 212         if (unlikely(left <= -period)) {
 213                 left = period;
 214                 local64_set(&hwc->period_left, left);
 215                 hwc->last_period = period;
 216                 overflow = 1;
 217         }
 218
 219         if (unlikely(left <= 0)) {
 220                 left += period;
 221                 local64_set(&hwc->period_left, left);
 222                 hwc->last_period = period;
 223                 overflow = 1;
 224         }
 225
 226         /*
 227          * Limit the maximum period to prevent the counter value
 228          * from overtaking the one we are about to program. In
 229          * effect we are reducing max_period to account for
 230          * interrupt latency (and we are being very conservative).
 231          */
 232         if (left > (max_period >> 1))
 233                 left = (max_period >> 1);
 234
 235         local64_set(&hwc->prev_count, (u64)-left);
 236
 237         perf_event_update_userpage(event);
 238
 239         return overflow;
 240 }
 241
 242 void riscv_pmu_start(struct perf_event *event, int flags)
 243 {
 244         struct hw_perf_event *hwc = &event->hw;
 245         struct riscv_pmu *rvpmu = to_riscv_pmu(event->pmu);
 246         uint64_t max_period = riscv_pmu_ctr_get_width_mask(event);
 247         u64 init_val;
 248
 249         if (flags & PERF_EF_RELOAD)
 250                 WARN_ON_ONCE(!(event->hw.state & PERF_HES_UPTODATE));
 251
 252         hwc->state = 0;
 253         riscv_pmu_event_set_period(event);
 254         init_val = local64_read(&hwc->prev_count) & max_period;
 255         rvpmu->ctr_start(event, init_val);
 256         perf_event_update_userpage(event);
 257 }
 258
 259 static int riscv_pmu_add(struct perf_event *event, int flags)
 260 {
 261         struct riscv_pmu *rvpmu = to_riscv_pmu(event->pmu);
 262         struct cpu_hw_events *cpuc = this_cpu_ptr(rvpmu->hw_events);
 263         struct hw_perf_event *hwc = &event->hw;
 264         int idx;
 265
 266         idx = rvpmu->ctr_get_idx(event);
 267         if (idx < 0)
 268                 return idx;
 269
 270         hwc->idx = idx;
 271         cpuc->events[idx] = event;
 272         cpuc->n_events++;
 273         hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
 274         if (flags & PERF_EF_START)
 275                 riscv_pmu_start(event, PERF_EF_RELOAD);
 276
 277         /* Propagate our changes to the userspace mapping. */
 278         perf_event_update_userpage(event);
 279
 280         return 0;
 281 }
 282
 283 static void riscv_pmu_del(struct perf_event *event, int flags)
 284 {
 285         struct riscv_pmu *rvpmu = to_riscv_pmu(event->pmu);
 286         struct cpu_hw_events *cpuc = this_cpu_ptr(rvpmu->hw_events);
 287         struct hw_perf_event *hwc = &event->hw;
 288
 289         riscv_pmu_stop(event, PERF_EF_UPDATE);
 290         cpuc->events[hwc->idx] = NULL;
 291         /* The firmware need to reset the counter mapping */
 292         if (rvpmu->ctr_stop)
 293                 rvpmu->ctr_stop(event, RISCV_PMU_STOP_FLAG_RESET);
 294         cpuc->n_events--;
 295         if (rvpmu->ctr_clear_idx)
 296                 rvpmu->ctr_clear_idx(event);
 297         perf_event_update_userpage(event);
 298         hwc->idx = -1;
 299 }
 300
 301 static void riscv_pmu_read(struct perf_event *event)
 302 {
 303         riscv_pmu_event_update(event);
 304 }
 305
 306 static int riscv_pmu_event_init(struct perf_event *event)
 307 {
 308         struct hw_perf_event *hwc = &event->hw;
 309         struct riscv_pmu *rvpmu = to_riscv_pmu(event->pmu);
 310         int mapped_event;
 311         u64 event_config = 0;
 312         uint64_t cmask;
 313
 314         /* driver does not support branch stack sampling */
 315         if (has_branch_stack(event))
 316                 return -EOPNOTSUPP;
 317
 318         hwc->flags = 0;
 319         mapped_event = rvpmu->event_map(event, &event_config);
 320         if (mapped_event < 0) {
 321                 pr_debug("event %x:%llx not supported\n", event->attr.type,
 322                          event->attr.config);
 323                 return mapped_event;
 324         }
 325
 326         /*
 327          * idx is set to -1 because the index of a general event should not be
 328          * decided until binding to some counter in pmu->add().
 329          * config will contain the information about counter CSR
 330          * the idx will contain the counter index
 331          */
 332         hwc->config = event_config;
 333         hwc->idx = -1;
 334         hwc->event_base = mapped_event;
 335
 336         if (rvpmu->event_init)
 337                 rvpmu->event_init(event);
 338
 339         if (!is_sampling_event(event)) {
 340                 /*
 341                  * For non-sampling runs, limit the sample_period to half
 342                  * of the counter width. That way, the new counter value
 343                  * is far less likely to overtake the previous one unless
 344                  * you have some serious IRQ latency issues.
 345                  */
 346                 cmask = riscv_pmu_ctr_get_width_mask(event);
 347                 hwc->sample_period  =  cmask >> 1;
 348                 hwc->last_period    = hwc->sample_period;
 349                 local64_set(&hwc->period_left, hwc->sample_period);
 350         }
 351
 352         return 0;
 353 }
 354
 355 static int riscv_pmu_event_idx(struct perf_event *event)
 356 {
 357         struct riscv_pmu *rvpmu = to_riscv_pmu(event->pmu);
 358
 359         if (!(event->hw.flags & PERF_EVENT_FLAG_USER_READ_CNT))
 360                 return 0;
 361
 362         if (rvpmu->csr_index)
 363                 return rvpmu->csr_index(event) + 1;
 364
 365         return 0;
 366 }
 367
 368 static void riscv_pmu_event_mapped(struct perf_event *event, struct mm_struct *mm)
 369 {
 370         struct riscv_pmu *rvpmu = to_riscv_pmu(event->pmu);
 371
 372         if (rvpmu->event_mapped) {
 373                 rvpmu->event_mapped(event, mm);
 374                 perf_event_update_userpage(event);
 375         }
 376 }
 377
 378 static void riscv_pmu_event_unmapped(struct perf_event *event, struct mm_struct *mm)
 379 {
 380         struct riscv_pmu *rvpmu = to_riscv_pmu(event->pmu);
 381
 382         if (rvpmu->event_unmapped) {
 383                 rvpmu->event_unmapped(event, mm);
 384                 perf_event_update_userpage(event);
 385         }
 386 }
 387
 388 struct riscv_pmu *riscv_pmu_alloc(void)
 389 {
 390         struct riscv_pmu *pmu;
 391         int cpuid, i;
 392         struct cpu_hw_events *cpuc;
 393
 394         pmu = kzalloc(sizeof(*pmu), GFP_KERNEL);
 395         if (!pmu)
 396                 goto out;
 397
 398         pmu->hw_events = alloc_percpu_gfp(struct cpu_hw_events, GFP_KERNEL);
 399         if (!pmu->hw_events) {
 400                 pr_info("failed to allocate per-cpu PMU data.\n");
 401                 goto out_free_pmu;
 402         }
 403
 404         for_each_possible_cpu(cpuid) {
 405                 cpuc = per_cpu_ptr(pmu->hw_events, cpuid);
 406                 cpuc->n_events = 0;
 407                 for (i = 0; i < RISCV_MAX_COUNTERS; i++)
 408                         cpuc->events[i] = NULL;
 409                 cpuc->snapshot_addr = NULL;
 410         }
 411         pmu->pmu = (struct pmu) {
 412                 .event_init     = riscv_pmu_event_init,
 413                 .event_mapped   = riscv_pmu_event_mapped,
 414                 .event_unmapped = riscv_pmu_event_unmapped,
 415                 .event_idx      = riscv_pmu_event_idx,
 416                 .add            = riscv_pmu_add,
 417                 .del            = riscv_pmu_del,
 418                 .start          = riscv_pmu_start,
 419                 .stop           = riscv_pmu_stop,
 420                 .read           = riscv_pmu_read,
 421         };
 422
 423         return pmu;
 424
 425 out_free_pmu:
 426         kfree(pmu);
 427 out:
 428         return NULL;
 429 }