iolog: enable replay_redirect on iolog replay
[fio.git] / iolog.c
diff --git a/iolog.c b/iolog.c
index 71afe86c29a1e50847128591be9bb10adbcea70b..686c7134b478ba0426bba60f2f4a715115bdea71 100644 (file)
--- a/iolog.c
+++ b/iolog.c
@@ -20,6 +20,8 @@
 #include "filelock.h"
 #include "smalloc.h"
 
+static int iolog_flush(struct io_log *log);
+
 static const char iolog_ver2[] = "fio version 2 iolog";
 
 void queue_io_piece(struct thread_data *td, struct io_piece *ipo)
@@ -344,7 +346,7 @@ static int read_iolog2(struct thread_data *td, FILE *f)
        unsigned long long offset;
        unsigned int bytes;
        int reads, writes, waits, fileno = 0, file_action = 0; /* stupid gcc */
-       char *fname, *act;
+       char *rfname, *fname, *act;
        char *str, *p;
        enum fio_ddir rw;
 
@@ -355,7 +357,7 @@ static int read_iolog2(struct thread_data *td, FILE *f)
         * for doing verifications.
         */
        str = malloc(4096);
-       fname = malloc(256+16);
+       rfname = fname = malloc(256+16);
        act = malloc(256+16);
 
        reads = writes = waits = 0;
@@ -363,8 +365,12 @@ static int read_iolog2(struct thread_data *td, FILE *f)
                struct io_piece *ipo;
                int r;
 
-               r = sscanf(p, "%256s %256s %llu %u", fname, act, &offset,
+               r = sscanf(p, "%256s %256s %llu %u", rfname, act, &offset,
                                                                        &bytes);
+
+               if (td->o.replay_redirect)
+                       fname = td->o.replay_redirect;
+
                if (r == 4) {
                        /*
                         * Check action first
@@ -449,7 +455,7 @@ static int read_iolog2(struct thread_data *td, FILE *f)
 
        free(str);
        free(act);
-       free(fname);
+       free(rfname);
 
        if (writes && read_only) {
                log_err("fio: <%s> skips replay of %d writes due to"
@@ -574,17 +580,41 @@ void setup_log(struct io_log **log, struct log_params *p,
               const char *filename)
 {
        struct io_log *l;
+       int i;
+       struct io_u_plat_entry *entry;
+       struct flist_head *list;
 
-       l = smalloc(sizeof(*l));
-       l->nr_samples = 0;
+       l = scalloc(1, sizeof(*l));
+       INIT_FLIST_HEAD(&l->io_logs);
        l->log_type = p->log_type;
        l->log_offset = p->log_offset;
        l->log_gz = p->log_gz;
        l->log_gz_store = p->log_gz_store;
        l->avg_msec = p->avg_msec;
+       l->hist_msec = p->hist_msec;
+       l->hist_coarseness = p->hist_coarseness;
        l->filename = strdup(filename);
        l->td = p->td;
 
+       /* Initialize histogram lists for each r/w direction,
+        * with initial io_u_plat of all zeros:
+        */
+       for (i = 0; i < DDIR_RWDIR_CNT; i++) {
+               list = &l->hist_window[i].list;
+               INIT_FLIST_HEAD(list);
+               entry = calloc(1, sizeof(struct io_u_plat_entry));
+               flist_add(&entry->list, list);
+       }
+
+       if (l->td && l->td->o.io_submit_mode != IO_MODE_OFFLOAD) {
+               struct io_logs *p;
+
+               p = calloc(1, sizeof(*l->pending));
+               p->max_samples = DEF_LOG_ENTRIES;
+               p->log = calloc(p->max_samples, log_entry_sz(l));
+               l->pending = p;
+       }
+
        if (l->log_offset)
                l->log_ddir_mask = LOG_OFFSET_SAMPLE_BIT;
 
@@ -593,7 +623,7 @@ void setup_log(struct io_log **log, struct log_params *p,
        if (l->log_gz && !p->td)
                l->log_gz = 0;
        else if (l->log_gz || l->log_gz_store) {
-               pthread_mutex_init(&l->chunk_lock, NULL);
+               mutex_init_pshared(&l->chunk_lock);
                p->td->flags |= TD_F_COMPRESS_LOG;
        }
 
@@ -628,11 +658,87 @@ static void clear_file_buffer(void *buf)
 
 void free_log(struct io_log *log)
 {
-       free(log->log);
+       while (!flist_empty(&log->io_logs)) {
+               struct io_logs *cur_log;
+
+               cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
+               flist_del_init(&cur_log->list);
+               free(cur_log->log);
+               sfree(cur_log);
+       }
+
+       if (log->pending) {
+               free(log->pending->log);
+               free(log->pending);
+               log->pending = NULL;
+       }
+
+       free(log->pending);
        free(log->filename);
        sfree(log);
 }
 
+inline unsigned long hist_sum(int j, int stride, unsigned int *io_u_plat,
+               unsigned int *io_u_plat_last)
+{
+       unsigned long sum;
+       int k;
+
+       if (io_u_plat_last) {
+               for (k = sum = 0; k < stride; k++)
+                       sum += io_u_plat[j + k] - io_u_plat_last[j + k];
+       } else {
+               for (k = sum = 0; k < stride; k++)
+                       sum += io_u_plat[j + k];
+       }
+
+       return sum;
+}
+
+static void flush_hist_samples(FILE *f, int hist_coarseness, void *samples,
+                              uint64_t sample_size)
+{
+       struct io_sample *s;
+       int log_offset;
+       uint64_t i, j, nr_samples;
+       struct io_u_plat_entry *entry, *entry_before;
+       unsigned int *io_u_plat;
+       unsigned int *io_u_plat_before;
+
+       int stride = 1 << hist_coarseness;
+       
+       if (!sample_size)
+               return;
+
+       s = __get_sample(samples, 0, 0);
+       log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0;
+
+       nr_samples = sample_size / __log_entry_sz(log_offset);
+
+       for (i = 0; i < nr_samples; i++) {
+               s = __get_sample(samples, log_offset, i);
+
+               entry = (struct io_u_plat_entry *) (uintptr_t) s->val;
+               io_u_plat = entry->io_u_plat;
+
+               entry_before = flist_first_entry(&entry->list, struct io_u_plat_entry, list);
+               io_u_plat_before = entry_before->io_u_plat;
+
+               fprintf(f, "%lu, %u, %u, ", (unsigned long) s->time,
+                                               io_sample_ddir(s), s->bs);
+               for (j = 0; j < FIO_IO_U_PLAT_NR - stride; j += stride) {
+                       fprintf(f, "%lu, ", hist_sum(j, stride, io_u_plat,
+                                               io_u_plat_before));
+               }
+               fprintf(f, "%lu\n", (unsigned long)
+                       hist_sum(FIO_IO_U_PLAT_NR - stride, stride, io_u_plat,
+                                       io_u_plat_before));
+
+               flist_del(&entry_before->list);
+               free(entry_before);
+       }
+}
+
 void flush_samples(FILE *f, void *samples, uint64_t sample_size)
 {
        struct io_sample *s;
@@ -673,7 +779,8 @@ struct iolog_flush_data {
        struct workqueue_work work;
        struct io_log *log;
        void *samples;
-       uint64_t nr_samples;
+       uint32_t nr_samples;
+       bool free;
 };
 
 #define GZ_CHUNK       131072
@@ -700,6 +807,7 @@ static int z_stream_init(z_stream *stream, int gz_hdr)
 {
        int wbits = 15;
 
+       memset(stream, 0, sizeof(*stream));
        stream->zalloc = Z_NULL;
        stream->zfree = Z_NULL;
        stream->opaque = Z_NULL;
@@ -734,7 +842,8 @@ static void finish_chunk(z_stream *stream, FILE *f,
 
        ret = inflateEnd(stream);
        if (ret != Z_OK)
-               log_err("fio: failed to end log inflation (%d)\n", ret);
+               log_err("fio: failed to end log inflation seq %d (%d)\n",
+                               iter->seq, ret);
 
        flush_samples(f, iter->buf, iter->buf_used);
        free(iter->buf);
@@ -751,7 +860,7 @@ static size_t inflate_chunk(struct iolog_compress *ic, int gz_hdr, FILE *f,
 {
        size_t ret;
 
-       dprint(FD_COMPRESS, "inflate chunk size=%lu, seq=%u",
+       dprint(FD_COMPRESS, "inflate chunk size=%lu, seq=%u\n",
                                (unsigned long) ic->len, ic->seq);
 
        if (ic->seq != iter->seq) {
@@ -798,7 +907,7 @@ static size_t inflate_chunk(struct iolog_compress *ic, int gz_hdr, FILE *f,
 
        ret = (void *) stream->next_in - ic->buf;
 
-       dprint(FD_COMPRESS, "inflated to size=%lu\n", (unsigned long) ret);
+       dprint(FD_COMPRESS, "inflated to size=%lu\n", (unsigned long) iter->buf_size);
 
        return ret;
 }
@@ -936,7 +1045,7 @@ int iolog_file_inflate(const char *file)
 
 #endif
 
-void flush_log(struct io_log *log, int do_append)
+void flush_log(struct io_log *log, bool do_append)
 {
        void *buf;
        FILE *f;
@@ -954,7 +1063,20 @@ void flush_log(struct io_log *log, int do_append)
 
        inflate_gz_chunks(log, f);
 
-       flush_samples(f, log->log, log->nr_samples * log_entry_sz(log));
+       while (!flist_empty(&log->io_logs)) {
+               struct io_logs *cur_log;
+
+               cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
+               flist_del_init(&cur_log->list);
+               
+               if (log->td && log == log->td->clat_hist_log)
+                       flush_hist_samples(f, log->hist_coarseness, cur_log->log,
+                                          log_sample_sz(log, cur_log));
+               else
+                       flush_samples(f, cur_log->log, log_sample_sz(log, cur_log));
+               
+               sfree(cur_log);
+       }
 
        fclose(f);
        clear_file_buffer(buf);
@@ -963,7 +1085,7 @@ void flush_log(struct io_log *log, int do_append)
 static int finish_log(struct thread_data *td, struct io_log *log, int trylock)
 {
        if (td->flags & TD_F_COMPRESS_LOG)
-               iolog_flush(log, 1);
+               iolog_flush(log);
 
        if (trylock) {
                if (fio_trylock_file(log->filename))
@@ -1005,7 +1127,7 @@ size_t log_chunk_sizes(struct io_log *log)
 
 static int gz_work(struct iolog_flush_data *data)
 {
-       struct iolog_compress *c;
+       struct iolog_compress *c = NULL;
        struct flist_head list;
        unsigned int seq;
        z_stream stream;
@@ -1014,6 +1136,7 @@ static int gz_work(struct iolog_flush_data *data)
 
        INIT_FLIST_HEAD(&list);
 
+       memset(&stream, 0, sizeof(stream));
        stream.zalloc = Z_NULL;
        stream.zfree = Z_NULL;
        stream.opaque = Z_NULL;
@@ -1029,9 +1152,13 @@ static int gz_work(struct iolog_flush_data *data)
        stream.next_in = (void *) data->samples;
        stream.avail_in = data->nr_samples * log_entry_sz(data->log);
 
-       dprint(FD_COMPRESS, "deflate input size=%lu, seq=%u\n",
-                               (unsigned long) stream.avail_in, seq);
+       dprint(FD_COMPRESS, "deflate input size=%lu, seq=%u, log=%s\n",
+                               (unsigned long) stream.avail_in, seq,
+                               data->log->filename);
        do {
+               if (c)
+                       dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq,
+                               (unsigned long) c->len);
                c = get_new_chunk(seq);
                stream.avail_out = GZ_CHUNK;
                stream.next_out = c->buf;
@@ -1051,9 +1178,26 @@ static int gz_work(struct iolog_flush_data *data)
        stream.avail_out = GZ_CHUNK - c->len;
 
        ret = deflate(&stream, Z_FINISH);
-       if (ret == Z_STREAM_END)
-               c->len = GZ_CHUNK - stream.avail_out;
-       else {
+       if (ret < 0) {
+               /*
+                * Z_BUF_ERROR is special, it just means we need more
+                * output space. We'll handle that below. Treat any other
+                * error as fatal.
+                */
+               if (ret != Z_BUF_ERROR) {
+                       log_err("fio: deflate log (%d)\n", ret);
+                       flist_del(&c->list);
+                       free_chunk(c);
+                       goto err;
+               }
+       }
+
+       total -= c->len;
+       c->len = GZ_CHUNK - stream.avail_out;
+       total += c->len;
+       dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq, (unsigned long) c->len);
+
+       if (ret != Z_STREAM_END) {
                do {
                        c = get_new_chunk(seq);
                        stream.avail_out = GZ_CHUNK;
@@ -1062,6 +1206,8 @@ static int gz_work(struct iolog_flush_data *data)
                        c->len = GZ_CHUNK - stream.avail_out;
                        total += c->len;
                        flist_add_tail(&c->list, &list);
+                       dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq,
+                               (unsigned long) c->len);
                } while (ret != Z_STREAM_END);
        }
 
@@ -1081,7 +1227,8 @@ static int gz_work(struct iolog_flush_data *data)
 
        ret = 0;
 done:
-       free(data);
+       if (data->free)
+               free(data);
        return ret;
 err:
        while (!flist_empty(&list)) {
@@ -1145,39 +1292,67 @@ void iolog_compress_exit(struct thread_data *td)
  * Queue work item to compress the existing log entries. We reset the
  * current log to a small size, and reference the existing log in the
  * data that we queue for compression. Once compression has been done,
- * this old log is freed. If called with wait == 1, will not return until
- * the log compression has completed.
+ * this old log is freed. If called with finish == true, will not return
+ * until the log compression has completed, and will flush all previous
+ * logs too
  */
-int iolog_flush(struct io_log *log, int wait)
+static int iolog_flush(struct io_log *log)
 {
        struct iolog_flush_data *data;
 
-       io_u_quiesce(log->td);
-
        data = malloc(sizeof(*data));
        if (!data)
                return 1;
 
        data->log = log;
+       data->free = false;
 
-       data->samples = log->log;
-       data->nr_samples = log->nr_samples;
+       while (!flist_empty(&log->io_logs)) {
+               struct io_logs *cur_log;
 
-       log->nr_samples = 0;
-       log->max_samples = DEF_LOG_ENTRIES;
-       log->log = malloc(log->max_samples * log_entry_sz(log));
+               cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
+               flist_del_init(&cur_log->list);
+
+               data->samples = cur_log->log;
+               data->nr_samples = cur_log->nr_samples;
+
+               sfree(cur_log);
 
-       if (!wait)
-               workqueue_enqueue(&log->td->log_compress_wq, &data->work);
-       else
                gz_work(data);
+       }
 
+       free(data);
        return 0;
 }
 
+int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
+{
+       struct iolog_flush_data *data;
+
+       data = malloc(sizeof(*data));
+       if (!data)
+               return 1;
+
+       data->log = log;
+
+       data->samples = cur_log->log;
+       data->nr_samples = cur_log->nr_samples;
+       data->free = true;
+
+       cur_log->nr_samples = cur_log->max_samples = 0;
+       cur_log->log = NULL;
+
+       workqueue_enqueue(&log->td->log_compress_wq, &data->work);
+       return 0;
+}
 #else
 
-int iolog_flush(struct io_log *log, int wait)
+static int iolog_flush(struct io_log *log)
+{
+       return 1;
+}
+
+int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
 {
        return 1;
 }
@@ -1193,6 +1368,29 @@ void iolog_compress_exit(struct thread_data *td)
 
 #endif
 
+struct io_logs *iolog_cur_log(struct io_log *log)
+{
+       if (flist_empty(&log->io_logs))
+               return NULL;
+
+       return flist_last_entry(&log->io_logs, struct io_logs, list);
+}
+
+uint64_t iolog_nr_samples(struct io_log *iolog)
+{
+       struct flist_head *entry;
+       uint64_t ret = 0;
+
+       flist_for_each(entry, &iolog->io_logs) {
+               struct io_logs *cur_log;
+
+               cur_log = flist_entry(entry, struct io_logs, list);
+               ret += cur_log->nr_samples;
+       }
+
+       return ret;
+}
+
 static int __write_log(struct thread_data *td, struct io_log *log, int try)
 {
        if (log)
@@ -1243,6 +1441,20 @@ static int write_clat_log(struct thread_data *td, int try, bool unit_log)
        return ret;
 }
 
+static int write_clat_hist_log(struct thread_data *td, int try, bool unit_log)
+{
+       int ret;
+
+       if (!unit_log)
+               return 0;
+
+       ret = __write_log(td, td->clat_hist_log, try);
+       if (!ret)
+               td->clat_hist_log = NULL;
+
+       return ret;
+}
+
 static int write_lat_log(struct thread_data *td, int try, bool unit_log)
 {
        int ret;
@@ -1277,8 +1489,9 @@ enum {
        SLAT_LOG_MASK   = 4,
        CLAT_LOG_MASK   = 8,
        IOPS_LOG_MASK   = 16,
+       CLAT_HIST_LOG_MASK = 32,
 
-       ALL_LOG_NR      = 5,
+       ALL_LOG_NR      = 6,
 };
 
 struct log_type {
@@ -1307,6 +1520,10 @@ static struct log_type log_types[] = {
                .mask   = IOPS_LOG_MASK,
                .fn     = write_iops_log,
        },
+       {
+               .mask   = CLAT_HIST_LOG_MASK,
+               .fn     = write_clat_hist_log,
+       }
 };
 
 void td_writeout_logs(struct thread_data *td, bool unit_logs)