drivers/perf/arm_spe_pmu.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Perf support for the Statistical Profiling Extension, introduced as
   4  * part of ARMv8.2.
   5  *
   6  * Copyright (C) 2016 ARM Limited
   7  *
   8  * Author: Will Deacon <[email protected]>
   9  */
  10
  11 #define PMUNAME                                 "arm_spe"
  12 #define DRVNAME                                 PMUNAME "_pmu"
  13 #define pr_fmt(fmt)                             DRVNAME ": " fmt
  14
  15 #include <linux/bitfield.h>
  16 #include <linux/bitops.h>
  17 #include <linux/bug.h>
  18 #include <linux/capability.h>
  19 #include <linux/cpuhotplug.h>
  20 #include <linux/cpumask.h>
  21 #include <linux/device.h>
  22 #include <linux/errno.h>
  23 #include <linux/interrupt.h>
  24 #include <linux/irq.h>
  25 #include <linux/kernel.h>
  26 #include <linux/list.h>
  27 #include <linux/module.h>
  28 #include <linux/of_address.h>
  29 #include <linux/of_device.h>
  30 #include <linux/perf_event.h>
  31 #include <linux/perf/arm_pmu.h>
  32 #include <linux/platform_device.h>
  33 #include <linux/printk.h>
  34 #include <linux/slab.h>
  35 #include <linux/smp.h>
  36 #include <linux/vmalloc.h>
  37
  38 #include <asm/barrier.h>
  39 #include <asm/cpufeature.h>
  40 #include <asm/mmu.h>
  41 #include <asm/sysreg.h>
  42
  43 /*
  44  * Cache if the event is allowed to trace Context information.
  45  * This allows us to perform the check, i.e, perfmon_capable(),
  46  * in the context of the event owner, once, during the event_init().
  47  */
  48 #define SPE_PMU_HW_FLAGS_CX                     0x00001
  49
  50 static_assert((PERF_EVENT_FLAG_ARCH & SPE_PMU_HW_FLAGS_CX) == SPE_PMU_HW_FLAGS_CX);
  51
  52 static void set_spe_event_has_cx(struct perf_event *event)
  53 {
  54         if (IS_ENABLED(CONFIG_PID_IN_CONTEXTIDR) && perfmon_capable())
  55                 event->hw.flags |= SPE_PMU_HW_FLAGS_CX;
  56 }
  57
  58 static bool get_spe_event_has_cx(struct perf_event *event)
  59 {
  60         return !!(event->hw.flags & SPE_PMU_HW_FLAGS_CX);
  61 }
  62
  63 #define ARM_SPE_BUF_PAD_BYTE                    0
  64
  65 struct arm_spe_pmu_buf {
  66         int                                     nr_pages;
  67         bool                                    snapshot;
  68         void                                    *base;
  69 };
  70
  71 struct arm_spe_pmu {
  72         struct pmu                              pmu;
  73         struct platform_device                  *pdev;
  74         cpumask_t                               supported_cpus;
  75         struct hlist_node                       hotplug_node;
  76
  77         int                                     irq; /* PPI */
  78         u16                                     pmsver;
  79         u16                                     min_period;
  80         u16                                     counter_sz;
  81
  82 #define SPE_PMU_FEAT_FILT_EVT                   (1UL << 0)
  83 #define SPE_PMU_FEAT_FILT_TYP                   (1UL << 1)
  84 #define SPE_PMU_FEAT_FILT_LAT                   (1UL << 2)
  85 #define SPE_PMU_FEAT_ARCH_INST                  (1UL << 3)
  86 #define SPE_PMU_FEAT_LDS                        (1UL << 4)
  87 #define SPE_PMU_FEAT_ERND                       (1UL << 5)
  88 #define SPE_PMU_FEAT_INV_FILT_EVT               (1UL << 6)
  89 #define SPE_PMU_FEAT_DEV_PROBED                 (1UL << 63)
  90         u64                                     features;
  91
  92         u16                                     max_record_sz;
  93         u16                                     align;
  94         struct perf_output_handle __percpu      *handle;
  95 };
  96
  97 #define to_spe_pmu(p) (container_of(p, struct arm_spe_pmu, pmu))
  98
  99 /* Convert a free-running index from perf into an SPE buffer offset */
 100 #define PERF_IDX2OFF(idx, buf)  ((idx) % ((buf)->nr_pages << PAGE_SHIFT))
 101
 102 /* Keep track of our dynamic hotplug state */
 103 static enum cpuhp_state arm_spe_pmu_online;
 104
 105 enum arm_spe_pmu_buf_fault_action {
 106         SPE_PMU_BUF_FAULT_ACT_SPURIOUS,
 107         SPE_PMU_BUF_FAULT_ACT_FATAL,
 108         SPE_PMU_BUF_FAULT_ACT_OK,
 109 };
 110
 111 /* This sysfs gunk was really good fun to write. */
 112 enum arm_spe_pmu_capabilities {
 113         SPE_PMU_CAP_ARCH_INST = 0,
 114         SPE_PMU_CAP_ERND,
 115         SPE_PMU_CAP_FEAT_MAX,
 116         SPE_PMU_CAP_CNT_SZ = SPE_PMU_CAP_FEAT_MAX,
 117         SPE_PMU_CAP_MIN_IVAL,
 118 };
 119
 120 static int arm_spe_pmu_feat_caps[SPE_PMU_CAP_FEAT_MAX] = {
 121         [SPE_PMU_CAP_ARCH_INST] = SPE_PMU_FEAT_ARCH_INST,
 122         [SPE_PMU_CAP_ERND]      = SPE_PMU_FEAT_ERND,
 123 };
 124
 125 static u32 arm_spe_pmu_cap_get(struct arm_spe_pmu *spe_pmu, int cap)
 126 {
 127         if (cap < SPE_PMU_CAP_FEAT_MAX)
 128                 return !!(spe_pmu->features & arm_spe_pmu_feat_caps[cap]);
 129
 130         switch (cap) {
 131         case SPE_PMU_CAP_CNT_SZ:
 132                 return spe_pmu->counter_sz;
 133         case SPE_PMU_CAP_MIN_IVAL:
 134                 return spe_pmu->min_period;
 135         default:
 136                 WARN(1, "unknown cap %d\n", cap);
 137         }
 138
 139         return 0;
 140 }
 141
 142 static ssize_t arm_spe_pmu_cap_show(struct device *dev,
 143                                     struct device_attribute *attr,
 144                                     char *buf)
 145 {
 146         struct arm_spe_pmu *spe_pmu = dev_get_drvdata(dev);
 147         struct dev_ext_attribute *ea =
 148                 container_of(attr, struct dev_ext_attribute, attr);
 149         int cap = (long)ea->var;
 150
 151         return sysfs_emit(buf, "%u\n", arm_spe_pmu_cap_get(spe_pmu, cap));
 152 }
 153
 154 #define SPE_EXT_ATTR_ENTRY(_name, _func, _var)                          \
 155         &((struct dev_ext_attribute[]) {                                \
 156                 { __ATTR(_name, S_IRUGO, _func, NULL), (void *)_var }   \
 157         })[0].attr.attr
 158
 159 #define SPE_CAP_EXT_ATTR_ENTRY(_name, _var)                             \
 160         SPE_EXT_ATTR_ENTRY(_name, arm_spe_pmu_cap_show, _var)
 161
 162 static struct attribute *arm_spe_pmu_cap_attr[] = {
 163         SPE_CAP_EXT_ATTR_ENTRY(arch_inst, SPE_PMU_CAP_ARCH_INST),
 164         SPE_CAP_EXT_ATTR_ENTRY(ernd, SPE_PMU_CAP_ERND),
 165         SPE_CAP_EXT_ATTR_ENTRY(count_size, SPE_PMU_CAP_CNT_SZ),
 166         SPE_CAP_EXT_ATTR_ENTRY(min_interval, SPE_PMU_CAP_MIN_IVAL),
 167         NULL,
 168 };
 169
 170 static const struct attribute_group arm_spe_pmu_cap_group = {
 171         .name   = "caps",
 172         .attrs  = arm_spe_pmu_cap_attr,
 173 };
 174
 175 /* User ABI */
 176 #define ATTR_CFG_FLD_ts_enable_CFG              config  /* PMSCR_EL1.TS */
 177 #define ATTR_CFG_FLD_ts_enable_LO               0
 178 #define ATTR_CFG_FLD_ts_enable_HI               0
 179 #define ATTR_CFG_FLD_pa_enable_CFG              config  /* PMSCR_EL1.PA */
 180 #define ATTR_CFG_FLD_pa_enable_LO               1
 181 #define ATTR_CFG_FLD_pa_enable_HI               1
 182 #define ATTR_CFG_FLD_pct_enable_CFG             config  /* PMSCR_EL1.PCT */
 183 #define ATTR_CFG_FLD_pct_enable_LO              2
 184 #define ATTR_CFG_FLD_pct_enable_HI              2
 185 #define ATTR_CFG_FLD_jitter_CFG                 config  /* PMSIRR_EL1.RND */
 186 #define ATTR_CFG_FLD_jitter_LO                  16
 187 #define ATTR_CFG_FLD_jitter_HI                  16
 188 #define ATTR_CFG_FLD_branch_filter_CFG          config  /* PMSFCR_EL1.B */
 189 #define ATTR_CFG_FLD_branch_filter_LO           32
 190 #define ATTR_CFG_FLD_branch_filter_HI           32
 191 #define ATTR_CFG_FLD_load_filter_CFG            config  /* PMSFCR_EL1.LD */
 192 #define ATTR_CFG_FLD_load_filter_LO             33
 193 #define ATTR_CFG_FLD_load_filter_HI             33
 194 #define ATTR_CFG_FLD_store_filter_CFG           config  /* PMSFCR_EL1.ST */
 195 #define ATTR_CFG_FLD_store_filter_LO            34
 196 #define ATTR_CFG_FLD_store_filter_HI            34
 197
 198 #define ATTR_CFG_FLD_event_filter_CFG           config1 /* PMSEVFR_EL1 */
 199 #define ATTR_CFG_FLD_event_filter_LO            0
 200 #define ATTR_CFG_FLD_event_filter_HI            63
 201
 202 #define ATTR_CFG_FLD_min_latency_CFG            config2 /* PMSLATFR_EL1.MINLAT */
 203 #define ATTR_CFG_FLD_min_latency_LO             0
 204 #define ATTR_CFG_FLD_min_latency_HI             11
 205
 206 #define ATTR_CFG_FLD_inv_event_filter_CFG       config3 /* PMSNEVFR_EL1 */
 207 #define ATTR_CFG_FLD_inv_event_filter_LO        0
 208 #define ATTR_CFG_FLD_inv_event_filter_HI        63
 209
 210 /* Why does everything I do descend into this? */
 211 #define __GEN_PMU_FORMAT_ATTR(cfg, lo, hi)                              \
 212         (lo) == (hi) ? #cfg ":" #lo "\n" : #cfg ":" #lo "-" #hi
 213
 214 #define _GEN_PMU_FORMAT_ATTR(cfg, lo, hi)                               \
 215         __GEN_PMU_FORMAT_ATTR(cfg, lo, hi)
 216
 217 #define GEN_PMU_FORMAT_ATTR(name)                                       \
 218         PMU_FORMAT_ATTR(name,                                           \
 219         _GEN_PMU_FORMAT_ATTR(ATTR_CFG_FLD_##name##_CFG,                 \
 220                              ATTR_CFG_FLD_##name##_LO,                  \
 221                              ATTR_CFG_FLD_##name##_HI))
 222
 223 #define _ATTR_CFG_GET_FLD(attr, cfg, lo, hi)                            \
 224         ((((attr)->cfg) >> lo) & GENMASK(hi - lo, 0))
 225
 226 #define ATTR_CFG_GET_FLD(attr, name)                                    \
 227         _ATTR_CFG_GET_FLD(attr,                                         \
 228                           ATTR_CFG_FLD_##name##_CFG,                    \
 229                           ATTR_CFG_FLD_##name##_LO,                     \
 230                           ATTR_CFG_FLD_##name##_HI)
 231
 232 GEN_PMU_FORMAT_ATTR(ts_enable);
 233 GEN_PMU_FORMAT_ATTR(pa_enable);
 234 GEN_PMU_FORMAT_ATTR(pct_enable);
 235 GEN_PMU_FORMAT_ATTR(jitter);
 236 GEN_PMU_FORMAT_ATTR(branch_filter);
 237 GEN_PMU_FORMAT_ATTR(load_filter);
 238 GEN_PMU_FORMAT_ATTR(store_filter);
 239 GEN_PMU_FORMAT_ATTR(event_filter);
 240 GEN_PMU_FORMAT_ATTR(inv_event_filter);
 241 GEN_PMU_FORMAT_ATTR(min_latency);
 242
 243 static struct attribute *arm_spe_pmu_formats_attr[] = {
 244         &format_attr_ts_enable.attr,
 245         &format_attr_pa_enable.attr,
 246         &format_attr_pct_enable.attr,
 247         &format_attr_jitter.attr,
 248         &format_attr_branch_filter.attr,
 249         &format_attr_load_filter.attr,
 250         &format_attr_store_filter.attr,
 251         &format_attr_event_filter.attr,
 252         &format_attr_inv_event_filter.attr,
 253         &format_attr_min_latency.attr,
 254         NULL,
 255 };
 256
 257 static umode_t arm_spe_pmu_format_attr_is_visible(struct kobject *kobj,
 258                                                   struct attribute *attr,
 259                                                   int unused)
 260         {
 261         struct device *dev = kobj_to_dev(kobj);
 262         struct arm_spe_pmu *spe_pmu = dev_get_drvdata(dev);
 263
 264         if (attr == &format_attr_inv_event_filter.attr && !(spe_pmu->features & SPE_PMU_FEAT_INV_FILT_EVT))
 265                 return 0;
 266
 267         return attr->mode;
 268 }
 269
 270 static const struct attribute_group arm_spe_pmu_format_group = {
 271         .name   = "format",
 272         .is_visible = arm_spe_pmu_format_attr_is_visible,
 273         .attrs  = arm_spe_pmu_formats_attr,
 274 };
 275
 276 static ssize_t cpumask_show(struct device *dev,
 277                             struct device_attribute *attr, char *buf)
 278 {
 279         struct arm_spe_pmu *spe_pmu = dev_get_drvdata(dev);
 280
 281         return cpumap_print_to_pagebuf(true, buf, &spe_pmu->supported_cpus);
 282 }
 283 static DEVICE_ATTR_RO(cpumask);
 284
 285 static struct attribute *arm_spe_pmu_attrs[] = {
 286         &dev_attr_cpumask.attr,
 287         NULL,
 288 };
 289
 290 static const struct attribute_group arm_spe_pmu_group = {
 291         .attrs  = arm_spe_pmu_attrs,
 292 };
 293
 294 static const struct attribute_group *arm_spe_pmu_attr_groups[] = {
 295         &arm_spe_pmu_group,
 296         &arm_spe_pmu_cap_group,
 297         &arm_spe_pmu_format_group,
 298         NULL,
 299 };
 300
 301 /* Convert between user ABI and register values */
 302 static u64 arm_spe_event_to_pmscr(struct perf_event *event)
 303 {
 304         struct perf_event_attr *attr = &event->attr;
 305         u64 reg = 0;
 306
 307         reg |= FIELD_PREP(PMSCR_EL1_TS, ATTR_CFG_GET_FLD(attr, ts_enable));
 308         reg |= FIELD_PREP(PMSCR_EL1_PA, ATTR_CFG_GET_FLD(attr, pa_enable));
 309         reg |= FIELD_PREP(PMSCR_EL1_PCT, ATTR_CFG_GET_FLD(attr, pct_enable));
 310
 311         if (!attr->exclude_user)
 312                 reg |= PMSCR_EL1_E0SPE;
 313
 314         if (!attr->exclude_kernel)
 315                 reg |= PMSCR_EL1_E1SPE;
 316
 317         if (get_spe_event_has_cx(event))
 318                 reg |= PMSCR_EL1_CX;
 319
 320         return reg;
 321 }
 322
 323 static void arm_spe_event_sanitise_period(struct perf_event *event)
 324 {
 325         struct arm_spe_pmu *spe_pmu = to_spe_pmu(event->pmu);
 326         u64 period = event->hw.sample_period;
 327         u64 max_period = PMSIRR_EL1_INTERVAL_MASK;
 328
 329         if (period < spe_pmu->min_period)
 330                 period = spe_pmu->min_period;
 331         else if (period > max_period)
 332                 period = max_period;
 333         else
 334                 period &= max_period;
 335
 336         event->hw.sample_period = period;
 337 }
 338
 339 static u64 arm_spe_event_to_pmsirr(struct perf_event *event)
 340 {
 341         struct perf_event_attr *attr = &event->attr;
 342         u64 reg = 0;
 343
 344         arm_spe_event_sanitise_period(event);
 345
 346         reg |= FIELD_PREP(PMSIRR_EL1_RND, ATTR_CFG_GET_FLD(attr, jitter));
 347         reg |= event->hw.sample_period;
 348
 349         return reg;
 350 }
 351
 352 static u64 arm_spe_event_to_pmsfcr(struct perf_event *event)
 353 {
 354         struct perf_event_attr *attr = &event->attr;
 355         u64 reg = 0;
 356
 357         reg |= FIELD_PREP(PMSFCR_EL1_LD, ATTR_CFG_GET_FLD(attr, load_filter));
 358         reg |= FIELD_PREP(PMSFCR_EL1_ST, ATTR_CFG_GET_FLD(attr, store_filter));
 359         reg |= FIELD_PREP(PMSFCR_EL1_B, ATTR_CFG_GET_FLD(attr, branch_filter));
 360
 361         if (reg)
 362                 reg |= PMSFCR_EL1_FT;
 363
 364         if (ATTR_CFG_GET_FLD(attr, event_filter))
 365                 reg |= PMSFCR_EL1_FE;
 366
 367         if (ATTR_CFG_GET_FLD(attr, inv_event_filter))
 368                 reg |= PMSFCR_EL1_FnE;
 369
 370         if (ATTR_CFG_GET_FLD(attr, min_latency))
 371                 reg |= PMSFCR_EL1_FL;
 372
 373         return reg;
 374 }
 375
 376 static u64 arm_spe_event_to_pmsevfr(struct perf_event *event)
 377 {
 378         struct perf_event_attr *attr = &event->attr;
 379         return ATTR_CFG_GET_FLD(attr, event_filter);
 380 }
 381
 382 static u64 arm_spe_event_to_pmsnevfr(struct perf_event *event)
 383 {
 384         struct perf_event_attr *attr = &event->attr;
 385         return ATTR_CFG_GET_FLD(attr, inv_event_filter);
 386 }
 387
 388 static u64 arm_spe_event_to_pmslatfr(struct perf_event *event)
 389 {
 390         struct perf_event_attr *attr = &event->attr;
 391         return FIELD_PREP(PMSLATFR_EL1_MINLAT, ATTR_CFG_GET_FLD(attr, min_latency));
 392 }
 393
 394 static void arm_spe_pmu_pad_buf(struct perf_output_handle *handle, int len)
 395 {
 396         struct arm_spe_pmu_buf *buf = perf_get_aux(handle);
 397         u64 head = PERF_IDX2OFF(handle->head, buf);
 398
 399         memset(buf->base + head, ARM_SPE_BUF_PAD_BYTE, len);
 400         if (!buf->snapshot)
 401                 perf_aux_output_skip(handle, len);
 402 }
 403
 404 static u64 arm_spe_pmu_next_snapshot_off(struct perf_output_handle *handle)
 405 {
 406         struct arm_spe_pmu_buf *buf = perf_get_aux(handle);
 407         struct arm_spe_pmu *spe_pmu = to_spe_pmu(handle->event->pmu);
 408         u64 head = PERF_IDX2OFF(handle->head, buf);
 409         u64 limit = buf->nr_pages * PAGE_SIZE;
 410
 411         /*
 412          * The trace format isn't parseable in reverse, so clamp
 413          * the limit to half of the buffer size in snapshot mode
 414          * so that the worst case is half a buffer of records, as
 415          * opposed to a single record.
 416          */
 417         if (head < limit >> 1)
 418                 limit >>= 1;
 419
 420         /*
 421          * If we're within max_record_sz of the limit, we must
 422          * pad, move the head index and recompute the limit.
 423          */
 424         if (limit - head < spe_pmu->max_record_sz) {
 425                 arm_spe_pmu_pad_buf(handle, limit - head);
 426                 handle->head = PERF_IDX2OFF(limit, buf);
 427                 limit = ((buf->nr_pages * PAGE_SIZE) >> 1) + handle->head;
 428         }
 429
 430         return limit;
 431 }
 432
 433 static u64 __arm_spe_pmu_next_off(struct perf_output_handle *handle)
 434 {
 435         struct arm_spe_pmu *spe_pmu = to_spe_pmu(handle->event->pmu);
 436         struct arm_spe_pmu_buf *buf = perf_get_aux(handle);
 437         const u64 bufsize = buf->nr_pages * PAGE_SIZE;
 438         u64 limit = bufsize;
 439         u64 head, tail, wakeup;
 440
 441         /*
 442          * The head can be misaligned for two reasons:
 443          *
 444          * 1. The hardware left PMBPTR pointing to the first byte after
 445          *    a record when generating a buffer management event.
 446          *
 447          * 2. We used perf_aux_output_skip to consume handle->size bytes
 448          *    and CIRC_SPACE was used to compute the size, which always
 449          *    leaves one entry free.
 450          *
 451          * Deal with this by padding to the next alignment boundary and
 452          * moving the head index. If we run out of buffer space, we'll
 453          * reduce handle->size to zero and end up reporting truncation.
 454          */
 455         head = PERF_IDX2OFF(handle->head, buf);
 456         if (!IS_ALIGNED(head, spe_pmu->align)) {
 457                 unsigned long delta = roundup(head, spe_pmu->align) - head;
 458
 459                 delta = min(delta, handle->size);
 460                 arm_spe_pmu_pad_buf(handle, delta);
 461                 head = PERF_IDX2OFF(handle->head, buf);
 462         }
 463
 464         /* If we've run out of free space, then nothing more to do */
 465         if (!handle->size)
 466                 goto no_space;
 467
 468         /* Compute the tail and wakeup indices now that we've aligned head */
 469         tail = PERF_IDX2OFF(handle->head + handle->size, buf);
 470         wakeup = PERF_IDX2OFF(handle->wakeup, buf);
 471
 472         /*
 473          * Avoid clobbering unconsumed data. We know we have space, so
 474          * if we see head == tail we know that the buffer is empty. If
 475          * head > tail, then there's nothing to clobber prior to
 476          * wrapping.
 477          */
 478         if (head < tail)
 479                 limit = round_down(tail, PAGE_SIZE);
 480
 481         /*
 482          * Wakeup may be arbitrarily far into the future. If it's not in
 483          * the current generation, either we'll wrap before hitting it,
 484          * or it's in the past and has been handled already.
 485          *
 486          * If there's a wakeup before we wrap, arrange to be woken up by
 487          * the page boundary following it. Keep the tail boundary if
 488          * that's lower.
 489          */
 490         if (handle->wakeup < (handle->head + handle->size) && head <= wakeup)
 491                 limit = min(limit, round_up(wakeup, PAGE_SIZE));
 492
 493         if (limit > head)
 494                 return limit;
 495
 496         arm_spe_pmu_pad_buf(handle, handle->size);
 497 no_space:
 498         perf_aux_output_flag(handle, PERF_AUX_FLAG_TRUNCATED);
 499         perf_aux_output_end(handle, 0);
 500         return 0;
 501 }
 502
 503 static u64 arm_spe_pmu_next_off(struct perf_output_handle *handle)
 504 {
 505         struct arm_spe_pmu_buf *buf = perf_get_aux(handle);
 506         struct arm_spe_pmu *spe_pmu = to_spe_pmu(handle->event->pmu);
 507         u64 limit = __arm_spe_pmu_next_off(handle);
 508         u64 head = PERF_IDX2OFF(handle->head, buf);
 509
 510         /*
 511          * If the head has come too close to the end of the buffer,
 512          * then pad to the end and recompute the limit.
 513          */
 514         if (limit && (limit - head < spe_pmu->max_record_sz)) {
 515                 arm_spe_pmu_pad_buf(handle, limit - head);
 516                 limit = __arm_spe_pmu_next_off(handle);
 517         }
 518
 519         return limit;
 520 }
 521
 522 static void arm_spe_perf_aux_output_begin(struct perf_output_handle *handle,
 523                                           struct perf_event *event)
 524 {
 525         u64 base, limit;
 526         struct arm_spe_pmu_buf *buf;
 527
 528         /* Start a new aux session */
 529         buf = perf_aux_output_begin(handle, event);
 530         if (!buf) {
 531                 event->hw.state |= PERF_HES_STOPPED;
 532                 /*
 533                  * We still need to clear the limit pointer, since the
 534                  * profiler might only be disabled by virtue of a fault.
 535                  */
 536                 limit = 0;
 537                 goto out_write_limit;
 538         }
 539
 540         limit = buf->snapshot ? arm_spe_pmu_next_snapshot_off(handle)
 541                               : arm_spe_pmu_next_off(handle);
 542         if (limit)
 543                 limit |= PMBLIMITR_EL1_E;
 544
 545         limit += (u64)buf->base;
 546         base = (u64)buf->base + PERF_IDX2OFF(handle->head, buf);
 547         write_sysreg_s(base, SYS_PMBPTR_EL1);
 548
 549 out_write_limit:
 550         write_sysreg_s(limit, SYS_PMBLIMITR_EL1);
 551 }
 552
 553 static void arm_spe_perf_aux_output_end(struct perf_output_handle *handle)
 554 {
 555         struct arm_spe_pmu_buf *buf = perf_get_aux(handle);
 556         u64 offset, size;
 557
 558         offset = read_sysreg_s(SYS_PMBPTR_EL1) - (u64)buf->base;
 559         size = offset - PERF_IDX2OFF(handle->head, buf);
 560
 561         if (buf->snapshot)
 562                 handle->head = offset;
 563
 564         perf_aux_output_end(handle, size);
 565 }
 566
 567 static void arm_spe_pmu_disable_and_drain_local(void)
 568 {
 569         /* Disable profiling at EL0 and EL1 */
 570         write_sysreg_s(0, SYS_PMSCR_EL1);
 571         isb();
 572
 573         /* Drain any buffered data */
 574         psb_csync();
 575         dsb(nsh);
 576
 577         /* Disable the profiling buffer */
 578         write_sysreg_s(0, SYS_PMBLIMITR_EL1);
 579         isb();
 580 }
 581
 582 /* IRQ handling */
 583 static enum arm_spe_pmu_buf_fault_action
 584 arm_spe_pmu_buf_get_fault_act(struct perf_output_handle *handle)
 585 {
 586         const char *err_str;
 587         u64 pmbsr;
 588         enum arm_spe_pmu_buf_fault_action ret;
 589
 590         /*
 591          * Ensure new profiling data is visible to the CPU and any external
 592          * aborts have been resolved.
 593          */
 594         psb_csync();
 595         dsb(nsh);
 596
 597         /* Ensure hardware updates to PMBPTR_EL1 are visible */
 598         isb();
 599
 600         /* Service required? */
 601         pmbsr = read_sysreg_s(SYS_PMBSR_EL1);
 602         if (!FIELD_GET(PMBSR_EL1_S, pmbsr))
 603                 return SPE_PMU_BUF_FAULT_ACT_SPURIOUS;
 604
 605         /*
 606          * If we've lost data, disable profiling and also set the PARTIAL
 607          * flag to indicate that the last record is corrupted.
 608          */
 609         if (FIELD_GET(PMBSR_EL1_DL, pmbsr))
 610                 perf_aux_output_flag(handle, PERF_AUX_FLAG_TRUNCATED |
 611                                              PERF_AUX_FLAG_PARTIAL);
 612
 613         /* Report collisions to userspace so that it can up the period */
 614         if (FIELD_GET(PMBSR_EL1_COLL, pmbsr))
 615                 perf_aux_output_flag(handle, PERF_AUX_FLAG_COLLISION);
 616
 617         /* We only expect buffer management events */
 618         switch (FIELD_GET(PMBSR_EL1_EC, pmbsr)) {
 619         case PMBSR_EL1_EC_BUF:
 620                 /* Handled below */
 621                 break;
 622         case PMBSR_EL1_EC_FAULT_S1:
 623         case PMBSR_EL1_EC_FAULT_S2:
 624                 err_str = "Unexpected buffer fault";
 625                 goto out_err;
 626         default:
 627                 err_str = "Unknown error code";
 628                 goto out_err;
 629         }
 630
 631         /* Buffer management event */
 632         switch (FIELD_GET(PMBSR_EL1_BUF_BSC_MASK, pmbsr)) {
 633         case PMBSR_EL1_BUF_BSC_FULL:
 634                 ret = SPE_PMU_BUF_FAULT_ACT_OK;
 635                 goto out_stop;
 636         default:
 637                 err_str = "Unknown buffer status code";
 638         }
 639
 640 out_err:
 641         pr_err_ratelimited("%s on CPU %d [PMBSR=0x%016llx, PMBPTR=0x%016llx, PMBLIMITR=0x%016llx]\n",
 642                            err_str, smp_processor_id(), pmbsr,
 643                            read_sysreg_s(SYS_PMBPTR_EL1),
 644                            read_sysreg_s(SYS_PMBLIMITR_EL1));
 645         ret = SPE_PMU_BUF_FAULT_ACT_FATAL;
 646
 647 out_stop:
 648         arm_spe_perf_aux_output_end(handle);
 649         return ret;
 650 }
 651
 652 static irqreturn_t arm_spe_pmu_irq_handler(int irq, void *dev)
 653 {
 654         struct perf_output_handle *handle = dev;
 655         struct perf_event *event = handle->event;
 656         enum arm_spe_pmu_buf_fault_action act;
 657
 658         if (!perf_get_aux(handle))
 659                 return IRQ_NONE;
 660
 661         act = arm_spe_pmu_buf_get_fault_act(handle);
 662         if (act == SPE_PMU_BUF_FAULT_ACT_SPURIOUS)
 663                 return IRQ_NONE;
 664
 665         /*
 666          * Ensure perf callbacks have completed, which may disable the
 667          * profiling buffer in response to a TRUNCATION flag.
 668          */
 669         irq_work_run();
 670
 671         switch (act) {
 672         case SPE_PMU_BUF_FAULT_ACT_FATAL:
 673                 /*
 674                  * If a fatal exception occurred then leaving the profiling
 675                  * buffer enabled is a recipe waiting to happen. Since
 676                  * fatal faults don't always imply truncation, make sure
 677                  * that the profiling buffer is disabled explicitly before
 678                  * clearing the syndrome register.
 679                  */
 680                 arm_spe_pmu_disable_and_drain_local();
 681                 break;
 682         case SPE_PMU_BUF_FAULT_ACT_OK:
 683                 /*
 684                  * We handled the fault (the buffer was full), so resume
 685                  * profiling as long as we didn't detect truncation.
 686                  * PMBPTR might be misaligned, but we'll burn that bridge
 687                  * when we get to it.
 688                  */
 689                 if (!(handle->aux_flags & PERF_AUX_FLAG_TRUNCATED)) {
 690                         arm_spe_perf_aux_output_begin(handle, event);
 691                         isb();
 692                 }
 693                 break;
 694         case SPE_PMU_BUF_FAULT_ACT_SPURIOUS:
 695                 /* We've seen you before, but GCC has the memory of a sieve. */
 696                 break;
 697         }
 698
 699         /* The buffer pointers are now sane, so resume profiling. */
 700         write_sysreg_s(0, SYS_PMBSR_EL1);
 701         return IRQ_HANDLED;
 702 }
 703
 704 static u64 arm_spe_pmsevfr_res0(u16 pmsver)
 705 {
 706         switch (pmsver) {
 707         case ID_AA64DFR0_EL1_PMSVer_IMP:
 708                 return PMSEVFR_EL1_RES0_IMP;
 709         case ID_AA64DFR0_EL1_PMSVer_V1P1:
 710                 return PMSEVFR_EL1_RES0_V1P1;
 711         case ID_AA64DFR0_EL1_PMSVer_V1P2:
 712         /* Return the highest version we support in default */
 713         default:
 714                 return PMSEVFR_EL1_RES0_V1P2;
 715         }
 716 }
 717
 718 /* Perf callbacks */
 719 static int arm_spe_pmu_event_init(struct perf_event *event)
 720 {
 721         u64 reg;
 722         struct perf_event_attr *attr = &event->attr;
 723         struct arm_spe_pmu *spe_pmu = to_spe_pmu(event->pmu);
 724
 725         /* This is, of course, deeply driver-specific */
 726         if (attr->type != event->pmu->type)
 727                 return -ENOENT;
 728
 729         if (event->cpu >= 0 &&
 730             !cpumask_test_cpu(event->cpu, &spe_pmu->supported_cpus))
 731                 return -ENOENT;
 732
 733         if (arm_spe_event_to_pmsevfr(event) & arm_spe_pmsevfr_res0(spe_pmu->pmsver))
 734                 return -EOPNOTSUPP;
 735
 736         if (arm_spe_event_to_pmsnevfr(event) & arm_spe_pmsevfr_res0(spe_pmu->pmsver))
 737                 return -EOPNOTSUPP;
 738
 739         if (attr->exclude_idle)
 740                 return -EOPNOTSUPP;
 741
 742         /*
 743          * Feedback-directed frequency throttling doesn't work when we
 744          * have a buffer of samples. We'd need to manually count the
 745          * samples in the buffer when it fills up and adjust the event
 746          * count to reflect that. Instead, just force the user to specify
 747          * a sample period.
 748          */
 749         if (attr->freq)
 750                 return -EINVAL;
 751
 752         reg = arm_spe_event_to_pmsfcr(event);
 753         if ((FIELD_GET(PMSFCR_EL1_FE, reg)) &&
 754             !(spe_pmu->features & SPE_PMU_FEAT_FILT_EVT))
 755                 return -EOPNOTSUPP;
 756
 757         if ((FIELD_GET(PMSFCR_EL1_FnE, reg)) &&
 758             !(spe_pmu->features & SPE_PMU_FEAT_INV_FILT_EVT))
 759                 return -EOPNOTSUPP;
 760
 761         if ((FIELD_GET(PMSFCR_EL1_FT, reg)) &&
 762             !(spe_pmu->features & SPE_PMU_FEAT_FILT_TYP))
 763                 return -EOPNOTSUPP;
 764
 765         if ((FIELD_GET(PMSFCR_EL1_FL, reg)) &&
 766             !(spe_pmu->features & SPE_PMU_FEAT_FILT_LAT))
 767                 return -EOPNOTSUPP;
 768
 769         set_spe_event_has_cx(event);
 770         reg = arm_spe_event_to_pmscr(event);
 771         if (!perfmon_capable() &&
 772             (reg & (PMSCR_EL1_PA | PMSCR_EL1_PCT)))
 773                 return -EACCES;
 774
 775         return 0;
 776 }
 777
 778 static void arm_spe_pmu_start(struct perf_event *event, int flags)
 779 {
 780         u64 reg;
 781         struct arm_spe_pmu *spe_pmu = to_spe_pmu(event->pmu);
 782         struct hw_perf_event *hwc = &event->hw;
 783         struct perf_output_handle *handle = this_cpu_ptr(spe_pmu->handle);
 784
 785         hwc->state = 0;
 786         arm_spe_perf_aux_output_begin(handle, event);
 787         if (hwc->state)
 788                 return;
 789
 790         reg = arm_spe_event_to_pmsfcr(event);
 791         write_sysreg_s(reg, SYS_PMSFCR_EL1);
 792
 793         reg = arm_spe_event_to_pmsevfr(event);
 794         write_sysreg_s(reg, SYS_PMSEVFR_EL1);
 795
 796         if (spe_pmu->features & SPE_PMU_FEAT_INV_FILT_EVT) {
 797                 reg = arm_spe_event_to_pmsnevfr(event);
 798                 write_sysreg_s(reg, SYS_PMSNEVFR_EL1);
 799         }
 800
 801         reg = arm_spe_event_to_pmslatfr(event);
 802         write_sysreg_s(reg, SYS_PMSLATFR_EL1);
 803
 804         if (flags & PERF_EF_RELOAD) {
 805                 reg = arm_spe_event_to_pmsirr(event);
 806                 write_sysreg_s(reg, SYS_PMSIRR_EL1);
 807                 isb();
 808                 reg = local64_read(&hwc->period_left);
 809                 write_sysreg_s(reg, SYS_PMSICR_EL1);
 810         }
 811
 812         reg = arm_spe_event_to_pmscr(event);
 813         isb();
 814         write_sysreg_s(reg, SYS_PMSCR_EL1);
 815 }
 816
 817 static void arm_spe_pmu_stop(struct perf_event *event, int flags)
 818 {
 819         struct arm_spe_pmu *spe_pmu = to_spe_pmu(event->pmu);
 820         struct hw_perf_event *hwc = &event->hw;
 821         struct perf_output_handle *handle = this_cpu_ptr(spe_pmu->handle);
 822
 823         /* If we're already stopped, then nothing to do */
 824         if (hwc->state & PERF_HES_STOPPED)
 825                 return;
 826
 827         /* Stop all trace generation */
 828         arm_spe_pmu_disable_and_drain_local();
 829
 830         if (flags & PERF_EF_UPDATE) {
 831                 /*
 832                  * If there's a fault pending then ensure we contain it
 833                  * to this buffer, since we might be on the context-switch
 834                  * path.
 835                  */
 836                 if (perf_get_aux(handle)) {
 837                         enum arm_spe_pmu_buf_fault_action act;
 838
 839                         act = arm_spe_pmu_buf_get_fault_act(handle);
 840                         if (act == SPE_PMU_BUF_FAULT_ACT_SPURIOUS)
 841                                 arm_spe_perf_aux_output_end(handle);
 842                         else
 843                                 write_sysreg_s(0, SYS_PMBSR_EL1);
 844                 }
 845
 846                 /*
 847                  * This may also contain ECOUNT, but nobody else should
 848                  * be looking at period_left, since we forbid frequency
 849                  * based sampling.
 850                  */
 851                 local64_set(&hwc->period_left, read_sysreg_s(SYS_PMSICR_EL1));
 852                 hwc->state |= PERF_HES_UPTODATE;
 853         }
 854
 855         hwc->state |= PERF_HES_STOPPED;
 856 }
 857
 858 static int arm_spe_pmu_add(struct perf_event *event, int flags)
 859 {
 860         int ret = 0;
 861         struct arm_spe_pmu *spe_pmu = to_spe_pmu(event->pmu);
 862         struct hw_perf_event *hwc = &event->hw;
 863         int cpu = event->cpu == -1 ? smp_processor_id() : event->cpu;
 864
 865         if (!cpumask_test_cpu(cpu, &spe_pmu->supported_cpus))
 866                 return -ENOENT;
 867
 868         hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED;
 869
 870         if (flags & PERF_EF_START) {
 871                 arm_spe_pmu_start(event, PERF_EF_RELOAD);
 872                 if (hwc->state & PERF_HES_STOPPED)
 873                         ret = -EINVAL;
 874         }
 875
 876         return ret;
 877 }
 878
 879 static void arm_spe_pmu_del(struct perf_event *event, int flags)
 880 {
 881         arm_spe_pmu_stop(event, PERF_EF_UPDATE);
 882 }
 883
 884 static void arm_spe_pmu_read(struct perf_event *event)
 885 {
 886 }
 887
 888 static void *arm_spe_pmu_setup_aux(struct perf_event *event, void **pages,
 889                                    int nr_pages, bool snapshot)
 890 {
 891         int i, cpu = event->cpu;
 892         struct page **pglist;
 893         struct arm_spe_pmu_buf *buf;
 894
 895         /* We need at least two pages for this to work. */
 896         if (nr_pages < 2)
 897                 return NULL;
 898
 899         /*
 900          * We require an even number of pages for snapshot mode, so that
 901          * we can effectively treat the buffer as consisting of two equal
 902          * parts and give userspace a fighting chance of getting some
 903          * useful data out of it.
 904          */
 905         if (snapshot && (nr_pages & 1))
 906                 return NULL;
 907
 908         if (cpu == -1)
 909                 cpu = raw_smp_processor_id();
 910
 911         buf = kzalloc_node(sizeof(*buf), GFP_KERNEL, cpu_to_node(cpu));
 912         if (!buf)
 913                 return NULL;
 914
 915         pglist = kcalloc(nr_pages, sizeof(*pglist), GFP_KERNEL);
 916         if (!pglist)
 917                 goto out_free_buf;
 918
 919         for (i = 0; i < nr_pages; ++i)
 920                 pglist[i] = virt_to_page(pages[i]);
 921
 922         buf->base = vmap(pglist, nr_pages, VM_MAP, PAGE_KERNEL);
 923         if (!buf->base)
 924                 goto out_free_pglist;
 925
 926         buf->nr_pages   = nr_pages;
 927         buf->snapshot   = snapshot;
 928
 929         kfree(pglist);
 930         return buf;
 931
 932 out_free_pglist:
 933         kfree(pglist);
 934 out_free_buf:
 935         kfree(buf);
 936         return NULL;
 937 }
 938
 939 static void arm_spe_pmu_free_aux(void *aux)
 940 {
 941         struct arm_spe_pmu_buf *buf = aux;
 942
 943         vunmap(buf->base);
 944         kfree(buf);
 945 }
 946
 947 /* Initialisation and teardown functions */
 948 static int arm_spe_pmu_perf_init(struct arm_spe_pmu *spe_pmu)
 949 {
 950         static atomic_t pmu_idx = ATOMIC_INIT(-1);
 951
 952         int idx;
 953         char *name;
 954         struct device *dev = &spe_pmu->pdev->dev;
 955
 956         spe_pmu->pmu = (struct pmu) {
 957                 .module = THIS_MODULE,
 958                 .capabilities   = PERF_PMU_CAP_EXCLUSIVE | PERF_PMU_CAP_ITRACE,
 959                 .attr_groups    = arm_spe_pmu_attr_groups,
 960                 /*
 961                  * We hitch a ride on the software context here, so that
 962                  * we can support per-task profiling (which is not possible
 963                  * with the invalid context as it doesn't get sched callbacks).
 964                  * This requires that userspace either uses a dummy event for
 965                  * perf_event_open, since the aux buffer is not setup until
 966                  * a subsequent mmap, or creates the profiling event in a
 967                  * disabled state and explicitly PERF_EVENT_IOC_ENABLEs it
 968                  * once the buffer has been created.
 969                  */
 970                 .task_ctx_nr    = perf_sw_context,
 971                 .event_init     = arm_spe_pmu_event_init,
 972                 .add            = arm_spe_pmu_add,
 973                 .del            = arm_spe_pmu_del,
 974                 .start          = arm_spe_pmu_start,
 975                 .stop           = arm_spe_pmu_stop,
 976                 .read           = arm_spe_pmu_read,
 977                 .setup_aux      = arm_spe_pmu_setup_aux,
 978                 .free_aux       = arm_spe_pmu_free_aux,
 979         };
 980
 981         idx = atomic_inc_return(&pmu_idx);
 982         name = devm_kasprintf(dev, GFP_KERNEL, "%s_%d", PMUNAME, idx);
 983         if (!name) {
 984                 dev_err(dev, "failed to allocate name for pmu %d\n", idx);
 985                 return -ENOMEM;
 986         }
 987
 988         return perf_pmu_register(&spe_pmu->pmu, name, -1);
 989 }
 990
 991 static void arm_spe_pmu_perf_destroy(struct arm_spe_pmu *spe_pmu)
 992 {
 993         perf_pmu_unregister(&spe_pmu->pmu);
 994 }
 995
 996 static void __arm_spe_pmu_dev_probe(void *info)
 997 {
 998         int fld;
 999         u64 reg;
1000         struct arm_spe_pmu *spe_pmu = info;
1001         struct device *dev = &spe_pmu->pdev->dev;
1002
1003         fld = cpuid_feature_extract_unsigned_field(read_cpuid(ID_AA64DFR0_EL1),
1004                                                    ID_AA64DFR0_EL1_PMSVer_SHIFT);
1005         if (!fld) {
1006                 dev_err(dev,
1007                         "unsupported ID_AA64DFR0_EL1.PMSVer [%d] on CPU %d\n",
1008                         fld, smp_processor_id());
1009                 return;
1010         }
1011         spe_pmu->pmsver = (u16)fld;
1012
1013         /* Read PMBIDR first to determine whether or not we have access */
1014         reg = read_sysreg_s(SYS_PMBIDR_EL1);
1015         if (FIELD_GET(PMBIDR_EL1_P, reg)) {
1016                 dev_err(dev,
1017                         "profiling buffer owned by higher exception level\n");
1018                 return;
1019         }
1020
1021         /* Minimum alignment. If it's out-of-range, then fail the probe */
1022         fld = FIELD_GET(PMBIDR_EL1_ALIGN, reg);
1023         spe_pmu->align = 1 << fld;
1024         if (spe_pmu->align > SZ_2K) {
1025                 dev_err(dev, "unsupported PMBIDR.Align [%d] on CPU %d\n",
1026                         fld, smp_processor_id());
1027                 return;
1028         }
1029
1030         /* It's now safe to read PMSIDR and figure out what we've got */
1031         reg = read_sysreg_s(SYS_PMSIDR_EL1);
1032         if (FIELD_GET(PMSIDR_EL1_FE, reg))
1033                 spe_pmu->features |= SPE_PMU_FEAT_FILT_EVT;
1034
1035         if (FIELD_GET(PMSIDR_EL1_FnE, reg))
1036                 spe_pmu->features |= SPE_PMU_FEAT_INV_FILT_EVT;
1037
1038         if (FIELD_GET(PMSIDR_EL1_FT, reg))
1039                 spe_pmu->features |= SPE_PMU_FEAT_FILT_TYP;
1040
1041         if (FIELD_GET(PMSIDR_EL1_FL, reg))
1042                 spe_pmu->features |= SPE_PMU_FEAT_FILT_LAT;
1043
1044         if (FIELD_GET(PMSIDR_EL1_ARCHINST, reg))
1045                 spe_pmu->features |= SPE_PMU_FEAT_ARCH_INST;
1046
1047         if (FIELD_GET(PMSIDR_EL1_LDS, reg))
1048                 spe_pmu->features |= SPE_PMU_FEAT_LDS;
1049
1050         if (FIELD_GET(PMSIDR_EL1_ERND, reg))
1051                 spe_pmu->features |= SPE_PMU_FEAT_ERND;
1052
1053         /* This field has a spaced out encoding, so just use a look-up */
1054         fld = FIELD_GET(PMSIDR_EL1_INTERVAL, reg);
1055         switch (fld) {
1056         case PMSIDR_EL1_INTERVAL_256:
1057                 spe_pmu->min_period = 256;
1058                 break;
1059         case PMSIDR_EL1_INTERVAL_512:
1060                 spe_pmu->min_period = 512;
1061                 break;
1062         case PMSIDR_EL1_INTERVAL_768:
1063                 spe_pmu->min_period = 768;
1064                 break;
1065         case PMSIDR_EL1_INTERVAL_1024:
1066                 spe_pmu->min_period = 1024;
1067                 break;
1068         case PMSIDR_EL1_INTERVAL_1536:
1069                 spe_pmu->min_period = 1536;
1070                 break;
1071         case PMSIDR_EL1_INTERVAL_2048:
1072                 spe_pmu->min_period = 2048;
1073                 break;
1074         case PMSIDR_EL1_INTERVAL_3072:
1075                 spe_pmu->min_period = 3072;
1076                 break;
1077         default:
1078                 dev_warn(dev, "unknown PMSIDR_EL1.Interval [%d]; assuming 8\n",
1079                          fld);
1080                 fallthrough;
1081         case PMSIDR_EL1_INTERVAL_4096:
1082                 spe_pmu->min_period = 4096;
1083         }
1084
1085         /* Maximum record size. If it's out-of-range, then fail the probe */
1086         fld = FIELD_GET(PMSIDR_EL1_MAXSIZE, reg);
1087         spe_pmu->max_record_sz = 1 << fld;
1088         if (spe_pmu->max_record_sz > SZ_2K || spe_pmu->max_record_sz < 16) {
1089                 dev_err(dev, "unsupported PMSIDR_EL1.MaxSize [%d] on CPU %d\n",
1090                         fld, smp_processor_id());
1091                 return;
1092         }
1093
1094         fld = FIELD_GET(PMSIDR_EL1_COUNTSIZE, reg);
1095         switch (fld) {
1096         default:
1097                 dev_warn(dev, "unknown PMSIDR_EL1.CountSize [%d]; assuming 2\n",
1098                          fld);
1099                 fallthrough;
1100         case PMSIDR_EL1_COUNTSIZE_12_BIT_SAT:
1101                 spe_pmu->counter_sz = 12;
1102                 break;
1103         case PMSIDR_EL1_COUNTSIZE_16_BIT_SAT:
1104                 spe_pmu->counter_sz = 16;
1105         }
1106
1107         dev_info(dev,
1108                  "probed SPEv1.%d for CPUs %*pbl [max_record_sz %u, align %u, features 0x%llx]\n",
1109                  spe_pmu->pmsver - 1, cpumask_pr_args(&spe_pmu->supported_cpus),
1110                  spe_pmu->max_record_sz, spe_pmu->align, spe_pmu->features);
1111
1112         spe_pmu->features |= SPE_PMU_FEAT_DEV_PROBED;
1113 }
1114
1115 static void __arm_spe_pmu_reset_local(void)
1116 {
1117         /*
1118          * This is probably overkill, as we have no idea where we're
1119          * draining any buffered data to...
1120          */
1121         arm_spe_pmu_disable_and_drain_local();
1122
1123         /* Reset the buffer base pointer */
1124         write_sysreg_s(0, SYS_PMBPTR_EL1);
1125         isb();
1126
1127         /* Clear any pending management interrupts */
1128         write_sysreg_s(0, SYS_PMBSR_EL1);
1129         isb();
1130 }
1131
1132 static void __arm_spe_pmu_setup_one(void *info)
1133 {
1134         struct arm_spe_pmu *spe_pmu = info;
1135
1136         __arm_spe_pmu_reset_local();
1137         enable_percpu_irq(spe_pmu->irq, IRQ_TYPE_NONE);
1138 }
1139
1140 static void __arm_spe_pmu_stop_one(void *info)
1141 {
1142         struct arm_spe_pmu *spe_pmu = info;
1143
1144         disable_percpu_irq(spe_pmu->irq);
1145         __arm_spe_pmu_reset_local();
1146 }
1147
1148 static int arm_spe_pmu_cpu_startup(unsigned int cpu, struct hlist_node *node)
1149 {
1150         struct arm_spe_pmu *spe_pmu;
1151
1152         spe_pmu = hlist_entry_safe(node, struct arm_spe_pmu, hotplug_node);
1153         if (!cpumask_test_cpu(cpu, &spe_pmu->supported_cpus))
1154                 return 0;
1155
1156         __arm_spe_pmu_setup_one(spe_pmu);
1157         return 0;
1158 }
1159
1160 static int arm_spe_pmu_cpu_teardown(unsigned int cpu, struct hlist_node *node)
1161 {
1162         struct arm_spe_pmu *spe_pmu;
1163
1164         spe_pmu = hlist_entry_safe(node, struct arm_spe_pmu, hotplug_node);
1165         if (!cpumask_test_cpu(cpu, &spe_pmu->supported_cpus))
1166                 return 0;
1167
1168         __arm_spe_pmu_stop_one(spe_pmu);
1169         return 0;
1170 }
1171
1172 static int arm_spe_pmu_dev_init(struct arm_spe_pmu *spe_pmu)
1173 {
1174         int ret;
1175         cpumask_t *mask = &spe_pmu->supported_cpus;
1176
1177         /* Make sure we probe the hardware on a relevant CPU */
1178         ret = smp_call_function_any(mask,  __arm_spe_pmu_dev_probe, spe_pmu, 1);
1179         if (ret || !(spe_pmu->features & SPE_PMU_FEAT_DEV_PROBED))
1180                 return -ENXIO;
1181
1182         /* Request our PPIs (note that the IRQ is still disabled) */
1183         ret = request_percpu_irq(spe_pmu->irq, arm_spe_pmu_irq_handler, DRVNAME,
1184                                  spe_pmu->handle);
1185         if (ret)
1186                 return ret;
1187
1188         /*
1189          * Register our hotplug notifier now so we don't miss any events.
1190          * This will enable the IRQ for any supported CPUs that are already
1191          * up.
1192          */
1193         ret = cpuhp_state_add_instance(arm_spe_pmu_online,
1194                                        &spe_pmu->hotplug_node);
1195         if (ret)
1196                 free_percpu_irq(spe_pmu->irq, spe_pmu->handle);
1197
1198         return ret;
1199 }
1200
1201 static void arm_spe_pmu_dev_teardown(struct arm_spe_pmu *spe_pmu)
1202 {
1203         cpuhp_state_remove_instance(arm_spe_pmu_online, &spe_pmu->hotplug_node);
1204         free_percpu_irq(spe_pmu->irq, spe_pmu->handle);
1205 }
1206
1207 /* Driver and device probing */
1208 static int arm_spe_pmu_irq_probe(struct arm_spe_pmu *spe_pmu)
1209 {
1210         struct platform_device *pdev = spe_pmu->pdev;
1211         int irq = platform_get_irq(pdev, 0);
1212
1213         if (irq < 0)
1214                 return -ENXIO;
1215
1216         if (!irq_is_percpu(irq)) {
1217                 dev_err(&pdev->dev, "expected PPI but got SPI (%d)\n", irq);
1218                 return -EINVAL;
1219         }
1220
1221         if (irq_get_percpu_devid_partition(irq, &spe_pmu->supported_cpus)) {
1222                 dev_err(&pdev->dev, "failed to get PPI partition (%d)\n", irq);
1223                 return -EINVAL;
1224         }
1225
1226         spe_pmu->irq = irq;
1227         return 0;
1228 }
1229
1230 static const struct of_device_id arm_spe_pmu_of_match[] = {
1231         { .compatible = "arm,statistical-profiling-extension-v1", .data = (void *)1 },
1232         { /* Sentinel */ },
1233 };
1234 MODULE_DEVICE_TABLE(of, arm_spe_pmu_of_match);
1235
1236 static const struct platform_device_id arm_spe_match[] = {
1237         { ARMV8_SPE_PDEV_NAME, 0},
1238         { }
1239 };
1240 MODULE_DEVICE_TABLE(platform, arm_spe_match);
1241
1242 static int arm_spe_pmu_device_probe(struct platform_device *pdev)
1243 {
1244         int ret;
1245         struct arm_spe_pmu *spe_pmu;
1246         struct device *dev = &pdev->dev;
1247
1248         /*
1249          * If kernelspace is unmapped when running at EL0, then the SPE
1250          * buffer will fault and prematurely terminate the AUX session.
1251          */
1252         if (arm64_kernel_unmapped_at_el0()) {
1253                 dev_warn_once(dev, "profiling buffer inaccessible. Try passing \"kpti=off\" on the kernel command line\n");
1254                 return -EPERM;
1255         }
1256
1257         spe_pmu = devm_kzalloc(dev, sizeof(*spe_pmu), GFP_KERNEL);
1258         if (!spe_pmu)
1259                 return -ENOMEM;
1260
1261         spe_pmu->handle = alloc_percpu(typeof(*spe_pmu->handle));
1262         if (!spe_pmu->handle)
1263                 return -ENOMEM;
1264
1265         spe_pmu->pdev = pdev;
1266         platform_set_drvdata(pdev, spe_pmu);
1267
1268         ret = arm_spe_pmu_irq_probe(spe_pmu);
1269         if (ret)
1270                 goto out_free_handle;
1271
1272         ret = arm_spe_pmu_dev_init(spe_pmu);
1273         if (ret)
1274                 goto out_free_handle;
1275
1276         ret = arm_spe_pmu_perf_init(spe_pmu);
1277         if (ret)
1278                 goto out_teardown_dev;
1279
1280         return 0;
1281
1282 out_teardown_dev:
1283         arm_spe_pmu_dev_teardown(spe_pmu);
1284 out_free_handle:
1285         free_percpu(spe_pmu->handle);
1286         return ret;
1287 }
1288
1289 static int arm_spe_pmu_device_remove(struct platform_device *pdev)
1290 {
1291         struct arm_spe_pmu *spe_pmu = platform_get_drvdata(pdev);
1292
1293         arm_spe_pmu_perf_destroy(spe_pmu);
1294         arm_spe_pmu_dev_teardown(spe_pmu);
1295         free_percpu(spe_pmu->handle);
1296         return 0;
1297 }
1298
1299 static struct platform_driver arm_spe_pmu_driver = {
1300         .id_table = arm_spe_match,
1301         .driver = {
1302                 .name           = DRVNAME,
1303                 .of_match_table = of_match_ptr(arm_spe_pmu_of_match),
1304                 .suppress_bind_attrs = true,
1305         },
1306         .probe  = arm_spe_pmu_device_probe,
1307         .remove = arm_spe_pmu_device_remove,
1308 };
1309
1310 static int __init arm_spe_pmu_init(void)
1311 {
1312         int ret;
1313
1314         ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, DRVNAME,
1315                                       arm_spe_pmu_cpu_startup,
1316                                       arm_spe_pmu_cpu_teardown);
1317         if (ret < 0)
1318                 return ret;
1319         arm_spe_pmu_online = ret;
1320
1321         ret = platform_driver_register(&arm_spe_pmu_driver);
1322         if (ret)
1323                 cpuhp_remove_multi_state(arm_spe_pmu_online);
1324
1325         return ret;
1326 }
1327
1328 static void __exit arm_spe_pmu_exit(void)
1329 {
1330         platform_driver_unregister(&arm_spe_pmu_driver);
1331         cpuhp_remove_multi_state(arm_spe_pmu_online);
1332 }
1333
1334 module_init(arm_spe_pmu_init);
1335 module_exit(arm_spe_pmu_exit);
1336
1337 MODULE_DESCRIPTION("Perf driver for the ARMv8.2 Statistical Profiling Extension");
1338 MODULE_AUTHOR("Will Deacon <[email protected]>");
1339 MODULE_LICENSE("GPL v2");