Cache layout improvements
[fio.git] / init.c
diff --git a/init.c b/init.c
index 4a3716bfeb5c058c235e9e529e73c4addef210b2..563dcb7510c048b2468e850e7e961de5f09a2af9 100644 (file)
--- a/init.c
+++ b/init.c
@@ -9,7 +9,9 @@
 #include <string.h>
 #include <errno.h>
 #include <sys/ipc.h>
+#ifndef FIO_NO_HAVE_SHM_H
 #include <sys/shm.h>
+#endif
 #include <sys/types.h>
 #include <sys/stat.h>
 
 
 #include "lib/getopt.h"
 
-#include "fio_version.h"
-
-#if FIO_PATCH > 0
-const char fio_version_string[] =      __fio_stringify(FIO_MAJOR) "."  \
-                                       __fio_stringify(FIO_MINOR) "."  \
-                                       __fio_stringify(FIO_PATCH);
-#else
-const char fio_version_string[] =      __fio_stringify(FIO_MAJOR) "."  \
-                                       __fio_stringify(FIO_MINOR);
-#endif
+const char fio_version_string[] = FIO_VERSION;
 
 #define FIO_RANDSEED           (0xb1899bedUL)
 
@@ -45,7 +38,7 @@ static struct thread_data def_thread;
 struct thread_data *threads = NULL;
 
 int exitall_on_terminate = 0;
-int terse_output = 0;
+int output_format = FIO_OUTPUT_NORMAL;
 int eta_print;
 unsigned long long mlock_size = 0;
 FILE *f_out = NULL;
@@ -74,21 +67,23 @@ static char cmd_optstr[256];
 static int did_arg;
 
 const fio_fp64_t def_percentile_list[FIO_IO_U_LIST_MAX_LEN] = {
-       { .u.f  =  1.0 },
-       { .u.f  =  5.0 },
-       { .u.f  = 10.0 },
-       { .u.f  = 20.0 },
-       { .u.f  = 30.0 },
-       { .u.f  = 40.0 },
-       { .u.f  = 50.0 },
-       { .u.f  = 60.0 },
-       { .u.f  = 70.0 },
-       { .u.f  = 80.0 },
-       { .u.f  = 90.0 },
-       { .u.f  = 95.0 },
-       { .u.f  = 99.0 },
-       { .u.f  = 99.5 },
-       { .u.f  = 99.9 },
+       { .u.f  =  1.00 },
+       { .u.f  =  5.00 },
+       { .u.f  = 10.00 },
+       { .u.f  = 20.00 },
+       { .u.f  = 30.00 },
+       { .u.f  = 40.00 },
+       { .u.f  = 50.00 },
+       { .u.f  = 60.00 },
+       { .u.f  = 70.00 },
+       { .u.f  = 80.00 },
+       { .u.f  = 90.00 },
+       { .u.f  = 95.00 },
+       { .u.f  = 99.00 },
+       { .u.f  = 99.50 },
+       { .u.f  = 99.90 },
+       { .u.f  = 99.95 },
+       { .u.f  = 99.99 },
 };
 
 #define FIO_CLIENT_FLAG                (1 << 16)
@@ -123,6 +118,11 @@ static struct option l_opts[FIO_NR_OPTIONS] = {
                .has_arg        = optional_argument,
                .val            = 'm' | FIO_CLIENT_FLAG,
        },
+       {
+               .name           = (char *) "output-format",
+               .has_arg        = optional_argument,
+               .val            = 'F' | FIO_CLIENT_FLAG,
+       },
        {
                .name           = (char *) "version",
                .has_arg        = no_argument,
@@ -221,6 +221,7 @@ static void free_shm(void)
 
                threads = NULL;
                file_hash_exit();
+               flow_exit();
                fio_debug_jobp = NULL;
                shmdt(tp);
                shmctl(shm_id, IPC_RMID, &sbuf);
@@ -255,7 +256,7 @@ static int setup_thread_area(void)
                shm_id = shmget(0, size, IPC_CREAT | 0600);
                if (shm_id != -1)
                        break;
-               if (errno != EINVAL) {
+               if (errno != EINVAL && errno != ENOMEM && errno != ENOSPC) {
                        perror("shmget");
                        break;
                }
@@ -277,6 +278,9 @@ static int setup_thread_area(void)
        fio_debug_jobp = (void *) hash + file_hash_size;
        *fio_debug_jobp = -1;
        file_hash_init(hash);
+
+       flow_init();
+
        return 0;
 }
 
@@ -315,6 +319,10 @@ static struct thread_data *get_new_job(int global, struct thread_data *parent,
        profile_add_hooks(td);
 
        td->thread_number = thread_number;
+
+       if (!parent || !parent->o.group_reporting)
+               stat_number++;
+
        return td;
 }
 
@@ -324,6 +332,7 @@ static void put_job(struct thread_data *td)
                return;
 
        profile_td_exit(td);
+       flow_exit_job(td);
 
        if (td->error)
                log_info("fio: %s\n", td->verror);
@@ -364,6 +373,8 @@ static int setup_rate(struct thread_data *td)
                ret = __setup_rate(td, DDIR_READ);
        if (td->o.rate[DDIR_WRITE] || td->o.rate_iops[DDIR_WRITE])
                ret |= __setup_rate(td, DDIR_WRITE);
+       if (td->o.rate[DDIR_TRIM] || td->o.rate_iops[DDIR_TRIM])
+               ret |= __setup_rate(td, DDIR_TRIM);
 
        return ret;
 }
@@ -372,7 +383,9 @@ static int fixed_block_size(struct thread_options *o)
 {
        return o->min_bs[DDIR_READ] == o->max_bs[DDIR_READ] &&
                o->min_bs[DDIR_WRITE] == o->max_bs[DDIR_WRITE] &&
-               o->min_bs[DDIR_READ] == o->min_bs[DDIR_WRITE];
+               o->min_bs[DDIR_TRIM] == o->max_bs[DDIR_TRIM] &&
+               o->min_bs[DDIR_READ] == o->min_bs[DDIR_WRITE] &&
+               o->min_bs[DDIR_READ] == o->min_bs[DDIR_TRIM];
 }
 
 /*
@@ -428,8 +441,14 @@ static int fixup_options(struct thread_data *td)
                o->min_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
        if (!o->max_bs[DDIR_WRITE])
                o->max_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
+       if (!o->min_bs[DDIR_TRIM])
+               o->min_bs[DDIR_TRIM] = o->bs[DDIR_TRIM];
+       if (!o->max_bs[DDIR_TRIM])
+               o->max_bs[DDIR_TRIM] = o->bs[DDIR_TRIM];
+
 
        o->rw_min_bs = min(o->min_bs[DDIR_READ], o->min_bs[DDIR_WRITE]);
+       o->rw_min_bs = min(o->min_bs[DDIR_TRIM], o->rw_min_bs);
 
        /*
         * For random IO, allow blockalign offset other than min_bs.
@@ -438,9 +457,12 @@ static int fixup_options(struct thread_data *td)
                o->ba[DDIR_READ] = o->min_bs[DDIR_READ];
        if (!o->ba[DDIR_WRITE] || !td_random(td))
                o->ba[DDIR_WRITE] = o->min_bs[DDIR_WRITE];
+       if (!o->ba[DDIR_TRIM] || !td_random(td))
+               o->ba[DDIR_TRIM] = o->min_bs[DDIR_TRIM];
 
        if ((o->ba[DDIR_READ] != o->min_bs[DDIR_READ] ||
-           o->ba[DDIR_WRITE] != o->min_bs[DDIR_WRITE]) &&
+           o->ba[DDIR_WRITE] != o->min_bs[DDIR_WRITE] ||
+           o->ba[DDIR_TRIM] != o->min_bs[DDIR_TRIM]) &&
            !o->norandommap) {
                log_err("fio: Any use of blockalign= turns off randommap\n");
                o->norandommap = 1;
@@ -493,15 +515,19 @@ static int fixup_options(struct thread_data *td)
        if (o->open_files > o->nr_files || !o->open_files)
                o->open_files = o->nr_files;
 
-       if (((o->rate[0] + o->rate[1]) && (o->rate_iops[0] + o->rate_iops[1]))||
-           ((o->ratemin[0] + o->ratemin[1]) && (o->rate_iops_min[0] +
-               o->rate_iops_min[1]))) {
+       if (((o->rate[DDIR_READ] + o->rate[DDIR_WRITE] + o->rate[DDIR_TRIM]) &&
+           (o->rate_iops[DDIR_READ] + o->rate_iops[DDIR_WRITE] + o->rate_iops[DDIR_TRIM])) ||
+           ((o->ratemin[DDIR_READ] + o->ratemin[DDIR_WRITE] + o->ratemin[DDIR_TRIM]) &&
+           (o->rate_iops_min[DDIR_READ] + o->rate_iops_min[DDIR_WRITE] + o->rate_iops_min[DDIR_TRIM]))) {
                log_err("fio: rate and rate_iops are mutually exclusive\n");
                ret = 1;
        }
-       if ((o->rate[0] < o->ratemin[0]) || (o->rate[1] < o->ratemin[1]) ||
-           (o->rate_iops[0] < o->rate_iops_min[0]) ||
-           (o->rate_iops[1] < o->rate_iops_min[1])) {
+       if ((o->rate[DDIR_READ] < o->ratemin[DDIR_READ]) ||
+           (o->rate[DDIR_WRITE] < o->ratemin[DDIR_WRITE]) ||
+           (o->rate[DDIR_TRIM] < o->ratemin[DDIR_TRIM]) ||
+           (o->rate_iops[DDIR_READ] < o->rate_iops_min[DDIR_READ]) ||
+           (o->rate_iops[DDIR_WRITE] < o->rate_iops_min[DDIR_WRITE]) ||
+           (o->rate_iops[DDIR_TRIM] < o->rate_iops_min[DDIR_TRIM])) {
                log_err("fio: minimum rate exceeds rate\n");
                ret = 1;
        }
@@ -550,6 +576,35 @@ static int fixup_options(struct thread_data *td)
        }
 #endif
 
+#ifdef WIN32
+       /*
+        * Windows doesn't support O_DIRECT or O_SYNC with the _open interface,
+        * so fail if we're passed those flags
+        */
+       if ((td->io_ops->flags & FIO_SYNCIO) && (td->o.odirect || td->o.sync_io)) {
+               log_err("fio: Windows does not support direct or non-buffered io with"
+                               " the synchronous ioengines. Use the 'windowsaio' ioengine"
+                               " with 'direct=1' and 'iodepth=1' instead.\n");
+               ret = 1;
+       }
+#endif
+
+       /*
+        * For fully compressible data, just zero them at init time.
+        * It's faster than repeatedly filling it.
+        */
+       if (td->o.compress_percentage == 100) {
+               td->o.zero_buffers = 1;
+               td->o.compress_percentage = 0;
+       }
+
+       /*
+        * Using a non-uniform random distribution excludes usage of
+        * a random map
+        */
+       if (td->o.random_distribution != FIO_RAND_DIST_RANDOM)
+               td->o.norandommap = 1;
+
        return ret;
 }
 
@@ -655,31 +710,6 @@ void td_fill_rand_seeds(struct thread_data *td)
        init_rand_seed(&td->buf_state, td->rand_seeds[7]);
 }
 
-/*
- * Initialize the various random states we need (random io, block size ranges,
- * read/write mix, etc).
- */
-static int init_random_state(struct thread_data *td)
-{
-       int fd;
-
-       fd = open("/dev/urandom", O_RDONLY);
-       if (fd == -1) {
-               td_verror(td, errno, "open");
-               return 1;
-       }
-
-       if (read(fd, td->rand_seeds, sizeof(td->rand_seeds)) <
-           (int) sizeof(td->rand_seeds)) {
-               td_verror(td, EIO, "read");
-               close(fd);
-               return 1;
-       }
-
-       close(fd);
-       td_fill_rand_seeds(td);
-       return 0;
-}
 
 /*
  * Initializes the ioengine configured for a job, if it has not been done so
@@ -737,6 +767,24 @@ int ioengine_load(struct thread_data *td)
        return 0;
 }
 
+static void init_flags(struct thread_data *td)
+{
+       struct thread_options *o = &td->o;
+
+       if (o->verify_backlog)
+               td->flags |= TD_F_VER_BACKLOG;
+       if (o->trim_backlog)
+               td->flags |= TD_F_TRIM_BACKLOG;
+       if (o->read_iolog_file)
+               td->flags |= TD_F_READ_IOLOG;
+       if (o->refill_buffers)
+               td->flags |= TD_F_REFILL_BUFFERS;
+       if (o->scramble_buffers)
+               td->flags |= TD_F_SCRAMBLE_BUFFERS;
+       if (o->verify != VERIFY_NONE)
+               td->flags |= TD_F_VER_NONE;
+}
+
 /*
  * Adds a job to the list of things todo. Sanitizes the various options
  * to make sure we don't have conflicts, and initializes various
@@ -745,7 +793,8 @@ int ioengine_load(struct thread_data *td)
 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
 {
        const char *ddir_str[] = { NULL, "read", "write", "rw", NULL,
-                                  "randread", "randwrite", "randrw" };
+                                  "randread", "randwrite", "randrw",
+                                  "trim", NULL, NULL, NULL, "randtrim" };
        unsigned int i;
        char fname[PATH_MAX];
        int numjobs, file_alloced;
@@ -756,6 +805,8 @@ static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
        if (td == &def_thread)
                return 0;
 
+       init_flags(td);
+
        /*
         * if we are just dumping the output command line, don't add the job
         */
@@ -796,6 +847,8 @@ static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
        if (fixup_options(td))
                goto err;
 
+       flow_init_job(td);
+
        /*
         * IO engines only need this for option callbacks, and the address may
         * change in subprocesses.
@@ -810,7 +863,7 @@ static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
                        f->real_file_size = -1ULL;
        }
 
-       td->mutex = fio_mutex_init(0);
+       td->mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
 
        td->ts.clat_percentiles = td->o.clat_percentiles;
        if (td->o.overwrite_plist)
@@ -818,10 +871,12 @@ static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
        else
                memcpy(td->ts.percentile_list, def_percentile_list, sizeof(def_percentile_list));
 
-       td->ts.clat_stat[0].min_val = td->ts.clat_stat[1].min_val = ULONG_MAX;
-       td->ts.slat_stat[0].min_val = td->ts.slat_stat[1].min_val = ULONG_MAX;
-       td->ts.lat_stat[0].min_val = td->ts.lat_stat[1].min_val = ULONG_MAX;
-       td->ts.bw_stat[0].min_val = td->ts.bw_stat[1].min_val = ULONG_MAX;
+       for (i = 0; i < DDIR_RWDIR_CNT; i++) {
+               td->ts.clat_stat[i].min_val = ULONG_MAX;
+               td->ts.slat_stat[i].min_val = ULONG_MAX;
+               td->ts.lat_stat[i].min_val = ULONG_MAX;
+               td->ts.bw_stat[i].min_val = ULONG_MAX;
+       }
        td->ddir_seq_nr = td->o.ddir_seq_nr;
 
        if ((td->o.stonewall || td->o.new_group) && prev_group_jobs) {
@@ -832,8 +887,10 @@ static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
        td->groupid = groupid;
        prev_group_jobs++;
 
-       if (init_random_state(td))
+       if (init_random_state(td, td->rand_seeds, sizeof(td->rand_seeds))) {
+               td_verror(td, errno, "init_random_state");
                goto err;
+       }
 
        if (setup_rate(td))
                goto err;
@@ -851,7 +908,7 @@ static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
        if (!td->o.name)
                td->o.name = strdup(jobname);
 
-       if (!terse_output) {
+       if (output_format == FIO_OUTPUT_NORMAL) {
                if (!job_add_num) {
                        if (!strcmp(td->io_ops->name, "cpuio")) {
                                log_info("%s: ioengine=cpu, cpuload=%u,"
@@ -859,18 +916,20 @@ static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
                                                        td->o.cpuload,
                                                        td->o.cpucycle);
                        } else {
-                               char *c1, *c2, *c3, *c4;
+                               char *c1, *c2, *c3, *c4, *c5, *c6;
 
                                c1 = to_kmg(td->o.min_bs[DDIR_READ]);
                                c2 = to_kmg(td->o.max_bs[DDIR_READ]);
                                c3 = to_kmg(td->o.min_bs[DDIR_WRITE]);
                                c4 = to_kmg(td->o.max_bs[DDIR_WRITE]);
+                               c5 = to_kmg(td->o.min_bs[DDIR_TRIM]);
+                               c6 = to_kmg(td->o.max_bs[DDIR_TRIM]);
 
-                               log_info("%s: (g=%d): rw=%s, bs=%s-%s/%s-%s,"
+                               log_info("%s: (g=%d): rw=%s, bs=%s-%s/%s-%s/%s-%s,"
                                         " ioengine=%s, iodepth=%u\n",
                                                td->o.name, td->groupid,
                                                ddir_str[td->o.td_ddir],
-                                               c1, c2, c3, c4,
+                                               c1, c2, c3, c4, c5, c6,
                                                td->io_ops->name,
                                                td->o.iodepth);
 
@@ -878,6 +937,8 @@ static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
                                free(c2);
                                free(c3);
                                free(c4);
+                               free(c5);
+                               free(c6);
                        }
                } else if (job_add_num == 1)
                        log_info("...\n");
@@ -1170,7 +1231,7 @@ static int fill_def_thread(void)
 
        fio_getaffinity(getpid(), &def_thread.o.cpumask);
        def_thread.o.timeout = def_timeout;
-
+       def_thread.o.error_dump = 1;
        /*
         * fill default options
         */
@@ -1180,18 +1241,19 @@ static int fill_def_thread(void)
 
 static void usage(const char *name)
 {
-       printf("fio %s\n", fio_version_string);
+       printf("%s\n", fio_version_string);
        printf("%s [options] [job options] <job file(s)>\n", name);
        printf("  --debug=options\tEnable debug logging. May be one/more of:\n"
                "\t\t\tprocess,file,io,mem,blktrace,verify,random,parse,\n"
                "\t\t\tdiskutil,job,mutex,profile,time,net\n");
        printf("  --output\t\tWrite output to file\n");
-       printf("  --timeout\t\tRuntime in seconds\n");
+       printf("  --runtime\t\tRuntime in seconds\n");
        printf("  --latency-log\t\tGenerate per-job latency logs\n");
        printf("  --bandwidth-log\tGenerate per-job bandwidth logs\n");
        printf("  --minimal\t\tMinimal (terse) output\n");
-       printf("  --version\t\tPrint version info and exit\n");
+       printf("  --output-format=x\tOutput format (terse,json,normal)\n");
        printf("  --terse-version=x\tSet terse version output format to 'x'\n");
+       printf("  --version\t\tPrint version info and exit\n");
        printf("  --help\t\tPrint this page\n");
        printf("  --cmdhelp=cmd\t\tPrint command help, \"all\" for all of"
                " them\n");
@@ -1383,7 +1445,17 @@ int parse_cmd_line(int argc, char *argv[])
                        f_err = f_out;
                        break;
                case 'm':
-                       terse_output = 1;
+                       output_format = FIO_OUTPUT_TERSE;
+                       break;
+               case 'F':
+                       if (!strcmp(optarg, "minimal") ||
+                           !strcmp(optarg, "terse") ||
+                           !strcmp(optarg, "csv"))
+                               output_format = FIO_OUTPUT_TERSE;
+                       else if (!strcmp(optarg, "json"))
+                               output_format = FIO_OUTPUT_JSON;
+                       else
+                               output_format = FIO_OUTPUT_NORMAL;
                        break;
                case 'h':
                        if (!cur_client) {
@@ -1411,13 +1483,14 @@ int parse_cmd_line(int argc, char *argv[])
                        break;
                case 'v':
                        if (!cur_client) {
-                               log_info("fio %s\n", fio_version_string);
+                               log_info("%s\n", fio_version_string);
                                do_exit++;
                        }
                        break;
                case 'V':
                        terse_version = atoi(optarg);
-                       if (!(terse_version == 2 || terse_version == 3)) {
+                       if (!(terse_version == 2 || terse_version == 3 ||
+                            terse_version == 4)) {
                                log_err("fio: bad terse version format\n");
                                exit_val = 1;
                                do_exit++;
@@ -1535,6 +1608,18 @@ int parse_cmd_line(int argc, char *argv[])
                                exit_val = 1;
                                break;
                        }
+                       /*
+                        * If the next argument exists and isn't an option,
+                        * assume it's a job file for this client only.
+                        */
+                       while (optind < argc) {
+                               if (!strncmp(argv[optind], "--", 2) ||
+                                   !strncmp(argv[optind], "-", 1))
+                                       break;
+
+                               fio_client_add_ini_file(cur_client, argv[optind]);
+                               optind++;
+                       }
                        break;
                default:
                        do_exit++;
@@ -1605,6 +1690,11 @@ int parse_options(int argc, char *argv[])
                                free(ini_file[i]);
                        }
                }
+       } else if (nr_clients) {
+               if (fill_def_thread())
+                       return 1;
+               if (fio_clients_send_ini(NULL))
+                       return 1;
        }
 
        free(ini_file);
@@ -1636,8 +1726,8 @@ int parse_options(int argc, char *argv[])
                fio_gtod_cpu = def_thread.o.gtod_cpu;
        }
 
-       if (!terse_output)
-               log_info("fio %s\n", fio_version_string);
+       if (output_format == FIO_OUTPUT_NORMAL)
+               log_info("%s\n", fio_version_string);
 
        return 0;
 }