#include <time.h>
#include <ctype.h>
#include <sched.h>
+#include <libaio.h>
+#include <math.h>
+#include <limits.h>
+#include <assert.h>
#include <sys/time.h>
#include <sys/types.h>
#include <sys/stat.h>
#include <semaphore.h>
#include <sys/ipc.h>
#include <sys/shm.h>
+#include <sys/ioctl.h>
#include <asm/unistd.h>
+#include <asm/types.h>
+#include <asm/bitops.h>
-#define MAX_JOBS (1024)
+#include "arch.h"
+#include "list.h"
+#include "md5.h"
-/*
- * assume we don't have _get either, if _set isn't defined
- */
-#ifndef __NR_ioprio_set
-
-#if defined(__i386__)
-#define __NR_ioprio_set 289
-#define __NR_ioprio_get 290
-#elif defined(__powerpc__) || defined(__powerpc64__)
-#define __NR_ioprio_set 273
-#define __NR_ioprio_get 274
-#elif defined(__x86_64__)
-#define __NR_ioprio_set 251
-#define __NR_ioprio_get 252
-#elif defined(__ia64__)
-#define __NR_ioprio_set 1274
-#define __NR_ioprio_get 1275
-#elif defined(__alpha__)
-#define __NR_ioprio_set 442
-#define __NR_ioprio_get 443
-#elif defined(__s390x__) || defined(__s390__)
-#define __NR_ioprio_set 282
-#define __NR_ioprio_get 283
-#else
-#error "Unsupported arch"
+#ifndef BLKGETSIZE64
+#define BLKGETSIZE64 _IOR(0x12,114,size_t)
#endif
-#endif
+#define MAX_JOBS (1024)
static int ioprio_set(int which, int who, int ioprio)
{
return syscall(__NR_ioprio_set, which, who, ioprio);
}
+/*
+ * we want fadvise64 really, but it's so tangled... later
+ */
+static int fadvise(int fd, loff_t offset, size_t len, int advice)
+{
+#if 0
+ return syscall(__NR_fadvise64, fd, offset, offset >> 32, len, advice);
+#else
+ return posix_fadvise(fd, (off_t) offset, len, advice);
+#endif
+}
+
enum {
IOPRIO_WHO_PROCESS = 1,
IOPRIO_WHO_PGRP,
#define MASK (4095)
#define DEF_BS (4096)
-#define DEF_TIMEOUT (30)
+#define DEF_TIMEOUT (0)
#define DEF_RATE_CYCLE (1000)
#define DEF_ODIRECT (1)
#define DEF_SEQUENTIAL (1)
-#define DEF_WRITESTAT (0)
#define DEF_RAND_REPEAT (1)
+#define DEF_OVERWRITE (0)
+#define DEF_CREATE (1)
+#define DEF_INVALIDATE (1)
+#define DEF_SYNCIO (0)
+#define DEF_RANDSEED (0xb1899bedUL)
+#define DEF_BWAVGTIME (500)
+#define DEF_CREATE_SER (1)
+#define DEF_CREATE_FSYNC (1)
+#define DEF_LOOPS (1)
+#define DEF_VERIFY (0)
+#define DEF_STONEWALL (0)
+#define DEF_NUMJOBS (1)
#define ALIGN(buf) (char *) (((unsigned long) (buf) + MASK) & ~(MASK))
-static int sequential = DEF_SEQUENTIAL;
-static int write_stat = DEF_WRITESTAT;
static int repeatable = DEF_RAND_REPEAT;
-static int timeout = DEF_TIMEOUT;
-static int odirect = DEF_ODIRECT;
-static int global_bs = DEF_BS;
+static int rate_quit = 1;
+static int write_lat_log;
+static int write_bw_log;
+static int exitall_on_terminate;
static int thread_number;
static char *ini_file;
-static int shm_id;
+static int max_jobs = MAX_JOBS;
+
+static char run_str[MAX_JOBS + 1];
-static cpu_set_t def_cpumask;
+static int shm_id;
enum {
DDIR_READ = 0,
DDIR_WRITE,
};
+/*
+ * thread life cycle
+ */
+enum {
+ TD_NOT_CREATED = 0,
+ TD_CREATED,
+ TD_RUNNING,
+ TD_VERIFYING,
+ TD_EXITED,
+ TD_REAPED,
+};
+
+enum {
+ MEM_MALLOC,
+ MEM_SHM,
+};
+
+/*
+ * The io unit
+ */
+struct io_u {
+ struct iocb iocb;
+ struct timeval start_time;
+ struct timeval issue_time;
+
+ char *buf;
+ unsigned int buflen;
+ unsigned long long offset;
+
+ struct list_head list;
+};
+
+struct io_stat {
+ unsigned long val;
+ unsigned long val_sq;
+ unsigned long max_val;
+ unsigned long min_val;
+ unsigned long samples;
+};
+
+struct io_sample {
+ unsigned long time;
+ unsigned long val;
+};
+
+struct io_log {
+ unsigned long nr_samples;
+ unsigned long max_samples;
+ struct io_sample *log;
+};
+
+struct io_piece {
+ struct list_head list;
+ unsigned long long offset;
+ unsigned int len;
+};
+
+#define FIO_HDR_MAGIC 0xf00baaef
+
+struct verify_header {
+ unsigned int fio_magic;
+ unsigned int len;
+ char md5_digest[MD5_HASH_WORDS * 4];
+};
+
+#define td_read(td) ((td)->ddir == DDIR_READ)
+#define td_write(td) ((td)->ddir == DDIR_WRITE)
+#define should_fsync(td) (td_write(td) && !(td)->odirect)
+
+#define BLOCKS_PER_MAP (8 * sizeof(long))
+#define TO_MAP_BLOCK(td, b) ((b) - ((td)->file_offset / (td)->min_bs))
+#define RAND_MAP_IDX(td, b) (TO_MAP_BLOCK(td, b) / BLOCKS_PER_MAP)
+#define RAND_MAP_BIT(td, b) (TO_MAP_BLOCK(td, b) & (BLOCKS_PER_MAP - 1))
+
struct thread_data {
char file_name[256];
+ char directory[256];
int thread_number;
int error;
int fd;
- int stat_fd;
pid_t pid;
+ char *orig_buffer;
volatile int terminate;
+ volatile int runstate;
+ volatile int old_runstate;
unsigned int ddir;
unsigned int ioprio;
unsigned int sequential;
unsigned int bs;
+ unsigned int min_bs;
+ unsigned int max_bs;
unsigned int odirect;
- unsigned int delay_sleep;
+ unsigned int thinktime;
unsigned int fsync_blocks;
+ unsigned int start_delay;
+ unsigned int timeout;
+ unsigned int use_aio;
+ unsigned int create_file;
+ unsigned int overwrite;
+ unsigned int invalidate_cache;
+ unsigned int bw_avg_time;
+ unsigned int create_serialize;
+ unsigned int create_fsync;
+ unsigned int loops;
+ unsigned long long file_size;
+ unsigned long long file_offset;
+ unsigned int sync_io;
+ unsigned int mem_type;
+ unsigned int verify;
+ unsigned int stonewall;
+ unsigned int numjobs;
cpu_set_t cpumask;
+ struct drand48_data bsrange_state;
+ struct drand48_data verify_state;
+
+ int shm_id;
+
+ off_t cur_off;
+
+ io_context_t aio_ctx;
+ unsigned int aio_depth;
+ struct io_event *aio_events;
+
+ unsigned int cur_depth;
+ struct list_head io_u_freelist;
+ struct list_head io_u_busylist;
+
unsigned int rate;
unsigned int ratemin;
unsigned int ratecycle;
unsigned long rate_usec_cycle;
long rate_pending_usleep;
- unsigned long rate_blocks;
+ unsigned long rate_bytes;
struct timeval lastrate;
- unsigned long max_latency; /* msec */
- unsigned long min_latency; /* msec */
unsigned long runtime; /* sec */
- unsigned long blocks;
+ unsigned long long io_size;
+
unsigned long io_blocks;
- unsigned long last_block;
+ unsigned long io_bytes;
+ unsigned long this_io_bytes;
+ unsigned long last_bytes;
sem_t mutex;
+
struct drand48_data random_state;
+ unsigned long *file_map;
+ unsigned int num_maps;
/*
- * bandwidth stat
+ * bandwidth and latency stats
*/
- unsigned long stat_time;
- unsigned long stat_time_last;
- unsigned long stat_blocks_last;
+ struct io_stat clat_stat; /* completion latency */
+ struct io_stat slat_stat; /* submission latency */
+
+ struct io_stat bw_stat; /* bandwidth stats */
+ unsigned long stat_io_bytes;
+ struct timeval stat_sample_time;
+
+ struct io_log *lat_log;
+ struct io_log *bw_log;
struct timeval start;
+ struct rusage ru_start;
+ struct rusage ru_end;
+
+ struct list_head io_hist_list;
};
static struct thread_data *threads;
+static struct thread_data def_thread;
static sem_t startup_sem;
-void sig_handler(int sig)
+static void sig_handler(int sig)
{
int i;
struct thread_data *td = &threads[i];
td->terminate = 1;
+ td->start_delay = 0;
}
}
-int init_random_state(struct thread_data *td)
+static int init_random_state(struct thread_data *td)
{
- unsigned long seed = 123;
-
- if (td->sequential)
- return 0;
-
- if (!repeatable) {
- int fd = open("/dev/random", O_RDONLY);
-
- if (fd == -1) {
- td->error = errno;
- return 1;
- }
+ unsigned long seed;
+ int fd, num_maps, blocks;
- if (read(fd, &seed, sizeof(seed)) < (int) sizeof(seed)) {
- td->error = EIO;
- close(fd);
- return 1;
- }
+ fd = open("/dev/random", O_RDONLY);
+ if (fd == -1) {
+ td->error = errno;
+ return 1;
+ }
+ if (read(fd, &seed, sizeof(seed)) < (int) sizeof(seed)) {
+ td->error = EIO;
close(fd);
+ return 1;
}
- srand48_r(seed, &td->random_state);
- return 0;
-}
-
-void shutdown_stat_file(struct thread_data *td)
-{
- if (td->stat_fd != -1) {
- fsync(td->stat_fd);
- close(td->stat_fd);
- }
-}
+ close(fd);
-int init_stat_file(struct thread_data *td)
-{
- char n[256];
+ srand48_r(seed, &td->bsrange_state);
+ srand48_r(seed, &td->verify_state);
- if (!write_stat)
+ if (td->sequential)
return 0;
- sprintf(n, "%s.stat", td->file_name);
- td->stat_fd = open(n, O_WRONLY | O_CREAT | O_TRUNC, 0644);
- if (td->stat_fd == -1) {
- td->error = errno;
- return 1;
- }
+ if (repeatable)
+ seed = DEF_RANDSEED;
+ blocks = (td->io_size + td->min_bs - 1) / td->min_bs;
+ num_maps = blocks / BLOCKS_PER_MAP;
+ td->file_map = malloc(num_maps * sizeof(long));
+ td->num_maps = num_maps;
+ memset(td->file_map, 0, num_maps * sizeof(long));
+
+ srand48_r(seed, &td->random_state);
return 0;
}
-unsigned long utime_since(struct timeval *s, struct timeval *e)
+static unsigned long utime_since(struct timeval *s, struct timeval *e)
{
double sec, usec;
return sec + usec;
}
-unsigned long mtime_since(struct timeval *s, struct timeval *e)
+static unsigned long utime_since_now(struct timeval *s)
+{
+ struct timeval t;
+
+ gettimeofday(&t, NULL);
+ return utime_since(s, &t);
+}
+
+static unsigned long mtime_since(struct timeval *s, struct timeval *e)
{
double sec, usec;
return sec + usec;
}
-unsigned long time_since(struct timeval *s, struct timeval *e)
+static unsigned long mtime_since_now(struct timeval *s)
{
- double sec, usec, ret;
+ struct timeval t;
- sec = e->tv_sec - s->tv_sec;
- usec = e->tv_usec - s->tv_usec;
- if (sec > 0 && usec < 0) {
- sec--;
- usec += 1000000;
+ gettimeofday(&t, NULL);
+ return mtime_since(s, &t);
+}
+
+static inline unsigned long msec_now(struct timeval *s)
+{
+ return s->tv_sec * 1000 + s->tv_usec / 1000;
+}
+
+static int random_map_free(struct thread_data *td, unsigned long long block)
+{
+ unsigned int idx = RAND_MAP_IDX(td, block);
+ unsigned int bit = RAND_MAP_BIT(td, block);
+
+ return (td->file_map[idx] & (1UL << bit)) == 0;
+}
+
+static int get_next_free_block(struct thread_data *td, unsigned long long *b)
+{
+ int i;
+
+ *b = 0;
+ i = 0;
+ while ((*b) * td->min_bs < td->io_size) {
+ if (td->file_map[i] != -1UL) {
+ *b += ffz(td->file_map[i]);
+ return 0;
+ }
+
+ *b += BLOCKS_PER_MAP;
+ i++;
}
- ret = sec + usec / (double) 1000000;
- if (ret < 0)
- ret = 0;
+ return 1;
+}
+
+static void mark_random_map(struct thread_data *td, struct io_u *io_u)
+{
+ unsigned long block = io_u->offset / td->min_bs;
+ unsigned int blocks = 0;
+
+ while (blocks < (io_u->buflen / td->min_bs)) {
+ int idx, bit;
+
+ if (!random_map_free(td, block))
+ break;
+
+ idx = RAND_MAP_IDX(td, block);
+ bit = RAND_MAP_BIT(td, block);
- return (unsigned long) ret;
+ assert(idx < td->num_maps);
+
+ td->file_map[idx] |= (1UL << bit);
+ block++;
+ blocks++;
+ }
+
+ if ((blocks * td->min_bs) < io_u->buflen)
+ io_u->buflen = blocks * td->min_bs;
}
-unsigned long get_next_offset(struct thread_data *td)
+static int get_next_offset(struct thread_data *td, unsigned long long *offset)
{
- unsigned long b;
+ unsigned long long b, rb;
long r;
if (!td->sequential) {
- lrand48_r(&td->random_state, &r);
- b = (1+(double) (td->blocks-1) * r / (RAND_MAX+1.0));
- } else {
- b = td->last_block;
- td->last_block++;
+ unsigned long max_blocks = td->io_size / td->min_bs;
+ int loops = 50;
+
+ do {
+ lrand48_r(&td->random_state, &r);
+ b = ((max_blocks - 1) * r / (RAND_MAX+1.0));
+ rb = b + (td->file_offset / td->min_bs);
+ loops--;
+ } while (!random_map_free(td, rb) && loops);
+
+ if (!loops) {
+ if (get_next_free_block(td, &b))
+ return 1;
+ }
+ } else
+ b = td->last_bytes / td->min_bs;
+
+ *offset = (b * td->min_bs) + td->file_offset;
+ if (*offset > td->file_size)
+ return 1;
+
+ return 0;
+}
+
+static unsigned int get_next_buflen(struct thread_data *td)
+{
+ unsigned int buflen;
+ long r;
+
+ if (td->min_bs == td->max_bs)
+ buflen = td->min_bs;
+ else {
+ lrand48_r(&td->bsrange_state, &r);
+ buflen = (1 + (double) (td->max_bs - 1) * r / (RAND_MAX + 1.0));
+ buflen = (buflen + td->min_bs - 1) & ~(td->min_bs - 1);
+ }
+
+ if (buflen > td->io_size - td->this_io_bytes)
+ buflen = td->io_size - td->this_io_bytes;
+
+ return buflen;
+}
+
+static inline void add_stat_sample(struct thread_data *td, struct io_stat *is,
+ unsigned long val)
+{
+ if (val > is->max_val)
+ is->max_val = val;
+ if (val < is->min_val)
+ is->min_val = val;
+
+ is->val += val;
+ is->val_sq += val * val;
+ is->samples++;
+}
+
+static void add_log_sample(struct thread_data *td, struct io_log *log,
+ unsigned long val)
+{
+ if (log->nr_samples == log->max_samples) {
+ int new_size = sizeof(struct io_sample) * log->max_samples * 2;
+
+ log->log = realloc(log->log, new_size);
+ log->max_samples <<= 1;
}
- return b * td->bs;
+ log->log[log->nr_samples].val = val;
+ log->log[log->nr_samples].time = mtime_since_now(&td->start);
+ log->nr_samples++;
+}
+
+static void add_clat_sample(struct thread_data *td, unsigned long msec)
+{
+ add_stat_sample(td, &td->clat_stat, msec);
+
+ if (td->lat_log)
+ add_log_sample(td, td->lat_log, msec);
+}
+
+static void add_slat_sample(struct thread_data *td, unsigned long msec)
+{
+ add_stat_sample(td, &td->slat_stat, msec);
}
-void add_stat_sample(struct thread_data *td, unsigned long msec)
+static void add_bw_sample(struct thread_data *td)
{
- char sample[256];
+ unsigned long spent = mtime_since_now(&td->stat_sample_time);
+ unsigned long rate;
- if (!td->stat_fd)
+ if (spent < td->bw_avg_time)
return;
-#if 0
- sprintf(sample, "%lu, %lu\n", td->io_blocks, msec);
- write(td->stat_fd, sample, strlen(sample));
-#else
- td->stat_time += msec;
- td->stat_time_last += msec;
- td->stat_blocks_last++;
+ rate = (td->this_io_bytes - td->stat_io_bytes) / spent;
+ add_stat_sample(td, &td->bw_stat, rate);
- if (td->stat_time_last >= 500) {
- unsigned long rate = td->stat_blocks_last * td->bs / (td->stat_time_last);
+ if (td->bw_log)
+ add_log_sample(td, td->bw_log, rate);
- td->stat_time_last = 0;
- td->stat_blocks_last = 0;
- sprintf(sample, "%lu, %lu\n", td->stat_time, rate);
- //sprintf(sample, "%lu, %lu\n", td->io_blocks, msec);
- write(td->stat_fd, sample, strlen(sample));
- }
-#endif
+ gettimeofday(&td->stat_sample_time, NULL);
+ td->stat_io_bytes = td->this_io_bytes;
+}
+
+/*
+ * busy looping version for the last few usec
+ */
+static void __usec_sleep(int usec)
+{
+ struct timeval start;
+
+ gettimeofday(&start, NULL);
+ while (utime_since_now(&start) < usec)
+ __asm__ __volatile__("rep;nop": : :"memory");
}
-void usec_sleep(int usec)
+static void usec_sleep(int usec)
{
struct timespec req = { .tv_sec = 0, .tv_nsec = usec * 1000 };
struct timespec rem;
do {
+ if (usec < 5000) {
+ __usec_sleep(usec);
+ break;
+ }
rem.tv_sec = rem.tv_nsec = 0;
nanosleep(&req, &rem);
if (!rem.tv_nsec)
break;
req.tv_nsec = rem.tv_nsec;
+ usec = rem.tv_nsec * 1000;
} while (1);
}
-void rate_throttle(struct thread_data *td, unsigned long time_spent)
+static void rate_throttle(struct thread_data *td, unsigned long time_spent,
+ unsigned int bytes)
{
+ unsigned long usec_cycle;
+
if (!td->rate)
return;
- if (time_spent < td->rate_usec_cycle) {
- unsigned long s = td->rate_usec_cycle - time_spent;
+ usec_cycle = td->rate_usec_cycle * (bytes / td->min_bs);
+
+ if (time_spent < usec_cycle) {
+ unsigned long s = usec_cycle - time_spent;
td->rate_pending_usleep += s;
if (td->rate_pending_usleep >= 100000) {
td->rate_pending_usleep = 0;
}
} else {
- long overtime = time_spent - td->rate_usec_cycle;
+ long overtime = time_spent - usec_cycle;
td->rate_pending_usleep -= overtime;
}
}
-int check_min_rate(struct thread_data *td, struct timeval *now)
+static int check_min_rate(struct thread_data *td, struct timeval *now)
{
- unsigned long spent = mtime_since(&td->start, now);
+ unsigned long spent;
unsigned long rate;
/*
* allow a 2 second settle period in the beginning
*/
- if (spent < 2000)
+ if (mtime_since(&td->start, now) < 2000)
return 0;
/*
* if rate blocks is set, sample is running
*/
- if (td->rate_blocks) {
+ if (td->rate_bytes) {
spent = mtime_since(&td->lastrate, now);
if (spent < td->ratecycle)
return 0;
- rate = ((td->io_blocks - td->rate_blocks) * td->bs) / spent;
+ rate = (td->this_io_bytes - td->rate_bytes) / spent;
if (rate < td->ratemin) {
printf("Client%d: min rate %d not met, got %ldKiB/sec\n", td->thread_number, td->ratemin, rate);
+ if (rate_quit)
+ sig_handler(0);
return 1;
}
}
- td->rate_blocks = td->io_blocks;
+ td->rate_bytes = td->this_io_bytes;
memcpy(&td->lastrate, now, sizeof(*now));
return 0;
}
-#define should_fsync(td) ((td)->ddir == DDIR_WRITE && !(td)->odirect)
-
-void do_thread_io(struct thread_data *td)
+static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
{
- struct timeval s, e;
- char *buffer, *ptr;
- unsigned long blocks, msec, usec;
-
- ptr = malloc(td->bs + MASK);
- buffer = ALIGN(ptr);
-
- gettimeofday(&td->start, NULL);
+ if (!td->timeout)
+ return 0;
+ if (mtime_since(&td->start, t) >= td->timeout * 1000)
+ return 1;
- if (td->ratemin)
- memcpy(&td->lastrate, &td->start, sizeof(td->start));
+ return 0;
+}
- for (blocks = 0; blocks < td->blocks; blocks++) {
- off_t offset = get_next_offset(td);
- int ret;
+static void fill_random_bytes(struct thread_data *td,
+ unsigned char *p, unsigned int len)
+{
+ unsigned int todo;
+ double r;
- if (td->terminate)
- break;
+ while (len) {
+ drand48_r(&td->verify_state, &r);
- if (lseek(td->fd, offset, SEEK_SET) == -1) {
- td->error = errno;
- break;
- }
+ /*
+ * lrand48_r seems to be broken and only fill the bottom
+ * 32-bits, even on 64-bit archs with 64-bit longs
+ */
+ todo = sizeof(r);
+ if (todo > len)
+ todo = len;
- if (td->delay_sleep)
- usec_sleep(td->delay_sleep);
+ memcpy(p, &r, todo);
- gettimeofday(&s, NULL);
+ len -= todo;
+ p += todo;
+ }
+}
- if (td->ddir == DDIR_READ)
- ret = read(td->fd, buffer, td->bs);
- else
- ret = write(td->fd, buffer, td->bs);
+static void hexdump(void *buffer, int len)
+{
+ unsigned char *p = buffer;
+ int i;
- if (ret < (int) td->bs) {
- if (ret == -1)
- td->error = errno;
- break;
- }
+ for (i = 0; i < len; i++)
+ printf("%02x", p[i]);
+ printf("\n");
+}
- td->io_blocks++;
+static int verify_io_u(struct io_u *io_u)
+{
+ struct verify_header *hdr = (struct verify_header *) io_u->buf;
+ unsigned char *p = (unsigned char *) io_u->buf;
+ struct md5_ctx md5_ctx;
+ int ret;
- if (should_fsync(td) && td->fsync_blocks &&
- (td->io_blocks % td->fsync_blocks) == 0)
- fsync(td->fd);
+ if (hdr->fio_magic != FIO_HDR_MAGIC)
+ return 1;
- gettimeofday(&e, NULL);
+ memset(&md5_ctx, 0, sizeof(md5_ctx));
+ p += sizeof(*hdr);
+ md5_update(&md5_ctx, p, hdr->len - sizeof(*hdr));
- usec = utime_since(&s, &e);
+ ret = memcmp(hdr->md5_digest, md5_ctx.hash, sizeof(md5_ctx.hash));
+ if (ret) {
+ hexdump(hdr->md5_digest, sizeof(hdr->md5_digest));
+ hexdump(md5_ctx.hash, sizeof(md5_ctx.hash));
+ }
- rate_throttle(td, usec);
+ return ret;
+}
- if (check_min_rate(td, &e)) {
- td->error = ENODATA;
- break;
- }
+/*
+ * fill body of io_u->buf with random data and add a header with the
+ * (eg) sha1sum of that data.
+ */
+static void populate_io_u(struct thread_data *td, struct io_u *io_u)
+{
+ struct md5_ctx md5_ctx;
+ struct verify_header hdr;
+ unsigned char *p = (unsigned char *) io_u->buf;
+
+ hdr.fio_magic = FIO_HDR_MAGIC;
+ hdr.len = io_u->buflen;
+ p += sizeof(hdr);
+ fill_random_bytes(td, p, io_u->buflen - sizeof(hdr));
+
+ memset(&md5_ctx, 0, sizeof(md5_ctx));
+ md5_update(&md5_ctx, p, io_u->buflen - sizeof(hdr));
+ memcpy(hdr.md5_digest, md5_ctx.hash, sizeof(md5_ctx.hash));
+ memcpy(io_u->buf, &hdr, sizeof(hdr));
+}
- msec = usec / 1000;
- add_stat_sample(td, msec);
+static void put_io_u(struct thread_data *td, struct io_u *io_u)
+{
+ list_del(&io_u->list);
+ list_add(&io_u->list, &td->io_u_freelist);
+ td->cur_depth--;
+}
- if (msec < td->min_latency)
- td->min_latency = msec;
- if (msec > td->max_latency)
- td->max_latency = msec;
- }
+#define queue_full(td) (list_empty(&(td)->io_u_freelist))
- if (should_fsync(td))
- fsync(td->fd);
+static struct io_u *__get_io_u(struct thread_data *td)
+{
+ struct io_u *io_u;
- gettimeofday(&e, NULL);
- td->runtime = mtime_since(&td->start, &e);
+ if (queue_full(td))
+ return NULL;
- free(ptr);
+ io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
+ list_del(&io_u->list);
+ list_add(&io_u->list, &td->io_u_busylist);
+ td->cur_depth++;
+ return io_u;
}
-
-void *thread_main(int shm_id, int offset, char *argv[])
+
+static struct io_u *get_io_u(struct thread_data *td)
{
- struct thread_data *td;
- void *data;
- struct stat st;
- int ret = 1, flags;
+ struct io_u *io_u;
- data = shmat(shm_id, NULL, 0);
- td = data + offset * sizeof(struct thread_data);
- td->pid = getpid();
+ io_u = __get_io_u(td);
+ if (!io_u)
+ return NULL;
- td->fd = -1;
+ if (get_next_offset(td, &io_u->offset)) {
+ put_io_u(td, io_u);
+ return NULL;
+ }
- if (sched_setaffinity(td->pid, sizeof(td->cpumask), &td->cpumask) == -1) {
- td->error = errno;
- goto err;
+ io_u->buflen = get_next_buflen(td);
+ if (!io_u->buflen) {
+ put_io_u(td, io_u);
+ return NULL;
}
- printf("Thread (%s) (pid=%u) (f=%s) started\n", td->ddir == DDIR_READ ? "read" : "write", td->pid, td->file_name);
- fflush(stdout);
+ if (io_u->buflen + io_u->offset > td->file_size)
+ io_u->buflen = td->file_size - io_u->offset;
- sprintf(argv[0], "fio%d", offset);
+ if (!td->sequential)
+ mark_random_map(td, io_u);
- flags = 0;
- if (td->odirect)
- flags |= O_DIRECT;
+ td->last_bytes += io_u->buflen;
- if (td->ddir == DDIR_READ)
- td->fd = open(td->file_name, flags | O_RDONLY);
- else
- td->fd = open(td->file_name, flags | O_WRONLY | O_CREAT | O_TRUNC, 0644);
+ if (td->verify)
+ populate_io_u(td, io_u);
- if (td->fd == -1) {
- td->error = errno;
- goto err;
+ if (td->use_aio) {
+ if (td_read(td))
+ io_prep_pread(&io_u->iocb, td->fd, io_u->buf, io_u->buflen, io_u->offset);
+ else
+ io_prep_pwrite(&io_u->iocb, td->fd, io_u->buf, io_u->buflen, io_u->offset);
}
- if (init_random_state(td))
- goto out;
- if (init_stat_file(td))
- goto out;
+ gettimeofday(&io_u->start_time, NULL);
+ return io_u;
+}
- if (td->ddir == DDIR_READ) {
- if (fstat(td->fd, &st) == -1) {
- td->error = errno;
- goto out;
- }
+static inline void td_set_runstate(struct thread_data *td, int runstate)
+{
+ td->old_runstate = td->runstate;
+ td->runstate = runstate;
+}
- td->blocks = st.st_size / td->bs;
- if (!td->blocks) {
- td->error = EINVAL;
- goto out;
- }
- } else
- td->blocks = 1024 * 1024 * 1024 / td->bs;
+static int get_next_verify(struct thread_data *td,
+ unsigned long long *offset, unsigned int *len)
+{
+ struct io_piece *ipo;
- if (td->ioprio) {
- if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
+ if (list_empty(&td->io_hist_list))
+ return 1;
+
+ ipo = list_entry(td->io_hist_list.next, struct io_piece, list);
+ list_del(&ipo->list);
+
+ *offset = ipo->offset;
+ *len = ipo->len;
+ free(ipo);
+ return 0;
+}
+
+static void prune_io_piece_log(struct thread_data *td)
+{
+ struct io_piece *ipo;
+
+ while (!list_empty(&td->io_hist_list)) {
+ ipo = list_entry(td->io_hist_list.next, struct io_piece, list);
+
+ list_del(&ipo->list);
+ free(ipo);
+ }
+}
+
+/*
+ * log a succesful write, so we can unwind the log for verify
+ */
+static void log_io_piece(struct thread_data *td, struct io_u *io_u)
+{
+ struct io_piece *ipo = malloc(sizeof(struct io_piece));
+ struct list_head *entry;
+
+ INIT_LIST_HEAD(&ipo->list);
+ ipo->offset = io_u->offset;
+ ipo->len = io_u->buflen;
+
+ /*
+ * for random io where the writes extend the file, it will typically
+ * be laid out with the block scattered as written. it's faster to
+ * read them in in that order again, so don't sort
+ */
+ if (td->sequential || !td->overwrite) {
+ list_add_tail(&ipo->list, &td->io_hist_list);
+ return;
+ }
+
+ /*
+ * for random io, sort the list so verify will run faster
+ */
+ entry = &td->io_hist_list;
+ while ((entry = entry->prev) != &td->io_hist_list) {
+ struct io_piece *__ipo = list_entry(entry, struct io_piece, list);
+
+ if (__ipo->offset < ipo->offset)
+ break;
+ }
+
+ list_add(&ipo->list, entry);
+}
+
+static void do_sync_verify(struct thread_data *td)
+{
+ struct timeval t;
+ struct io_u *io_u = NULL;
+ int ret;
+
+ td_set_runstate(td, TD_VERIFYING);
+
+ io_u = __get_io_u(td);
+
+ if (!td->odirect) {
+ if (fadvise(td->fd, td->file_offset, td->io_size, POSIX_FADV_DONTNEED) < 0) {
+ td->error = errno;
+ goto out;
+ }
+ }
+
+ do {
+ if (td->terminate)
+ break;
+
+ gettimeofday(&t, NULL);
+ if (runtime_exceeded(td, &t))
+ break;
+
+ if (get_next_verify(td, &io_u->offset, &io_u->buflen))
+ break;
+
+ if (td->cur_off != io_u->offset) {
+ if (lseek(td->fd, io_u->offset, SEEK_SET) == -1) {
+ td->error = errno;
+ break;
+ }
+ }
+
+ ret = read(td->fd, io_u->buf, io_u->buflen);
+ if (ret < (int) io_u->buflen) {
+ if (ret == -1) {
+ td->error = errno;
+ break;
+ } else if (!ret)
+ break;
+ else
+ io_u->buflen = ret;
+ }
+
+ if (verify_io_u(io_u))
+ break;
+
+ td->cur_off = io_u->offset + io_u->buflen;
+ } while (1);
+
+out:
+ td_set_runstate(td, TD_RUNNING);
+ put_io_u(td, io_u);
+}
+
+static void do_sync_io(struct thread_data *td)
+{
+ unsigned long msec, usec;
+ struct io_u *io_u = NULL;
+ struct timeval e;
+
+ while (td->this_io_bytes < td->io_size) {
+ int ret;
+
+ if (td->terminate)
+ break;
+
+ io_u = get_io_u(td);
+ if (!io_u)
+ break;
+
+ if (td->cur_off != io_u->offset) {
+ if (lseek(td->fd, io_u->offset, SEEK_SET) == -1) {
+ td->error = errno;
+ break;
+ }
+ }
+
+ if (td_read(td))
+ ret = read(td->fd, io_u->buf, io_u->buflen);
+ else
+ ret = write(td->fd, io_u->buf, io_u->buflen);
+
+ if (ret < (int) io_u->buflen) {
+ if (ret == -1)
+ td->error = errno;
+ break;
+ }
+
+ if (td_write(td))
+ log_io_piece(td, io_u);
+
+ td->io_blocks++;
+ td->io_bytes += io_u->buflen;
+ td->this_io_bytes += io_u->buflen;
+ td->cur_off = io_u->offset + io_u->buflen;
+
+ gettimeofday(&e, NULL);
+
+ usec = utime_since(&io_u->start_time, &e);
+
+ rate_throttle(td, usec, io_u->buflen);
+
+ if (check_min_rate(td, &e)) {
+ td->error = ENODATA;
+ break;
+ }
+
+ msec = usec / 1000;
+ add_clat_sample(td, msec);
+ add_bw_sample(td);
+
+ if (runtime_exceeded(td, &e))
+ break;
+
+ put_io_u(td, io_u);
+ io_u = NULL;
+
+ if (td->thinktime)
+ usec_sleep(td->thinktime);
+
+ if (should_fsync(td) && td->fsync_blocks &&
+ (td->io_blocks % td->fsync_blocks) == 0)
+ fsync(td->fd);
+ }
+
+ if (io_u)
+ put_io_u(td, io_u);
+
+ if (should_fsync(td))
+ fsync(td->fd);
+}
+
+static int io_u_getevents(struct thread_data *td, int min, int max,
+ struct timespec *t)
+{
+ int r;
+
+ do {
+ r = io_getevents(td->aio_ctx, min, max, td->aio_events, t);
+ if (r != -EAGAIN && r != -EINTR)
+ break;
+ } while (1);
+
+ return r;
+}
+
+static int io_u_queue(struct thread_data *td, struct io_u *io_u)
+{
+ struct iocb *iocb = &io_u->iocb;
+ int ret;
+
+ do {
+ ret = io_submit(td->aio_ctx, 1, &iocb);
+ if (ret == 1)
+ return 0;
+ else if (ret == -EAGAIN)
+ usleep(100);
+ else if (ret == -EINTR)
+ continue;
+ else
+ break;
+ } while (1);
+
+ return ret;
+}
+
+#define iocb_time(iocb) ((unsigned long) (iocb)->data)
+#define ev_to_iou(ev) (struct io_u *) ((unsigned long) (ev)->obj)
+
+static int ios_completed(struct thread_data *td, int nr)
+{
+ unsigned long msec;
+ struct io_u *io_u;
+ struct timeval e;
+ int i, bytes_done;
+
+ gettimeofday(&e, NULL);
+
+ for (i = 0, bytes_done = 0; i < nr; i++) {
+ io_u = ev_to_iou(td->aio_events + i);
+
+ td->io_blocks++;
+ td->io_bytes += io_u->buflen;
+ td->this_io_bytes += io_u->buflen;
+
+ msec = mtime_since(&io_u->issue_time, &e);
+
+ add_clat_sample(td, msec);
+ add_bw_sample(td);
+
+ if (td_write(td))
+ log_io_piece(td, io_u);
+
+ bytes_done += io_u->buflen;
+ put_io_u(td, io_u);
+ }
+
+ return bytes_done;
+}
+
+static void cleanup_pending_aio(struct thread_data *td)
+{
+ struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
+ struct list_head *entry, *n;
+ struct io_u *io_u;
+ int r;
+
+ /*
+ * get immediately available events, if any
+ */
+ r = io_u_getevents(td, 0, td->cur_depth, &ts);
+ if (r > 0)
+ ios_completed(td, r);
+
+ /*
+ * now cancel remaining active events
+ */
+ list_for_each_safe(entry, n, &td->io_u_busylist) {
+ io_u = list_entry(entry, struct io_u, list);
+
+ r = io_cancel(td->aio_ctx, &io_u->iocb, td->aio_events);
+ if (!r)
+ put_io_u(td, io_u);
+ }
+
+ if (td->cur_depth) {
+ r = io_u_getevents(td, td->cur_depth, td->cur_depth, NULL);
+ if (r > 0)
+ ios_completed(td, r);
+ }
+}
+
+static int async_do_verify(struct thread_data *td, struct io_u **io_u)
+{
+ struct io_u *v_io_u = *io_u;
+ int ret = 0;
+
+ if (v_io_u) {
+ ret = verify_io_u(v_io_u);
+ put_io_u(td, v_io_u);
+ *io_u = NULL;
+ }
+
+ return ret;
+}
+
+static void do_async_verify(struct thread_data *td)
+{
+ struct timeval t;
+ struct io_u *io_u, *v_io_u = NULL;
+ int ret;
+
+ td_set_runstate(td, TD_VERIFYING);
+
+ do {
+ if (td->terminate)
+ break;
+
+ gettimeofday(&t, NULL);
+ if (runtime_exceeded(td, &t))
+ break;
+
+ io_u = __get_io_u(td);
+ if (!io_u)
+ break;
+
+ if (get_next_verify(td, &io_u->offset, &io_u->buflen)) {
+ put_io_u(td, io_u);
+ break;
+ }
+
+ io_prep_pread(&io_u->iocb, td->fd, io_u->buf, io_u->buflen, io_u->offset);
+ ret = io_u_queue(td, io_u);
+ if (ret) {
+ put_io_u(td, io_u);
+ td->error = ret;
+ break;
+ }
+
+ /*
+ * we have one pending to verify, do that while the next
+ * we are doing io on the next one
+ */
+ if (async_do_verify(td, &v_io_u))
+ break;
+
+ ret = io_u_getevents(td, 1, 1, NULL);
+ if (ret != 1) {
+ if (ret < 0)
+ td->error = ret;
+ break;
+ }
+
+ v_io_u = ev_to_iou(td->aio_events);
+
+ td->cur_off = v_io_u->offset + v_io_u->buflen;
+
+ /*
+ * if we can't submit more io, we need to verify now
+ */
+ if (queue_full(td) && async_do_verify(td, &v_io_u))
+ break;
+
+ } while (1);
+
+ async_do_verify(td, &v_io_u);
+
+ if (td->cur_depth)
+ cleanup_pending_aio(td);
+
+ td_set_runstate(td, TD_RUNNING);
+}
+
+static void do_async_io(struct thread_data *td)
+{
+ struct timeval s, e;
+ unsigned long usec;
+
+ while (td->this_io_bytes < td->io_size) {
+ struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
+ struct timespec *timeout;
+ int ret, min_evts = 0;
+ struct io_u *io_u;
+ unsigned int bytes_done;
+
+ if (td->terminate)
+ break;
+
+ io_u = get_io_u(td);
+ if (!io_u)
+ break;
+
+ memcpy(&s, &io_u->start_time, sizeof(s));
+
+ ret = io_u_queue(td, io_u);
+ if (ret) {
+ put_io_u(td, io_u);
+ td->error = ret;
+ break;
+ }
+
+ gettimeofday(&io_u->issue_time, NULL);
+ add_slat_sample(td, mtime_since(&io_u->start_time, &io_u->issue_time));
+ if (td->cur_depth < td->aio_depth) {
+ timeout = &ts;
+ min_evts = 0;
+ } else {
+ timeout = NULL;
+ min_evts = 1;
+ }
+
+ ret = io_u_getevents(td, min_evts, td->cur_depth, timeout);
+ if (ret < 0) {
+ td->error = ret;
+ break;
+ } else if (!ret)
+ continue;
+
+ bytes_done = ios_completed(td, ret);
+
+ /*
+ * the rate is batched for now, it should work for batches
+ * of completions except the very first one which may look
+ * a little bursty
+ */
+ gettimeofday(&e, NULL);
+ usec = utime_since(&s, &e);
+
+ rate_throttle(td, usec, bytes_done);
+
+ if (check_min_rate(td, &e)) {
+ td->error = ENODATA;
+ break;
+ }
+
+ if (runtime_exceeded(td, &e))
+ break;
+
+ if (td->thinktime)
+ usec_sleep(td->thinktime);
+
+ if (should_fsync(td) && td->fsync_blocks &&
+ (td->io_blocks % td->fsync_blocks) == 0)
+ fsync(td->fd);
+ }
+
+ if (td->cur_depth)
+ cleanup_pending_aio(td);
+
+ if (should_fsync(td))
+ fsync(td->fd);
+}
+
+static void cleanup_aio(struct thread_data *td)
+{
+ io_destroy(td->aio_ctx);
+
+ if (td->aio_events)
+ free(td->aio_events);
+}
+
+static int init_aio(struct thread_data *td)
+{
+ if (io_queue_init(td->aio_depth, &td->aio_ctx)) {
+ td->error = errno;
+ return 1;
+ }
+
+ td->aio_events = malloc(td->aio_depth * sizeof(struct io_event));
+ return 0;
+}
+
+static void cleanup_io_u(struct thread_data *td)
+{
+ struct list_head *entry, *n;
+ struct io_u *io_u;
+
+ list_for_each_safe(entry, n, &td->io_u_freelist) {
+ io_u = list_entry(entry, struct io_u, list);
+
+ list_del(&io_u->list);
+ free(io_u);
+ }
+
+ if (td->mem_type == MEM_MALLOC)
+ free(td->orig_buffer);
+ else if (td->mem_type == MEM_SHM) {
+ struct shmid_ds sbuf;
+
+ shmdt(td->orig_buffer);
+ shmctl(td->shm_id, IPC_RMID, &sbuf);
+ }
+}
+
+static int init_io_u(struct thread_data *td)
+{
+ struct io_u *io_u;
+ int i, max_units, mem_size;
+ char *p;
+
+ if (!td->use_aio)
+ max_units = 1;
+ else
+ max_units = td->aio_depth;
+
+ mem_size = td->max_bs * max_units + MASK;
+
+ if (td->mem_type == MEM_MALLOC)
+ td->orig_buffer = malloc(mem_size);
+ else if (td->mem_type == MEM_SHM) {
+ td->shm_id = shmget(IPC_PRIVATE, mem_size, IPC_CREAT | 0600);
+ if (td->shm_id < 0) {
+ td->error = errno;
+ perror("shmget");
+ return 1;
+ }
+
+ td->orig_buffer = shmat(td->shm_id, NULL, 0);
+ if (td->orig_buffer == (void *) -1) {
+ td->error = errno;
+ perror("shmat");
+ return 1;
+ }
+ }
+
+ INIT_LIST_HEAD(&td->io_u_freelist);
+ INIT_LIST_HEAD(&td->io_u_busylist);
+ INIT_LIST_HEAD(&td->io_hist_list);
+
+ p = ALIGN(td->orig_buffer);
+ for (i = 0; i < max_units; i++) {
+ io_u = malloc(sizeof(*io_u));
+ memset(io_u, 0, sizeof(*io_u));
+ INIT_LIST_HEAD(&io_u->list);
+
+ io_u->buf = p + td->max_bs * i;
+ list_add(&io_u->list, &td->io_u_freelist);
+ }
+
+ return 0;
+}
+
+static void setup_log(struct io_log **log)
+{
+ struct io_log *l = malloc(sizeof(*l));
+
+ l->nr_samples = 0;
+ l->max_samples = 1024;
+ l->log = malloc(l->max_samples * sizeof(struct io_sample));
+ *log = l;
+}
+
+static void finish_log(struct thread_data *td, struct io_log *log, char *name)
+{
+ char file_name[128];
+ FILE *f;
+ unsigned int i;
+
+ sprintf(file_name, "client%d_%s.log", td->thread_number, name);
+ f = fopen(file_name, "w");
+ if (!f) {
+ perror("fopen log");
+ return;
+ }
+
+ for (i = 0; i < log->nr_samples; i++)
+ fprintf(f, "%lu, %lu\n", log->log[i].time, log->log[i].val);
+
+ fclose(f);
+ free(log->log);
+ free(log);
+}
+
+static int create_file(struct thread_data *td)
+{
+ unsigned long long left;
+ unsigned int bs;
+ char *b;
+ int r;
+
+ /*
+ * unless specifically asked for overwrite, let normal io extend it
+ */
+ if (td_write(td) && !td->overwrite)
+ return 0;
+
+ if (!td->file_size) {
+ fprintf(stderr, "Need size for create\n");
+ td->error = EINVAL;
+ return 1;
+ }
+
+ printf("Client%d: Laying out IO file\n", td->thread_number);
+
+ td->fd = open(td->file_name, O_WRONLY | O_CREAT | O_TRUNC, 0644);
+ if (td->fd < 0) {
+ td->error = errno;
+ return 1;
+ }
+
+ if (ftruncate(td->fd, td->file_size) == -1) {
+ td->error = errno;
+ return 1;
+ }
+
+ td->io_size = td->file_size;
+ b = malloc(td->max_bs);
+ memset(b, 0, td->max_bs);
+
+ left = td->file_size;
+ while (left) {
+ bs = td->max_bs;
+ if (bs > left)
+ bs = left;
+
+ r = write(td->fd, b, bs);
+
+ if (r == (int) bs) {
+ left -= bs;
+ continue;
+ } else {
+ if (r < 0)
+ td->error = errno;
+ else
+ td->error = EIO;
+
+ break;
+ }
+ }
+
+ if (td->create_fsync)
+ fsync(td->fd);
+
+ close(td->fd);
+ td->fd = -1;
+ free(b);
+ return 0;
+}
+
+static int file_exists(struct thread_data *td)
+{
+ struct stat st;
+
+ if (stat(td->file_name, &st) != -1)
+ return 1;
+
+ return errno != ENOENT;
+}
+
+static int get_file_size(struct thread_data *td)
+{
+ size_t bytes = 0;
+ struct stat st;
+
+ if (fstat(td->fd, &st) == -1) {
+ td->error = errno;
+ return 1;
+ }
+
+ /*
+ * if block device, get size via BLKGETSIZE64 ioctl. try that as well
+ * if this is a link, fall back to st.st_size if it fails
+ */
+ if (S_ISBLK(st.st_mode) || S_ISLNK(st.st_mode)) {
+ if (ioctl(td->fd, BLKGETSIZE64, &bytes)) {
+ if (S_ISBLK(st.st_mode)) {
+ td->error = errno;
+ return 1;
+ } else
+ bytes = st.st_size;
+ }
+ } else
+ bytes = st.st_size;
+
+ if (td_read(td)) {
+ if (td->file_size > bytes)
+ bytes = td->file_size;
+ } else {
+ if (!td->file_size)
+ td->file_size = 1024 * 1024 * 1024;
+
+ bytes = td->file_size;
+ }
+
+ if (td->file_offset > bytes) {
+ fprintf(stderr, "Client%d: offset larger than length\n", td->thread_number);
+ return 1;
+ }
+
+ td->io_size = bytes - td->file_offset;
+ if (td->io_size == 0) {
+ fprintf(stderr, "Client%d: no io blocks\n", td->thread_number);
+ td->error = EINVAL;
+ return 1;
+ }
+
+ return 0;
+}
+
+static int setup_file(struct thread_data *td)
+{
+ int flags = 0;
+
+ if (!file_exists(td)) {
+ if (!td->create_file) {
+ td->error = ENOENT;
+ return 1;
+ }
+ if (create_file(td))
+ return 1;
+ }
+
+ if (td->odirect)
+ flags |= O_DIRECT;
+
+ if (td_read(td))
+ td->fd = open(td->file_name, flags | O_RDONLY);
+ else {
+ if (!td->overwrite)
+ flags |= O_TRUNC;
+ if (td->sync_io)
+ flags |= O_SYNC;
+ if (td->verify)
+ flags |= O_RDWR;
+ else
+ flags |= O_WRONLY;
+
+ td->fd = open(td->file_name, flags | O_CREAT, 0600);
+ }
+
+ if (td->fd == -1) {
+ td->error = errno;
+ return 1;
+ }
+
+ if (get_file_size(td))
+ return 1;
+
+ if (td_write(td) && ftruncate(td->fd, td->file_size) == -1) {
+ td->error = errno;
+ return 1;
+ }
+
+ if (td->invalidate_cache) {
+ if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_DONTNEED) < 0) {
td->error = errno;
- goto out;
+ return 1;
+ }
+ }
+
+ return 0;
+}
+
+static void clear_io_state(struct thread_data *td)
+{
+ if (!td->use_aio)
+ lseek(td->fd, SEEK_SET, 0);
+
+ td->cur_off = 0;
+ td->last_bytes = 0;
+ td->stat_io_bytes = 0;
+ td->this_io_bytes = 0;
+
+ if (td->file_map)
+ memset(td->file_map, 0, td->num_maps * sizeof(long));
+}
+
+static void *thread_main(int shm_id, int offset, char *argv[])
+{
+ struct thread_data *td;
+ int ret = 1;
+ void *data;
+
+ setsid();
+
+ data = shmat(shm_id, NULL, 0);
+ if (data == (void *) -1) {
+ perror("shmat");
+ return NULL;
+ }
+
+ td = data + offset * sizeof(struct thread_data);
+ td->pid = getpid();
+
+ if (init_io_u(td))
+ goto err;
+
+ if (sched_setaffinity(td->pid, sizeof(td->cpumask), &td->cpumask) == -1) {
+ td->error = errno;
+ goto err;
+ }
+
+ sprintf(argv[0], "fio%d", offset);
+
+ if (td->use_aio && init_aio(td))
+ goto err;
+
+ if (td->ioprio) {
+ if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
+ td->error = errno;
+ goto err;
}
}
sem_post(&startup_sem);
sem_wait(&td->mutex);
- do_thread_io(td);
+
+ if (!td->create_serialize && setup_file(td))
+ goto err;
+
+ if (init_random_state(td))
+ goto err;
+
+ gettimeofday(&td->start, NULL);
+
+ getrusage(RUSAGE_SELF, &td->ru_start);
+
+ while (td->loops--) {
+ gettimeofday(&td->stat_sample_time, NULL);
+
+ if (td->ratemin)
+ memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
+
+ clear_io_state(td);
+ prune_io_piece_log(td);
+
+ if (!td->use_aio)
+ do_sync_io(td);
+ else
+ do_async_io(td);
+
+ if (td->error)
+ break;
+
+ if (!td->verify)
+ continue;
+
+ clear_io_state(td);
+
+ if (!td->use_aio)
+ do_sync_verify(td);
+ else
+ do_async_verify(td);
+
+ if (td->error)
+ break;
+ }
+
+ td->runtime = mtime_since_now(&td->start);
+ getrusage(RUSAGE_SELF, &td->ru_end);
ret = 0;
-out:
- shutdown_stat_file(td);
+ if (td->bw_log)
+ finish_log(td, td->bw_log, "bw");
+ if (td->lat_log)
+ finish_log(td, td->lat_log, "lat");
+
+ if (exitall_on_terminate)
+ sig_handler(0);
+
err:
- if (td->fd != -1)
+ if (td->fd != -1) {
close(td->fd);
- if (ret)
+ td->fd = -1;
+ }
+ if (td->use_aio)
+ cleanup_aio(td);
+ cleanup_io_u(td);
+ if (ret) {
sem_post(&startup_sem);
+ sem_wait(&td->mutex);
+ }
+ td_set_runstate(td, TD_EXITED);
shmdt(data);
return NULL;
}
-void free_shm(void)
+static void free_shm(void)
+{
+ struct shmid_ds sbuf;
+
+ if (threads) {
+ shmdt(threads);
+ threads = NULL;
+ shmctl(shm_id, IPC_RMID, &sbuf);
+ }
+}
+
+static int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
+ double *mean, double *dev)
{
- shmdt(threads);
+ double n;
+
+ if (is->samples == 0)
+ return 0;
+
+ *min = is->min_val;
+ *max = is->max_val;
+
+ n = (double) is->samples;
+ *mean = (double) is->val / n;
+ *dev = sqrt(((double) is->val_sq - (*mean * *mean) / n) / (n - 1));
+ return 1;
}
-void show_thread_status(struct thread_data *td)
+static void show_thread_status(struct thread_data *td)
{
int prio, prio_class;
- unsigned long bw = 0;
+ unsigned long min, max, bw = 0, ctx;
+ double mean, dev, usr_cpu, sys_cpu;
+
+ if (!td->io_bytes && !td->error)
+ return;
if (td->runtime)
- bw = (td->io_blocks * td->bs) / td->runtime;
+ bw = td->io_bytes / td->runtime;
prio = td->ioprio & 0xff;
prio_class = td->ioprio >> IOPRIO_CLASS_SHIFT;
- printf("thread%d (%s): err=%2d, prio=%1d/%1d maxl=%5lumsec, io=%6luMiB, bw=%6luKiB/sec\n", td->thread_number, td->ddir == DDIR_READ ? " read": "write", td->error, prio_class, prio, td->max_latency, td->io_blocks * td->bs >> 20, bw);
-}
+ printf("Client%d: err=%2d, io=%6luMiB, bw=%6luKiB/s, runt=%6lumsec\n", td->thread_number, td->error, td->io_bytes >> 20, bw, td->runtime);
-void usage(char *progname)
-{
- printf("%s: <-s 0/1> <-b kb> <-t sec> <-w 0/1> <-c r,w,r...> file0... fileN\n", progname);
+ if (calc_lat(&td->slat_stat, &min, &max, &mean, &dev))
+ printf(" slat (msec): min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev);
+ if (calc_lat(&td->clat_stat, &min, &max, &mean, &dev))
+ printf(" clat (msec): min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev);
+ if (calc_lat(&td->bw_stat, &min, &max, &mean, &dev))
+ printf(" bw (KiB/s) : min=%5lu, max=%5lu, avg=%5.02f, dev=%5.02f\n", min, max, mean, dev);
+
+ if (td->runtime) {
+ unsigned long t;
+
+ t = mtime_since(&td->ru_start.ru_utime, &td->ru_end.ru_utime);
+ usr_cpu = (double) t * 100 / (double) td->runtime;
+
+ t = mtime_since(&td->ru_start.ru_stime, &td->ru_end.ru_stime);
+ sys_cpu = (double) t * 100 / (double) td->runtime;
+ } else {
+ usr_cpu = 0;
+ sys_cpu = 0;
+ }
+
+ ctx = td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
+
+ printf(" cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%lu\n", usr_cpu, sys_cpu, ctx);
}
-int setup_rate(struct thread_data *td)
+static int setup_rate(struct thread_data *td)
{
int nr_reads_per_sec;
return -1;
}
- nr_reads_per_sec = td->rate * 1024 / td->bs;
+ nr_reads_per_sec = (td->rate * 1024) / td->min_bs;
td->rate_usec_cycle = 1000000 / nr_reads_per_sec;
td->rate_pending_usleep = 0;
return 0;
}
-struct thread_data *get_new_job(void)
+static struct thread_data *get_new_job(int global, struct thread_data *parent)
{
struct thread_data *td;
- if (thread_number >= MAX_JOBS)
+ if (global)
+ return &def_thread;
+ if (thread_number >= max_jobs)
return NULL;
td = &threads[thread_number++];
+ memset(td, 0, sizeof(*td));
+ sprintf(td->directory, ".");
+
+ td->fd = -1;
td->thread_number = thread_number;
- td->ddir = DDIR_READ;
- td->bs = global_bs;
- td->odirect = 1;
- td->delay_sleep = 0;
- td->rate = 0;
- td->ratecycle = DEF_RATE_CYCLE;
- td->sequential = sequential;
- td->ioprio = 0;
- memcpy(&td->cpumask, &def_cpumask, sizeof(td->cpumask));
+
+ td->ddir = parent->ddir;
+ td->ioprio = parent->ioprio;
+ td->sequential = parent->sequential;
+ td->bs = parent->bs;
+ td->min_bs = parent->min_bs;
+ td->max_bs = parent->max_bs;
+ td->odirect = parent->odirect;
+ td->thinktime = parent->thinktime;
+ td->fsync_blocks = parent->fsync_blocks;
+ td->start_delay = parent->start_delay;
+ td->timeout = parent->timeout;
+ td->use_aio = parent->use_aio;
+ td->create_file = parent->create_file;
+ td->overwrite = parent->overwrite;
+ td->invalidate_cache = parent->invalidate_cache;
+ td->file_size = parent->file_size;
+ td->file_offset = parent->file_offset;
+ td->rate = parent->rate;
+ td->ratemin = parent->ratemin;
+ td->ratecycle = parent->ratecycle;
+ td->aio_depth = parent->aio_depth;
+ td->sync_io = parent->sync_io;
+ td->mem_type = parent->mem_type;
+ td->bw_avg_time = parent->bw_avg_time;
+ td->create_serialize = parent->create_serialize;
+ td->create_fsync = parent->create_fsync;
+ td->loops = parent->loops;
+ td->verify = parent->verify;
+ td->stonewall = parent->stonewall;
+ td->numjobs = parent->numjobs;
+ memcpy(&td->cpumask, &parent->cpumask, sizeof(td->cpumask));
return td;
}
thread_number--;
}
-int add_job(struct thread_data *td, const char *filename, int prioclass,
- int prio)
+static int add_job(struct thread_data *td, const char *jobname, int prioclass,
+ int prio)
{
- strcpy(td->file_name, filename);
- td->stat_fd = -1;
+ int numjobs;
+
+ if (td == &def_thread)
+ return 0;
+
+ sprintf(td->file_name, "%s/%s.%d", td->directory, jobname, td->thread_number);
sem_init(&td->mutex, 1, 0);
- td->min_latency = 10000000;
td->ioprio = (prioclass << IOPRIO_CLASS_SHIFT) | prio;
+ td->clat_stat.min_val = ULONG_MAX;
+ td->slat_stat.min_val = ULONG_MAX;
+ td->bw_stat.min_val = ULONG_MAX;
+
+ run_str[td->thread_number - 1] = 'P';
+
+ if (td->use_aio && !td->aio_depth)
+ td->aio_depth = 1;
+
+ if (td->min_bs == -1U)
+ td->min_bs = td->bs;
+ if (td->max_bs == -1U)
+ td->max_bs = td->bs;
+ if (td_read(td))
+ td->verify = 0;
+
if (setup_rate(td))
- return -1;
+ goto err;
+
+ if (write_lat_log)
+ setup_log(&td->lat_log);
+ if (write_bw_log)
+ setup_log(&td->bw_log);
+
+ printf("Client%d: rw=%d, prio=%d/%d, seq=%d, odir=%d, bs=%d-%d, rate=%d, aio=%d, aio_depth=%d\n", td->thread_number, td->ddir, prioclass, prio, td->sequential, td->odirect, td->min_bs, td->max_bs, td->rate, td->use_aio, td->aio_depth);
- printf("Client%d: file=%s, rw=%d, prio=%d, seq=%d, odir=%d, bs=%d, rate=%d\n", td->thread_number, filename, td->ddir, td->ioprio, td->sequential, td->odirect, td->bs, td->rate);
+ /*
+ * recurse add identical jobs, clear numjobs and stonewall options
+ * as they don't apply to sub-jobs
+ */
+ numjobs = td->numjobs;
+ while (--numjobs) {
+ struct thread_data *td_new = get_new_job(0, td);
+
+ if (!td_new)
+ break;
+
+ td_new->numjobs = 1;
+ td_new->stonewall = 0;
+
+ if (add_job(td_new, jobname, prioclass, prio))
+ break;
+ }
return 0;
+err:
+ put_job(td);
+ return -1;
}
static void fill_cpu_mask(cpu_set_t cpumask, int cpu)
}
}
-void fill_option(const char *input, char *output)
+unsigned long get_mult(char c)
{
- int i;
-
- i = 0;
- while (input[i] != ',' && input[i] != '}' && input[i] != '\0') {
- output[i] = input[i];
- i++;
+ switch (c) {
+ case 'k':
+ case 'K':
+ return 1024;
+ case 'm':
+ case 'M':
+ return 1024 * 1024;
+ case 'g':
+ case 'G':
+ return 1024 * 1024 * 1024;
+ default:
+ return 1;
}
-
- output[i] = '\0';
}
/*
- * job key words:
- *
- * file=
- * bs=
- * rw=
- * direct=
+ * convert string after '=' into decimal value, noting any size suffix
*/
-void parse_jobs_cmd(int argc, char *argv[], int index)
+static int str_cnv(char *p, unsigned long long *val)
{
- struct thread_data *td;
- unsigned int prio, prioclass, cpu;
- char *string, *filename, *p, *c;
- int i;
-
- string = malloc(256);
- filename = malloc(256);
-
- for (i = index; i < argc; i++) {
- p = argv[i];
+ char *str;
+ int len;
- c = strpbrk(p, "{");
- if (!c)
- break;
+ str = strstr(p, "=");
+ if (!str)
+ return 1;
- filename[0] = 0;
+ str++;
+ len = strlen(str);
- td = get_new_job();
- if (!td)
- break;
+ *val = strtoul(str, NULL, 10);
+ if (*val == ULONG_MAX && errno == ERANGE)
+ return 1;
- prioclass = 2;
- prio = 4;
+ *val *= get_mult(str[len - 2]);
+ return 0;
+}
- c = strstr(p, "rw=");
- if (c) {
- c += 3;
- if (*c == '0')
- td->ddir = DDIR_READ;
- else
- td->ddir = DDIR_WRITE;
- }
+static int check_strcnv(char *p, char *name, unsigned long long *val)
+{
+ if (!strstr(p, name))
+ return 1;
- c = strstr(p, "prio=");
- if (c) {
- c += 5;
- prio = *c - '0';
- }
+ return str_cnv(p, val);
+}
- c = strstr(p, "prioclass=");
- if (c) {
- c += 10;
- prioclass = *c - '0';
- }
+static int check_str(char *p, char *name, char *option)
+{
+ char *s = strstr(p, name);
- c = strstr(p, "file=");
- if (c) {
- c += 5;
- fill_option(c, filename);
- }
+ if (!s)
+ return 1;
- c = strstr(p, "bs=");
- if (c) {
- c += 3;
- fill_option(c, string);
- td->bs = strtoul(string, NULL, 10);
- td->bs <<= 10;
- }
+ s += strlen(name);
+ if (strstr(s, option))
+ return 0;
- c = strstr(p, "direct=");
- if (c) {
- c += 7;
- if (*c != '0')
- td->odirect = 1;
- else
- td->odirect = 0;
- }
+ return 1;
+}
- c = strstr(p, "delay=");
- if (c) {
- c += 6;
- fill_option(c, string);
- td->delay_sleep = strtoul(string, NULL, 10);
- }
+static int check_strstore(char *p, char *name, char *dest)
+{
+ char *s = strstr(p, name);
- c = strstr(p, "rate=");
- if (c) {
- c += 5;
- fill_option(c, string);
- td->rate = strtoul(string, NULL, 10);
- }
+ if (!s)
+ return 1;
- c = strstr(p, "ratemin=");
- if (c) {
- c += 8;
- fill_option(c, string);
- td->ratemin = strtoul(string, NULL, 10);
- }
+ s = strstr(p, "=");
+ if (!s)
+ return 1;
- c = strstr(p, "ratecycle=");
- if (c) {
- c += 10;
- fill_option(c, string);
- td->ratecycle = strtoul(string, NULL, 10);
- }
+ s++;
+ while (isblank(*s))
+ s++;
- c = strstr(p, "cpumask=");
- if (c) {
- c += 8;
- fill_option(c, string);
- cpu = strtoul(string, NULL, 10);
- fill_cpu_mask(td->cpumask, cpu);
- }
+ strcpy(dest, s);
+ return 0;
+}
- c = strstr(p, "fsync=");
- if (c) {
- c += 6;
- fill_option(c, string);
- td->fsync_blocks = strtoul(string, NULL, 10);
- }
+static int check_range(char *p, char *name, unsigned long *s, unsigned long *e)
+{
+ char str[128];
+ char s1, s2;
- c = strstr(p, "random");
- if (c)
- td->sequential = 0;
- c = strstr(p, "sequential");
- if (c)
- td->sequential = 1;
+ sprintf(str, "%s=%%lu%%c-%%lu%%c", name);
+ if (sscanf(p, str, s, &s1, e, &s2) == 4) {
+ *s *= get_mult(s1);
+ *e *= get_mult(s2);
+ return 0;
+ }
- if (add_job(td, filename, prioclass, prio))
- put_job(td);
+ sprintf(str, "%s = %%lu%%c-%%lu%%c", name);
+ if (sscanf(p, str, s, &s1, e, &s2) == 4) {
+ *s *= get_mult(s1);
+ *e *= get_mult(s2);
+ return 0;
}
- free(string);
- free(filename);
+ sprintf(str, "%s=%%lu-%%lu", name);
+ if (sscanf(p, str, s, e) == 2)
+ return 0;
+
+ sprintf(str, "%s = %%lu-%%lu", name);
+ if (sscanf(p, str, s, e) == 2)
+ return 0;
+
+ return 1;
+
}
-int check_int(char *p, char *name, unsigned int *val)
+static int check_int(char *p, char *name, unsigned int *val)
{
char str[128];
return 1;
}
-int is_empty(char *line)
+static int is_empty_or_comment(char *line)
{
unsigned int i;
- for (i = 0; i < strlen(line); i++)
+ for (i = 0; i < strlen(line); i++) {
+ if (line[i] == ';')
+ return 1;
if (!isspace(line[i]) && !iscntrl(line[i]))
return 0;
+ }
return 1;
}
-int parse_jobs_ini(char *file)
+static int parse_jobs_ini(char *file)
{
- unsigned int prioclass, prio, cpu;
+ unsigned int prioclass, prio, cpu, global;
+ unsigned long long ull;
+ unsigned long ul1, ul2;
struct thread_data *td;
char *string, *name;
fpos_t off;
name = malloc(256);
while ((p = fgets(string, 4096, f)) != NULL) {
+ if (is_empty_or_comment(p))
+ continue;
if (sscanf(p, "[%s]", name) != 1)
continue;
+ global = !strncmp(name, "global", 6);
+
name[strlen(name) - 1] = '\0';
- td = get_new_job();
+ td = get_new_job(global, &def_thread);
if (!td)
break;
fgetpos(f, &off);
while ((p = fgets(string, 4096, f)) != NULL) {
- if (is_empty(p))
- break;
- if (!check_int(p, "bs", &td->bs)) {
- td->bs <<= 10;
- fgetpos(f, &off);
+ if (is_empty_or_comment(p))
continue;
- }
+ if (strstr(p, "["))
+ break;
if (!check_int(p, "rw", &td->ddir)) {
fgetpos(f, &off);
continue;
fgetpos(f, &off);
continue;
}
- if (!check_int(p, "delay", &td->delay_sleep)) {
+ if (!check_int(p, "thinktime", &td->thinktime)) {
fgetpos(f, &off);
continue;
}
fgetpos(f, &off);
continue;
}
- if (!strcmp(p, "sequential")) {
+ if (!check_int(p, "startdelay", &td->start_delay)) {
+ fgetpos(f, &off);
+ continue;
+ }
+ if (!check_int(p, "timeout", &td->timeout)) {
+ fgetpos(f, &off);
+ continue;
+ }
+ if (!check_int(p, "invalidate",&td->invalidate_cache)) {
+ fgetpos(f, &off);
+ continue;
+ }
+ if (!check_int(p, "aio_depth", &td->aio_depth)) {
+ fgetpos(f, &off);
+ continue;
+ }
+ if (!check_int(p, "sync", &td->sync_io)) {
+ fgetpos(f, &off);
+ continue;
+ }
+ if (!check_int(p, "bwavgtime", &td->bw_avg_time)) {
+ fgetpos(f, &off);
+ continue;
+ }
+ if (!check_int(p, "create_serialize", &td->create_serialize)) {
+ fgetpos(f, &off);
+ continue;
+ }
+ if (!check_int(p, "create_fsync", &td->create_fsync)) {
+ fgetpos(f, &off);
+ continue;
+ }
+ if (!check_int(p, "loops", &td->loops)) {
+ fgetpos(f, &off);
+ continue;
+ }
+ if (!check_int(p, "verify", &td->verify)) {
+ fgetpos(f, &off);
+ continue;
+ }
+ if (!check_int(p, "numjobs", &td->numjobs)) {
+ fgetpos(f, &off);
+ continue;
+ }
+ if (!check_range(p, "bsrange", &ul1, &ul2)) {
+ if (ul1 & 511)
+ printf("bad min block size, must be a multiple of 512\n");
+ else
+ td->min_bs = ul1;
+ if (ul2 & 511)
+ printf("bad max block size, must be a multiple of 512\n");
+ else
+ td->max_bs = ul2;
+ fgetpos(f, &off);
+ continue;
+ }
+ if (!check_strcnv(p, "bs", &ull)) {
+ if (ull & 511)
+ printf("bad block size, must be a multiple of 512\n");
+ else
+ td->bs = ull;
+ fgetpos(f, &off);
+ continue;
+ }
+ if (!check_strcnv(p, "size", &td->file_size)) {
+ fgetpos(f, &off);
+ continue;
+ }
+ if (!check_strcnv(p, "offset", &td->file_offset)) {
+ fgetpos(f, &off);
+ continue;
+ }
+ if (!check_strstore(p, "directory", td->directory)) {
+ printf("got dir %s\n", td->directory);
+ fgetpos(f, &off);
+ continue;
+ }
+ if (!check_str(p, "mem", "malloc")) {
+ td->mem_type = MEM_MALLOC;
+ fgetpos(f, &off);
+ continue;
+ }
+ if (!check_str(p, "mem", "shm")) {
+ td->mem_type = MEM_SHM;
+ fgetpos(f, &off);
+ continue;
+ }
+ if (!strncmp(p, "sequential", 10)) {
td->sequential = 1;
fgetpos(f, &off);
continue;
}
- if (!strcmp(p, "random")) {
+ if (!strncmp(p, "random", 6)) {
td->sequential = 0;
fgetpos(f, &off);
continue;
}
+ if (!strncmp(p, "aio", 3)) {
+ td->use_aio = 1;
+ fgetpos(f, &off);
+ continue;
+ }
+ if (!strncmp(p, "create", 6)) {
+ td->create_file = 1;
+ fgetpos(f, &off);
+ continue;
+ }
+ if (!strncmp(p, "overwrite", 9)) {
+ td->overwrite = 1;
+ fgetpos(f, &off);
+ continue;
+ }
+ if (!strncmp(p, "exitall", 7)) {
+ exitall_on_terminate = 1;
+ fgetpos(f, &off);
+ continue;
+ }
+ if (!strncmp(p, "stonewall", 9)) {
+ td->stonewall = 1;
+ fgetpos(f, &off);
+ continue;
+ }
+ printf("Client%d: bad option %s\n",td->thread_number,p);
}
fsetpos(f, &off);
if (add_job(td, name, prioclass, prio))
- put_job(td);
+ break;
}
free(string);
free(name);
+ fclose(f);
return 0;
}
-int parse_options(int argc, char *argv[])
+static int parse_options(int argc, char *argv[])
{
int i;
switch (*parm) {
case 's':
parm++;
- sequential = !!atoi(parm);
+ def_thread.sequential = !!atoi(parm);
break;
case 'b':
parm++;
- global_bs = atoi(parm);
- global_bs <<= 10;
- if (!global_bs) {
+ def_thread.bs = atoi(parm);
+ def_thread.bs <<= 10;
+ if (!def_thread.bs) {
printf("bad block size\n");
- global_bs = DEF_BS;
+ def_thread.bs = DEF_BS;
}
break;
case 't':
parm++;
- timeout = atoi(parm);
- if (!timeout) {
- printf("bad timeout\n");
- timeout = DEF_TIMEOUT;
- }
- break;
- case 'w':
- parm++;
- write_stat = !!atoi(parm);
+ def_thread.timeout = atoi(parm);
break;
case 'r':
parm++;
repeatable = !!atoi(parm);
break;
+ case 'R':
+ parm++;
+ rate_quit = !!atoi(parm);
+ break;
case 'o':
parm++;
- odirect = !!atoi(parm);
+ def_thread.odirect = !!atoi(parm);
break;
case 'f':
if (i + 1 >= argc) {
break;
}
ini_file = strdup(argv[i+1]);
+ i++;
+ break;
+ case 'l':
+ write_lat_log = 1;
+ break;
+ case 'w':
+ write_bw_log = 1;
break;
default:
printf("bad option %s\n", argv[i]);
return i;
}
-int main(int argc, char *argv[])
+static void print_thread_status(struct thread_data *td, int nr_running,
+ int t_rate, int m_rate)
{
- static unsigned long max_run[2], min_run[2], total_blocks[2];
- static unsigned long max_bw[2], min_bw[2], maxl[2], minl[2];
- static unsigned long read_mb, write_mb, read_agg, write_agg;
- int i;
+ printf("Threads now running: %d", nr_running);
+ if (m_rate || t_rate)
+ printf(", commitrate %d/%dKiB/sec", t_rate, m_rate);
+ printf(" : [%s]\r", run_str);
+ fflush(stdout);
+}
- shm_id = shmget(0, MAX_JOBS * sizeof(struct thread_data), IPC_CREAT | 0600);
- if (shm_id == -1) {
- perror("shmget");
- return 1;
- }
+static void check_str_update(struct thread_data *td, int n, int t, int m)
+{
+ char c = run_str[td->thread_number - 1];
- threads = shmat(shm_id, NULL, 0);
- if (threads == (void *) -1 ) {
- perror("shmat");
- return 1;
+ if (td->runstate == td->old_runstate)
+ return;
+
+ switch (td->runstate) {
+ case TD_REAPED:
+ c = '_';
+ break;
+ case TD_EXITED:
+ c = 'E';
+ break;
+ case TD_RUNNING:
+ if (td_read(td)) {
+ if (td->sequential)
+ c = 'R';
+ else
+ c = 'r';
+ } else {
+ if (td->sequential)
+ c = 'W';
+ else
+ c = 'w';
+ }
+ break;
+ case TD_VERIFYING:
+ c = 'V';
+ break;
+ case TD_CREATED:
+ c = 'C';
+ break;
+ case TD_NOT_CREATED:
+ c = 'P';
+ break;
+ default:
+ printf("state %d\n", td->runstate);
}
- atexit(free_shm);
+ run_str[td->thread_number - 1] = c;
+ print_thread_status(td, n, t, m);
+ td->old_runstate = td->runstate;
+}
- if (sched_getaffinity(getpid(), sizeof(def_cpumask), &def_cpumask) == -1) {
- perror("sched_getaffinity");
- return 1;
- }
+static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
+{
+ int i;
- i = parse_options(argc, argv);
+ /*
+ * reap exited threads (TD_EXITED -> TD_REAPED)
+ */
+ for (i = 0; i < thread_number; i++) {
+ struct thread_data *td = &threads[i];
- if (ini_file) {
- if (parse_jobs_ini(ini_file))
- return 1;
- } else
- parse_jobs_cmd(argc, argv, i);
+ check_str_update(td, *nr_running, *t_rate, *m_rate);
- if (!thread_number) {
- printf("Nothing to do\n");
- return 1;
+ if (td->runstate != TD_EXITED)
+ continue;
+
+ td_set_runstate(td, TD_REAPED);
+ waitpid(td->pid, NULL, 0);
+ (*nr_running)--;
+ (*m_rate) -= td->ratemin;
+ (*t_rate) -= td->rate;
+ check_str_update(td, *nr_running, *t_rate, *m_rate);
+
+ if (td->terminate)
+ continue;
}
+}
- printf("%s: %s, bs=%uKiB, timeo=%u, write_stat=%u, odirect=%d\n", argv[0], sequential ? "sequential" : "random", global_bs >> 10, timeout, write_stat, odirect);
+static void run_threads(char *argv[])
+{
+ struct timeval genesis;
+ struct thread_data *td;
+ unsigned long spent;
+ int i, todo, nr_running, m_rate, t_rate, nr_started;
printf("Starting %d threads\n", thread_number);
fflush(stdout);
+ signal(SIGINT, sig_handler);
+
+ todo = thread_number;
+ nr_running = 0;
+ nr_started = 0;
+ m_rate = t_rate = 0;
+
for (i = 0; i < thread_number; i++) {
- sem_init(&startup_sem, 1, 1);
+ td = &threads[i];
+
+ if (!td->create_serialize)
+ continue;
- if (fork())
- sem_wait(&startup_sem);
- else {
- thread_main(shm_id, i, argv);
- exit(0);
+ /*
+ * do file setup here so it happens sequentially,
+ * we don't want X number of threads getting their
+ * client data interspersed on disk
+ */
+ if (setup_file(td)) {
+ td_set_runstate(td, TD_REAPED);
+ todo--;
}
}
- signal(SIGALRM, sig_handler);
- alarm(timeout);
+ gettimeofday(&genesis, NULL);
- printf("Kicking off %d threads\n", thread_number);
- fflush(stdout);
- for (i = 0; i < thread_number; i++) {
- struct thread_data *td = &threads[i];
+ while (todo) {
+ /*
+ * create threads (TD_NOT_CREATED -> TD_CREATED)
+ */
+ for (i = 0; i < thread_number; i++) {
+ td = &threads[i];
+
+ if (td->runstate != TD_NOT_CREATED)
+ continue;
+
+ /*
+ * never got a chance to start, killed by other
+ * thread for some reason
+ */
+ if (td->terminate) {
+ todo--;
+ continue;
+ }
+
+ if (td->start_delay) {
+ spent = mtime_since_now(&genesis);
+
+ if (td->start_delay * 1000 > spent)
+ continue;
+ }
+
+ if (td->stonewall && (nr_started || nr_running))
+ break;
+
+ td_set_runstate(td, TD_CREATED);
+ check_str_update(td, nr_running, t_rate, m_rate);
+ sem_init(&startup_sem, 1, 1);
+ todo--;
+ nr_started++;
+
+ if (fork())
+ sem_wait(&startup_sem);
+ else {
+ thread_main(shm_id, i, argv);
+ exit(0);
+ }
+ }
+
+ /*
+ * start created threads (TD_CREATED -> TD_RUNNING)
+ */
+ for (i = 0; i < thread_number; i++) {
+ struct thread_data *td = &threads[i];
+
+ if (td->runstate != TD_CREATED)
+ continue;
+
+ td_set_runstate(td, TD_RUNNING);
+ nr_running++;
+ nr_started--;
+ m_rate += td->ratemin;
+ t_rate += td->rate;
+ check_str_update(td, nr_running, t_rate, m_rate);
+ sem_post(&td->mutex);
+ }
+
+ for (i = 0; i < thread_number; i++) {
+ struct thread_data *td = &threads[i];
+
+ if (td->runstate != TD_RUNNING &&
+ td->runstate != TD_VERIFYING)
+ continue;
- sem_post(&td->mutex);
+ check_str_update(td, nr_running, t_rate, m_rate);
+ }
+
+ reap_threads(&nr_running, &t_rate, &m_rate);
+
+ if (todo)
+ usleep(100000);
}
- for (i = 0; i < thread_number; i++) {
- struct thread_data *td = &threads[i];
+ while (nr_running) {
+ reap_threads(&nr_running, &t_rate, &m_rate);
+ usleep(10000);
+ }
+}
- waitpid(td->pid, NULL, 0);
+int setup_thread_area(void)
+{
+ /*
+ * 1024 is too much on some machines, scale max_jobs if
+ * we get a failure that looks like too large a shm segment
+ */
+ do {
+ int s = max_jobs * sizeof(struct thread_data);
+
+ shm_id = shmget(0, s, IPC_CREAT | 0600);
+ if (shm_id != -1)
+ break;
+ if (errno != EINVAL) {
+ perror("shmget");
+ break;
+ }
+
+ max_jobs >>= 1;
+ } while (max_jobs);
+
+ if (shm_id == -1)
+ return 1;
+
+ threads = shmat(shm_id, NULL, 0);
+ if (threads == (void *) -1) {
+ perror("shmat");
+ return 1;
+ }
+
+ atexit(free_shm);
+ return 0;
+}
+
+int main(int argc, char *argv[])
+{
+ static unsigned long max_run[2], min_run[2];
+ static unsigned long max_bw[2], min_bw[2];
+ static unsigned long io_mb[2], agg[2];
+ int i;
+
+ if (setup_thread_area())
+ return 1;
+
+ if (sched_getaffinity(getpid(), sizeof(cpu_set_t), &def_thread.cpumask) == -1) {
+ perror("sched_getaffinity");
+ return 1;
+ }
+
+ /*
+ * fill globals
+ */
+ def_thread.ddir = DDIR_READ;
+ def_thread.bs = DEF_BS;
+ def_thread.min_bs = -1;
+ def_thread.max_bs = -1;
+ def_thread.odirect = DEF_ODIRECT;
+ def_thread.ratecycle = DEF_RATE_CYCLE;
+ def_thread.sequential = DEF_SEQUENTIAL;
+ def_thread.timeout = DEF_TIMEOUT;
+ def_thread.create_file = DEF_CREATE;
+ def_thread.overwrite = DEF_OVERWRITE;
+ def_thread.invalidate_cache = DEF_INVALIDATE;
+ def_thread.sync_io = DEF_SYNCIO;
+ def_thread.mem_type = MEM_MALLOC;
+ def_thread.bw_avg_time = DEF_BWAVGTIME;
+ def_thread.create_serialize = DEF_CREATE_SER;
+ def_thread.create_fsync = DEF_CREATE_FSYNC;
+ def_thread.loops = DEF_LOOPS;
+ def_thread.verify = DEF_VERIFY;
+ def_thread.stonewall = DEF_STONEWALL;
+ def_thread.numjobs = DEF_NUMJOBS;
+
+ i = parse_options(argc, argv);
+
+ if (!ini_file) {
+ printf("Need job file\n");
+ return 1;
+ }
+
+ if (parse_jobs_ini(ini_file))
+ return 1;
+
+ if (!thread_number) {
+ printf("Nothing to do\n");
+ return 1;
}
+ run_threads(argv);
+
min_bw[0] = min_run[0] = ~0UL;
min_bw[1] = min_run[1] = ~0UL;
- minl[0] = minl[1] = ~0UL;
+ io_mb[0] = io_mb[1] = 0;
+ agg[0] = agg[1] = 0;
for (i = 0; i < thread_number; i++) {
struct thread_data *td = &threads[i];
unsigned long bw = 0;
- if (td->error)
- goto show_stat;
-
- if (td->runtime < min_run[td->ddir])
- min_run[td->ddir] = td->runtime;
- if (td->runtime > max_run[td->ddir])
- max_run[td->ddir] = td->runtime;
-
- if (td->runtime)
- bw = (td->io_blocks * td->bs) / td->runtime;
- if (bw < min_bw[td->ddir])
- min_bw[td->ddir] = bw;
- if (bw > max_bw[td->ddir])
- max_bw[td->ddir] = bw;
- if (td->max_latency < minl[td->ddir])
- minl[td->ddir] = td->max_latency;
- if (td->max_latency > maxl[td->ddir])
- maxl[td->ddir] = td->max_latency;
-
- total_blocks[td->ddir] += td->io_blocks;
-
- if (td->ddir == DDIR_READ) {
- read_mb += (td->bs * td->io_blocks) >> 20;
- if (td->runtime)
- read_agg += (td->io_blocks * td->bs) / td->runtime;
- }
- if (td->ddir == DDIR_WRITE) {
- write_mb += (td->bs * td->io_blocks) >> 20;
+ if (!td->error) {
+ if (td->runtime < min_run[td->ddir])
+ min_run[td->ddir] = td->runtime;
+ if (td->runtime > max_run[td->ddir])
+ max_run[td->ddir] = td->runtime;
+
if (td->runtime)
- write_agg += (td->io_blocks * td->bs) / td->runtime;
+ bw = td->io_bytes / td->runtime;
+ if (bw < min_bw[td->ddir])
+ min_bw[td->ddir] = bw;
+ if (bw > max_bw[td->ddir])
+ max_bw[td->ddir] = bw;
+
+ io_mb[td->ddir] += td->io_bytes >> 20;
}
-show_stat:
show_thread_status(td);
}
+
+ if (max_run[0])
+ agg[0] = (io_mb[0] * 1024 * 1000) / max_run[0];
+ if (max_run[1])
+ agg[1] = (io_mb[1] * 1024 * 1000) / max_run[1];
- printf("Run status:\n");
+ printf("\nRun status:\n");
if (max_run[DDIR_READ])
- printf(" READ: io=%luMiB, aggrb=%lu, minl=%lu, maxl=%lu, minb=%lu, maxb=%lu, mint=%lumsec, maxt=%lumsec\n", read_mb, read_agg, minl[0], maxl[0], min_bw[0], max_bw[0], min_run[0], max_run[0]);
+ printf(" READ: io=%luMiB, aggrb=%lu, minb=%lu, maxb=%lu, mint=%lumsec, maxt=%lumsec\n", io_mb[0], agg[0], min_bw[0], max_bw[0], min_run[0], max_run[0]);
if (max_run[DDIR_WRITE])
- printf(" WRITE: io=%luMiB, aggrb=%lu, minl=%lu, maxl=%lu, minb=%lu, maxb=%lu, mint=%lumsec, maxt=%lumsec\n", write_mb, write_agg, minl[1], maxl[1], min_bw[1], max_bw[1], min_run[1], max_run[1]);
+ printf(" WRITE: io=%luMiB, aggrb=%lu, minb=%lu, maxb=%lu, mint=%lumsec, maxt=%lumsec\n", io_mb[1], agg[1], min_bw[1], max_bw[1], min_run[1], max_run[1]);
+
return 0;
}
projects / disktools.git / blobdiff