Back to home page

OSCL-LXR

 
 

    


0001 // SPDX-License-Identifier: GPL-2.0-only
0002 /*
0003  * HiSilicon SoC Hardware event counters support
0004  *
0005  * Copyright (C) 2017 HiSilicon Limited
0006  * Author: Anurup M <anurup.m@huawei.com>
0007  *         Shaokun Zhang <zhangshaokun@hisilicon.com>
0008  *
0009  * This code is based on the uncore PMUs like arm-cci and arm-ccn.
0010  */
0011 #include <linux/bitmap.h>
0012 #include <linux/bitops.h>
0013 #include <linux/bug.h>
0014 #include <linux/err.h>
0015 #include <linux/errno.h>
0016 #include <linux/interrupt.h>
0017 
0018 #include <asm/cputype.h>
0019 #include <asm/local64.h>
0020 
0021 #include "hisi_uncore_pmu.h"
0022 
0023 #define HISI_GET_EVENTID(ev) (ev->hw.config_base & 0xff)
0024 #define HISI_MAX_PERIOD(nr) (GENMASK_ULL((nr) - 1, 0))
0025 
0026 /*
0027  * PMU format attributes
0028  */
0029 ssize_t hisi_format_sysfs_show(struct device *dev,
0030                    struct device_attribute *attr, char *buf)
0031 {
0032     struct dev_ext_attribute *eattr;
0033 
0034     eattr = container_of(attr, struct dev_ext_attribute, attr);
0035 
0036     return sysfs_emit(buf, "%s\n", (char *)eattr->var);
0037 }
0038 EXPORT_SYMBOL_GPL(hisi_format_sysfs_show);
0039 
0040 /*
0041  * PMU event attributes
0042  */
0043 ssize_t hisi_event_sysfs_show(struct device *dev,
0044                   struct device_attribute *attr, char *page)
0045 {
0046     struct dev_ext_attribute *eattr;
0047 
0048     eattr = container_of(attr, struct dev_ext_attribute, attr);
0049 
0050     return sysfs_emit(page, "config=0x%lx\n", (unsigned long)eattr->var);
0051 }
0052 EXPORT_SYMBOL_GPL(hisi_event_sysfs_show);
0053 
0054 /*
0055  * sysfs cpumask attributes. For uncore PMU, we only have a single CPU to show
0056  */
0057 ssize_t hisi_cpumask_sysfs_show(struct device *dev,
0058                 struct device_attribute *attr, char *buf)
0059 {
0060     struct hisi_pmu *hisi_pmu = to_hisi_pmu(dev_get_drvdata(dev));
0061 
0062     return sysfs_emit(buf, "%d\n", hisi_pmu->on_cpu);
0063 }
0064 EXPORT_SYMBOL_GPL(hisi_cpumask_sysfs_show);
0065 
0066 static bool hisi_validate_event_group(struct perf_event *event)
0067 {
0068     struct perf_event *sibling, *leader = event->group_leader;
0069     struct hisi_pmu *hisi_pmu = to_hisi_pmu(event->pmu);
0070     /* Include count for the event */
0071     int counters = 1;
0072 
0073     if (!is_software_event(leader)) {
0074         /*
0075          * We must NOT create groups containing mixed PMUs, although
0076          * software events are acceptable
0077          */
0078         if (leader->pmu != event->pmu)
0079             return false;
0080 
0081         /* Increment counter for the leader */
0082         if (leader != event)
0083             counters++;
0084     }
0085 
0086     for_each_sibling_event(sibling, event->group_leader) {
0087         if (is_software_event(sibling))
0088             continue;
0089         if (sibling->pmu != event->pmu)
0090             return false;
0091         /* Increment counter for each sibling */
0092         counters++;
0093     }
0094 
0095     /* The group can not count events more than the counters in the HW */
0096     return counters <= hisi_pmu->num_counters;
0097 }
0098 
0099 int hisi_uncore_pmu_get_event_idx(struct perf_event *event)
0100 {
0101     struct hisi_pmu *hisi_pmu = to_hisi_pmu(event->pmu);
0102     unsigned long *used_mask = hisi_pmu->pmu_events.used_mask;
0103     u32 num_counters = hisi_pmu->num_counters;
0104     int idx;
0105 
0106     idx = find_first_zero_bit(used_mask, num_counters);
0107     if (idx == num_counters)
0108         return -EAGAIN;
0109 
0110     set_bit(idx, used_mask);
0111 
0112     return idx;
0113 }
0114 EXPORT_SYMBOL_GPL(hisi_uncore_pmu_get_event_idx);
0115 
0116 ssize_t hisi_uncore_pmu_identifier_attr_show(struct device *dev,
0117                          struct device_attribute *attr,
0118                          char *page)
0119 {
0120     struct hisi_pmu *hisi_pmu = to_hisi_pmu(dev_get_drvdata(dev));
0121 
0122     return sysfs_emit(page, "0x%08x\n", hisi_pmu->identifier);
0123 }
0124 EXPORT_SYMBOL_GPL(hisi_uncore_pmu_identifier_attr_show);
0125 
0126 static void hisi_uncore_pmu_clear_event_idx(struct hisi_pmu *hisi_pmu, int idx)
0127 {
0128     clear_bit(idx, hisi_pmu->pmu_events.used_mask);
0129 }
0130 
0131 static irqreturn_t hisi_uncore_pmu_isr(int irq, void *data)
0132 {
0133     struct hisi_pmu *hisi_pmu = data;
0134     struct perf_event *event;
0135     unsigned long overflown;
0136     int idx;
0137 
0138     overflown = hisi_pmu->ops->get_int_status(hisi_pmu);
0139     if (!overflown)
0140         return IRQ_NONE;
0141 
0142     /*
0143      * Find the counter index which overflowed if the bit was set
0144      * and handle it.
0145      */
0146     for_each_set_bit(idx, &overflown, hisi_pmu->num_counters) {
0147         /* Write 1 to clear the IRQ status flag */
0148         hisi_pmu->ops->clear_int_status(hisi_pmu, idx);
0149         /* Get the corresponding event struct */
0150         event = hisi_pmu->pmu_events.hw_events[idx];
0151         if (!event)
0152             continue;
0153 
0154         hisi_uncore_pmu_event_update(event);
0155         hisi_uncore_pmu_set_event_period(event);
0156     }
0157 
0158     return IRQ_HANDLED;
0159 }
0160 
0161 int hisi_uncore_pmu_init_irq(struct hisi_pmu *hisi_pmu,
0162                  struct platform_device *pdev)
0163 {
0164     int irq, ret;
0165 
0166     irq = platform_get_irq(pdev, 0);
0167     if (irq < 0)
0168         return irq;
0169 
0170     ret = devm_request_irq(&pdev->dev, irq, hisi_uncore_pmu_isr,
0171                    IRQF_NOBALANCING | IRQF_NO_THREAD,
0172                    dev_name(&pdev->dev), hisi_pmu);
0173     if (ret < 0) {
0174         dev_err(&pdev->dev,
0175             "Fail to request IRQ: %d ret: %d.\n", irq, ret);
0176         return ret;
0177     }
0178 
0179     hisi_pmu->irq = irq;
0180 
0181     return 0;
0182 }
0183 EXPORT_SYMBOL_GPL(hisi_uncore_pmu_init_irq);
0184 
0185 int hisi_uncore_pmu_event_init(struct perf_event *event)
0186 {
0187     struct hw_perf_event *hwc = &event->hw;
0188     struct hisi_pmu *hisi_pmu;
0189 
0190     if (event->attr.type != event->pmu->type)
0191         return -ENOENT;
0192 
0193     /*
0194      * We do not support sampling as the counters are all
0195      * shared by all CPU cores in a CPU die(SCCL). Also we
0196      * do not support attach to a task(per-process mode)
0197      */
0198     if (is_sampling_event(event) || event->attach_state & PERF_ATTACH_TASK)
0199         return -EOPNOTSUPP;
0200 
0201     /*
0202      *  The uncore counters not specific to any CPU, so cannot
0203      *  support per-task
0204      */
0205     if (event->cpu < 0)
0206         return -EINVAL;
0207 
0208     /*
0209      * Validate if the events in group does not exceed the
0210      * available counters in hardware.
0211      */
0212     if (!hisi_validate_event_group(event))
0213         return -EINVAL;
0214 
0215     hisi_pmu = to_hisi_pmu(event->pmu);
0216     if (event->attr.config > hisi_pmu->check_event)
0217         return -EINVAL;
0218 
0219     if (hisi_pmu->on_cpu == -1)
0220         return -EINVAL;
0221     /*
0222      * We don't assign an index until we actually place the event onto
0223      * hardware. Use -1 to signify that we haven't decided where to put it
0224      * yet.
0225      */
0226     hwc->idx        = -1;
0227     hwc->config_base    = event->attr.config;
0228 
0229     /* Enforce to use the same CPU for all events in this PMU */
0230     event->cpu = hisi_pmu->on_cpu;
0231 
0232     return 0;
0233 }
0234 EXPORT_SYMBOL_GPL(hisi_uncore_pmu_event_init);
0235 
0236 /*
0237  * Set the counter to count the event that we're interested in,
0238  * and enable interrupt and counter.
0239  */
0240 static void hisi_uncore_pmu_enable_event(struct perf_event *event)
0241 {
0242     struct hisi_pmu *hisi_pmu = to_hisi_pmu(event->pmu);
0243     struct hw_perf_event *hwc = &event->hw;
0244 
0245     hisi_pmu->ops->write_evtype(hisi_pmu, hwc->idx,
0246                     HISI_GET_EVENTID(event));
0247 
0248     if (hisi_pmu->ops->enable_filter)
0249         hisi_pmu->ops->enable_filter(event);
0250 
0251     hisi_pmu->ops->enable_counter_int(hisi_pmu, hwc);
0252     hisi_pmu->ops->enable_counter(hisi_pmu, hwc);
0253 }
0254 
0255 /*
0256  * Disable counter and interrupt.
0257  */
0258 static void hisi_uncore_pmu_disable_event(struct perf_event *event)
0259 {
0260     struct hisi_pmu *hisi_pmu = to_hisi_pmu(event->pmu);
0261     struct hw_perf_event *hwc = &event->hw;
0262 
0263     hisi_pmu->ops->disable_counter(hisi_pmu, hwc);
0264     hisi_pmu->ops->disable_counter_int(hisi_pmu, hwc);
0265 
0266     if (hisi_pmu->ops->disable_filter)
0267         hisi_pmu->ops->disable_filter(event);
0268 }
0269 
0270 void hisi_uncore_pmu_set_event_period(struct perf_event *event)
0271 {
0272     struct hisi_pmu *hisi_pmu = to_hisi_pmu(event->pmu);
0273     struct hw_perf_event *hwc = &event->hw;
0274 
0275     /*
0276      * The HiSilicon PMU counters support 32 bits or 48 bits, depending on
0277      * the PMU. We reduce it to 2^(counter_bits - 1) to account for the
0278      * extreme interrupt latency. So we could hopefully handle the overflow
0279      * interrupt before another 2^(counter_bits - 1) events occur and the
0280      * counter overtakes its previous value.
0281      */
0282     u64 val = BIT_ULL(hisi_pmu->counter_bits - 1);
0283 
0284     local64_set(&hwc->prev_count, val);
0285     /* Write start value to the hardware event counter */
0286     hisi_pmu->ops->write_counter(hisi_pmu, hwc, val);
0287 }
0288 EXPORT_SYMBOL_GPL(hisi_uncore_pmu_set_event_period);
0289 
0290 void hisi_uncore_pmu_event_update(struct perf_event *event)
0291 {
0292     struct hisi_pmu *hisi_pmu = to_hisi_pmu(event->pmu);
0293     struct hw_perf_event *hwc = &event->hw;
0294     u64 delta, prev_raw_count, new_raw_count;
0295 
0296     do {
0297         /* Read the count from the counter register */
0298         new_raw_count = hisi_pmu->ops->read_counter(hisi_pmu, hwc);
0299         prev_raw_count = local64_read(&hwc->prev_count);
0300     } while (local64_cmpxchg(&hwc->prev_count, prev_raw_count,
0301                  new_raw_count) != prev_raw_count);
0302     /*
0303      * compute the delta
0304      */
0305     delta = (new_raw_count - prev_raw_count) &
0306         HISI_MAX_PERIOD(hisi_pmu->counter_bits);
0307     local64_add(delta, &event->count);
0308 }
0309 EXPORT_SYMBOL_GPL(hisi_uncore_pmu_event_update);
0310 
0311 void hisi_uncore_pmu_start(struct perf_event *event, int flags)
0312 {
0313     struct hisi_pmu *hisi_pmu = to_hisi_pmu(event->pmu);
0314     struct hw_perf_event *hwc = &event->hw;
0315 
0316     if (WARN_ON_ONCE(!(hwc->state & PERF_HES_STOPPED)))
0317         return;
0318 
0319     WARN_ON_ONCE(!(hwc->state & PERF_HES_UPTODATE));
0320     hwc->state = 0;
0321     hisi_uncore_pmu_set_event_period(event);
0322 
0323     if (flags & PERF_EF_RELOAD) {
0324         u64 prev_raw_count =  local64_read(&hwc->prev_count);
0325 
0326         hisi_pmu->ops->write_counter(hisi_pmu, hwc, prev_raw_count);
0327     }
0328 
0329     hisi_uncore_pmu_enable_event(event);
0330     perf_event_update_userpage(event);
0331 }
0332 EXPORT_SYMBOL_GPL(hisi_uncore_pmu_start);
0333 
0334 void hisi_uncore_pmu_stop(struct perf_event *event, int flags)
0335 {
0336     struct hw_perf_event *hwc = &event->hw;
0337 
0338     hisi_uncore_pmu_disable_event(event);
0339     WARN_ON_ONCE(hwc->state & PERF_HES_STOPPED);
0340     hwc->state |= PERF_HES_STOPPED;
0341 
0342     if (hwc->state & PERF_HES_UPTODATE)
0343         return;
0344 
0345     /* Read hardware counter and update the perf counter statistics */
0346     hisi_uncore_pmu_event_update(event);
0347     hwc->state |= PERF_HES_UPTODATE;
0348 }
0349 EXPORT_SYMBOL_GPL(hisi_uncore_pmu_stop);
0350 
0351 int hisi_uncore_pmu_add(struct perf_event *event, int flags)
0352 {
0353     struct hisi_pmu *hisi_pmu = to_hisi_pmu(event->pmu);
0354     struct hw_perf_event *hwc = &event->hw;
0355     int idx;
0356 
0357     hwc->state = PERF_HES_STOPPED | PERF_HES_UPTODATE;
0358 
0359     /* Get an available counter index for counting */
0360     idx = hisi_pmu->ops->get_event_idx(event);
0361     if (idx < 0)
0362         return idx;
0363 
0364     event->hw.idx = idx;
0365     hisi_pmu->pmu_events.hw_events[idx] = event;
0366 
0367     if (flags & PERF_EF_START)
0368         hisi_uncore_pmu_start(event, PERF_EF_RELOAD);
0369 
0370     return 0;
0371 }
0372 EXPORT_SYMBOL_GPL(hisi_uncore_pmu_add);
0373 
0374 void hisi_uncore_pmu_del(struct perf_event *event, int flags)
0375 {
0376     struct hisi_pmu *hisi_pmu = to_hisi_pmu(event->pmu);
0377     struct hw_perf_event *hwc = &event->hw;
0378 
0379     hisi_uncore_pmu_stop(event, PERF_EF_UPDATE);
0380     hisi_uncore_pmu_clear_event_idx(hisi_pmu, hwc->idx);
0381     perf_event_update_userpage(event);
0382     hisi_pmu->pmu_events.hw_events[hwc->idx] = NULL;
0383 }
0384 EXPORT_SYMBOL_GPL(hisi_uncore_pmu_del);
0385 
0386 void hisi_uncore_pmu_read(struct perf_event *event)
0387 {
0388     /* Read hardware counter and update the perf counter statistics */
0389     hisi_uncore_pmu_event_update(event);
0390 }
0391 EXPORT_SYMBOL_GPL(hisi_uncore_pmu_read);
0392 
0393 void hisi_uncore_pmu_enable(struct pmu *pmu)
0394 {
0395     struct hisi_pmu *hisi_pmu = to_hisi_pmu(pmu);
0396     bool enabled = !bitmap_empty(hisi_pmu->pmu_events.used_mask,
0397                     hisi_pmu->num_counters);
0398 
0399     if (!enabled)
0400         return;
0401 
0402     hisi_pmu->ops->start_counters(hisi_pmu);
0403 }
0404 EXPORT_SYMBOL_GPL(hisi_uncore_pmu_enable);
0405 
0406 void hisi_uncore_pmu_disable(struct pmu *pmu)
0407 {
0408     struct hisi_pmu *hisi_pmu = to_hisi_pmu(pmu);
0409 
0410     hisi_pmu->ops->stop_counters(hisi_pmu);
0411 }
0412 EXPORT_SYMBOL_GPL(hisi_uncore_pmu_disable);
0413 
0414 
0415 /*
0416  * The Super CPU Cluster (SCCL) and CPU Cluster (CCL) IDs can be
0417  * determined from the MPIDR_EL1, but the encoding varies by CPU:
0418  *
0419  * - For MT variants of TSV110:
0420  *   SCCL is Aff2[7:3], CCL is Aff2[2:0]
0421  *
0422  * - For other MT parts:
0423  *   SCCL is Aff3[7:0], CCL is Aff2[7:0]
0424  *
0425  * - For non-MT parts:
0426  *   SCCL is Aff2[7:0], CCL is Aff1[7:0]
0427  */
0428 static void hisi_read_sccl_and_ccl_id(int *scclp, int *cclp)
0429 {
0430     u64 mpidr = read_cpuid_mpidr();
0431     int aff3 = MPIDR_AFFINITY_LEVEL(mpidr, 3);
0432     int aff2 = MPIDR_AFFINITY_LEVEL(mpidr, 2);
0433     int aff1 = MPIDR_AFFINITY_LEVEL(mpidr, 1);
0434     bool mt = mpidr & MPIDR_MT_BITMASK;
0435     int sccl, ccl;
0436 
0437     if (mt && read_cpuid_part_number() == HISI_CPU_PART_TSV110) {
0438         sccl = aff2 >> 3;
0439         ccl = aff2 & 0x7;
0440     } else if (mt) {
0441         sccl = aff3;
0442         ccl = aff2;
0443     } else {
0444         sccl = aff2;
0445         ccl = aff1;
0446     }
0447 
0448     if (scclp)
0449         *scclp = sccl;
0450     if (cclp)
0451         *cclp = ccl;
0452 }
0453 
0454 /*
0455  * Check whether the CPU is associated with this uncore PMU
0456  */
0457 static bool hisi_pmu_cpu_is_associated_pmu(struct hisi_pmu *hisi_pmu)
0458 {
0459     int sccl_id, ccl_id;
0460 
0461     /* If SCCL_ID is -1, the PMU is in a SICL and has no CPU affinity */
0462     if (hisi_pmu->sccl_id == -1)
0463         return true;
0464 
0465     if (hisi_pmu->ccl_id == -1) {
0466         /* If CCL_ID is -1, the PMU only shares the same SCCL */
0467         hisi_read_sccl_and_ccl_id(&sccl_id, NULL);
0468 
0469         return sccl_id == hisi_pmu->sccl_id;
0470     }
0471 
0472     hisi_read_sccl_and_ccl_id(&sccl_id, &ccl_id);
0473 
0474     return sccl_id == hisi_pmu->sccl_id && ccl_id == hisi_pmu->ccl_id;
0475 }
0476 
0477 int hisi_uncore_pmu_online_cpu(unsigned int cpu, struct hlist_node *node)
0478 {
0479     struct hisi_pmu *hisi_pmu = hlist_entry_safe(node, struct hisi_pmu,
0480                              node);
0481 
0482     if (!hisi_pmu_cpu_is_associated_pmu(hisi_pmu))
0483         return 0;
0484 
0485     cpumask_set_cpu(cpu, &hisi_pmu->associated_cpus);
0486 
0487     /* If another CPU is already managing this PMU, simply return. */
0488     if (hisi_pmu->on_cpu != -1)
0489         return 0;
0490 
0491     /* Use this CPU in cpumask for event counting */
0492     hisi_pmu->on_cpu = cpu;
0493 
0494     /* Overflow interrupt also should use the same CPU */
0495     WARN_ON(irq_set_affinity(hisi_pmu->irq, cpumask_of(cpu)));
0496 
0497     return 0;
0498 }
0499 EXPORT_SYMBOL_GPL(hisi_uncore_pmu_online_cpu);
0500 
0501 int hisi_uncore_pmu_offline_cpu(unsigned int cpu, struct hlist_node *node)
0502 {
0503     struct hisi_pmu *hisi_pmu = hlist_entry_safe(node, struct hisi_pmu,
0504                              node);
0505     cpumask_t pmu_online_cpus;
0506     unsigned int target;
0507 
0508     if (!cpumask_test_and_clear_cpu(cpu, &hisi_pmu->associated_cpus))
0509         return 0;
0510 
0511     /* Nothing to do if this CPU doesn't own the PMU */
0512     if (hisi_pmu->on_cpu != cpu)
0513         return 0;
0514 
0515     /* Give up ownership of the PMU */
0516     hisi_pmu->on_cpu = -1;
0517 
0518     /* Choose a new CPU to migrate ownership of the PMU to */
0519     cpumask_and(&pmu_online_cpus, &hisi_pmu->associated_cpus,
0520             cpu_online_mask);
0521     target = cpumask_any_but(&pmu_online_cpus, cpu);
0522     if (target >= nr_cpu_ids)
0523         return 0;
0524 
0525     perf_pmu_migrate_context(&hisi_pmu->pmu, cpu, target);
0526     /* Use this CPU for event counting */
0527     hisi_pmu->on_cpu = target;
0528     WARN_ON(irq_set_affinity(hisi_pmu->irq, cpumask_of(target)));
0529 
0530     return 0;
0531 }
0532 EXPORT_SYMBOL_GPL(hisi_uncore_pmu_offline_cpu);
0533 
0534 void hisi_pmu_init(struct pmu *pmu, const char *name,
0535         const struct attribute_group **attr_groups, struct module *module)
0536 {
0537     pmu->name               = name;
0538     pmu->module             = module;
0539     pmu->task_ctx_nr        = perf_invalid_context;
0540     pmu->event_init         = hisi_uncore_pmu_event_init;
0541     pmu->pmu_enable         = hisi_uncore_pmu_enable;
0542     pmu->pmu_disable        = hisi_uncore_pmu_disable;
0543     pmu->add                = hisi_uncore_pmu_add;
0544     pmu->del                = hisi_uncore_pmu_del;
0545     pmu->start              = hisi_uncore_pmu_start;
0546     pmu->stop               = hisi_uncore_pmu_stop;
0547     pmu->read               = hisi_uncore_pmu_read;
0548     pmu->attr_groups        = attr_groups;
0549 }
0550 EXPORT_SYMBOL_GPL(hisi_pmu_init);
0551 
0552 MODULE_LICENSE("GPL v2");