Add support for reserved keywords
[fio.git] / fio.c
diff --git a/fio.c b/fio.c
index 62fab49..dfd15e0 100644 (file)
--- a/fio.c
+++ b/fio.c
 #include "fio.h"
 #include "hash.h"
 #include "smalloc.h"
+#include "verify.h"
+#include "diskutil.h"
 
 unsigned long page_mask;
 unsigned long page_size;
-#define ALIGN(buf)     \
+
+#define PAGE_ALIGN(buf)        \
        (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
 
 int groupid = 0;
@@ -170,28 +173,22 @@ static void set_sig_handlers(void)
        sigaction(SIGQUIT, &act, NULL);
 }
 
-static inline int should_check_rate(struct thread_data *td)
-{
-       struct thread_options *o = &td->o;
-
-       /*
-        * If some rate setting was given, we need to check it
-        */
-       if (o->rate || o->ratemin || o->rate_iops || o->rate_iops_min)
-               return 1;
-
-       return 0;
-}
-
 /*
  * Check if we are above the minimum rate given.
  */
-static int check_min_rate(struct thread_data *td, struct timeval *now)
+static int __check_min_rate(struct thread_data *td, struct timeval *now,
+                           enum td_ddir ddir)
 {
        unsigned long long bytes = 0;
        unsigned long iops = 0;
        unsigned long spent;
        unsigned long rate;
+       unsigned int ratemin = 0;
+       unsigned int rate_iops = 0;
+       unsigned int rate_iops_min = 0;
+
+       if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir])
+               return 0;
 
        /*
         * allow a 2 second settle period in the beginning
@@ -199,38 +196,35 @@ static int check_min_rate(struct thread_data *td, struct timeval *now)
        if (mtime_since(&td->start, now) < 2000)
                return 0;
 
-       if (td_read(td)) {
-               iops += td->io_blocks[DDIR_READ];
-               bytes += td->this_io_bytes[DDIR_READ];
-       }
-       if (td_write(td)) {
-               iops += td->io_blocks[DDIR_WRITE];
-               bytes += td->this_io_bytes[DDIR_WRITE];
-       }
+       iops += td->io_blocks[ddir];
+       bytes += td->this_io_bytes[ddir];
+       ratemin += td->o.ratemin[ddir];
+       rate_iops += td->o.rate_iops[ddir];
+       rate_iops_min += td->o.rate_iops_min[ddir];
 
        /*
         * if rate blocks is set, sample is running
         */
-       if (td->rate_bytes || td->rate_blocks) {
-               spent = mtime_since(&td->lastrate, now);
+       if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) {
+               spent = mtime_since(&td->lastrate[ddir], now);
                if (spent < td->o.ratecycle)
                        return 0;
 
-               if (td->o.rate) {
+               if (td->o.rate[ddir]) {
                        /*
                         * check bandwidth specified rate
                         */
-                       if (bytes < td->rate_bytes) {
+                       if (bytes < td->rate_bytes[ddir]) {
                                log_err("%s: min rate %u not met\n", td->o.name,
-                                                               td->o.ratemin);
+                                                               ratemin);
                                return 1;
                        } else {
-                               rate = (bytes - td->rate_bytes) / spent;
-                               if (rate < td->o.ratemin ||
-                                   bytes < td->rate_bytes) {
+                               rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
+                               if (rate < ratemin ||
+                                   bytes < td->rate_bytes[ddir]) {
                                        log_err("%s: min rate %u not met, got"
-                                               " %luKiB/sec\n", td->o.name,
-                                                       td->o.ratemin, rate);
+                                               " %luKB/sec\n", td->o.name,
+                                                       ratemin, rate);
                                        return 1;
                                }
                        }
@@ -238,29 +232,41 @@ static int check_min_rate(struct thread_data *td, struct timeval *now)
                        /*
                         * checks iops specified rate
                         */
-                       if (iops < td->o.rate_iops) {
+                       if (iops < rate_iops) {
                                log_err("%s: min iops rate %u not met\n",
-                                               td->o.name, td->o.rate_iops);
+                                               td->o.name, rate_iops);
                                return 1;
                        } else {
-                               rate = (iops - td->rate_blocks) / spent;
-                               if (rate < td->o.rate_iops_min ||
-                                   iops < td->rate_blocks) {
+                               rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
+                               if (rate < rate_iops_min ||
+                                   iops < td->rate_blocks[ddir]) {
                                        log_err("%s: min iops rate %u not met,"
                                                " got %lu\n", td->o.name,
-                                                       td->o.rate_iops_min,
-                                                       rate);
+                                                       rate_iops_min, rate);
                                }
                        }
                }
        }
 
-       td->rate_bytes = bytes;
-       td->rate_blocks = iops;
-       memcpy(&td->lastrate, now, sizeof(*now));
+       td->rate_bytes[ddir] = bytes;
+       td->rate_blocks[ddir] = iops;
+       memcpy(&td->lastrate[ddir], now, sizeof(*now));
        return 0;
 }
 
+static int check_min_rate(struct thread_data *td, struct timeval *now,
+                         unsigned long *bytes_done)
+{
+       int ret = 0;
+
+       if (bytes_done[0])
+               ret |= __check_min_rate(td, now, 0);
+       if (bytes_done[1])
+               ret |= __check_min_rate(td, now, 1);
+
+       return ret;
+}
+
 static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
 {
        if (!td->o.timeout)
@@ -284,7 +290,7 @@ static void cleanup_pending_aio(struct thread_data *td)
        /*
         * get immediately available events, if any
         */
-       r = io_u_queued_complete(td, 0);
+       r = io_u_queued_complete(td, 0, NULL);
        if (r < 0)
                return;
 
@@ -312,7 +318,7 @@ static void cleanup_pending_aio(struct thread_data *td)
        }
 
        if (td->cur_depth)
-               r = io_u_queued_complete(td, td->cur_depth);
+               r = io_u_queued_complete(td, td->cur_depth, NULL);
 }
 
 /*
@@ -342,7 +348,7 @@ requeue:
                put_io_u(td, io_u);
                return 1;
        } else if (ret == FIO_Q_QUEUED) {
-               if (io_u_queued_complete(td, 1) < 0)
+               if (io_u_queued_complete(td, 1, NULL) < 0)
                        return 1;
        } else if (ret == FIO_Q_COMPLETED) {
                if (io_u->error) {
@@ -350,7 +356,7 @@ requeue:
                        return 1;
                }
 
-               if (io_u_sync_complete(td, io_u) < 0)
+               if (io_u_sync_complete(td, io_u, NULL) < 0)
                        return 1;
        } else if (ret == FIO_Q_BUSY) {
                if (td_io_commit(td))
@@ -367,6 +373,51 @@ static inline void update_tv_cache(struct thread_data *td)
                fio_gettime(&td->tv_cache, NULL);
 }
 
+static int break_on_this_error(struct thread_data *td, int *retptr)
+{
+       int ret = *retptr;
+
+       if (ret < 0 || td->error) {
+               int err;
+
+               if (!td->o.continue_on_error)
+                       return 1;
+
+               if (ret < 0)
+                       err = -ret;
+               else
+                       err = td->error;
+
+               if (td_non_fatal_error(err)) {
+                       /*
+                        * Continue with the I/Os in case of
+                        * a non fatal error.
+                        */
+                       update_error_count(td, err);
+                       td_clear_error(td);
+                       *retptr = 0;
+                       return 0;
+               } else if (td->o.fill_device && err == ENOSPC) {
+                       /*
+                        * We expect to hit this error if
+                        * fill_device option is set.
+                        */
+                       td_clear_error(td);
+                       td->terminate = 1;
+                       return 1;
+               } else {
+                       /*
+                        * Stop the I/O in case of a fatal
+                        * error.
+                        */
+                       update_error_count(td, err);
+                       return 1;
+               }
+       }
+
+       return 0;
+}
+
 /*
  * The main verify engine. Runs over the writes we previously submitted,
  * reads the blocks back in, and checks the crc/md5 of the data.
@@ -383,7 +434,7 @@ static void do_verify(struct thread_data *td)
         * read from disk.
         */
        for_each_file(td, f, i) {
-               if (!(f->flags & FIO_FILE_OPEN))
+               if (!fio_file_open(f))
                        continue;
                if (fio_io_sync(td, f))
                        break;
@@ -400,18 +451,17 @@ static void do_verify(struct thread_data *td)
        while (!td->terminate) {
                int ret2, full;
 
-               io_u = __get_io_u(td);
-               if (!io_u)
-                       break;
-
                update_tv_cache(td);
 
                if (runtime_exceeded(td, &td->tv_cache)) {
-                       put_io_u(td, io_u);
                        td->terminate = 1;
                        break;
                }
 
+               io_u = __get_io_u(td);
+               if (!io_u)
+                       break;
+
                if (get_next_verify(td, io_u)) {
                        put_io_u(td, io_u);
                        break;
@@ -422,14 +472,18 @@ static void do_verify(struct thread_data *td)
                        break;
                }
 
-               io_u->end_io = verify_io_u;
+               if (td->o.verify_async)
+                       io_u->end_io = verify_io_u_async;
+               else
+                       io_u->end_io = verify_io_u;
 
                ret = td_io_queue(td, io_u);
                switch (ret) {
                case FIO_Q_COMPLETED:
-                       if (io_u->error)
+                       if (io_u->error) {
                                ret = -io_u->error;
-                       else if (io_u->resid) {
+                               clear_io_u(td, io_u);
+                       } else if (io_u->resid) {
                                int bytes = io_u->xfer_buflen - io_u->resid;
                                struct fio_file *f = io_u->file;
 
@@ -454,7 +508,7 @@ static void do_verify(struct thread_data *td)
                                requeue_io_u(td, &io_u);
                        } else {
 sync_done:
-                               ret = io_u_sync_complete(td, io_u);
+                               ret = io_u_sync_complete(td, io_u, NULL);
                                if (ret < 0)
                                        break;
                        }
@@ -473,7 +527,7 @@ sync_done:
                        break;
                }
 
-               if (ret < 0 || td->error)
+               if (break_on_this_error(td, &ret))
                        break;
 
                /*
@@ -492,7 +546,7 @@ sync_done:
                                 * and do the verification on them through
                                 * the callback handler
                                 */
-                               if (io_u_queued_complete(td, min_events) < 0) {
+                               if (io_u_queued_complete(td, min_events, NULL) < 0) {
                                        ret = -1;
                                        break;
                                }
@@ -506,7 +560,7 @@ sync_done:
                min_events = td->cur_depth;
 
                if (min_events)
-                       ret = io_u_queued_complete(td, min_events);
+                       ret = io_u_queued_complete(td, min_events, NULL);
        } else
                cleanup_pending_aio(td);
 
@@ -519,7 +573,6 @@ sync_done:
  */
 static void do_io(struct thread_data *td)
 {
-       unsigned long usec;
        unsigned int i;
        int ret = 0;
 
@@ -530,7 +583,7 @@ static void do_io(struct thread_data *td)
 
        while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) {
                struct timeval comp_time;
-               long bytes_done = 0;
+               unsigned long bytes_done[2] = { 0, 0 };
                int min_evts = 0;
                struct io_u *io_u;
                int ret2, full;
@@ -538,24 +591,26 @@ static void do_io(struct thread_data *td)
                if (td->terminate)
                        break;
 
-               io_u = get_io_u(td);
-               if (!io_u)
-                       break;
-
                update_tv_cache(td);
 
                if (runtime_exceeded(td, &td->tv_cache)) {
-                       put_io_u(td, io_u);
                        td->terminate = 1;
                        break;
                }
 
+               io_u = get_io_u(td);
+               if (!io_u)
+                       break;
+
                /*
                 * Add verification end_io handler, if asked to verify
                 * a previously written file.
                 */
                if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ) {
-                       io_u->end_io = verify_io_u;
+                       if (td->o.verify_async)
+                               io_u->end_io = verify_io_u_async;
+                       else
+                               io_u->end_io = verify_io_u;
                        td_set_runstate(td, TD_VERIFYING);
                } else if (in_ramp_time(td))
                        td_set_runstate(td, TD_RAMP);
@@ -565,9 +620,10 @@ static void do_io(struct thread_data *td)
                ret = td_io_queue(td, io_u);
                switch (ret) {
                case FIO_Q_COMPLETED:
-                       if (io_u->error)
+                       if (io_u->error) {
                                ret = -io_u->error;
-                       else if (io_u->resid) {
+                               clear_io_u(td, io_u);
+                       } else if (io_u->resid) {
                                int bytes = io_u->xfer_buflen - io_u->resid;
                                struct fio_file *f = io_u->file;
 
@@ -592,12 +648,13 @@ static void do_io(struct thread_data *td)
                                requeue_io_u(td, &io_u);
                        } else {
 sync_done:
-                               if (should_check_rate(td))
+                               if (__should_check_rate(td, 0) ||
+                                   __should_check_rate(td, 1))
                                        fio_gettime(&comp_time, NULL);
 
-                               bytes_done = io_u_sync_complete(td, io_u);
-                               if (bytes_done < 0)
-                                       ret = bytes_done;
+                               ret = io_u_sync_complete(td, io_u, bytes_done);
+                               if (ret < 0)
+                                       break;
                        }
                        break;
                case FIO_Q_QUEUED:
@@ -621,7 +678,7 @@ sync_done:
                        break;
                }
 
-               if (ret < 0 || td->error)
+               if (break_on_this_error(td, &ret))
                        break;
 
                /*
@@ -633,34 +690,25 @@ sync_done:
                        if (full && !min_evts)
                                min_evts = 1;
 
-                       if (should_check_rate(td))
+                       if (__should_check_rate(td, 0) ||
+                           __should_check_rate(td, 1))
                                fio_gettime(&comp_time, NULL);
 
                        do {
-                               ret = io_u_queued_complete(td, min_evts);
-                               if (ret <= 0)
+                               ret = io_u_queued_complete(td, min_evts, bytes_done);
+                               if (ret < 0)
                                        break;
 
-                               bytes_done += ret;
                        } while (full && (td->cur_depth > td->o.iodepth_low));
                }
 
                if (ret < 0)
                        break;
-               if (!bytes_done)
+               if (!(bytes_done[0] + bytes_done[1]))
                        continue;
 
-               /*
-                * the rate is batched for now, it should work for batches
-                * of completions except the very first one which may look
-                * a little bursty
-                */
-               if (!in_ramp_time(td) && should_check_rate(td)) {
-                       usec = utime_since(&td->tv_cache, &comp_time);
-
-                       rate_throttle(td, usec, bytes_done);
-
-                       if (check_min_rate(td, &comp_time)) {
+               if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
+                       if (check_min_rate(td, &comp_time, bytes_done)) {
                                if (exitall_on_terminate)
                                        terminate_threads(td->groupid);
                                td_verror(td, EIO, "check_min_rate");
@@ -694,13 +742,13 @@ sync_done:
 
                i = td->cur_depth;
                if (i)
-                       ret = io_u_queued_complete(td, i);
+                       ret = io_u_queued_complete(td, i, NULL);
 
                if (should_fsync(td) && td->o.end_fsync) {
                        td_set_runstate(td, TD_FSYNCING);
 
                        for_each_file(td, f, i) {
-                               if (!(f->flags & FIO_FILE_OPEN))
+                               if (!fio_file_open(f))
                                        continue;
                                fio_io_sync(td, f);
                        }
@@ -757,8 +805,8 @@ static int init_io_u(struct thread_data *td)
        if (allocate_io_mem(td))
                return 1;
 
-       if (td->o.odirect)
-               p = ALIGN(td->orig_buffer);
+       if (td->o.odirect || td->o.mem_align)
+               p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align;
        else
                p = td->orig_buffer;
 
@@ -778,9 +826,11 @@ static int init_io_u(struct thread_data *td)
                io_u = ptr;
                memset(io_u, 0, sizeof(*io_u));
                INIT_FLIST_HEAD(&io_u->list);
+               dprint(FD_MEM, "io_u alloc %p, index %u\n", io_u, i);
 
                if (!(td->io_ops->flags & FIO_NOIO)) {
                        io_u->buf = p + max_bs * i;
+                       dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
 
                        if (td_write(td) && !td->o.refill_buffers)
                                io_u_fill_buffer(td, io_u, max_bs);
@@ -876,9 +926,8 @@ static void reset_io_counters(struct thread_data *td)
        td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0;
        td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
        td->zone_bytes = 0;
-       td->rate_bytes = 0;
-       td->rate_blocks = 0;
-       td->rw_end_set[0] = td->rw_end_set[1] = 0;
+       td->rate_bytes[0] = td->rate_bytes[1] = 0;
+       td->rate_blocks[0] = td->rate_blocks[1] = 0;
 
        td->last_was_sync = 0;
 
@@ -922,7 +971,7 @@ static void clear_io_state(struct thread_data *td)
 
        close_files(td);
        for_each_file(td, f, i)
-               f->flags &= ~FIO_FILE_DONE;
+               fio_file_clear_done(f);
 }
 
 static int exec_string(const char *string)
@@ -949,6 +998,7 @@ static void *thread_main(void *data)
 {
        unsigned long long runtime[2], elapsed;
        struct thread_data *td = data;
+       pthread_condattr_t attr;
        int clear_state;
 
        if (!td->o.use_thread)
@@ -963,8 +1013,14 @@ static void *thread_main(void *data)
        INIT_FLIST_HEAD(&td->io_u_requeues);
        INIT_FLIST_HEAD(&td->io_log_list);
        INIT_FLIST_HEAD(&td->io_hist_list);
+       INIT_FLIST_HEAD(&td->verify_list);
+       pthread_mutex_init(&td->io_u_lock, NULL);
        td->io_hist_tree = RB_ROOT;
 
+       pthread_condattr_init(&attr);
+       pthread_cond_init(&td->verify_cond, &attr);
+       pthread_cond_init(&td->free_cond, &attr);
+
        td_set_runstate(td, TD_INITIALIZED);
        dprint(FD_MUTEX, "up startup_mutex\n");
        fio_mutex_up(startup_mutex);
@@ -988,7 +1044,10 @@ static void *thread_main(void *data)
        if (init_io_u(td))
                goto err;
 
-       if (td->o.cpumask_set && fio_setaffinity(td) == -1) {
+       if (td->o.verify_async && verify_async_init(td))
+               goto err;
+
+       if (td->o.cpumask_set && fio_setaffinity(td->pid, td->o.cpumask) == -1) {
                td_verror(td, errno, "cpu_set_affinity");
                goto err;
        }
@@ -999,7 +1058,7 @@ static void *thread_main(void *data)
         */
        if (td->o.gtod_cpu) {
                fio_cpu_clear(&td->o.cpumask, td->o.gtod_cpu);
-               if (fio_setaffinity(td) == -1) {
+               if (fio_setaffinity(td->pid, td->o.cpumask) == -1) {
                        td_verror(td, errno, "cpu_set_affinity");
                        goto err;
                }
@@ -1046,10 +1105,13 @@ static void *thread_main(void *data)
        clear_state = 0;
        while (keep_running(td)) {
                fio_gettime(&td->start, NULL);
-               memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
+               memcpy(&td->ts.stat_sample_time[0], &td->start,
+                               sizeof(td->start));
+               memcpy(&td->ts.stat_sample_time[1], &td->start,
+                               sizeof(td->start));
                memcpy(&td->tv_cache, &td->start, sizeof(td->start));
 
-               if (td->o.ratemin)
+               if (td->o.ratemin[0] || td->o.ratemin[1])
                        memcpy(&td->lastrate, &td->ts.stat_sample_time,
                                                        sizeof(td->lastrate));
 
@@ -1063,21 +1125,11 @@ static void *thread_main(void *data)
                clear_state = 1;
 
                if (td_read(td) && td->io_bytes[DDIR_READ]) {
-                       if (td->rw_end_set[DDIR_READ])
-                               elapsed = utime_since(&td->start,
-                                                     &td->rw_end[DDIR_READ]);
-                       else
-                               elapsed = utime_since_now(&td->start);
-
+                       elapsed = utime_since_now(&td->start);
                        runtime[DDIR_READ] += elapsed;
                }
                if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
-                       if (td->rw_end_set[DDIR_WRITE])
-                               elapsed = utime_since(&td->start,
-                                                     &td->rw_end[DDIR_WRITE]);
-                       else
-                               elapsed = utime_since_now(&td->start);
-
+                       elapsed = utime_since_now(&td->start);
                        runtime[DDIR_WRITE] += elapsed;
                }
 
@@ -1151,6 +1203,9 @@ err:
                td_verror(td, ret, "fio_cpuset_exit");
        }
 
+       if (td->o.verify_async)
+               verify_async_exit(td);
+
        /*
         * do this very late, it will log file closing as well
         */
@@ -1265,8 +1320,8 @@ static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
                continue;
 reaped:
                (*nr_running)--;
-               (*m_rate) -= td->o.ratemin;
-               (*t_rate) -= td->o.rate;
+               (*m_rate) -= (td->o.ratemin[0] + td->o.ratemin[1]);
+               (*t_rate) -= (td->o.rate[0] + td->o.rate[1]);
                if (!td->pid)
                        pending--;
 
@@ -1386,8 +1441,10 @@ static void run_threads(void)
                         * its own files. so close them, if we opened them
                         * for creation
                         */
-                       for_each_file(td, f, i)
-                               td_io_close_file(td, f);
+                       for_each_file(td, f, i) {
+                               if (fio_file_open(f))
+                                       td_io_close_file(td, f);
+                       }
                }
 
                init_disk_util(td);
@@ -1465,7 +1522,13 @@ static void run_threads(void)
                                        *fio_debug_jobp = pid;
                        }
                        dprint(FD_MUTEX, "wait on startup_mutex\n");
-                       fio_mutex_down(startup_mutex);
+                       if (fio_mutex_down_timeout(startup_mutex, 10)) {
+                               log_err("fio: job startup hung? exiting.\n");
+                               terminate_threads(TERMINATE_ALL);
+                               fio_abort = 1;
+                               nr_started--;
+                               break;
+                       }
                        dprint(FD_MUTEX, "done waiting on startup_mutex\n");
                }
 
@@ -1521,8 +1584,8 @@ static void run_threads(void)
                                td_set_runstate(td, TD_RUNNING);
                        nr_running++;
                        nr_started--;
-                       m_rate += td->o.ratemin;
-                       t_rate += td->o.rate;
+                       m_rate += td->o.ratemin[0] + td->o.ratemin[1];
+                       t_rate += td->o.rate[0] + td->o.rate[1];
                        todo--;
                        fio_mutex_up(td->mutex);
                }
@@ -1555,12 +1618,6 @@ int main(int argc, char *argv[])
        if (!getenv("LC_NUMERIC"))
                setlocale(LC_NUMERIC, "en_US");
 
-       if (parse_options(argc, argv))
-               return 1;
-
-       if (!thread_number)
-               return 0;
-
        ps = sysconf(_SC_PAGESIZE);
        if (ps < 0) {
                log_err("Failed to get page size\n");
@@ -1570,6 +1627,14 @@ int main(int argc, char *argv[])
        page_size = ps;
        page_mask = ps - 1;
 
+       fio_keywords_init();
+
+       if (parse_options(argc, argv))
+               return 1;
+
+       if (!thread_number)
+               return 0;
+
        if (write_bw_log) {
                setup_log(&agg_io_log[DDIR_READ]);
                setup_log(&agg_io_log[DDIR_WRITE]);