backend: fix potential divide-by-zero
[fio.git] / backend.c
index bf9d066e012abc8a44a938fcbf319a68a81d6d64..c1c8f96774850a755681dffa710d0ee1550fe16b 100644 (file)
--- a/backend.c
+++ b/backend.c
 #include "server.h"
 #include "lib/getrusage.h"
 #include "idletime.h"
+#include "err.h"
 
 static pthread_t disk_util_thread;
 static struct fio_mutex *disk_thread_mutex;
 static struct fio_mutex *startup_mutex;
-static struct fio_mutex *writeout_mutex;
 static struct flist_head *cgroup_list;
 static char *cgroup_mnt;
 static int exit_value;
@@ -182,7 +182,11 @@ static int __check_min_rate(struct thread_data *td, struct timeval *now,
                                                                ratemin);
                                return 1;
                        } else {
-                               rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
+                               if (spent)
+                                       rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
+                               else
+                                       rate = 0;
+
                                if (rate < ratemin ||
                                    bytes < td->rate_bytes[ddir]) {
                                        log_err("%s: min rate %u not met, got"
@@ -345,7 +349,7 @@ static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
                return 0;
        if (!td->o.timeout)
                return 0;
-       if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000)
+       if (utime_since(&td->epoch, t) >= td->o.timeout)
                return 1;
 
        return 0;
@@ -478,6 +482,12 @@ static void do_verify(struct thread_data *td, uint64_t verify_bytes)
                                break;
 
                        while ((io_u = get_io_u(td)) != NULL) {
+                               if (IS_ERR(io_u)) {
+                                       io_u = NULL;
+                                       ret = FIO_Q_BUSY;
+                                       goto reap;
+                               }
+
                                /*
                                 * We are only interested in the places where
                                 * we wrote or trimmed IOs. Turn those into
@@ -574,6 +584,7 @@ sync_done:
                 * completed io_u's first. Note that we can get BUSY even
                 * without IO queued, if the system is resource starved.
                 */
+reap:
                full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
                if (full || !td->o.iodepth_batch_complete) {
                        min_events = min(td->o.iodepth_batch_complete,
@@ -615,6 +626,19 @@ sync_done:
        dprint(FD_VERIFY, "exiting loop\n");
 }
 
+static unsigned int exceeds_number_ios(struct thread_data *td)
+{
+       unsigned long long number_ios;
+
+       if (!td->o.number_ios)
+               return 0;
+
+       number_ios = ddir_rw_sum(td->this_io_blocks);
+       number_ios += td->io_u_queued + td->io_u_in_flight;
+
+       return number_ios >= td->o.number_ios;
+}
+
 static int io_bytes_exceeded(struct thread_data *td)
 {
        unsigned long long bytes;
@@ -628,7 +652,7 @@ static int io_bytes_exceeded(struct thread_data *td)
        else
                bytes = td->this_io_bytes[DDIR_TRIM];
 
-       return bytes >= td->o.size;
+       return bytes >= td->o.size || exceeds_number_ios(td);
 }
 
 /*
@@ -642,7 +666,7 @@ static uint64_t do_io(struct thread_data *td)
        uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
        unsigned int i;
        int ret = 0;
-       uint64_t bytes_issued = 0;
+       uint64_t total_bytes, bytes_issued = 0;
 
        if (in_ramp_time(td))
                td_set_runstate(td, TD_RAMP);
@@ -651,6 +675,16 @@ static uint64_t do_io(struct thread_data *td)
 
        lat_target_init(td);
 
+       /*
+        * If verify_backlog is enabled, we'll run the verify in this
+        * handler as well. For that case, we may need up to twice the
+        * amount of bytes.
+        */
+       total_bytes = td->o.size;
+       if (td->o.verify != VERIFY_NONE &&
+          (td_write(td) && td->o.verify_backlog))
+               total_bytes += td->o.size;
+
        while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
                (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td) ||
                td->o.time_based) {
@@ -678,11 +712,18 @@ static uint64_t do_io(struct thread_data *td)
                if (flow_threshold_exceeded(td))
                        continue;
 
-               if (bytes_issued >= (uint64_t) td->o.size)
+               if (bytes_issued >= total_bytes)
                        break;
 
                io_u = get_io_u(td);
-               if (!io_u) {
+               if (IS_ERR_OR_NULL(io_u)) {
+                       int err = PTR_ERR(io_u);
+
+                       io_u = NULL;
+                       if (err == -EBUSY) {
+                               ret = FIO_Q_BUSY;
+                               goto reap;
+                       }
                        if (td->o.latency_target)
                                goto reap;
                        break;
@@ -714,6 +755,11 @@ static uint64_t do_io(struct thread_data *td)
                else
                        td_set_runstate(td, TD_RUNNING);
 
+               /*
+                * Always log IO before it's issued, so we know the specific
+                * order of it. The logged unit will track when the IO has
+                * completed.
+                */
                if (td_write(td) && io_u->ddir == DDIR_WRITE &&
                    td->o.do_verify &&
                    td->o.verify != VERIFY_NONE &&
@@ -1068,7 +1114,8 @@ static int switch_ioscheduler(struct thread_data *td)
        /*
         * Read back and check that the selected scheduler is now the default.
         */
-       ret = fread(tmp, 1, sizeof(tmp), f);
+       memset(tmp, 0, sizeof(tmp));
+       ret = fread(tmp, 1, sizeof(tmp) - 1, f);
        if (ferror(f) || ret < 0) {
                td_verror(td, errno, "fread");
                fclose(f);
@@ -1097,6 +1144,8 @@ static int keep_running(struct thread_data *td)
                td->o.loops--;
                return 1;
        }
+       if (exceeds_number_ios(td))
+               return 0;
 
        if (td->o.size != -1ULL && ddir_rw_sum(td->io_bytes) < td->o.size) {
                uint64_t diff;
@@ -1109,6 +1158,9 @@ static int keep_running(struct thread_data *td)
                if (diff < td_max_bs(td))
                        return 0;
 
+               if (fio_files_done(td))
+                       return 0;
+
                return 1;
        }
 
@@ -1163,6 +1215,12 @@ static uint64_t do_dry_run(struct thread_data *td)
                        td->ts.total_io_u[io_u->ddir]++;
                }
 
+               if (td_write(td) && io_u->ddir == DDIR_WRITE &&
+                   td->o.do_verify &&
+                   td->o.verify != VERIFY_NONE &&
+                   !td->o.experimental_verify)
+                       log_io_piece(td, io_u);
+
                ret = io_u_sync_complete(td, io_u, bytes_done);
                (void) ret;
        }
@@ -1220,13 +1278,6 @@ static void *thread_main(void *data)
        fio_mutex_down(td->mutex);
        dprint(FD_MUTEX, "done waiting on td->mutex\n");
 
-       /*
-        * the ->mutex mutex is now no longer used, close it to avoid
-        * eating a file descriptor
-        */
-       fio_mutex_remove(td->mutex);
-       td->mutex = NULL;
-
        /*
         * A new gid requires privilege, so we need to do this before setting
         * the uid.
@@ -1252,6 +1303,15 @@ static void *thread_main(void *data)
         * allocations.
         */
        if (o->cpumask_set) {
+               if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) {
+                       ret = fio_cpus_split(&o->cpumask, td->thread_number - 1);
+                       if (!ret) {
+                               log_err("fio: no CPUs set\n");
+                               log_err("fio: Try increasing number of available CPUs\n");
+                               td_verror(td, EINVAL, "cpus_split");
+                               goto err;
+                       }
+               }
                ret = fio_setaffinity(td->pid, o->cpumask);
                if (ret == -1) {
                        td_verror(td, errno, "cpu_set_affinity");
@@ -1434,44 +1494,8 @@ static void *thread_main(void *data)
 
        fio_unpin_memory(td);
 
-       fio_mutex_down(writeout_mutex);
-       if (td->bw_log) {
-               if (o->bw_log_file) {
-                       finish_log_named(td, td->bw_log,
-                                               o->bw_log_file, "bw");
-               } else
-                       finish_log(td, td->bw_log, "bw");
-       }
-       if (td->lat_log) {
-               if (o->lat_log_file) {
-                       finish_log_named(td, td->lat_log,
-                                               o->lat_log_file, "lat");
-               } else
-                       finish_log(td, td->lat_log, "lat");
-       }
-       if (td->slat_log) {
-               if (o->lat_log_file) {
-                       finish_log_named(td, td->slat_log,
-                                               o->lat_log_file, "slat");
-               } else
-                       finish_log(td, td->slat_log, "slat");
-       }
-       if (td->clat_log) {
-               if (o->lat_log_file) {
-                       finish_log_named(td, td->clat_log,
-                                               o->lat_log_file, "clat");
-               } else
-                       finish_log(td, td->clat_log, "clat");
-       }
-       if (td->iops_log) {
-               if (o->iops_log_file) {
-                       finish_log_named(td, td->iops_log,
-                                               o->iops_log_file, "iops");
-               } else
-                       finish_log(td, td->iops_log, "iops");
-       }
+       fio_writeout_logs(td);
 
-       fio_mutex_up(writeout_mutex);
        if (o->exec_postrun)
                exec_string(o, o->exec_postrun, (const char *)"postrun");
 
@@ -1506,6 +1530,9 @@ err:
        fio_mutex_remove(td->rusage_sem);
        td->rusage_sem = NULL;
 
+       fio_mutex_remove(td->mutex);
+       td->mutex = NULL;
+
        td_set_runstate(td, TD_EXITED);
        return (void *) (uintptr_t) td->error;
 }
@@ -1653,8 +1680,8 @@ static void do_usleep(unsigned int usecs)
 static void run_threads(void)
 {
        struct thread_data *td;
-       unsigned long spent;
        unsigned int i, todo, nr_running, m_rate, t_rate, nr_started;
+       uint64_t spent;
 
        if (fio_gtod_offload && fio_start_gtod_thread())
                return;
@@ -1752,9 +1779,9 @@ static void run_threads(void)
                        }
 
                        if (td->o.start_delay) {
-                               spent = mtime_since_genesis();
+                               spent = utime_since_genesis();
 
-                               if (td->o.start_delay * 1000 > spent)
+                               if (td->o.start_delay > spent)
                                        continue;
                        }
 
@@ -1973,9 +2000,6 @@ int fio_backend(void)
        startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED);
        if (startup_mutex == NULL)
                return 1;
-       writeout_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED);
-       if (writeout_mutex == NULL)
-               return 1;
 
        set_genesis_time();
        stat_init();
@@ -2006,7 +2030,6 @@ int fio_backend(void)
        sfree(cgroup_mnt);
 
        fio_mutex_remove(startup_mutex);
-       fio_mutex_remove(writeout_mutex);
        fio_mutex_remove(disk_thread_mutex);
        stat_exit();
        return exit_value;