Merge branch 'arm-detect-pmull' of https://github.com/sitsofe/fio
[fio.git] / engines / cpu.c
index 09872508e12129d1408040eadb40055c0ebb8181..ce74dbcea8758e776b03b02d41ebb1489cc490c7 100644 (file)
@@ -8,11 +8,26 @@
 #include "../fio.h"
 #include "../optgroup.h"
 
+// number of 32 bit integers to sort
+size_t qsort_size = (256 * (1ULL << 10)); // 256KB
+
+struct mwc {
+       uint32_t w;
+       uint32_t z;
+};
+
+enum stress_mode {
+       FIO_CPU_NOOP = 0,
+       FIO_CPU_QSORT = 1,
+};
+
 struct cpu_options {
        void *pad;
        unsigned int cpuload;
        unsigned int cpucycle;
+       enum stress_mode cpumode;
        unsigned int exit_io_done;
+       int32_t *qsort_data;
 };
 
 static struct fio_option options[] = {
@@ -25,6 +40,26 @@ static struct fio_option options[] = {
                .category = FIO_OPT_C_ENGINE,
                .group  = FIO_OPT_G_INVALID,
        },
+       {
+               .name     = "cpumode",
+               .lname    = "cpumode",
+               .type     = FIO_OPT_STR,
+               .help     = "Stress mode",
+               .off1     = offsetof(struct cpu_options, cpumode),
+               .def      = "noop",
+               .posval = {
+                         { .ival = "noop",
+                           .oval = FIO_CPU_NOOP,
+                           .help = "NOOP instructions",
+                         },
+                         { .ival = "qsort",
+                           .oval = FIO_CPU_QSORT,
+                           .help = "QSORT computation",
+                         },
+               },
+               .category = FIO_OPT_C_ENGINE,
+               .group    = FIO_OPT_G_INVALID,
+       },
        {
                .name   = "cpuchunks",
                .lname  = "CPU chunk",
@@ -52,6 +87,91 @@ static struct fio_option options[] = {
        },
 };
 
+/*
+ *  mwc32()
+ *      Multiply-with-carry random numbers
+ *      fast pseudo random number generator, see
+ *      http://www.cse.yorku.ca/~oz/marsaglia-rng.html
+ */
+uint32_t mwc32(struct mwc *mwc)
+{
+        mwc->z = 36969 * (mwc->z & 65535) + (mwc->z >> 16);
+        mwc->w = 18000 * (mwc->w & 65535) + (mwc->w >> 16);
+        return (mwc->z << 16) + mwc->w;
+}
+
+/*
+ *  stress_qsort_cmp_1()
+ *     qsort comparison - sort on int32 values
+ */
+static int stress_qsort_cmp_1(const void *p1, const void *p2)
+{
+       const int32_t *i1 = (const int32_t *)p1;
+       const int32_t *i2 = (const int32_t *)p2;
+
+       if (*i1 > *i2)
+               return 1;
+       else if (*i1 < *i2)
+               return -1;
+       else
+               return 0;
+}
+
+/*
+ *  stress_qsort_cmp_2()
+ *     qsort comparison - reverse sort on int32 values
+ */
+static int stress_qsort_cmp_2(const void *p1, const void *p2)
+{
+       return stress_qsort_cmp_1(p2, p1);
+}
+
+/*
+ *  stress_qsort_cmp_3()
+ *     qsort comparison - sort on int8 values
+ */
+static int stress_qsort_cmp_3(const void *p1, const void *p2)
+{
+       const int8_t *i1 = (const int8_t *)p1;
+       const int8_t *i2 = (const int8_t *)p2;
+
+       /* Force re-ordering on 8 bit value */
+       return *i1 - *i2;
+}
+
+static int do_qsort(struct thread_data *td)
+{
+       struct thread_options *o = &td->o;
+       struct cpu_options *co = td->eo;
+       struct timespec start, now;
+
+       fio_get_mono_time(&start);
+
+       /* Sort "random" data */
+       qsort(co->qsort_data, qsort_size, sizeof(*(co->qsort_data)), stress_qsort_cmp_1);
+
+       /* Reverse sort */
+       qsort(co->qsort_data, qsort_size, sizeof(*(co->qsort_data)), stress_qsort_cmp_2);
+
+       /* And re-order by byte compare */
+       qsort((uint8_t *)co->qsort_data, qsort_size * 4, sizeof(uint8_t), stress_qsort_cmp_3);
+
+       /* Reverse sort this again */
+       qsort(co->qsort_data, qsort_size, sizeof(*(co->qsort_data)), stress_qsort_cmp_2);
+       fio_get_mono_time(&now);
+
+       /* Adjusting cpucycle automatically to be as close as possible to the
+        * expected cpuload The time to execute do_qsort() may change over time
+        * as per : - the job concurrency - the cpu clock adjusted by the power
+        * management After every do_qsort() call, the next thinktime is
+        * adjusted regarding the last run performance
+        */
+       co->cpucycle = utime_since(&start, &now);
+       o->thinktime = ((unsigned long long) co->cpucycle *
+                               (100 - co->cpuload)) / co->cpuload;
+
+       return 0;
+}
 
 static enum fio_q_status fio_cpuio_queue(struct thread_data *td,
                                         struct io_u fio_unused *io_u)
@@ -63,14 +183,69 @@ static enum fio_q_status fio_cpuio_queue(struct thread_data *td,
                return FIO_Q_BUSY;
        }
 
-       usec_spin(co->cpucycle);
+       switch (co->cpumode) {
+       case FIO_CPU_NOOP:
+               usec_spin(co->cpucycle);
+               break;
+       case FIO_CPU_QSORT:
+               do_qsort(td);
+               break;
+       }
+
        return FIO_Q_COMPLETED;
 }
 
+static int noop_init(struct thread_data *td)
+{
+       struct cpu_options *co = td->eo;
+
+       log_info("%s (noop): ioengine=%s, cpuload=%u, cpucycle=%u\n",
+               td->o.name, td->io_ops->name, co->cpuload, co->cpucycle);
+       return 0;
+}
+
+static int qsort_cleanup(struct thread_data *td)
+{
+       struct cpu_options *co = td->eo;
+
+       if (co->qsort_data) {
+               free(co->qsort_data);
+               co->qsort_data = NULL;
+       }
+
+       return 0;
+}
+
+static int qsort_init(struct thread_data *td)
+{
+       /* Setting up a default entropy */
+       struct mwc mwc = { 521288629UL, 362436069UL };
+       struct cpu_options *co = td->eo;
+       int32_t *ptr;
+       int i;
+
+       co->qsort_data = calloc(qsort_size, sizeof(*co->qsort_data));
+       if (co->qsort_data == NULL) {
+               td_verror(td, ENOMEM, "qsort_init");
+               return 1;
+       }
+
+       /* This is expensive, init the memory once */
+       for (ptr = co->qsort_data, i = 0; i < qsort_size; i++)
+               *ptr++ = mwc32(&mwc);
+
+       log_info("%s (qsort): ioengine=%s, cpuload=%u, cpucycle=%u\n",
+               td->o.name, td->io_ops->name, co->cpuload, co->cpucycle);
+
+       return 0;
+}
+
 static int fio_cpuio_init(struct thread_data *td)
 {
        struct thread_options *o = &td->o;
        struct cpu_options *co = td->eo;
+       int td_previous_state;
+       char *msg;
 
        if (!co->cpuload) {
                td_vmsg(td, EINVAL, "cpu thread needs rate (cpuload=)","cpuio");
@@ -80,21 +255,59 @@ static int fio_cpuio_init(struct thread_data *td)
        if (co->cpuload > 100)
                co->cpuload = 100;
 
+       /* Saving the current thread state */
+       td_previous_state = td->runstate;
+
+       /* Reporting that we are preparing the engine
+        * This is useful as the qsort() calibration takes time
+        * This prevents the job from starting before init is completed
+        */
+       td_set_runstate(td, TD_SETTING_UP);
+
        /*
         * set thinktime_sleep and thinktime_spin appropriately
         */
        o->thinktime_blocks = 1;
+       o->thinktime_blocks_type = THINKTIME_BLOCKS_TYPE_COMPLETE;
        o->thinktime_spin = 0;
-       o->thinktime = (co->cpucycle * (100 - co->cpuload)) / co->cpuload;
+       o->thinktime = ((unsigned long long) co->cpucycle *
+                               (100 - co->cpuload)) / co->cpuload;
 
        o->nr_files = o->open_files = 1;
 
-       log_info("%s: ioengine=%s, cpuload=%u, cpucycle=%u\n",
-               td->o.name, td->io_ops->name, co->cpuload, co->cpucycle);
+       switch (co->cpumode) {
+       case FIO_CPU_NOOP:
+               noop_init(td);
+               break;
+       case FIO_CPU_QSORT:
+               qsort_init(td);
+               break;
+       default:
+               if (asprintf(&msg, "bad cpu engine mode: %d", co->cpumode) < 0)
+                       msg = NULL;
+               td_vmsg(td, EINVAL, msg ? : "(?)", __func__);
+               free(msg);
+               return 1;
+       }
 
+       /* Let's restore the previous state. */
+       td_set_runstate(td, td_previous_state);
        return 0;
 }
 
+static void fio_cpuio_cleanup(struct thread_data *td)
+{
+       struct cpu_options *co = td->eo;
+
+       switch (co->cpumode) {
+       case FIO_CPU_NOOP:
+               break;
+       case FIO_CPU_QSORT:
+               qsort_cleanup(td);
+               break;
+       }
+}
+
 static int fio_cpuio_open(struct thread_data fio_unused *td,
                          struct fio_file fio_unused *f)
 {
@@ -102,12 +315,13 @@ static int fio_cpuio_open(struct thread_data fio_unused *td,
 }
 
 static struct ioengine_ops ioengine = {
-       .name           = "cpuio",
-       .version        = FIO_IOOPS_VERSION,
-       .queue          = fio_cpuio_queue,
-       .init           = fio_cpuio_init,
-       .open_file      = fio_cpuio_open,
-       .flags          = FIO_SYNCIO | FIO_DISKLESSIO | FIO_NOIO,
+       .name                   = "cpuio",
+       .version                = FIO_IOOPS_VERSION,
+       .queue                  = fio_cpuio_queue,
+       .init                   = fio_cpuio_init,
+       .cleanup                = fio_cpuio_cleanup,
+       .open_file              = fio_cpuio_open,
+       .flags                  = FIO_SYNCIO | FIO_DISKLESSIO | FIO_NOIO,
        .options                = options,
        .option_struct_size     = sizeof(struct cpu_options),
 };