#include <errno.h>
#include <signal.h>
#include <time.h>
-#include <ctype.h>
-#include <sched.h>
-#include <libaio.h>
#include <math.h>
-#include <limits.h>
-#include <sys/time.h>
+#include <assert.h>
+#include <dirent.h>
+#include <libgen.h>
#include <sys/types.h>
#include <sys/stat.h>
#include <sys/wait.h>
-#include <semaphore.h>
#include <sys/ipc.h>
#include <sys/shm.h>
-#include <asm/unistd.h>
+#include <sys/ioctl.h>
+#include <sys/mman.h>
-#include "list.h"
-
-#define MAX_JOBS (1024)
-
-/*
- * 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"
-#endif
-#endif
-
-#ifndef __NR_fadvise64
-#if defined(__i386__)
-#define __NR_fadvise64 250
-#elif defined(__powerpc__) || defined(__powerpc64__)
-#define __NR_fadvise64 233
-#elif defined(__x86_64__)
-#define __NR_fadvise64 221
-#elif defined(__ia64__)
-#define __NR_fadvise64 1234
-#elif defined(__alpha__)
-#define __NR_fadvise64 413
-#elif defined(__s390x__) || defined(__s390__)
-#define __NR_fadvise64 253
-#else
-#error "Unsupported arch"
-#endif
-#endif
-
-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,
- IOPRIO_WHO_USER,
-};
-
-#define IOPRIO_CLASS_SHIFT 13
+#include "fio.h"
+#include "os.h"
#define MASK (4095)
-#define DEF_BS (4096)
-#define DEF_TIMEOUT (30)
-#define DEF_RATE_CYCLE (1000)
-#define DEF_ODIRECT (1)
-#define DEF_SEQUENTIAL (1)
-#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 ALIGN(buf) (char *) (((unsigned long) (buf) + MASK) & ~(MASK))
-static int repeatable = DEF_RAND_REPEAT;
-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;
+int groupid = 0;
+int thread_number = 0;
+char run_str[MAX_JOBS + 1];
+int shm_id = 0;
+static LIST_HEAD(disk_list);
-static int max_jobs = MAX_JOBS;
-
-static char run_str[MAX_JOBS + 1];
-
-static int shm_id;
-
-enum {
- DDIR_READ = 0,
- DDIR_WRITE,
-};
+static void update_io_ticks(void);
+static void disk_util_timer_arm(void);
/*
* thread life cycle
enum {
TD_NOT_CREATED = 0,
TD_CREATED,
- TD_STARTED,
+ 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;
-};
-
-#define td_read(td) ((td)->ddir == DDIR_READ)
-#define should_fsync(td) (!td_read(td) && !(td)->odirect)
-
-struct thread_data {
- char file_name[256];
- int thread_number;
- int error;
- int fd;
- pid_t pid;
- char *orig_buffer;
- volatile int terminate;
- volatile int 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 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;
- cpu_set_t cpumask;
-
- struct drand48_data bsrange_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_kb;
- struct timeval lastrate;
-
- unsigned long runtime; /* sec */
- unsigned long kb;
- unsigned long io_blocks;
- unsigned long io_kb;
- unsigned long last_kb;
- sem_t mutex;
- struct drand48_data random_state;
-
- /*
- * bandwidth and latency stats
- */
- 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_kb;
- struct timeval stat_sample_time;
-
- struct io_log *lat_log;
- struct io_log *bw_log;
-
- struct timeval start;
-};
-
-static struct thread_data *threads;
-static struct thread_data def_thread;
+#define should_fsync(td) (td_write(td) && !(td)->odirect)
static sem_t startup_sem;
-static void sig_handler(int sig)
+#define TERMINATE_ALL (-1)
+
+static void terminate_threads(int group_id)
{
int i;
for (i = 0; i < thread_number; i++) {
struct thread_data *td = &threads[i];
- td->terminate = 1;
- td->start_delay = 0;
+ if (group_id == TERMINATE_ALL || groupid == td->groupid) {
+ td->terminate = 1;
+ td->start_delay = 0;
+ }
}
}
-static int init_random_state(struct thread_data *td)
+static void sig_handler(int sig)
{
- unsigned long seed = DEF_RANDSEED;
-
- srand48_r(seed, &td->bsrange_state);
-
- if (td->sequential)
- return 0;
-
- if (!repeatable) {
- int 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;
- }
-
- close(fd);
+ switch (sig) {
+ case SIGALRM:
+ update_io_ticks();
+ disk_util_timer_arm();
+ break;
+ default:
+ printf("\nfio: terminating on signal\n");
+ fflush(stdout);
+ terminate_threads(TERMINATE_ALL);
+ break;
}
-
- srand48_r(seed, &td->random_state);
- return 0;
}
static unsigned long utime_since(struct timeval *s, struct timeval *e)
return sec + usec;
}
+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 s->tv_sec * 1000 + s->tv_usec / 1000;
}
-static unsigned long long get_next_offset(struct thread_data *td)
+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++;
+ }
+
+ 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)) {
+ unsigned int idx, bit;
+
+ if (!random_map_free(td, block))
+ break;
+
+ idx = RAND_MAP_IDX(td, block);
+ bit = RAND_MAP_BIT(td, block);
+
+ 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;
+}
+
+static int get_next_offset(struct thread_data *td, unsigned long long *offset)
{
- unsigned long long kb;
+ unsigned long long b, rb;
long r;
if (!td->sequential) {
- lrand48_r(&td->random_state, &r);
- kb = (1+(double) (td->kb-1) * r / (RAND_MAX+1.0));
+ 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
- kb = td->last_kb;
+ b = td->last_bytes / td->min_bs;
- return (kb << 10) + td->file_offset;
+ *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)
buflen = (buflen + td->min_bs - 1) & ~(td->min_bs - 1);
}
- if (buflen > ((td->kb - td->io_kb) << 10))
- buflen = (td->kb - td->io_kb) << 10;
+ if (buflen > td->io_size - td->this_io_bytes)
+ buflen = td->io_size - td->this_io_bytes;
- td->last_kb += buflen >> 10;
return buflen;
}
-static inline void add_stat_sample(struct thread_data *td, struct io_stat *is,
- unsigned long val)
+static inline void add_stat_sample(struct io_stat *is, unsigned long val)
{
if (val > is->max_val)
is->max_val = val;
is->samples++;
}
-static void add_log_sample(struct thread_data *td, struct io_log *log,
+static void add_log_sample(struct thread_data *td, struct io_log *iolog,
unsigned long val)
{
- if (log->nr_samples == log->max_samples) {
- int new_size = sizeof(struct io_sample) * log->max_samples * 2;
+ if (iolog->nr_samples == iolog->max_samples) {
+ int new_size = sizeof(struct io_sample) * iolog->max_samples*2;
- log->log = realloc(log->log, new_size);
- log->max_samples <<= 1;
+ iolog->log = realloc(iolog->log, new_size);
+ iolog->max_samples <<= 1;
}
- log->log[log->nr_samples].val = val;
- log->log[log->nr_samples].time = mtime_since_now(&td->start);
- log->nr_samples++;
+ iolog->log[iolog->nr_samples].val = val;
+ iolog->log[iolog->nr_samples].time = mtime_since_now(&td->start);
+ iolog->nr_samples++;
}
static void add_clat_sample(struct thread_data *td, unsigned long msec)
{
- add_stat_sample(td, &td->clat_stat, msec);
+ add_stat_sample(&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);
+ add_stat_sample(&td->slat_stat, msec);
}
-static void add_bw_sample(struct thread_data *td, unsigned long msec)
+static void add_bw_sample(struct thread_data *td)
{
unsigned long spent = mtime_since_now(&td->stat_sample_time);
unsigned long rate;
if (spent < td->bw_avg_time)
return;
- rate = ((td->io_kb - td->stat_io_kb) * 1024) / spent;
- add_stat_sample(td, &td->bw_stat, rate);
+ rate = (td->this_io_bytes - td->stat_io_bytes) / spent;
+ add_stat_sample(&td->bw_stat, rate);
if (td->bw_log)
add_log_sample(td, td->bw_log, rate);
gettimeofday(&td->stat_sample_time, NULL);
- td->stat_io_kb = td->io_kb;
+ td->stat_io_bytes = td->this_io_bytes;
+}
+
+/*
+ * busy looping version for the last few usec
+ */
+static void __usec_sleep(unsigned int usec)
+{
+ struct timeval start;
+
+ gettimeofday(&start, NULL);
+ while (utime_since_now(&start) < usec)
+ nop;
}
-static void usec_sleep(int usec)
+static void usec_sleep(struct thread_data *td, unsigned long usec)
{
- struct timespec req = { .tv_sec = 0, .tv_nsec = usec * 1000 };
- struct timespec rem;
+ struct timespec req, rem;
+
+ req.tv_sec = usec / 1000000;
+ req.tv_nsec = usec * 1000 - req.tv_sec * 1000000;
do {
+ if (usec < 5000) {
+ __usec_sleep(usec);
+ break;
+ }
+
rem.tv_sec = rem.tv_nsec = 0;
- nanosleep(&req, &rem);
- if (!rem.tv_nsec)
+ if (nanosleep(&req, &rem) < 0)
+ break;
+
+ if ((rem.tv_sec + rem.tv_nsec) == 0)
break;
req.tv_nsec = rem.tv_nsec;
- } while (1);
+ req.tv_sec = rem.tv_sec;
+
+ usec = rem.tv_sec * 1000000 + rem.tv_nsec / 1000;
+ } while (!td->terminate);
}
static void rate_throttle(struct thread_data *td, unsigned long time_spent,
td->rate_pending_usleep += s;
if (td->rate_pending_usleep >= 100000) {
- usec_sleep(td->rate_pending_usleep);
+ usec_sleep(td, td->rate_pending_usleep);
td->rate_pending_usleep = 0;
}
} else {
/*
* if rate blocks is set, sample is running
*/
- if (td->rate_kb) {
+ if (td->rate_bytes) {
spent = mtime_since(&td->lastrate, now);
if (spent < td->ratecycle)
return 0;
- rate = ((td->io_kb - td->rate_kb) * 1024) / 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);
+ terminate_threads(td->groupid);
return 1;
}
}
- td->rate_kb = td->io_kb;
+ td->rate_bytes = td->this_io_bytes;
memcpy(&td->lastrate, now, sizeof(*now));
return 0;
}
static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
{
- if (mtime_since(&td->start, t) >= td->timeout * 1000)
+ if (!td->timeout)
+ return 0;
+ if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
return 1;
return 0;
}
+static void fill_random_bytes(struct thread_data *td,
+ unsigned char *p, unsigned int len)
+{
+ unsigned int todo;
+ double r;
+
+ while (len) {
+ drand48_r(&td->verify_state, &r);
+
+ /*
+ * 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;
+
+ memcpy(p, &r, todo);
+
+ len -= todo;
+ p += todo;
+ }
+}
+
+static void hexdump(void *buffer, int len)
+{
+ unsigned char *p = buffer;
+ int i;
+
+ for (i = 0; i < len; i++)
+ printf("%02x", p[i]);
+ printf("\n");
+}
+
+static int verify_io_u_crc32(struct verify_header *hdr, struct io_u *io_u)
+{
+ unsigned char *p = (unsigned char *) io_u->buf;
+ unsigned long c;
+
+ p += sizeof(*hdr);
+ c = crc32(p, hdr->len - sizeof(*hdr));
+
+ return c != hdr->crc32;
+}
+
+static int verify_io_u_md5(struct verify_header *hdr, struct io_u *io_u)
+{
+ unsigned char *p = (unsigned char *) io_u->buf;
+ struct md5_ctx md5_ctx;
+ int ret;
+
+ memset(&md5_ctx, 0, sizeof(md5_ctx));
+ p += sizeof(*hdr);
+ md5_update(&md5_ctx, p, hdr->len - sizeof(*hdr));
+
+ 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));
+ }
+
+ return ret;
+}
+
+static int verify_io_u(struct io_u *io_u)
+{
+ struct verify_header *hdr = (struct verify_header *) io_u->buf;
+ int ret;
+
+ if (hdr->fio_magic != FIO_HDR_MAGIC)
+ return 1;
+
+ if (hdr->verify_type == VERIFY_MD5)
+ ret = verify_io_u_md5(hdr, io_u);
+ else if (hdr->verify_type == VERIFY_CRC32)
+ ret = verify_io_u_crc32(hdr, io_u);
+ else {
+ fprintf(stderr, "Bad verify type %d\n", hdr->verify_type);
+ ret = 1;
+ }
+
+ return ret;
+}
+
+static void fill_crc32(struct verify_header *hdr, void *p, unsigned int len)
+{
+ hdr->crc32 = crc32(p, len);
+}
+
+static void fill_md5(struct verify_header *hdr, void *p, unsigned int len)
+{
+ struct md5_ctx md5_ctx;
+
+ memset(&md5_ctx, 0, sizeof(md5_ctx));
+ md5_update(&md5_ctx, p, len);
+ memcpy(hdr->md5_digest, md5_ctx.hash, sizeof(md5_ctx.hash));
+}
+
/*
* 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 io_u *io_u)
+static void populate_io_u(struct thread_data *td, struct io_u *io_u)
{
+ unsigned char *p = (unsigned char *) io_u->buf;
+ struct verify_header hdr;
+
+ hdr.fio_magic = FIO_HDR_MAGIC;
+ hdr.len = io_u->buflen;
+ p += sizeof(hdr);
+ fill_random_bytes(td, p, io_u->buflen - sizeof(hdr));
+
+ if (td->verify == VERIFY_MD5) {
+ fill_md5(&hdr, p, io_u->buflen - sizeof(hdr));
+ hdr.verify_type = VERIFY_MD5;
+ } else {
+ fill_crc32(&hdr, p, io_u->buflen - sizeof(hdr));
+ hdr.verify_type = VERIFY_CRC32;
+ }
+
+ memcpy(io_u->buf, &hdr, sizeof(hdr));
}
static void put_io_u(struct thread_data *td, struct io_u *io_u)
td->cur_depth--;
}
-static struct io_u *get_io_u(struct thread_data *td)
+#define queue_full(td) (list_empty(&(td)->io_u_freelist))
+
+static struct io_u *__get_io_u(struct thread_data *td)
{
struct io_u *io_u;
- unsigned int len;
- unsigned long long off;
-
- if (list_empty(&td->io_u_freelist))
- return NULL;
- off = get_next_offset(td);
- len = get_next_buflen(td);
- if (!len)
+ if (queue_full(td))
return NULL;
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;
+}
- io_u->offset = off;
- io_u->buflen = len;
+static struct io_u *get_io_u(struct thread_data *td)
+{
+ struct io_u *io_u;
- if (!td_read(td) && td->verify)
- populate_io_u(io_u);
+ io_u = __get_io_u(td);
+ if (!io_u)
+ return NULL;
- 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 (get_next_offset(td, &io_u->offset)) {
+ put_io_u(td, io_u);
+ return NULL;
+ }
+
+ io_u->buflen = get_next_buflen(td);
+ if (!io_u->buflen) {
+ put_io_u(td, io_u);
+ return NULL;
+ }
+
+ if (io_u->buflen + io_u->offset > td->file_size)
+ io_u->buflen = td->file_size - io_u->offset;
+
+ if (!io_u->buflen) {
+ put_io_u(td, io_u);
+ return NULL;
}
+ if (!td->sequential)
+ mark_random_map(td, io_u);
+
+ td->last_bytes += io_u->buflen;
+
+ if (td->verify != VERIFY_NONE)
+ populate_io_u(td, io_u);
+
+ if (td->io_prep)
+ td->io_prep(td, io_u, td_read(td));
+
gettimeofday(&io_u->start_time, NULL);
- td->cur_depth++;
return io_u;
}
-static int do_sync_verify(struct thread_data *td)
+static inline void td_set_runstate(struct thread_data *td, int runstate)
{
- return 0;
+ td->old_runstate = td->runstate;
+ td->runstate = runstate;
}
-static void do_sync_io(struct thread_data *td)
+static int get_next_verify(struct thread_data *td,
+ unsigned long long *offset, unsigned int *len)
{
- unsigned long msec, usec;
- struct timeval e;
+ struct io_piece *ipo;
- td->cur_off = 0;
-
- for (td->io_kb = 0; td->io_kb < td->kb;) {
- struct io_u *io_u;
- int ret;
+ if (list_empty(&td->io_hist_list))
+ return 1;
- if (td->terminate)
- break;
+ ipo = list_entry(td->io_hist_list.next, struct io_piece, list);
+ list_del(&ipo->list);
- io_u = get_io_u(td);
- if (!io_u)
- break;
+ *offset = ipo->offset;
+ *len = ipo->len;
+ free(ipo);
+ return 0;
+}
- if (td->cur_off != io_u->offset) {
- if (lseek(td->fd, io_u->offset, SEEK_SET) == -1) {
- td->error = errno;
- break;
- }
- }
+static void prune_io_piece_log(struct thread_data *td)
+{
+ struct io_piece *ipo;
- 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);
+ while (!list_empty(&td->io_hist_list)) {
+ ipo = list_entry(td->io_hist_list.next, struct io_piece, list);
- if (ret < (int) io_u->buflen) {
- if (ret == -1)
- td->error = errno;
- break;
- }
+ list_del(&ipo->list);
+ free(ipo);
+ }
+}
- td->io_blocks++;
- td->io_kb += io_u->buflen >> 10;
- td->cur_off = io_u->offset + io_u->buflen;
+/*
+ * 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;
- gettimeofday(&e, NULL);
+ INIT_LIST_HEAD(&ipo->list);
+ ipo->offset = io_u->offset;
+ ipo->len = io_u->buflen;
- usec = utime_since(&io_u->start_time, &e);
+ /*
+ * 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;
+ }
- rate_throttle(td, usec, io_u->buflen);
+ /*
+ * 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 (check_min_rate(td, &e)) {
- td->error = ENODATA;
+ if (__ipo->offset < ipo->offset)
break;
- }
+ }
- msec = usec / 1000;
- add_clat_sample(td, msec);
- add_bw_sample(td, msec);
+ list_add(&ipo->list, entry);
+}
- if (runtime_exceeded(td, &e))
- break;
+static void do_sync_verify(struct thread_data *td)
+{
+ struct timeval t;
+ struct io_u *io_u = NULL;
+ int ret;
- put_io_u(td, io_u);
+ td_set_runstate(td, TD_VERIFYING);
- if (td->thinktime)
- usec_sleep(td->thinktime);
+ io_u = __get_io_u(td);
- if (should_fsync(td) && td->fsync_blocks &&
- (td->io_blocks % td->fsync_blocks) == 0)
- fsync(td->fd);
+ if (!td->odirect) {
+ if (!td->use_mmap) {
+ if (fadvise(td->fd, td->file_offset, td->io_size, POSIX_FADV_DONTNEED) < 0) {
+ td_verror(td, errno);
+ goto out;
+ }
+ } else {
+ if (madvise(td->mmap, td->io_size, MADV_DONTNEED)) {
+ td_verror(td, errno);
+ goto out;
+ }
+ }
}
- if (should_fsync(td))
+ 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_verror(td, errno);
+ break;
+ }
+ }
+
+ ret = read(td->fd, io_u->buf, io_u->buflen);
+ if (ret < (int) io_u->buflen) {
+ if (ret == -1) {
+ td_verror(td, 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 int __do_sync_mmap(struct thread_data *td, struct io_u *io_u)
+{
+ unsigned long long real_off = io_u->offset - td->file_offset;
+
+ if (td_read(td))
+ memcpy(io_u->buf, td->mmap + real_off, io_u->buflen);
+ else
+ memcpy(td->mmap + real_off, io_u->buf, io_u->buflen);
+
+ /*
+ * not really direct, but should drop the pages from the cache
+ */
+ if (td->odirect) {
+ msync(td->mmap + real_off, io_u->buflen, MS_SYNC);
+ madvise(td->mmap + real_off, io_u->buflen, MADV_DONTNEED);
+ }
+
+ return io_u->buflen;
+}
+
+static int __do_sync_rw(struct thread_data *td, struct io_u *io_u)
+{
+ if (td->cur_off != io_u->offset) {
+ if (lseek(td->fd, io_u->offset, SEEK_SET) == -1) {
+ td_verror(td, errno);
+ return 1;
+ }
+ }
+
+ if (td_read(td))
+ return read(td->fd, io_u->buf, io_u->buflen);
+ else
+ return write(td->fd, io_u->buf, io_u->buflen);
+}
+
+static void sync_td(struct thread_data *td)
+{
+ if (!td->use_mmap)
fsync(td->fd);
+ else
+ msync(td->mmap, td->file_size, MS_SYNC);
}
-static int io_u_queue(struct thread_data *td, struct io_u *io_u)
+static void do_sync_io(struct thread_data *td)
{
- struct iocb *iocb = &io_u->iocb;
- int ret;
+ unsigned long msec, usec;
+ struct io_u *io_u = NULL;
+ struct timeval e;
- 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;
+ 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->use_mmap)
+ ret = __do_sync_rw(td, io_u);
else
+ ret = __do_sync_mmap(td, io_u);
+
+ if (ret < (int) io_u->buflen) {
+ if (ret == -1)
+ td_verror(td, errno);
break;
- } while (1);
+ }
- return ret;
+ 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_verror(td, ENOMEM);
+ 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, td->thinktime);
+
+ if (should_fsync(td) && td->fsync_blocks &&
+ (td->io_blocks % td->fsync_blocks) == 0)
+ sync_td(td);
+ }
+
+ if (io_u)
+ put_io_u(td, io_u);
+
+ if (should_fsync(td))
+ sync_td(td);
+}
+
+static int io_u_getevents(struct thread_data *td, int min, int max,
+ struct timespec *t)
+{
+ return td->io_getevents(td, min, max, t);
+}
+
+static int io_u_queue(struct thread_data *td, struct io_u *io_u)
+{
+ return td->io_queue(td, io_u);
}
#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)
{
gettimeofday(&e, NULL);
for (i = 0, bytes_done = 0; i < nr; i++) {
- io_u = ev_to_iou(td->aio_events + i);
+ io_u = td->io_event(td, i);
td->io_blocks++;
- td->io_kb += io_u->buflen >> 10;
- bytes_done += io_u->buflen;
+ 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, 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);
}
/*
* get immediately available events, if any
*/
- r = io_getevents(td->aio_ctx, 0, td->cur_depth, td->aio_events, &ts);
+ r = io_u_getevents(td, 0, td->cur_depth, &ts);
if (r > 0)
ios_completed(td, r);
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);
+ r = td->io_cancel(td, io_u);
if (!r)
put_io_u(td, io_u);
}
if (td->cur_depth) {
- r = io_getevents(td->aio_ctx, td->cur_depth, td->cur_depth, td->aio_events, NULL);
+ r = io_u_getevents(td, td->cur_depth, td->cur_depth, NULL);
if (r > 0)
ios_completed(td, r);
}
}
-static int do_async_verify(struct thread_data *td)
+static int async_do_verify(struct thread_data *td, struct io_u **io_u)
{
- return 0;
+ 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;
+ }
+
+ td->io_prep(td, io_u, 1);
+
+ ret = io_u_queue(td, io_u);
+ if (ret) {
+ put_io_u(td, io_u);
+ td_verror(td, 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_verror(td, ret);
+ break;
+ }
+
+ v_io_u = td->io_event(td, 0);
+
+ 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;
- for (td->io_kb = 0; td->io_kb < td->kb;) {
+ 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;
ret = io_u_queue(td, io_u);
if (ret) {
put_io_u(td, io_u);
- td->error = ret;
+ td_verror(td, ret);
break;
}
min_evts = 1;
}
- ret = io_getevents(td->aio_ctx, min_evts, td->cur_depth, td->aio_events, timeout);
+ ret = io_u_getevents(td, min_evts, td->cur_depth, timeout);
if (ret < 0) {
- td->error = errno;
+ td_verror(td, ret);
break;
} else if (!ret)
continue;
rate_throttle(td, usec, bytes_done);
if (check_min_rate(td, &e)) {
- td->error = ENODATA;
+ td_verror(td, ENOMEM);
break;
}
break;
if (td->thinktime)
- usec_sleep(td->thinktime);
+ usec_sleep(td, td->thinktime);
if (should_fsync(td) && td->fsync_blocks &&
(td->io_blocks % td->fsync_blocks) == 0)
static void cleanup_aio(struct thread_data *td)
{
- io_destroy(td->aio_ctx);
-
- if (td->aio_events)
- free(td->aio_events);
+ if (td->io_engine == FIO_LIBAIO)
+ fio_libaio_cleanup(td);
+ else if (td->io_engine == FIO_POSIXAIO)
+ fio_posixaio_cleanup(td);
}
static int init_aio(struct thread_data *td)
{
- if (io_queue_init(td->aio_depth, &td->aio_ctx)) {
- td->error = errno;
+ if (td->io_engine == FIO_LIBAIO)
+ return fio_libaio_init(td);
+ else if (td->io_engine == FIO_POSIXAIO)
+ return fio_posixaio_init(td);
+ else {
+ fprintf(stderr, "bad io_engine %d\n", td->io_engine);
return 1;
}
-
- td->aio_events = malloc(td->aio_depth * sizeof(struct io_event));
- return 0;
}
static void cleanup_io_u(struct thread_data *td)
shmdt(td->orig_buffer);
shmctl(td->shm_id, IPC_RMID, &sbuf);
- }
+ } else if (td->mem_type == MEM_MMAP)
+ munmap(td->orig_buffer, td->orig_buffer_size);
+ else
+ fprintf(stderr, "Bad memory type %d\n", td->mem_type);
+
+ td->orig_buffer = NULL;
}
static int init_io_u(struct thread_data *td)
{
struct io_u *io_u;
- int i, max_units, mem_size;
+ int i, max_units;
char *p;
- if (!td->use_aio)
+ if (td->io_engine == FIO_SYNCIO)
max_units = 1;
else
max_units = td->aio_depth;
- mem_size = td->max_bs * max_units + MASK;
+ td->orig_buffer_size = td->max_bs * max_units + MASK;
if (td->mem_type == MEM_MALLOC)
- td->orig_buffer = malloc(mem_size);
+ td->orig_buffer = malloc(td->orig_buffer_size);
else if (td->mem_type == MEM_SHM) {
- td->shm_id = shmget(IPC_PRIVATE, mem_size, IPC_CREAT | 0600);
+ td->shm_id = shmget(IPC_PRIVATE, td->orig_buffer_size, IPC_CREAT | 0600);
if (td->shm_id < 0) {
- td->error = errno;
+ td_verror(td, errno);
perror("shmget");
return 1;
}
td->orig_buffer = shmat(td->shm_id, NULL, 0);
if (td->orig_buffer == (void *) -1) {
- td->error = errno;
+ td_verror(td, errno);
perror("shmat");
+ td->orig_buffer = NULL;
+ return 1;
+ }
+ } else if (td->mem_type == MEM_MMAP) {
+ td->orig_buffer = mmap(NULL, td->orig_buffer_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | OS_MAP_ANON, 0, 0);
+ if (td->orig_buffer == MAP_FAILED) {
+ td_verror(td, errno);
+ perror("mmap");
+ td->orig_buffer = NULL;
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++) {
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;
- 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)
+static int create_file(struct thread_data *td, unsigned long long size,
+ int extend)
{
unsigned long long left;
+ unsigned int bs;
+ int r, oflags;
char *b;
- int r, bs;
/*
* unless specifically asked for overwrite, let normal io extend it
*/
- if (!td_read(td) && !td->overwrite)
+ if (td_write(td) && !td->overwrite)
return 0;
- if (!td->file_size) {
+ if (!size) {
fprintf(stderr, "Need size for create\n");
- td->error = EINVAL;
+ td_verror(td, EINVAL);
return 1;
}
- printf("Client%d: Laying out IO file\n", td->thread_number);
+ if (!extend) {
+ oflags = O_CREAT | O_TRUNC;
+ printf("Client%d: Laying out IO file (%LuMiB)\n", td->thread_number, size >> 20);
+ } else {
+ oflags = O_APPEND;
+ printf("Client%d: Extending IO file (%Lu -> %LuMiB)\n", td->thread_number, (td->file_size - size) >> 20, td->file_size >> 20);
+ }
- td->fd = open(td->file_name, O_WRONLY | O_CREAT | O_TRUNC, 0644);
+ td->fd = open(td->file_name, O_WRONLY | oflags, 0644);
if (td->fd < 0) {
- td->error = errno;
+ td_verror(td, errno);
return 1;
}
- td->kb = td->file_size >> 10;
+ if (!extend && ftruncate(td->fd, td->file_size) == -1) {
+ td_verror(td, 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) {
+ left = size;
+ while (left && !td->terminate) {
bs = td->max_bs;
if (bs > left)
bs = left;
r = write(td->fd, b, bs);
- if (r == bs) {
+ if (r == (int) bs) {
left -= bs;
continue;
} else {
if (r < 0)
- td->error = errno;
+ td_verror(td, errno);
else
- td->error = EIO;
+ td_verror(td, EIO);
break;
}
}
- if (td->create_fsync)
+ if (td->terminate)
+ unlink(td->file_name);
+ else if (td->create_fsync)
fsync(td->fd);
close(td->fd);
return 0;
}
-static int file_exists(struct thread_data *td)
+static int file_size(struct thread_data *td)
{
struct stat st;
- if (stat(td->file_name, &st) != -1)
+ if (fstat(td->fd, &st) == -1) {
+ td_verror(td, errno);
return 1;
+ }
- return errno != ENOENT;
+ if (!td->file_size)
+ td->file_size = st.st_size;
+
+ return 0;
}
-static int setup_file(struct thread_data *td)
+static int bdev_size(struct thread_data *td)
{
- struct stat st;
- int flags = 0;
+ size_t bytes;
- if (!file_exists(td)) {
- if (!td->create_file) {
- td->error = ENOENT;
- return 1;
- }
- if (create_file(td))
- return 1;
+ if (ioctl(td->fd, BLKGETSIZE64, &bytes) < 0) {
+ td_verror(td, errno);
+ return 1;
}
- if (td->odirect)
- flags |= O_DIRECT;
+ /*
+ * no extend possibilities, so limit size to device size if too large
+ */
+ if (!td->file_size || td->file_size > bytes)
+ td->file_size = bytes;
- 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;
+ return 0;
+}
+
+static int get_file_size(struct thread_data *td)
+{
+ int ret;
- td->fd = open(td->file_name, flags | O_WRONLY | O_CREAT, 0600);
+ if (td->filetype == FIO_TYPE_FILE)
+ ret = file_size(td);
+ else
+ ret = bdev_size(td);
+
+ if (ret)
+ return ret;
+
+ if (td->file_offset > td->file_size) {
+ fprintf(stderr, "Client%d: offset larger than length (%Lu > %Lu)\n", td->thread_number, td->file_offset, td->file_size);
+ return 1;
}
- if (td->fd == -1) {
- td->error = errno;
+ td->io_size = td->file_size - td->file_offset;
+ if (td->io_size == 0) {
+ fprintf(stderr, "Client%d: no io blocks\n", td->thread_number);
+ td_verror(td, EINVAL);
return 1;
}
- if (td_read(td)) {
- if (fstat(td->fd, &st) == -1) {
- td->error = errno;
- return 1;
- }
+ return 0;
+}
- if (td->file_size > st.st_size)
- st.st_size = td->file_size;
- } else {
- if (!td->file_size)
- td->file_size = 1024 * 1024 * 1024;
+static int setup_file_mmap(struct thread_data *td)
+{
+ int flags;
- st.st_size = td->file_size;
+ if (td_read(td))
+ flags = PROT_READ;
+ else {
+ flags = PROT_WRITE;
+
+ if (td->verify != VERIFY_NONE)
+ flags |= PROT_READ;
}
- td->kb = (st.st_size - td->file_offset) / 1024;
- if (!td->kb) {
- fprintf(stderr, "Client%d: no io blocks\n", td->thread_number);
- td->error = EINVAL;
+ td->mmap = mmap(NULL, td->file_size, flags, MAP_SHARED, td->fd, td->file_offset);
+ if (td->mmap == MAP_FAILED) {
+ td->mmap = NULL;
+ td_verror(td, errno);
return 1;
}
if (td->invalidate_cache) {
- if (fadvise(td->fd, 0, st.st_size, POSIX_FADV_DONTNEED) < 0) {
- td->error = errno;
+ if (madvise(td->mmap, td->file_size, MADV_DONTNEED) < 0) {
+ td_verror(td, errno);
+ return 1;
+ }
+ }
+
+ if (td->sequential) {
+ if (madvise(td->mmap, td->file_size, MADV_SEQUENTIAL) < 0) {
+ td_verror(td, errno);
+ return 1;
+ }
+ } else {
+ if (madvise(td->mmap, td->file_size, MADV_RANDOM) < 0) {
+ td_verror(td, errno);
return 1;
}
}
return 0;
}
-static void *thread_main(int shm_id, int offset, char *argv[])
+static int setup_file_plain(struct thread_data *td)
{
- struct thread_data *td;
- int ret = 1;
- void *data;
-
- setsid();
-
- data = shmat(shm_id, NULL, 0);
- if (data == (void *) -1) {
- perror("shmat");
- return NULL;
+ if (td->invalidate_cache) {
+ if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_DONTNEED) < 0) {
+ td_verror(td, errno);
+ return 1;
+ }
}
- 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;
+ if (td->sequential) {
+ if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_SEQUENTIAL) < 0) {
+ td_verror(td, errno);
+ return 1;
+ }
+ } else {
+ if (fadvise(td->fd, td->file_offset, td->file_size, POSIX_FADV_RANDOM) < 0) {
+ td_verror(td, errno);
+ return 1;
+ }
}
- sprintf(argv[0], "fio%d", offset);
-
- if (td->use_aio && init_aio(td))
- goto err;
+ return 0;
+}
- if (init_random_state(td))
- goto err;
+static int setup_file(struct thread_data *td)
+{
+ struct stat st;
+ int flags = 0;
- if (td->ioprio) {
- if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
- td->error = errno;
- goto err;
+ if (stat(td->file_name, &st) == -1) {
+ if (errno != ENOENT) {
+ td_verror(td, errno);
+ return 1;
+ }
+ if (!td->create_file) {
+ td_verror(td, ENOENT);
+ return 1;
+ }
+ if (create_file(td, td->file_size, 0))
+ return 1;
+ } else if (td->filetype == FIO_TYPE_FILE) {
+ if (st.st_size < td->file_size) {
+ if (create_file(td, td->file_size - st.st_size, 1))
+ return 1;
}
}
- sem_post(&startup_sem);
- sem_wait(&td->mutex);
-
- if (!td->create_serialize && setup_file(td))
- goto err;
-
- gettimeofday(&td->start, NULL);
+ if (td->odirect)
+ flags |= O_DIRECT;
- if (td->ratemin)
- memcpy(&td->lastrate, &td->start, sizeof(td->start));
+ if (td_read(td))
+ td->fd = open(td->file_name, flags | O_RDONLY);
+ else {
+ if (td->filetype == FIO_TYPE_FILE) {
+ if (!td->overwrite)
+ flags |= O_TRUNC;
- memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
+ flags |= O_CREAT;
+ }
+ if (td->sync_io)
+ flags |= O_SYNC;
- while (td->loops--) {
- if (!td->use_aio) {
- do_sync_io(td);
+ flags |= O_RDWR;
- if (td->verify) {
- if (do_sync_verify(td))
- break;
- }
- } else {
- do_async_io(td);
+ td->fd = open(td->file_name, flags, 0600);
+ }
- if (td->verify) {
- if (do_async_verify(td))
- break;
- }
- }
+ if (td->fd == -1) {
+ td_verror(td, errno);
+ return 1;
}
- td->runtime = mtime_since_now(&td->start);
- ret = 0;
+ if (get_file_size(td))
+ return 1;
- if (td->bw_log)
- finish_log(td, td->bw_log, "bw");
- if (td->lat_log)
- finish_log(td, td->lat_log, "lat");
+ if (!td->use_mmap)
+ return setup_file_plain(td);
+ else
+ return setup_file_mmap(td);
+}
- if (exitall_on_terminate)
- sig_handler(0);
+static int check_dev_match(dev_t dev, char *path)
+{
+ unsigned int major, minor;
+ char line[256], *p;
+ FILE *f;
-err:
- if (td->fd != -1) {
- close(td->fd);
- td->fd = -1;
+ f = fopen(path, "r");
+ if (!f) {
+ perror("open path");
+ return 1;
}
- if (td->use_aio)
- cleanup_aio(td);
- cleanup_io_u(td);
- if (ret) {
- sem_post(&startup_sem);
- sem_wait(&td->mutex);
+
+ p = fgets(line, sizeof(line), f);
+ if (!p) {
+ fclose(f);
+ return 1;
}
- td->runstate = TD_EXITED;
- shmdt(data);
- return NULL;
-}
-static void free_shm(void)
-{
- struct shmid_ds sbuf;
+ if (sscanf(p, "%u:%u", &major, &minor) != 2) {
+ fclose(f);
+ return 1;
+ }
- if (threads) {
- shmdt(threads);
- threads = NULL;
- shmctl(shm_id, IPC_RMID, &sbuf);
+ if (((major << 8) | minor) == dev) {
+ fclose(f);
+ return 0;
}
+
+ fclose(f);
+ return 1;
}
-static int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
- double *mean, double *dev)
+static char *find_block_dir(dev_t dev, char *path)
{
- double n;
+ struct dirent *dir;
+ char *found = NULL;
+ struct stat st;
+ DIR *D;
- if (is->samples == 0)
- return 0;
+ D = opendir(path);
+ if (!D)
+ return NULL;
- *min = is->min_val;
- *max = is->max_val;
+ while ((dir = readdir(D)) != NULL) {
+ char full_path[256];
- n = (double) is->samples;
- *mean = (double) is->val / n;
- *dev = sqrt(((double) is->val_sq - (*mean * *mean) / n) / (n - 1));
- return 1;
-}
+ if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
+ continue;
+ if (!strcmp(dir->d_name, "device"))
+ continue;
-static void show_thread_status(struct thread_data *td)
-{
- int prio, prio_class;
- unsigned long min, max, bw = 0;
- double mean, dev;
+ sprintf(full_path, "%s/%s", path, dir->d_name);
- if (!td->io_kb && !td->error)
- return;
+ if (!strcmp(dir->d_name, "dev")) {
+ if (!check_dev_match(dev, full_path)) {
+ found = path;
+ break;
+ }
+ }
- if (td->runtime)
- bw = td->io_kb * 1024 / td->runtime;
+ if (stat(full_path, &st) == -1) {
+ perror("stat");
+ break;
+ }
- prio = td->ioprio & 0xff;
- prio_class = td->ioprio >> IOPRIO_CLASS_SHIFT;
+ if (!S_ISDIR(st.st_mode) || S_ISLNK(st.st_mode))
+ continue;
- printf("Client%d: err=%2d, io=%6luMiB, bw=%6luKiB/s, runt=%6lumsec\n", td->thread_number, td->error, td->io_kb >> 10, bw, td->runtime);
+ if ((found = find_block_dir(dev, full_path)) != NULL)
+ break;
+ }
- 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);
+ closedir(D);
+ return found;
}
-static int setup_rate(struct thread_data *td)
+static int get_io_ticks(struct disk_util *du, struct disk_util_stat *dus)
{
- int nr_reads_per_sec;
+ unsigned in_flight;
+ char line[256];
+ FILE *f;
+ char *p;
- if (!td->rate)
- return 0;
+ f = fopen(du->path, "r");
+ if (!f)
+ return 1;
- if (td->rate < td->ratemin) {
- fprintf(stderr, "min rate larger than nominal rate\n");
- return -1;
+ p = fgets(line, sizeof(line), f);
+ if (!p) {
+ fclose(f);
+ return 1;
+ }
+
+ if (sscanf(p, "%u %u %llu %u %u %u %llu %u %u %u %u\n", &dus->ios[0], &dus->merges[0], &dus->sectors[0], &dus->ticks[0], &dus->ios[1], &dus->merges[1], &dus->sectors[1], &dus->ticks[1], &in_flight, &dus->io_ticks, &dus->time_in_queue) != 11) {
+ fclose(f);
+ return 1;
}
- nr_reads_per_sec = td->rate * 1024 / td->min_bs;
- td->rate_usec_cycle = 1000000 / nr_reads_per_sec;
- td->rate_pending_usleep = 0;
+ fclose(f);
return 0;
}
-static struct thread_data *get_new_job(int global)
+static void update_io_tick_disk(struct disk_util *du)
{
- struct thread_data *td;
-
- if (global)
- return &def_thread;
- if (thread_number >= max_jobs)
- return NULL;
+ struct disk_util_stat __dus, *dus, *ldus;
+ struct timeval t;
- td = &threads[thread_number++];
- memset(td, 0, sizeof(*td));
+ if (get_io_ticks(du, &__dus))
+ return;
- td->fd = -1;
- td->thread_number = thread_number;
-
- td->ddir = def_thread.ddir;
- td->ioprio = def_thread.ioprio;
- td->sequential = def_thread.sequential;
- td->bs = def_thread.bs;
- td->min_bs = def_thread.min_bs;
- td->max_bs = def_thread.max_bs;
- td->odirect = def_thread.odirect;
- td->thinktime = def_thread.thinktime;
- td->fsync_blocks = def_thread.fsync_blocks;
- td->start_delay = def_thread.start_delay;
- td->timeout = def_thread.timeout;
- td->use_aio = def_thread.use_aio;
- td->create_file = def_thread.create_file;
- td->overwrite = def_thread.overwrite;
- td->invalidate_cache = def_thread.invalidate_cache;
- td->file_size = def_thread.file_size;
- td->file_offset = def_thread.file_offset;
- td->rate = def_thread.rate;
- td->ratemin = def_thread.ratemin;
- td->ratecycle = def_thread.ratecycle;
- td->aio_depth = def_thread.aio_depth;
- td->sync_io = def_thread.sync_io;
- td->mem_type = def_thread.mem_type;
- td->bw_avg_time = def_thread.bw_avg_time;
- td->create_serialize = def_thread.create_serialize;
- td->create_fsync = def_thread.create_fsync;
- td->loops = def_thread.loops;
- td->verify = def_thread.verify;
- memcpy(&td->cpumask, &def_thread.cpumask, sizeof(td->cpumask));
-
- return td;
-}
-
-static void put_job(struct thread_data *td)
-{
- memset(&threads[td->thread_number - 1], 0, sizeof(*td));
- thread_number--;
-}
-
-static int add_job(struct thread_data *td, const char *filename, int prioclass,
- int prio)
-{
- if (td == &def_thread)
- return 0;
+ dus = &du->dus;
+ ldus = &du->last_dus;
- strcpy(td->file_name, filename);
- sem_init(&td->mutex, 1, 0);
- td->ioprio = (prioclass << IOPRIO_CLASS_SHIFT) | prio;
+ dus->sectors[0] += (__dus.sectors[0] - ldus->sectors[0]);
+ dus->sectors[1] += (__dus.sectors[1] - ldus->sectors[1]);
+ dus->ios[0] += (__dus.ios[0] - ldus->ios[0]);
+ dus->ios[1] += (__dus.ios[1] - ldus->ios[1]);
+ dus->merges[0] += (__dus.merges[0] - ldus->merges[0]);
+ dus->merges[1] += (__dus.merges[1] - ldus->merges[1]);
+ dus->ticks[0] += (__dus.ticks[0] - ldus->ticks[0]);
+ dus->ticks[1] += (__dus.ticks[1] - ldus->ticks[1]);
+ dus->io_ticks += (__dus.io_ticks - ldus->io_ticks);
+ dus->time_in_queue += (__dus.time_in_queue - ldus->time_in_queue);
- td->clat_stat.min_val = ULONG_MAX;
- td->slat_stat.min_val = ULONG_MAX;
- td->bw_stat.min_val = ULONG_MAX;
+ gettimeofday(&t, NULL);
+ du->msec += mtime_since(&du->time, &t);
+ memcpy(&du->time, &t, sizeof(t));
+ memcpy(ldus, &__dus, sizeof(__dus));
+}
- run_str[td->thread_number - 1] = 'P';
+static void update_io_ticks(void)
+{
+ struct list_head *entry;
+ struct disk_util *du;
- if (td->use_aio && !td->aio_depth)
- td->aio_depth = 1;
+ list_for_each(entry, &disk_list) {
+ du = list_entry(entry, struct disk_util, list);
+ update_io_tick_disk(du);
+ }
+}
- if (td->min_bs == -1)
- td->min_bs = td->bs;
- if (td->max_bs == -1)
- td->max_bs = td->bs;
+static int disk_util_exists(dev_t dev)
+{
+ struct list_head *entry;
+ struct disk_util *du;
- if (setup_rate(td))
- return -1;
+ list_for_each(entry, &disk_list) {
+ du = list_entry(entry, struct disk_util, list);
- if (write_lat_log)
- setup_log(&td->lat_log);
- if (write_bw_log)
- setup_log(&td->bw_log);
+ if (du->dev == dev)
+ return 1;
+ }
- printf("Client%d: file=%s, rw=%d, prio=%d/%d, seq=%d, odir=%d, bs=%d-%d, rate=%d, aio=%d, aio_depth=%d\n", td->thread_number, filename, td->ddir, prioclass, prio, td->sequential, td->odirect, td->min_bs, td->max_bs, td->rate, td->use_aio, td->aio_depth);
return 0;
}
-static void fill_cpu_mask(cpu_set_t cpumask, int cpu)
+static void disk_util_add(dev_t dev, char *path)
{
- unsigned int i;
+ struct disk_util *du = malloc(sizeof(*du));
- CPU_ZERO(&cpumask);
+ memset(du, 0, sizeof(*du));
+ INIT_LIST_HEAD(&du->list);
+ sprintf(du->path, "%s/stat", path);
+ du->name = strdup(basename(path));
+ du->dev = dev;
- for (i = 0; i < sizeof(int) * 8; i++) {
- if ((1 << i) & cpu)
- CPU_SET(i, &cpumask);
- }
+ gettimeofday(&du->time, NULL);
+ get_io_ticks(du, &du->last_dus);
+
+ list_add_tail(&du->list, &disk_list);
}
-unsigned long get_mult(char c)
+static void init_disk_util(struct thread_data *td)
{
- 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;
+ struct stat st;
+ char foo[256], tmp[256];
+ dev_t dev;
+ char *p, *dir;
+
+ if (!td->do_disk_util)
+ return;
+
+ if (!stat(td->file_name, &st)) {
+ if (S_ISBLK(st.st_mode))
+ dev = st.st_rdev;
+ else
+ dev = st.st_dev;
+ } else {
+ /*
+ * must be a file, open "." in that path
+ */
+ strcpy(foo, td->file_name);
+ p = dirname(foo);
+ if (stat(p, &st)) {
+ perror("disk util stat");
+ return;
+ }
+
+ dev = st.st_dev;
}
-}
-/*
- * convert string after '=' into decimal value, noting any size suffix
- */
-static int str_cnv(char *p, unsigned long long *val)
-{
- char *str;
- int len;
+ if (disk_util_exists(dev))
+ return;
+
+ sprintf(foo, "/sys/block");
+ dir = find_block_dir(dev, foo);
+ if (!dir)
+ return;
- str = strstr(p, "=");
- if (!str)
- return 1;
+ /*
+ * if this is inside a partition dir, jump back to parent
+ */
+ sprintf(tmp, "%s/queue", dir);
+ if (stat(tmp, &st)) {
+ p = dirname(dir);
+ sprintf(tmp, "%s/queue", p);
+ if (stat(tmp, &st)) {
+ fprintf(stderr, "unknown sysfs layout\n");
+ return;
+ }
+ sprintf(foo, "%s", p);
+ }
- str++;
- len = strlen(str);
+ disk_util_add(dev, foo);
+}
- *val = strtoul(str, NULL, 10);
- if (*val == ULONG_MAX && errno == ERANGE)
- return 1;
+static void disk_util_timer_arm(void)
+{
+ struct itimerval itimer;
- *val *= get_mult(str[len - 2]);
- return 0;
+ itimer.it_value.tv_sec = 0;
+ itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000;
+ setitimer(ITIMER_REAL, &itimer, NULL);
}
-static int check_strcnv(char *p, char *name, unsigned long long *val)
+static void clear_io_state(struct thread_data *td)
{
- if (!strstr(p, name))
- return 1;
+ if (td->io_engine == FIO_SYNCIO)
+ lseek(td->fd, SEEK_SET, 0);
+
+ td->cur_off = 0;
+ td->last_bytes = 0;
+ td->stat_io_bytes = 0;
+ td->this_io_bytes = 0;
- return str_cnv(p, val);
+ if (td->file_map)
+ memset(td->file_map, 0, td->num_maps * sizeof(long));
}
-static int check_str(char *p, char *name, char *option)
+static void update_rusage_stat(struct thread_data *td)
{
- char *s = strstr(p, name);
+ if (!td->runtime)
+ return;
- if (!s)
- return 1;
+ getrusage(RUSAGE_SELF, &td->ru_end);
- s += strlen(name);
- if (strstr(s, option))
- return 0;
+ td->usr_time += mtime_since(&td->ru_start.ru_utime, &td->ru_end.ru_utime);
+ td->sys_time += mtime_since(&td->ru_start.ru_stime, &td->ru_end.ru_stime);
+ td->ctx += td->ru_end.ru_nvcsw + td->ru_end.ru_nivcsw - (td->ru_start.ru_nvcsw + td->ru_start.ru_nivcsw);
- return 1;
+
+ memcpy(&td->ru_start, &td->ru_end, sizeof(td->ru_end));
}
-static int check_range(char *p, char *name, unsigned long *s, unsigned long *e)
+static void *thread_main(void *data)
{
- char str[128];
- char s1, s2;
+ struct thread_data *td = data;
+ int ret = 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;
+ setsid();
+ td->pid = getpid();
+
+ if (init_io_u(td))
+ goto err;
+
+ if (fio_setaffinity(td) == -1) {
+ td_verror(td, errno);
+ goto err;
}
- 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 ((td->io_engine != FIO_SYNCIO) && init_aio(td))
+ goto err;
+
+ if (td->ioprio) {
+ if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
+ td_verror(td, errno);
+ goto err;
+ }
}
- sprintf(str, "%s=%%lu-%%lu", name);
- if (sscanf(p, str, s, e) == 2)
- return 0;
+ sem_post(&startup_sem);
+ sem_wait(&td->mutex);
- sprintf(str, "%s = %%lu-%%lu", name);
- if (sscanf(p, str, s, e) == 2)
- return 0;
+ if (!td->create_serialize && setup_file(td))
+ goto err;
- return 1;
+ if (init_random_state(td))
+ goto err;
-}
+ gettimeofday(&td->epoch, NULL);
-static int check_int(char *p, char *name, unsigned int *val)
-{
- char str[128];
+ while (td->loops--) {
+ getrusage(RUSAGE_SELF, &td->ru_start);
+ gettimeofday(&td->start, NULL);
+ memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
- sprintf(str, "%s=%%d", name);
- if (sscanf(p, str, val) == 1)
- return 0;
+ if (td->ratemin)
+ memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
- sprintf(str, "%s = %%d", name);
- if (sscanf(p, str, val) == 1)
- return 0;
+ clear_io_state(td);
+ prune_io_piece_log(td);
- return 1;
-}
+ if (td->io_engine == FIO_SYNCIO)
+ do_sync_io(td);
+ else
+ do_async_io(td);
-static int is_empty_or_comment(char *line)
-{
- unsigned int i;
+ if (td->error || td->terminate)
+ break;
- for (i = 0; i < strlen(line); i++) {
- if (line[i] == ';')
- return 1;
- if (!isspace(line[i]) && !iscntrl(line[i]))
- return 0;
+ td->runtime += mtime_since_now(&td->start);
+ update_rusage_stat(td);
+
+ if (td->verify == VERIFY_NONE)
+ continue;
+
+ clear_io_state(td);
+
+ if (td->io_engine == FIO_SYNCIO)
+ do_sync_verify(td);
+ else
+ do_async_verify(td);
+
+ if (td->error || td->terminate)
+ break;
}
- return 1;
+ ret = 0;
+
+ 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)
+ terminate_threads(td->groupid);
+
+err:
+ if (td->fd != -1) {
+ close(td->fd);
+ td->fd = -1;
+ }
+ if (td->mmap)
+ munmap(td->mmap, td->file_size);
+ if (td->io_engine != FIO_SYNCIO)
+ cleanup_aio(td);
+ cleanup_io_u(td);
+ if (ret) {
+ sem_post(&startup_sem);
+ sem_wait(&td->mutex);
+ }
+ td_set_runstate(td, TD_EXITED);
+ return NULL;
+
}
-static int parse_jobs_ini(char *file)
+static void *fork_main(int shmid, int offset)
{
- unsigned int prioclass, prio, cpu, global;
- unsigned long long ull;
- unsigned long ul1, ul2;
struct thread_data *td;
- char *string, *name;
- fpos_t off;
- FILE *f;
- char *p;
+ void *data;
- f = fopen(file, "r");
- if (!f) {
- perror("fopen");
- return 1;
+ data = shmat(shmid, NULL, 0);
+ if (data == (void *) -1) {
+ perror("shmat");
+ return NULL;
}
- string = malloc(4096);
- name = malloc(256);
+ td = data + offset * sizeof(struct thread_data);
+ thread_main(td);
+ shmdt(data);
+ return NULL;
+}
- while ((p = fgets(string, 4096, f)) != NULL) {
- if (is_empty_or_comment(p))
- continue;
- if (sscanf(p, "[%s]", name) != 1)
- continue;
+static int calc_lat(struct io_stat *is, unsigned long *min, unsigned long *max,
+ double *mean, double *dev)
+{
+ double n;
- global = !strncmp(name, "global", 6);
+ if (is->samples == 0)
+ return 0;
- name[strlen(name) - 1] = '\0';
+ *min = is->min_val;
+ *max = is->max_val;
- td = get_new_job(global);
- if (!td)
- break;
+ n = (double) is->samples;
+ *mean = (double) is->val / n;
+ *dev = sqrt(((double) is->val_sq - (*mean * *mean) / n) / (n - 1));
+ return 1;
+}
- prioclass = 2;
- prio = 4;
+static void show_thread_status(struct thread_data *td,
+ struct group_run_stats *rs)
+{
+ int prio, prio_class;
+ unsigned long min, max, bw = 0;
+ double mean, dev, usr_cpu, sys_cpu;
- fgetpos(f, &off);
- while ((p = fgets(string, 4096, f)) != NULL) {
- if (is_empty_or_comment(p))
- continue;
- if (strstr(p, "["))
- break;
- if (!check_int(p, "rw", &td->ddir)) {
- fgetpos(f, &off);
- continue;
- }
- if (!check_int(p, "prio", &prio)) {
- fgetpos(f, &off);
- continue;
- }
- if (!check_int(p, "prioclass", &prioclass)) {
- fgetpos(f, &off);
- continue;
- }
- if (!check_int(p, "direct", &td->odirect)) {
- fgetpos(f, &off);
- continue;
- }
- if (!check_int(p, "rate", &td->rate)) {
- fgetpos(f, &off);
- continue;
- }
- if (!check_int(p, "ratemin", &td->ratemin)) {
- fgetpos(f, &off);
- continue;
- }
- if (!check_int(p, "ratecycle", &td->ratecycle)) {
- fgetpos(f, &off);
- continue;
- }
- if (!check_int(p, "thinktime", &td->thinktime)) {
- fgetpos(f, &off);
- continue;
- }
- if (!check_int(p, "cpumask", &cpu)) {
- fill_cpu_mask(td->cpumask, cpu);
- fgetpos(f, &off);
- continue;
- }
- if (!check_int(p, "fsync", &td->fsync_blocks)) {
- fgetpos(f, &off);
- continue;
- }
- 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_range(p, "bsrange", &ul1, &ul2)) {
- td->min_bs = ul1;
- td->max_bs = ul2;
- fgetpos(f, &off);
- continue;
- }
- if (!check_strcnv(p, "bs", &ull)) {
- 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_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 (!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;
- }
- printf("Client%d: bad option %s\n",td->thread_number,p);
- }
- fsetpos(f, &off);
+ if (!td->io_bytes && !td->error)
+ return;
- if (add_job(td, name, prioclass, prio))
- put_job(td);
- }
+ if (td->runtime)
+ bw = td->io_bytes / td->runtime;
- free(string);
- free(name);
- fclose(f);
- return 0;
-}
+ prio = td->ioprio & 0xff;
+ prio_class = td->ioprio >> IOPRIO_CLASS_SHIFT;
-static int parse_options(int argc, char *argv[])
-{
- int i;
+ printf("Client%d (g=%d): err=%2d, io=%6luMiB, bw=%6luKiB/s, runt=%6lumsec\n", td->thread_number, td->groupid, td->error, td->io_bytes >> 20, bw, td->runtime);
- for (i = 1; i < argc; i++) {
- char *parm = argv[i];
+ 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)) {
+ double p_of_agg;
- if (parm[0] != '-')
- break;
+ p_of_agg = mean * 100 / (double) rs->agg[td->ddir];
+ printf(" bw (KiB/s) : min=%5lu, max=%5lu, per=%3.2f%%, avg=%5.02f, dev=%5.02f\n", min, max, p_of_agg, mean, dev);
+ }
- parm++;
- switch (*parm) {
- case 's':
- parm++;
- def_thread.sequential = !!atoi(parm);
- break;
- case 'b':
- parm++;
- def_thread.bs = atoi(parm);
- def_thread.bs <<= 10;
- if (!def_thread.bs) {
- printf("bad block size\n");
- def_thread.bs = DEF_BS;
- }
- break;
- case 't':
- 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++;
- def_thread.odirect = !!atoi(parm);
- break;
- case 'f':
- if (i + 1 >= argc) {
- printf("-f needs file as arg\n");
- 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]);
- break;
- }
+ if (td->runtime) {
+ usr_cpu = (double) td->usr_time * 100 / (double) td->runtime;
+ sys_cpu = (double) td->sys_time * 100 / (double) td->runtime;
+ } else {
+ usr_cpu = 0;
+ sys_cpu = 0;
}
- return i;
+ printf(" cpu : usr=%3.2f%%, sys=%3.2f%%, ctx=%lu\n", usr_cpu, sys_cpu, td->ctx);
}
-static void print_thread_status(struct thread_data *td, int nr_running,
- int t_rate, int m_rate)
+static void print_thread_status(int nr_running, int t_rate, int m_rate)
{
printf("Threads now running: %d", nr_running);
if (m_rate || t_rate)
fflush(stdout);
}
+static void check_str_update(struct thread_data *td, int n, int t, int m)
+{
+ char c = run_str[td->thread_number - 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);
+ }
+
+ run_str[td->thread_number - 1] = c;
+ print_thread_status(n, t, m);
+ td->old_runstate = td->runstate;
+}
+
static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
{
int i;
for (i = 0; i < thread_number; i++) {
struct thread_data *td = &threads[i];
+ check_str_update(td, *nr_running, *t_rate, *m_rate);
+
if (td->runstate != TD_EXITED)
continue;
- td->runstate = TD_REAPED;
- run_str[td->thread_number - 1] = '_';
- waitpid(td->pid, NULL, 0);
+ td_set_runstate(td, TD_REAPED);
+
+ if (td->use_thread) {
+ long ret;
+
+ if (pthread_join(td->thread, (void *) &ret))
+ perror("thread_join");
+ } else
+ waitpid(td->pid, NULL, 0);
+
(*nr_running)--;
(*m_rate) -= td->ratemin;
(*t_rate) -= td->rate;
-
- if (td->terminate)
- continue;
-
- print_thread_status(td, *nr_running, *t_rate, *m_rate);
+ check_str_update(td, *nr_running, *t_rate, *m_rate);
}
}
-static void run_threads(char *argv[])
+static void run_threads(void)
{
struct timeval genesis;
struct thread_data *td;
unsigned long spent;
- int i, todo, nr_running, m_rate, t_rate;
+ int i, todo, nr_running, m_rate, t_rate, nr_started;
printf("Starting %d threads\n", thread_number);
fflush(stdout);
signal(SIGINT, sig_handler);
+ signal(SIGALRM, sig_handler);
todo = thread_number;
nr_running = 0;
+ nr_started = 0;
m_rate = t_rate = 0;
for (i = 0; i < thread_number; i++) {
td = &threads[i];
+ init_disk_util(td);
+
if (!td->create_serialize)
continue;
* client data interspersed on disk
*/
if (setup_file(td)) {
- td->runstate = TD_REAPED;
+ td_set_runstate(td, TD_REAPED);
todo--;
}
}
continue;
}
- td->runstate = TD_CREATED;
- run_str[td->thread_number - 1] = 'C';
+ 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);
+ if (td->use_thread) {
+ if (pthread_create(&td->thread, NULL, thread_main, td)) {
+ perror("thread_create");
+ nr_started--;
+ }
+ } else {
+ if (fork())
+ sem_wait(&startup_sem);
+ else {
+ fork_main(shm_id, i);
+ exit(0);
+ }
}
}
/*
- * start created threads (TD_CREATED -> TD_STARTED)
+ * start created threads (TD_CREATED -> TD_RUNNING)
*/
for (i = 0; i < thread_number; i++) {
- struct thread_data *td = &threads[i];
+ td = &threads[i];
if (td->runstate != TD_CREATED)
continue;
- td->runstate = TD_STARTED;
- run_str[td->thread_number - 1] = '+';
+ 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++) {
+ td = &threads[i];
- print_thread_status(td, nr_running, t_rate, m_rate);
+ if (td->runstate != TD_RUNNING &&
+ td->runstate != TD_VERIFYING)
+ continue;
+
+ check_str_update(td, nr_running, t_rate, m_rate);
}
reap_threads(&nr_running, &t_rate, &m_rate);
reap_threads(&nr_running, &t_rate, &m_rate);
usleep(10000);
}
+
+ update_io_ticks();
}
-int setup_thread_area(void)
+static void show_group_stats(struct group_run_stats *rs, int id)
{
- /*
- * 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);
+ printf("\nRun status group %d:\n", id);
- shm_id = shmget(0, s, IPC_CREAT | 0600);
- if (shm_id != -1)
- break;
- if (errno != EINVAL) {
- perror("shmget");
- break;
- }
+ if (rs->max_run[DDIR_READ])
+ printf(" READ: io=%luMiB, aggrb=%lu, minb=%lu, maxb=%lu, mint=%lumsec, maxt=%lumsec\n", rs->io_mb[0], rs->agg[0], rs->min_bw[0], rs->max_bw[0], rs->min_run[0], rs->max_run[0]);
+ if (rs->max_run[DDIR_WRITE])
+ printf(" WRITE: io=%luMiB, aggrb=%lu, minb=%lu, maxb=%lu, mint=%lumsec, maxt=%lumsec\n", rs->io_mb[1], rs->agg[1], rs->min_bw[1], rs->max_bw[1], rs->min_run[1], rs->max_run[1]);
+}
- max_jobs >>= 1;
- } while (max_jobs);
+static void show_disk_util(void)
+{
+ struct disk_util_stat *dus;
+ struct list_head *entry;
+ struct disk_util *du;
+ double util;
- if (shm_id == -1)
- return 1;
+ printf("\nDisk stats (read/write):\n");
- threads = shmat(shm_id, NULL, 0);
- if (threads == (void *) -1) {
- perror("shmat");
- return 1;
- }
+ list_for_each(entry, &disk_list) {
+ du = list_entry(entry, struct disk_util, list);
+ dus = &du->dus;
- atexit(free_shm);
- return 0;
+ util = (double) 100 * du->dus.io_ticks / (double) du->msec;
+ if (util > 100.0)
+ util = 100.0;
+
+ printf(" %s: ios=%u/%u, merge=%u/%u, ticks=%u/%u, in_queue=%u, util=%3.2f%%\n", du->name, dus->ios[0], dus->ios[1], dus->merges[0], dus->merges[1], dus->ticks[0], dus->ticks[1], dus->time_in_queue, util);
+ }
}
-int main(int argc, char *argv[])
+static void show_run_stats(void)
{
- 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];
+ struct group_run_stats *runstats, *rs;
+ struct thread_data *td;
int i;
- if (setup_thread_area())
- return 1;
+ runstats = malloc(sizeof(struct group_run_stats) * (groupid + 1));
- if (sched_getaffinity(getpid(), sizeof(cpu_set_t), &def_thread.cpumask) == -1) {
- perror("sched_getaffinity");
- return 1;
- }
+ for (i = 0; i < groupid + 1; i++) {
+ rs = &runstats[i];
- /*
- * 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;
-
- i = parse_options(argc, argv);
-
- if (!ini_file) {
- printf("Need job file\n");
- return 1;
+ memset(rs, 0, sizeof(*rs));
+ rs->min_bw[0] = rs->min_run[0] = ~0UL;
+ rs->min_bw[1] = rs->min_run[1] = ~0UL;
}
- if (parse_jobs_ini(ini_file))
- return 1;
+ for (i = 0; i < thread_number; i++) {
+ unsigned long bw = 0;
- if (!thread_number) {
- printf("Nothing to do\n");
- return 1;
+ td = &threads[i];
+
+ if (td->error) {
+ printf("Client%d: %s\n", td->thread_number, td->verror);
+ continue;
+ }
+
+ rs = &runstats[td->groupid];
+
+ if (td->runtime < rs->min_run[td->ddir])
+ rs->min_run[td->ddir] = td->runtime;
+ if (td->runtime > rs->max_run[td->ddir])
+ rs->max_run[td->ddir] = td->runtime;
+
+ if (td->runtime)
+ bw = td->io_bytes / td->runtime;
+ if (bw < rs->min_bw[td->ddir])
+ rs->min_bw[td->ddir] = bw;
+ if (bw > rs->max_bw[td->ddir])
+ rs->max_bw[td->ddir] = bw;
+
+ rs->io_mb[td->ddir] += td->io_bytes >> 20;
}
- run_threads(argv);
+ for (i = 0; i < groupid + 1; i++) {
+ rs = &runstats[i];
+
+ if (rs->max_run[0])
+ rs->agg[0] = (rs->io_mb[0]*1024*1000) / rs->max_run[0];
+ if (rs->max_run[1])
+ rs->agg[1] = (rs->io_mb[1]*1024*1000) / rs->max_run[1];
+ }
- min_bw[0] = min_run[0] = ~0UL;
- min_bw[1] = min_run[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;
+ td = &threads[i];
+ rs = &runstats[td->groupid];
- 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;
+ show_thread_status(td, rs);
+ }
- if (td->runtime)
- bw = td->io_kb * 1024 / td->runtime;
- if (bw < min_bw[td->ddir])
- min_bw[td->ddir] = bw;
- if (bw > max_bw[td->ddir])
- max_bw[td->ddir] = bw;
+ for (i = 0; i < groupid + 1; i++)
+ show_group_stats(&runstats[i], i);
- io_mb[td->ddir] += td->io_kb >> 10;
- }
+ show_disk_util();
+}
- show_thread_status(td);
+int main(int argc, char *argv[])
+{
+ memset(run_str, 0, sizeof(run_str));
+
+ if (parse_options(argc, argv))
+ return 1;
+
+ if (!thread_number) {
+ printf("Nothing to do\n");
+ return 1;
}
-
- 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("\nRun status:\n");
- if (max_run[DDIR_READ])
- 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, 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]);
+
+ disk_util_timer_arm();
+
+ run_threads();
+ show_run_stats();
return 0;
}