verify: add new experimental mode that requires no meta data
[fio.git] / backend.c
index ce0a0098299294279ea4bf16dfd88b7079d9466e..507faa9992e7325d4c80de368e43f2a779893739 100644 (file)
--- a/backend.c
+++ b/backend.c
@@ -34,7 +34,9 @@
 #include <sys/stat.h>
 #include <sys/wait.h>
 #include <sys/ipc.h>
+#ifndef FIO_NO_HAVE_SHM_H
 #include <sys/shm.h>
+#endif
 #include <sys/mman.h>
 
 #include "fio.h"
@@ -62,11 +64,13 @@ struct io_log *agg_io_log[DDIR_RWDIR_CNT];
 
 int groupid = 0;
 unsigned int thread_number = 0;
+unsigned int stat_number = 0;
 unsigned int nr_process = 0;
 unsigned int nr_thread = 0;
 int shm_id = 0;
 int temp_stall_ts;
 unsigned long done_secs = 0;
+volatile int disk_util_exit = 0;
 
 #define PAGE_ALIGN(buf)        \
        (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask)
@@ -107,6 +111,14 @@ static void set_sig_handlers(void)
        act.sa_flags = SA_RESTART;
        sigaction(SIGTERM, &act, NULL);
 
+/* Windows uses SIGBREAK as a quit signal from other applications */
+#ifdef WIN32
+       memset(&act, 0, sizeof(act));
+       act.sa_handler = sig_int;
+       act.sa_flags = SA_RESTART;
+       sigaction(SIGBREAK, &act, NULL);
+#endif
+
        memset(&act, 0, sizeof(act));
        act.sa_handler = sig_show_status;
        act.sa_flags = SA_RESTART;
@@ -337,17 +349,17 @@ static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
        int ret = *retptr;
 
        if (ret < 0 || td->error) {
-               int err;
+               int err = td->error;
+               enum error_type_bit eb;
 
                if (ret < 0)
                        err = -ret;
-               else
-                       err = td->error;
 
-               if (!(td->o.continue_on_error & td_error_type(ddir, err)))
+               eb = td_error_type(ddir, err);
+               if (!(td->o.continue_on_error & (1 << eb)))
                        return 1;
 
-               if (td_non_fatal_error(err)) {
+               if (td_non_fatal_error(td, eb, err)) {
                        /*
                         * Continue with the I/Os in case of
                         * a non fatal error.
@@ -381,11 +393,12 @@ static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
  * The main verify engine. Runs over the writes we previously submitted,
  * reads the blocks back in, and checks the crc/md5 of the data.
  */
-static void do_verify(struct thread_data *td)
+static void do_verify(struct thread_data *td, uint64_t verify_bytes)
 {
        struct fio_file *f;
        struct io_u *io_u;
        int ret, min_events;
+       uint64_t io_bytes;
        unsigned int i;
 
        dprint(FD_VERIFY, "starting loop\n");
@@ -409,7 +422,9 @@ static void do_verify(struct thread_data *td)
        td_set_runstate(td, TD_VERIFYING);
 
        io_u = NULL;
+       io_bytes = 0;
        while (!td->terminate) {
+               enum fio_ddir ddir;
                int ret2, full;
 
                update_tv_cache(td);
@@ -425,18 +440,27 @@ static void do_verify(struct thread_data *td)
                if (flow_threshold_exceeded(td))
                        continue;
 
-               io_u = __get_io_u(td);
-               if (!io_u)
-                       break;
+               if (!td->o.experimental_verify) {
+                       io_u = __get_io_u(td);
+                       if (!io_u)
+                               break;
 
-               if (get_next_verify(td, io_u)) {
-                       put_io_u(td, io_u);
-                       break;
-               }
+                       if (get_next_verify(td, io_u)) {
+                               put_io_u(td, io_u);
+                               break;
+                       }
 
-               if (td_io_prep(td, io_u)) {
-                       put_io_u(td, io_u);
-                       break;
+                       if (td_io_prep(td, io_u)) {
+                               put_io_u(td, io_u);
+                               break;
+                       }
+               } else {
+                       io_u = get_io_u(td);
+                       if (!io_u)
+                               break;
+
+                       if (io_u->buflen + io_bytes > verify_bytes)
+                               break;
                }
 
                if (td->o.verify_async)
@@ -444,6 +468,8 @@ static void do_verify(struct thread_data *td)
                else
                        io_u->end_io = verify_io_u;
 
+               ddir = io_u->ddir;
+
                ret = td_io_queue(td, io_u);
                switch (ret) {
                case FIO_Q_COMPLETED:
@@ -465,6 +491,7 @@ static void do_verify(struct thread_data *td)
                                io_u->xfer_buflen = io_u->resid;
                                io_u->xfer_buf += bytes;
                                io_u->offset += bytes;
+                               io_bytes += bytes;
 
                                if (ddir_rw(io_u->ddir))
                                        td->ts.short_io_u[io_u->ddir]++;
@@ -480,6 +507,7 @@ sync_done:
                                if (ret < 0)
                                        break;
                        }
+                       io_bytes += io_u->xfer_buflen;
                        continue;
                case FIO_Q_QUEUED:
                        break;
@@ -495,7 +523,7 @@ sync_done:
                        break;
                }
 
-               if (break_on_this_error(td, io_u->ddir, &ret))
+               if (break_on_this_error(td, ddir, &ret))
                        break;
 
                /*
@@ -514,15 +542,18 @@ sync_done:
                                min_events = 1;
 
                        do {
+                               unsigned long bytes = 0;
+
                                /*
                                 * Reap required number of io units, if any,
                                 * and do the verification on them through
                                 * the callback handler
                                 */
-                               if (io_u_queued_complete(td, min_events, NULL) < 0) {
+                               if (io_u_queued_complete(td, min_events, &bytes) < 0) {
                                        ret = -1;
                                        break;
                                }
+                               io_bytes += bytes;
                        } while (full && (td->cur_depth > td->o.iodepth_low));
                }
                if (ret < 0)
@@ -582,7 +613,7 @@ static void do_io(struct thread_data *td)
                int ret2, full;
                enum fio_ddir ddir;
 
-               if (td->terminate)
+               if (td->terminate || td->done)
                        break;
 
                update_tv_cache(td);
@@ -717,7 +748,7 @@ sync_done:
 
                if (ret < 0)
                        break;
-               if (!ddir_rw_sum(bytes_done))
+               if (!ddir_rw_sum(bytes_done) && !(td->io_ops->flags & FIO_NOIO))
                        continue;
 
                if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
@@ -791,6 +822,10 @@ static void cleanup_io_u(struct thread_data *td)
                io_u = flist_entry(entry, struct io_u, list);
 
                flist_del(&io_u->list);
+
+               if (td->io_ops->io_u_free)
+                       td->io_ops->io_u_free(td, io_u);
+
                fio_memfree(io_u, sizeof(*io_u));
        }
 
@@ -802,6 +837,7 @@ static int init_io_u(struct thread_data *td)
        struct io_u *io_u;
        unsigned int max_bs, min_write;
        int cl_align, i, max_units;
+       int data_xfer = 1;
        char *p;
 
        max_units = td->o.iodepth;
@@ -811,6 +847,9 @@ static int init_io_u(struct thread_data *td)
        td->orig_buffer_size = (unsigned long long) max_bs
                                        * (unsigned long long) max_units;
 
+       if ((td->io_ops->flags & FIO_NOIO) || !(td_read(td) || td_write(td)))
+               data_xfer = 0;
+
        if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) {
                unsigned long bs;
 
@@ -823,7 +862,7 @@ static int init_io_u(struct thread_data *td)
                return 1;
        }
 
-       if (allocate_io_mem(td))
+       if (data_xfer && allocate_io_mem(td))
                return 1;
 
        if (td->o.odirect || td->o.mem_align ||
@@ -851,7 +890,7 @@ static int init_io_u(struct thread_data *td)
                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)) {
+               if (data_xfer) {
                        io_u->buf = p;
                        dprint(FD_MEM, "io_u %p, mem %p\n", io_u, io_u->buf);
 
@@ -869,6 +908,16 @@ static int init_io_u(struct thread_data *td)
                io_u->index = i;
                io_u->flags = IO_U_F_FREE;
                flist_add(&io_u->list, &td->io_u_freelist);
+
+               if (td->io_ops->io_u_init) {
+                       int ret = td->io_ops->io_u_init(td, io_u);
+
+                       if (ret) {
+                               log_err("fio: failed to init engine data: %d\n", ret);
+                               return 1;
+                       }
+               }
+
                p += max_bs;
        }
 
@@ -981,6 +1030,8 @@ static void *thread_main(void *data)
        } else
                td->pid = gettid();
 
+       fio_local_clock_init(td->o.use_thread);
+
        dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid);
 
        INIT_FLIST_HEAD(&td->io_u_freelist);
@@ -990,6 +1041,7 @@ static void *thread_main(void *data)
        INIT_FLIST_HEAD(&td->io_hist_list);
        INIT_FLIST_HEAD(&td->verify_list);
        INIT_FLIST_HEAD(&td->trim_list);
+       INIT_FLIST_HEAD(&td->next_rand_list);
        pthread_mutex_init(&td->io_u_lock, NULL);
        td->io_hist_tree = RB_ROOT;
 
@@ -1039,6 +1091,49 @@ static void *thread_main(void *data)
                goto err;
        }
 
+#ifdef CONFIG_LIBNUMA
+       /* numa node setup */
+       if (td->o.numa_cpumask_set || td->o.numa_memmask_set) {
+               int ret;
+
+               if (numa_available() < 0) {
+                       td_verror(td, errno, "Does not support NUMA API\n");
+                       goto err;
+               }
+
+               if (td->o.numa_cpumask_set) {
+                       ret = numa_run_on_node_mask(td->o.numa_cpunodesmask);
+                       if (ret == -1) {
+                               td_verror(td, errno, \
+                                       "numa_run_on_node_mask failed\n");
+                               goto err;
+                       }
+               }
+
+               if (td->o.numa_memmask_set) {
+
+                       switch (td->o.numa_mem_mode) {
+                       case MPOL_INTERLEAVE:
+                               numa_set_interleave_mask(td->o.numa_memnodesmask);
+                               break;
+                       case MPOL_BIND:
+                               numa_set_membind(td->o.numa_memnodesmask);
+                               break;
+                       case MPOL_LOCAL:
+                               numa_set_localalloc();
+                               break;
+                       case MPOL_PREFERRED:
+                               numa_set_preferred(td->o.numa_mem_prefer_node);
+                               break;
+                       case MPOL_DEFAULT:
+                       default:
+                               break;
+                       }
+
+               }
+       }
+#endif
+
        /*
         * May alter parameters that init_io_u() will use, so we need to
         * do this first.
@@ -1095,6 +1190,8 @@ static void *thread_main(void *data)
 
        clear_state = 0;
        while (keep_running(td)) {
+               uint64_t write_bytes;
+
                fio_gettime(&td->start, NULL);
                memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
                memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
@@ -1115,7 +1212,9 @@ static void *thread_main(void *data)
 
                prune_io_piece_log(td);
 
+               write_bytes = td->io_bytes[DDIR_WRITE];
                do_io(td);
+               write_bytes = td->io_bytes[DDIR_WRITE] - write_bytes;
 
                clear_state = 1;
 
@@ -1144,7 +1243,7 @@ static void *thread_main(void *data)
 
                fio_gettime(&td->start, NULL);
 
-               do_verify(td);
+               do_verify(td, write_bytes);
 
                td->ts.runtime[DDIR_READ] += utime_since_now(&td->start);
 
@@ -1214,8 +1313,8 @@ err:
                verify_async_exit(td);
 
        close_and_free_files(td);
-       close_ioengine(td);
        cleanup_io_u(td);
+       close_ioengine(td);
        cgroup_shutdown(td, &cgroup_mnt);
 
        if (td->o.cpumask_set) {
@@ -1326,7 +1425,7 @@ static void reap_threads(unsigned int *nr_running, unsigned int *t_rate,
                        if (WIFSIGNALED(status)) {
                                int sig = WTERMSIG(status);
 
-                               if (sig != SIGTERM)
+                               if (sig != SIGTERM && sig != SIGUSR2)
                                        log_err("fio: pid=%d, got signal=%d\n",
                                                        (int) td->pid, sig);
                                td->sig = sig;
@@ -1606,6 +1705,13 @@ void wait_for_disk_thread_exit(void)
        fio_mutex_down(disk_thread_mutex);
 }
 
+static void free_disk_util(void)
+{
+       disk_util_start_exit();
+       wait_for_disk_thread_exit();
+       disk_util_prune_entries();
+}
+
 static void *disk_thread_main(void *data)
 {
        int ret = 0;