linux/kernel/trace/trace_hwlat.c
<<
>>
Prefs
   1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * trace_hwlat.c - A simple Hardware Latency detector.
   4 *
   5 * Use this tracer to detect large system latencies induced by the behavior of
   6 * certain underlying system hardware or firmware, independent of Linux itself.
   7 * The code was developed originally to detect the presence of SMIs on Intel
   8 * and AMD systems, although there is no dependency upon x86 herein.
   9 *
  10 * The classical example usage of this tracer is in detecting the presence of
  11 * SMIs or System Management Interrupts on Intel and AMD systems. An SMI is a
  12 * somewhat special form of hardware interrupt spawned from earlier CPU debug
  13 * modes in which the (BIOS/EFI/etc.) firmware arranges for the South Bridge
  14 * LPC (or other device) to generate a special interrupt under certain
  15 * circumstances, for example, upon expiration of a special SMI timer device,
  16 * due to certain external thermal readings, on certain I/O address accesses,
  17 * and other situations. An SMI hits a special CPU pin, triggers a special
  18 * SMI mode (complete with special memory map), and the OS is unaware.
  19 *
  20 * Although certain hardware-inducing latencies are necessary (for example,
  21 * a modern system often requires an SMI handler for correct thermal control
  22 * and remote management) they can wreak havoc upon any OS-level performance
  23 * guarantees toward low-latency, especially when the OS is not even made
  24 * aware of the presence of these interrupts. For this reason, we need a
  25 * somewhat brute force mechanism to detect these interrupts. In this case,
  26 * we do it by hogging all of the CPU(s) for configurable timer intervals,
  27 * sampling the built-in CPU timer, looking for discontiguous readings.
  28 *
  29 * WARNING: This implementation necessarily introduces latencies. Therefore,
  30 *          you should NEVER use this tracer while running in a production
  31 *          environment requiring any kind of low-latency performance
  32 *          guarantee(s).
  33 *
  34 * Copyright (C) 2008-2009 Jon Masters, Red Hat, Inc. <jcm@redhat.com>
  35 * Copyright (C) 2013-2016 Steven Rostedt, Red Hat, Inc. <srostedt@redhat.com>
  36 *
  37 * Includes useful feedback from Clark Williams <williams@redhat.com>
  38 *
  39 */
  40#include <linux/kthread.h>
  41#include <linux/tracefs.h>
  42#include <linux/uaccess.h>
  43#include <linux/cpumask.h>
  44#include <linux/delay.h>
  45#include <linux/sched/clock.h>
  46#include "trace.h"
  47
  48static struct trace_array       *hwlat_trace;
  49
  50#define U64STR_SIZE             22                      /* 20 digits max */
  51
  52#define BANNER                  "hwlat_detector: "
  53#define DEFAULT_SAMPLE_WINDOW   1000000                 /* 1s */
  54#define DEFAULT_SAMPLE_WIDTH    500000                  /* 0.5s */
  55#define DEFAULT_LAT_THRESHOLD   10                      /* 10us */
  56
  57static struct dentry *hwlat_sample_width;       /* sample width us */
  58static struct dentry *hwlat_sample_window;      /* sample window us */
  59static struct dentry *hwlat_thread_mode;        /* hwlat thread mode */
  60
  61enum {
  62        MODE_NONE = 0,
  63        MODE_ROUND_ROBIN,
  64        MODE_PER_CPU,
  65        MODE_MAX
  66};
  67static char *thread_mode_str[] = { "none", "round-robin", "per-cpu" };
  68
  69/* Save the previous tracing_thresh value */
  70static unsigned long save_tracing_thresh;
  71
  72/* runtime kthread data */
  73struct hwlat_kthread_data {
  74        struct task_struct      *kthread;
  75        /* NMI timestamp counters */
  76        u64                     nmi_ts_start;
  77        u64                     nmi_total_ts;
  78        int                     nmi_count;
  79        int                     nmi_cpu;
  80};
  81
  82struct hwlat_kthread_data hwlat_single_cpu_data;
  83DEFINE_PER_CPU(struct hwlat_kthread_data, hwlat_per_cpu_data);
  84
  85/* Tells NMIs to call back to the hwlat tracer to record timestamps */
  86bool trace_hwlat_callback_enabled;
  87
  88/* If the user changed threshold, remember it */
  89static u64 last_tracing_thresh = DEFAULT_LAT_THRESHOLD * NSEC_PER_USEC;
  90
  91/* Individual latency samples are stored here when detected. */
  92struct hwlat_sample {
  93        u64                     seqnum;         /* unique sequence */
  94        u64                     duration;       /* delta */
  95        u64                     outer_duration; /* delta (outer loop) */
  96        u64                     nmi_total_ts;   /* Total time spent in NMIs */
  97        struct timespec64       timestamp;      /* wall time */
  98        int                     nmi_count;      /* # NMIs during this sample */
  99        int                     count;          /* # of iterations over thresh */
 100};
 101
 102/* keep the global state somewhere. */
 103static struct hwlat_data {
 104
 105        struct mutex lock;              /* protect changes */
 106
 107        u64     count;                  /* total since reset */
 108
 109        u64     sample_window;          /* total sampling window (on+off) */
 110        u64     sample_width;           /* active sampling portion of window */
 111
 112        int     thread_mode;            /* thread mode */
 113
 114} hwlat_data = {
 115        .sample_window          = DEFAULT_SAMPLE_WINDOW,
 116        .sample_width           = DEFAULT_SAMPLE_WIDTH,
 117        .thread_mode            = MODE_ROUND_ROBIN
 118};
 119
 120static struct hwlat_kthread_data *get_cpu_data(void)
 121{
 122        if (hwlat_data.thread_mode == MODE_PER_CPU)
 123                return this_cpu_ptr(&hwlat_per_cpu_data);
 124        else
 125                return &hwlat_single_cpu_data;
 126}
 127
 128static bool hwlat_busy;
 129
 130static void trace_hwlat_sample(struct hwlat_sample *sample)
 131{
 132        struct trace_array *tr = hwlat_trace;
 133        struct trace_event_call *call = &event_hwlat;
 134        struct trace_buffer *buffer = tr->array_buffer.buffer;
 135        struct ring_buffer_event *event;
 136        struct hwlat_entry *entry;
 137
 138        event = trace_buffer_lock_reserve(buffer, TRACE_HWLAT, sizeof(*entry),
 139                                          tracing_gen_ctx());
 140        if (!event)
 141                return;
 142        entry   = ring_buffer_event_data(event);
 143        entry->seqnum                   = sample->seqnum;
 144        entry->duration                 = sample->duration;
 145        entry->outer_duration           = sample->outer_duration;
 146        entry->timestamp                = sample->timestamp;
 147        entry->nmi_total_ts             = sample->nmi_total_ts;
 148        entry->nmi_count                = sample->nmi_count;
 149        entry->count                    = sample->count;
 150
 151        if (!call_filter_check_discard(call, entry, buffer, event))
 152                trace_buffer_unlock_commit_nostack(buffer, event);
 153}
 154
 155/* Macros to encapsulate the time capturing infrastructure */
 156#define time_type       u64
 157#define time_get()      trace_clock_local()
 158#define time_to_us(x)   div_u64(x, 1000)
 159#define time_sub(a, b)  ((a) - (b))
 160#define init_time(a, b) (a = b)
 161#define time_u64(a)     a
 162
 163void trace_hwlat_callback(bool enter)
 164{
 165        struct hwlat_kthread_data *kdata = get_cpu_data();
 166
 167        if (!kdata->kthread)
 168                return;
 169
 170        /*
 171         * Currently trace_clock_local() calls sched_clock() and the
 172         * generic version is not NMI safe.
 173         */
 174        if (!IS_ENABLED(CONFIG_GENERIC_SCHED_CLOCK)) {
 175                if (enter)
 176                        kdata->nmi_ts_start = time_get();
 177                else
 178                        kdata->nmi_total_ts += time_get() - kdata->nmi_ts_start;
 179        }
 180
 181        if (enter)
 182                kdata->nmi_count++;
 183}
 184
 185/*
 186 * hwlat_err - report a hwlat error.
 187 */
 188#define hwlat_err(msg) ({                                                       \
 189        struct trace_array *tr = hwlat_trace;                                   \
 190                                                                                \
 191        trace_array_printk_buf(tr->array_buffer.buffer, _THIS_IP_, msg);        \
 192})
 193
 194/**
 195 * get_sample - sample the CPU TSC and look for likely hardware latencies
 196 *
 197 * Used to repeatedly capture the CPU TSC (or similar), looking for potential
 198 * hardware-induced latency. Called with interrupts disabled and with
 199 * hwlat_data.lock held.
 200 */
 201static int get_sample(void)
 202{
 203        struct hwlat_kthread_data *kdata = get_cpu_data();
 204        struct trace_array *tr = hwlat_trace;
 205        struct hwlat_sample s;
 206        time_type start, t1, t2, last_t2;
 207        s64 diff, outer_diff, total, last_total = 0;
 208        u64 sample = 0;
 209        u64 thresh = tracing_thresh;
 210        u64 outer_sample = 0;
 211        int ret = -1;
 212        unsigned int count = 0;
 213
 214        do_div(thresh, NSEC_PER_USEC); /* modifies interval value */
 215
 216        kdata->nmi_total_ts = 0;
 217        kdata->nmi_count = 0;
 218        /* Make sure NMIs see this first */
 219        barrier();
 220
 221        trace_hwlat_callback_enabled = true;
 222
 223        init_time(last_t2, 0);
 224        start = time_get(); /* start timestamp */
 225        outer_diff = 0;
 226
 227        do {
 228
 229                t1 = time_get();        /* we'll look for a discontinuity */
 230                t2 = time_get();
 231
 232                if (time_u64(last_t2)) {
 233                        /* Check the delta from outer loop (t2 to next t1) */
 234                        outer_diff = time_to_us(time_sub(t1, last_t2));
 235                        /* This shouldn't happen */
 236                        if (outer_diff < 0) {
 237                                hwlat_err(BANNER "time running backwards\n");
 238                                goto out;
 239                        }
 240                        if (outer_diff > outer_sample)
 241                                outer_sample = outer_diff;
 242                }
 243                last_t2 = t2;
 244
 245                total = time_to_us(time_sub(t2, start)); /* sample width */
 246
 247                /* Check for possible overflows */
 248                if (total < last_total) {
 249                        hwlat_err("Time total overflowed\n");
 250                        break;
 251                }
 252                last_total = total;
 253
 254                /* This checks the inner loop (t1 to t2) */
 255                diff = time_to_us(time_sub(t2, t1));     /* current diff */
 256
 257                if (diff > thresh || outer_diff > thresh) {
 258                        if (!count)
 259                                ktime_get_real_ts64(&s.timestamp);
 260                        count++;
 261                }
 262
 263                /* This shouldn't happen */
 264                if (diff < 0) {
 265                        hwlat_err(BANNER "time running backwards\n");
 266                        goto out;
 267                }
 268
 269                if (diff > sample)
 270                        sample = diff; /* only want highest value */
 271
 272        } while (total <= hwlat_data.sample_width);
 273
 274        barrier(); /* finish the above in the view for NMIs */
 275        trace_hwlat_callback_enabled = false;
 276        barrier(); /* Make sure nmi_total_ts is no longer updated */
 277
 278        ret = 0;
 279
 280        /* If we exceed the threshold value, we have found a hardware latency */
 281        if (sample > thresh || outer_sample > thresh) {
 282                u64 latency;
 283
 284                ret = 1;
 285
 286                /* We read in microseconds */
 287                if (kdata->nmi_total_ts)
 288                        do_div(kdata->nmi_total_ts, NSEC_PER_USEC);
 289
 290                hwlat_data.count++;
 291                s.seqnum = hwlat_data.count;
 292                s.duration = sample;
 293                s.outer_duration = outer_sample;
 294                s.nmi_total_ts = kdata->nmi_total_ts;
 295                s.nmi_count = kdata->nmi_count;
 296                s.count = count;
 297                trace_hwlat_sample(&s);
 298
 299                latency = max(sample, outer_sample);
 300
 301                /* Keep a running maximum ever recorded hardware latency */
 302                if (latency > tr->max_latency) {
 303                        tr->max_latency = latency;
 304                        latency_fsnotify(tr);
 305                }
 306        }
 307
 308out:
 309        return ret;
 310}
 311
 312static struct cpumask save_cpumask;
 313
 314static void move_to_next_cpu(void)
 315{
 316        struct cpumask *current_mask = &save_cpumask;
 317        struct trace_array *tr = hwlat_trace;
 318        int next_cpu;
 319
 320        /*
 321         * If for some reason the user modifies the CPU affinity
 322         * of this thread, then stop migrating for the duration
 323         * of the current test.
 324         */
 325        if (!cpumask_equal(current_mask, current->cpus_ptr))
 326                goto change_mode;
 327
 328        get_online_cpus();
 329        cpumask_and(current_mask, cpu_online_mask, tr->tracing_cpumask);
 330        next_cpu = cpumask_next(raw_smp_processor_id(), current_mask);
 331        put_online_cpus();
 332
 333        if (next_cpu >= nr_cpu_ids)
 334                next_cpu = cpumask_first(current_mask);
 335
 336        if (next_cpu >= nr_cpu_ids) /* Shouldn't happen! */
 337                goto change_mode;
 338
 339        cpumask_clear(current_mask);
 340        cpumask_set_cpu(next_cpu, current_mask);
 341
 342        sched_setaffinity(0, current_mask);
 343        return;
 344
 345 change_mode:
 346        hwlat_data.thread_mode = MODE_NONE;
 347        pr_info(BANNER "cpumask changed while in round-robin mode, switching to mode none\n");
 348}
 349
 350/*
 351 * kthread_fn - The CPU time sampling/hardware latency detection kernel thread
 352 *
 353 * Used to periodically sample the CPU TSC via a call to get_sample. We
 354 * disable interrupts, which does (intentionally) introduce latency since we
 355 * need to ensure nothing else might be running (and thus preempting).
 356 * Obviously this should never be used in production environments.
 357 *
 358 * Executes one loop interaction on each CPU in tracing_cpumask sysfs file.
 359 */
 360static int kthread_fn(void *data)
 361{
 362        u64 interval;
 363
 364        while (!kthread_should_stop()) {
 365
 366                if (hwlat_data.thread_mode == MODE_ROUND_ROBIN)
 367                        move_to_next_cpu();
 368
 369                local_irq_disable();
 370                get_sample();
 371                local_irq_enable();
 372
 373                mutex_lock(&hwlat_data.lock);
 374                interval = hwlat_data.sample_window - hwlat_data.sample_width;
 375                mutex_unlock(&hwlat_data.lock);
 376
 377                do_div(interval, USEC_PER_MSEC); /* modifies interval value */
 378
 379                /* Always sleep for at least 1ms */
 380                if (interval < 1)
 381                        interval = 1;
 382
 383                if (msleep_interruptible(interval))
 384                        break;
 385        }
 386
 387        return 0;
 388}
 389
 390/*
 391 * stop_stop_kthread - Inform the hardware latency sampling/detector kthread to stop
 392 *
 393 * This kicks the running hardware latency sampling/detector kernel thread and
 394 * tells it to stop sampling now. Use this on unload and at system shutdown.
 395 */
 396static void stop_single_kthread(void)
 397{
 398        struct hwlat_kthread_data *kdata = get_cpu_data();
 399        struct task_struct *kthread;
 400
 401        get_online_cpus();
 402        kthread = kdata->kthread;
 403
 404        if (!kthread)
 405                goto out_put_cpus;
 406
 407        kthread_stop(kthread);
 408        kdata->kthread = NULL;
 409
 410out_put_cpus:
 411        put_online_cpus();
 412}
 413
 414
 415/*
 416 * start_single_kthread - Kick off the hardware latency sampling/detector kthread
 417 *
 418 * This starts the kernel thread that will sit and sample the CPU timestamp
 419 * counter (TSC or similar) and look for potential hardware latencies.
 420 */
 421static int start_single_kthread(struct trace_array *tr)
 422{
 423        struct hwlat_kthread_data *kdata = get_cpu_data();
 424        struct cpumask *current_mask = &save_cpumask;
 425        struct task_struct *kthread;
 426        int next_cpu;
 427
 428        get_online_cpus();
 429        if (kdata->kthread)
 430                goto out_put_cpus;
 431
 432        kthread = kthread_create(kthread_fn, NULL, "hwlatd");
 433        if (IS_ERR(kthread)) {
 434                pr_err(BANNER "could not start sampling thread\n");
 435                put_online_cpus();
 436                return -ENOMEM;
 437        }
 438
 439        /* Just pick the first CPU on first iteration */
 440        cpumask_and(current_mask, cpu_online_mask, tr->tracing_cpumask);
 441
 442        if (hwlat_data.thread_mode == MODE_ROUND_ROBIN) {
 443                next_cpu = cpumask_first(current_mask);
 444                cpumask_clear(current_mask);
 445                cpumask_set_cpu(next_cpu, current_mask);
 446
 447        }
 448
 449        sched_setaffinity(kthread->pid, current_mask);
 450
 451        kdata->kthread = kthread;
 452        wake_up_process(kthread);
 453
 454out_put_cpus:
 455        put_online_cpus();
 456        return 0;
 457}
 458
 459/*
 460 * stop_cpu_kthread - Stop a hwlat cpu kthread
 461 */
 462static void stop_cpu_kthread(unsigned int cpu)
 463{
 464        struct task_struct *kthread;
 465
 466        kthread = per_cpu(hwlat_per_cpu_data, cpu).kthread;
 467        if (kthread)
 468                kthread_stop(kthread);
 469        per_cpu(hwlat_per_cpu_data, cpu).kthread = NULL;
 470}
 471
 472/*
 473 * stop_per_cpu_kthreads - Inform the hardware latency sampling/detector kthread to stop
 474 *
 475 * This kicks the running hardware latency sampling/detector kernel threads and
 476 * tells it to stop sampling now. Use this on unload and at system shutdown.
 477 */
 478static void stop_per_cpu_kthreads(void)
 479{
 480        unsigned int cpu;
 481
 482        get_online_cpus();
 483        for_each_online_cpu(cpu)
 484                stop_cpu_kthread(cpu);
 485        put_online_cpus();
 486}
 487
 488/*
 489 * start_cpu_kthread - Start a hwlat cpu kthread
 490 */
 491static int start_cpu_kthread(unsigned int cpu)
 492{
 493        struct task_struct *kthread;
 494        char comm[24];
 495
 496        snprintf(comm, 24, "hwlatd/%d", cpu);
 497
 498        kthread = kthread_create_on_cpu(kthread_fn, NULL, cpu, comm);
 499        if (IS_ERR(kthread)) {
 500                pr_err(BANNER "could not start sampling thread\n");
 501                return -ENOMEM;
 502        }
 503
 504        per_cpu(hwlat_per_cpu_data, cpu).kthread = kthread;
 505        wake_up_process(kthread);
 506
 507        return 0;
 508}
 509
 510#ifdef CONFIG_HOTPLUG_CPU
 511static void hwlat_hotplug_workfn(struct work_struct *dummy)
 512{
 513        struct trace_array *tr = hwlat_trace;
 514        unsigned int cpu = smp_processor_id();
 515
 516        mutex_lock(&trace_types_lock);
 517        mutex_lock(&hwlat_data.lock);
 518        get_online_cpus();
 519
 520        if (!hwlat_busy || hwlat_data.thread_mode != MODE_PER_CPU)
 521                goto out_unlock;
 522
 523        if (!cpumask_test_cpu(cpu, tr->tracing_cpumask))
 524                goto out_unlock;
 525
 526        start_cpu_kthread(cpu);
 527
 528out_unlock:
 529        put_online_cpus();
 530        mutex_unlock(&hwlat_data.lock);
 531        mutex_unlock(&trace_types_lock);
 532}
 533
 534static DECLARE_WORK(hwlat_hotplug_work, hwlat_hotplug_workfn);
 535
 536/*
 537 * hwlat_cpu_init - CPU hotplug online callback function
 538 */
 539static int hwlat_cpu_init(unsigned int cpu)
 540{
 541        schedule_work_on(cpu, &hwlat_hotplug_work);
 542        return 0;
 543}
 544
 545/*
 546 * hwlat_cpu_die - CPU hotplug offline callback function
 547 */
 548static int hwlat_cpu_die(unsigned int cpu)
 549{
 550        stop_cpu_kthread(cpu);
 551        return 0;
 552}
 553
 554static void hwlat_init_hotplug_support(void)
 555{
 556        int ret;
 557
 558        ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "trace/hwlat:online",
 559                                hwlat_cpu_init, hwlat_cpu_die);
 560        if (ret < 0)
 561                pr_warn(BANNER "Error to init cpu hotplug support\n");
 562
 563        return;
 564}
 565#else /* CONFIG_HOTPLUG_CPU */
 566static void hwlat_init_hotplug_support(void)
 567{
 568        return;
 569}
 570#endif /* CONFIG_HOTPLUG_CPU */
 571
 572/*
 573 * start_per_cpu_kthreads - Kick off the hardware latency sampling/detector kthreads
 574 *
 575 * This starts the kernel threads that will sit on potentially all cpus and
 576 * sample the CPU timestamp counter (TSC or similar) and look for potential
 577 * hardware latencies.
 578 */
 579static int start_per_cpu_kthreads(struct trace_array *tr)
 580{
 581        struct cpumask *current_mask = &save_cpumask;
 582        unsigned int cpu;
 583        int retval;
 584
 585        get_online_cpus();
 586        /*
 587         * Run only on CPUs in which hwlat is allowed to run.
 588         */
 589        cpumask_and(current_mask, cpu_online_mask, tr->tracing_cpumask);
 590
 591        for_each_online_cpu(cpu)
 592                per_cpu(hwlat_per_cpu_data, cpu).kthread = NULL;
 593
 594        for_each_cpu(cpu, current_mask) {
 595                retval = start_cpu_kthread(cpu);
 596                if (retval)
 597                        goto out_error;
 598        }
 599        put_online_cpus();
 600
 601        return 0;
 602
 603out_error:
 604        put_online_cpus();
 605        stop_per_cpu_kthreads();
 606        return retval;
 607}
 608
 609static void *s_mode_start(struct seq_file *s, loff_t *pos)
 610{
 611        int mode = *pos;
 612
 613        mutex_lock(&hwlat_data.lock);
 614
 615        if (mode >= MODE_MAX)
 616                return NULL;
 617
 618        return pos;
 619}
 620
 621static void *s_mode_next(struct seq_file *s, void *v, loff_t *pos)
 622{
 623        int mode = ++(*pos);
 624
 625        if (mode >= MODE_MAX)
 626                return NULL;
 627
 628        return pos;
 629}
 630
 631static int s_mode_show(struct seq_file *s, void *v)
 632{
 633        loff_t *pos = v;
 634        int mode = *pos;
 635
 636        if (mode == hwlat_data.thread_mode)
 637                seq_printf(s, "[%s]", thread_mode_str[mode]);
 638        else
 639                seq_printf(s, "%s", thread_mode_str[mode]);
 640
 641        if (mode != MODE_MAX)
 642                seq_puts(s, " ");
 643
 644        return 0;
 645}
 646
 647static void s_mode_stop(struct seq_file *s, void *v)
 648{
 649        seq_puts(s, "\n");
 650        mutex_unlock(&hwlat_data.lock);
 651}
 652
 653static const struct seq_operations thread_mode_seq_ops = {
 654        .start          = s_mode_start,
 655        .next           = s_mode_next,
 656        .show           = s_mode_show,
 657        .stop           = s_mode_stop
 658};
 659
 660static int hwlat_mode_open(struct inode *inode, struct file *file)
 661{
 662        return seq_open(file, &thread_mode_seq_ops);
 663};
 664
 665static void hwlat_tracer_start(struct trace_array *tr);
 666static void hwlat_tracer_stop(struct trace_array *tr);
 667
 668/**
 669 * hwlat_mode_write - Write function for "mode" entry
 670 * @filp: The active open file structure
 671 * @ubuf: The user buffer that contains the value to write
 672 * @cnt: The maximum number of bytes to write to "file"
 673 * @ppos: The current position in @file
 674 *
 675 * This function provides a write implementation for the "mode" interface
 676 * to the hardware latency detector. hwlatd has different operation modes.
 677 * The "none" sets the allowed cpumask for a single hwlatd thread at the
 678 * startup and lets the scheduler handle the migration. The default mode is
 679 * the "round-robin" one, in which a single hwlatd thread runs, migrating
 680 * among the allowed CPUs in a round-robin fashion. The "per-cpu" mode
 681 * creates one hwlatd thread per allowed CPU.
 682 */
 683static ssize_t hwlat_mode_write(struct file *filp, const char __user *ubuf,
 684                                 size_t cnt, loff_t *ppos)
 685{
 686        struct trace_array *tr = hwlat_trace;
 687        const char *mode;
 688        char buf[64];
 689        int ret, i;
 690
 691        if (cnt >= sizeof(buf))
 692                return -EINVAL;
 693
 694        if (copy_from_user(buf, ubuf, cnt))
 695                return -EFAULT;
 696
 697        buf[cnt] = 0;
 698
 699        mode = strstrip(buf);
 700
 701        ret = -EINVAL;
 702
 703        /*
 704         * trace_types_lock is taken to avoid concurrency on start/stop
 705         * and hwlat_busy.
 706         */
 707        mutex_lock(&trace_types_lock);
 708        if (hwlat_busy)
 709                hwlat_tracer_stop(tr);
 710
 711        mutex_lock(&hwlat_data.lock);
 712
 713        for (i = 0; i < MODE_MAX; i++) {
 714                if (strcmp(mode, thread_mode_str[i]) == 0) {
 715                        hwlat_data.thread_mode = i;
 716                        ret = cnt;
 717                }
 718        }
 719
 720        mutex_unlock(&hwlat_data.lock);
 721
 722        if (hwlat_busy)
 723                hwlat_tracer_start(tr);
 724        mutex_unlock(&trace_types_lock);
 725
 726        *ppos += cnt;
 727
 728
 729
 730        return ret;
 731}
 732
 733/*
 734 * The width parameter is read/write using the generic trace_min_max_param
 735 * method. The *val is protected by the hwlat_data lock and is upper
 736 * bounded by the window parameter.
 737 */
 738static struct trace_min_max_param hwlat_width = {
 739        .lock           = &hwlat_data.lock,
 740        .val            = &hwlat_data.sample_width,
 741        .max            = &hwlat_data.sample_window,
 742        .min            = NULL,
 743};
 744
 745/*
 746 * The window parameter is read/write using the generic trace_min_max_param
 747 * method. The *val is protected by the hwlat_data lock and is lower
 748 * bounded by the width parameter.
 749 */
 750static struct trace_min_max_param hwlat_window = {
 751        .lock           = &hwlat_data.lock,
 752        .val            = &hwlat_data.sample_window,
 753        .max            = NULL,
 754        .min            = &hwlat_data.sample_width,
 755};
 756
 757static const struct file_operations thread_mode_fops = {
 758        .open           = hwlat_mode_open,
 759        .read           = seq_read,
 760        .llseek         = seq_lseek,
 761        .release        = seq_release,
 762        .write          = hwlat_mode_write
 763};
 764/**
 765 * init_tracefs - A function to initialize the tracefs interface files
 766 *
 767 * This function creates entries in tracefs for "hwlat_detector".
 768 * It creates the hwlat_detector directory in the tracing directory,
 769 * and within that directory is the count, width and window files to
 770 * change and view those values.
 771 */
 772static int init_tracefs(void)
 773{
 774        int ret;
 775        struct dentry *top_dir;
 776
 777        ret = tracing_init_dentry();
 778        if (ret)
 779                return -ENOMEM;
 780
 781        top_dir = tracefs_create_dir("hwlat_detector", NULL);
 782        if (!top_dir)
 783                return -ENOMEM;
 784
 785        hwlat_sample_window = tracefs_create_file("window", 0640,
 786                                                  top_dir,
 787                                                  &hwlat_window,
 788                                                  &trace_min_max_fops);
 789        if (!hwlat_sample_window)
 790                goto err;
 791
 792        hwlat_sample_width = tracefs_create_file("width", 0644,
 793                                                 top_dir,
 794                                                 &hwlat_width,
 795                                                 &trace_min_max_fops);
 796        if (!hwlat_sample_width)
 797                goto err;
 798
 799        hwlat_thread_mode = trace_create_file("mode", 0644,
 800                                              top_dir,
 801                                              NULL,
 802                                              &thread_mode_fops);
 803        if (!hwlat_thread_mode)
 804                goto err;
 805
 806        return 0;
 807
 808 err:
 809        tracefs_remove(top_dir);
 810        return -ENOMEM;
 811}
 812
 813static void hwlat_tracer_start(struct trace_array *tr)
 814{
 815        int err;
 816
 817        if (hwlat_data.thread_mode == MODE_PER_CPU)
 818                err = start_per_cpu_kthreads(tr);
 819        else
 820                err = start_single_kthread(tr);
 821        if (err)
 822                pr_err(BANNER "Cannot start hwlat kthread\n");
 823}
 824
 825static void hwlat_tracer_stop(struct trace_array *tr)
 826{
 827        if (hwlat_data.thread_mode == MODE_PER_CPU)
 828                stop_per_cpu_kthreads();
 829        else
 830                stop_single_kthread();
 831}
 832
 833static int hwlat_tracer_init(struct trace_array *tr)
 834{
 835        /* Only allow one instance to enable this */
 836        if (hwlat_busy)
 837                return -EBUSY;
 838
 839        hwlat_trace = tr;
 840
 841        hwlat_data.count = 0;
 842        tr->max_latency = 0;
 843        save_tracing_thresh = tracing_thresh;
 844
 845        /* tracing_thresh is in nsecs, we speak in usecs */
 846        if (!tracing_thresh)
 847                tracing_thresh = last_tracing_thresh;
 848
 849        if (tracer_tracing_is_on(tr))
 850                hwlat_tracer_start(tr);
 851
 852        hwlat_busy = true;
 853
 854        return 0;
 855}
 856
 857static void hwlat_tracer_reset(struct trace_array *tr)
 858{
 859        hwlat_tracer_stop(tr);
 860
 861        /* the tracing threshold is static between runs */
 862        last_tracing_thresh = tracing_thresh;
 863
 864        tracing_thresh = save_tracing_thresh;
 865        hwlat_busy = false;
 866}
 867
 868static struct tracer hwlat_tracer __read_mostly =
 869{
 870        .name           = "hwlat",
 871        .init           = hwlat_tracer_init,
 872        .reset          = hwlat_tracer_reset,
 873        .start          = hwlat_tracer_start,
 874        .stop           = hwlat_tracer_stop,
 875        .allow_instances = true,
 876};
 877
 878__init static int init_hwlat_tracer(void)
 879{
 880        int ret;
 881
 882        mutex_init(&hwlat_data.lock);
 883
 884        ret = register_tracer(&hwlat_tracer);
 885        if (ret)
 886                return ret;
 887
 888        hwlat_init_hotplug_support();
 889
 890        init_tracefs();
 891
 892        return 0;
 893}
 894late_initcall(init_hwlat_tracer);
 895