Improve logging accuracy
[fio.git] / iolog.c
diff --git a/iolog.c b/iolog.c
index 9bcf0d8e2982e37ce5d48fbd64e9653872e54ea0..71afe86c29a1e50847128591be9bb10adbcea70b 100644 (file)
--- a/iolog.c
+++ b/iolog.c
@@ -6,10 +6,19 @@
 #include <stdlib.h>
 #include <libgen.h>
 #include <assert.h>
+#include <sys/types.h>
+#include <sys/stat.h>
+#include <unistd.h>
+#ifdef CONFIG_ZLIB
+#include <zlib.h>
+#endif
+
 #include "flist.h"
 #include "fio.h"
 #include "verify.h"
 #include "trim.h"
+#include "filelock.h"
+#include "smalloc.h"
 
 static const char iolog_ver2[] = "fio version 2 iolog";
 
@@ -19,19 +28,14 @@ void queue_io_piece(struct thread_data *td, struct io_piece *ipo)
        td->total_io_size += ipo->len;
 }
 
-void log_io_u(struct thread_data *td, struct io_u *io_u)
+void log_io_u(const struct thread_data *td, const struct io_u *io_u)
 {
-       const char *act[] = { "read", "write", "sync", "datasync",
-                               "sync_file_range", "wait", "trim" };
-
-       assert(io_u->ddir <= 6);
-
        if (!td->o.write_iolog_file)
                return;
 
        fprintf(td->iolog_f, "%s %s %llu %lu\n", io_u->file->file_name,
-                                               act[io_u->ddir], io_u->offset,
-                                               io_u->buflen);
+                                               io_ddir_name(io_u->ddir),
+                                               io_u->offset, io_u->buflen);
 }
 
 void log_file(struct thread_data *td, struct fio_file *f,
@@ -56,20 +60,36 @@ void log_file(struct thread_data *td, struct fio_file *f,
 
 static void iolog_delay(struct thread_data *td, unsigned long delay)
 {
-       unsigned long usec = utime_since_now(&td->last_issue);
+       uint64_t usec = utime_since_now(&td->last_issue);
+       uint64_t this_delay;
+       struct timeval tv;
+
+       if (delay < td->time_offset) {
+               td->time_offset = 0;
+               return;
+       }
 
+       delay -= td->time_offset;
        if (delay < usec)
                return;
 
        delay -= usec;
 
-       /*
-        * less than 100 usec delay, just regard it as noise
-        */
-       if (delay < 100)
-               return;
+       fio_gettime(&tv, NULL);
+       while (delay && !td->terminate) {
+               this_delay = delay;
+               if (this_delay > 500000)
+                       this_delay = 500000;
+
+               usec_sleep(td, this_delay);
+               delay -= this_delay;
+       }
 
-       usec_sleep(td, delay);
+       usec = utime_since_now(&tv);
+       if (usec > delay)
+               td->time_offset = usec - delay;
+       else
+               td->time_offset = 0;
 }
 
 static int ipo_special(struct thread_data *td, struct io_piece *ipo)
@@ -96,7 +116,7 @@ static int ipo_special(struct thread_data *td, struct io_piece *ipo)
                td_io_close_file(td, f);
                break;
        case FIO_LOG_UNLINK_FILE:
-               unlink(f->file_name);
+               td_io_unlink_file(td, f);
                break;
        default:
                log_err("fio: bad file action %d\n", ipo->file_action);
@@ -114,7 +134,7 @@ int read_iolog_get(struct thread_data *td, struct io_u *io_u)
        while (!flist_empty(&td->io_log_list)) {
                int ret;
 
-               ipo = flist_entry(td->io_log_list.next, struct io_piece, list);
+               ipo = flist_first_entry(&td->io_log_list, struct io_piece, list);
                flist_del(&ipo->list);
                remove_trim_entry(td, ipo);
 
@@ -167,7 +187,7 @@ void prune_io_piece_log(struct thread_data *td)
        }
 
        while (!flist_empty(&td->io_hist_list)) {
-               ipo = flist_entry(td->io_hist_list.next, struct io_piece, list);
+               ipo = flist_first_entry(&td->io_hist_list, struct io_piece, list);
                flist_del(&ipo->list);
                remove_trim_entry(td, ipo);
                td->io_hist_len--;
@@ -188,6 +208,10 @@ void log_io_piece(struct thread_data *td, struct io_u *io_u)
        ipo->file = io_u->file;
        ipo->offset = io_u->offset;
        ipo->len = io_u->buflen;
+       ipo->numberio = io_u->numberio;
+       ipo->flags = IP_F_IN_FLIGHT;
+
+       io_u->ipo = ipo;
 
        if (io_u_should_trim(td, io_u)) {
                flist_add_tail(&ipo->trim_list, &td->trim_list);
@@ -208,7 +232,7 @@ void log_io_piece(struct thread_data *td, struct io_u *io_u)
         * drop the old one, which we rely on the rb insert/lookup for
         * handling.
         */
-       if ((!td_random(td) || !td->o.overwrite) &&
+       if (((!td->o.verifysort) || !td_random(td) || !td->o.overwrite) &&
              (file_randommap(td, ipo->file) || td->o.verify == VERIFY_NONE)) {
                INIT_FLIST_HEAD(&ipo->list);
                flist_add_tail(&ipo->list, &td->io_hist_list);
@@ -226,6 +250,7 @@ restart:
        p = &td->io_hist_tree.rb_node;
        parent = NULL;
        while (*p) {
+               int overlap = 0;
                parent = *p;
 
                __ipo = rb_entry(parent, struct io_piece, rb_node);
@@ -233,11 +258,18 @@ restart:
                        p = &(*p)->rb_left;
                else if (ipo->file > __ipo->file)
                        p = &(*p)->rb_right;
-               else if (ipo->offset < __ipo->offset)
+               else if (ipo->offset < __ipo->offset) {
                        p = &(*p)->rb_left;
-               else if (ipo->offset > __ipo->offset)
+                       overlap = ipo->offset + ipo->len > __ipo->offset;
+               }
+               else if (ipo->offset > __ipo->offset) {
                        p = &(*p)->rb_right;
-               else {
+                       overlap = __ipo->offset + __ipo->len > ipo->offset;
+               }
+               else
+                       overlap = 1;
+
+               if (overlap) {
                        dprint(FD_IO, "iolog: overlap %llu/%lu, %llu/%lu",
                                __ipo->offset, __ipo->len,
                                ipo->offset, ipo->len);
@@ -255,6 +287,45 @@ restart:
        td->io_hist_len++;
 }
 
+void unlog_io_piece(struct thread_data *td, struct io_u *io_u)
+{
+       struct io_piece *ipo = io_u->ipo;
+
+       if (td->ts.nr_block_infos) {
+               uint32_t *info = io_u_block_info(td, io_u);
+               if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) {
+                       if (io_u->ddir == DDIR_TRIM)
+                               *info = BLOCK_INFO_SET_STATE(*info,
+                                               BLOCK_STATE_TRIM_FAILURE);
+                       else if (io_u->ddir == DDIR_WRITE)
+                               *info = BLOCK_INFO_SET_STATE(*info,
+                                               BLOCK_STATE_WRITE_FAILURE);
+               }
+       }
+
+       if (!ipo)
+               return;
+
+       if (ipo->flags & IP_F_ONRB)
+               rb_erase(&ipo->rb_node, &td->io_hist_tree);
+       else if (ipo->flags & IP_F_ONLIST)
+               flist_del(&ipo->list);
+
+       free(ipo);
+       io_u->ipo = NULL;
+       td->io_hist_len--;
+}
+
+void trim_io_piece(struct thread_data *td, const struct io_u *io_u)
+{
+       struct io_piece *ipo = io_u->ipo;
+
+       if (!ipo)
+               return;
+
+       ipo->len = io_u->xfer_buflen - io_u->resid;
+}
+
 void write_iolog_close(struct thread_data *td)
 {
        fflush(td->iolog_f);
@@ -319,8 +390,7 @@ static int read_iolog2(struct thread_data *td, FILE *f)
                } else if (r == 2) {
                        rw = DDIR_INVAL;
                        if (!strcmp(act, "add")) {
-                               td->o.nr_files++;
-                               fileno = add_file(td, fname);
+                               fileno = add_file(td, fname, 0, 1);
                                file_action = FIO_LOG_ADD_FILE;
                                continue;
                        } else if (!strcmp(act, "open")) {
@@ -367,10 +437,11 @@ static int read_iolog2(struct thread_data *td, FILE *f)
                } else {
                        ipo->offset = offset;
                        ipo->len = bytes;
-                       if (bytes > td->o.max_bs[rw])
+                       if (rw != DDIR_INVAL && bytes > td->o.max_bs[rw])
                                td->o.max_bs[rw] = bytes;
                        ipo->fileno = fileno;
                        ipo->file_action = file_action;
+                       td->o.size += bytes;
                }
 
                queue_io_piece(td, ipo);
@@ -480,73 +551,802 @@ int init_iolog(struct thread_data *td)
        int ret = 0;
 
        if (td->o.read_iolog_file) {
+               int need_swap;
+
                /*
                 * Check if it's a blktrace file and load that if possible.
                 * Otherwise assume it's a normal log file and load that.
                 */
-               if (is_blktrace(td->o.read_iolog_file))
-                       ret = load_blktrace(td, td->o.read_iolog_file);
+               if (is_blktrace(td->o.read_iolog_file, &need_swap))
+                       ret = load_blktrace(td, td->o.read_iolog_file, need_swap);
                else
                        ret = init_iolog_read(td);
        } else if (td->o.write_iolog_file)
                ret = init_iolog_write(td);
 
+       if (ret)
+               td_verror(td, EINVAL, "failed initializing iolog");
+
        return ret;
 }
 
-void setup_log(struct io_log **log, unsigned long avg_msec, int log_type)
+void setup_log(struct io_log **log, struct log_params *p,
+              const char *filename)
 {
-       struct io_log *l = malloc(sizeof(*l));
+       struct io_log *l;
 
-       memset(l, 0, sizeof(*l));
+       l = smalloc(sizeof(*l));
        l->nr_samples = 0;
-       l->max_samples = 1024;
-       l->log_type = log_type;
-       l->log = malloc(l->max_samples * sizeof(struct io_sample));
-       l->avg_msec = avg_msec;
+       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->filename = strdup(filename);
+       l->td = p->td;
+
+       if (l->log_offset)
+               l->log_ddir_mask = LOG_OFFSET_SAMPLE_BIT;
+
+       INIT_FLIST_HEAD(&l->chunk_list);
+
+       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);
+               p->td->flags |= TD_F_COMPRESS_LOG;
+       }
+
        *log = l;
 }
 
-void __finish_log(struct io_log *log, const char *name)
+#ifdef CONFIG_SETVBUF
+static void *set_file_buffer(FILE *f)
 {
-       unsigned int i;
+       size_t size = 1048576;
+       void *buf;
+
+       buf = malloc(size);
+       setvbuf(f, buf, _IOFBF, size);
+       return buf;
+}
+
+static void clear_file_buffer(void *buf)
+{
+       free(buf);
+}
+#else
+static void *set_file_buffer(FILE *f)
+{
+       return NULL;
+}
+
+static void clear_file_buffer(void *buf)
+{
+}
+#endif
+
+void free_log(struct io_log *log)
+{
+       free(log->log);
+       free(log->filename);
+       sfree(log);
+}
+
+void flush_samples(FILE *f, void *samples, uint64_t sample_size)
+{
+       struct io_sample *s;
+       int log_offset;
+       uint64_t i, nr_samples;
+
+       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);
+
+               if (!log_offset) {
+                       fprintf(f, "%lu, %lu, %u, %u\n",
+                                       (unsigned long) s->time,
+                                       (unsigned long) s->val,
+                                       io_sample_ddir(s), s->bs);
+               } else {
+                       struct io_sample_offset *so = (void *) s;
+
+                       fprintf(f, "%lu, %lu, %u, %u, %llu\n",
+                                       (unsigned long) s->time,
+                                       (unsigned long) s->val,
+                                       io_sample_ddir(s), s->bs,
+                                       (unsigned long long) so->offset);
+               }
+       }
+}
+
+#ifdef CONFIG_ZLIB
+
+struct iolog_flush_data {
+       struct workqueue_work work;
+       struct io_log *log;
+       void *samples;
+       uint64_t nr_samples;
+};
+
+#define GZ_CHUNK       131072
+
+static struct iolog_compress *get_new_chunk(unsigned int seq)
+{
+       struct iolog_compress *c;
+
+       c = malloc(sizeof(*c));
+       INIT_FLIST_HEAD(&c->list);
+       c->buf = malloc(GZ_CHUNK);
+       c->len = 0;
+       c->seq = seq;
+       return c;
+}
+
+static void free_chunk(struct iolog_compress *ic)
+{
+       free(ic->buf);
+       free(ic);
+}
+
+static int z_stream_init(z_stream *stream, int gz_hdr)
+{
+       int wbits = 15;
+
+       stream->zalloc = Z_NULL;
+       stream->zfree = Z_NULL;
+       stream->opaque = Z_NULL;
+       stream->next_in = Z_NULL;
+
+       /*
+        * zlib magic - add 32 for auto-detection of gz header or not,
+        * if we decide to store files in a gzip friendly format.
+        */
+       if (gz_hdr)
+               wbits += 32;
+
+       if (inflateInit2(stream, wbits) != Z_OK)
+               return 1;
+
+       return 0;
+}
+
+struct inflate_chunk_iter {
+       unsigned int seq;
+       int err;
+       void *buf;
+       size_t buf_size;
+       size_t buf_used;
+       size_t chunk_sz;
+};
+
+static void finish_chunk(z_stream *stream, FILE *f,
+                        struct inflate_chunk_iter *iter)
+{
+       int ret;
+
+       ret = inflateEnd(stream);
+       if (ret != Z_OK)
+               log_err("fio: failed to end log inflation (%d)\n", ret);
+
+       flush_samples(f, iter->buf, iter->buf_used);
+       free(iter->buf);
+       iter->buf = NULL;
+       iter->buf_size = iter->buf_used = 0;
+}
+
+/*
+ * Iterative chunk inflation. Handles cases where we cross into a new
+ * sequence, doing flush finish of previous chunk if needed.
+ */
+static size_t inflate_chunk(struct iolog_compress *ic, int gz_hdr, FILE *f,
+                           z_stream *stream, struct inflate_chunk_iter *iter)
+{
+       size_t ret;
+
+       dprint(FD_COMPRESS, "inflate chunk size=%lu, seq=%u",
+                               (unsigned long) ic->len, ic->seq);
+
+       if (ic->seq != iter->seq) {
+               if (iter->seq)
+                       finish_chunk(stream, f, iter);
+
+               z_stream_init(stream, gz_hdr);
+               iter->seq = ic->seq;
+       }
+
+       stream->avail_in = ic->len;
+       stream->next_in = ic->buf;
+
+       if (!iter->buf_size) {
+               iter->buf_size = iter->chunk_sz;
+               iter->buf = malloc(iter->buf_size);
+       }
+
+       while (stream->avail_in) {
+               size_t this_out = iter->buf_size - iter->buf_used;
+               int err;
+
+               stream->avail_out = this_out;
+               stream->next_out = iter->buf + iter->buf_used;
+
+               err = inflate(stream, Z_NO_FLUSH);
+               if (err < 0) {
+                       log_err("fio: failed inflating log: %d\n", err);
+                       iter->err = err;
+                       break;
+               }
+
+               iter->buf_used += this_out - stream->avail_out;
+
+               if (!stream->avail_out) {
+                       iter->buf_size += iter->chunk_sz;
+                       iter->buf = realloc(iter->buf, iter->buf_size);
+                       continue;
+               }
+
+               if (err == Z_STREAM_END)
+                       break;
+       }
+
+       ret = (void *) stream->next_in - ic->buf;
+
+       dprint(FD_COMPRESS, "inflated to size=%lu\n", (unsigned long) ret);
+
+       return ret;
+}
+
+/*
+ * Inflate stored compressed chunks, or write them directly to the log
+ * file if so instructed.
+ */
+static int inflate_gz_chunks(struct io_log *log, FILE *f)
+{
+       struct inflate_chunk_iter iter = { .chunk_sz = log->log_gz, };
+       z_stream stream;
+
+       while (!flist_empty(&log->chunk_list)) {
+               struct iolog_compress *ic;
+
+               ic = flist_first_entry(&log->chunk_list, struct iolog_compress, list);
+               flist_del(&ic->list);
+
+               if (log->log_gz_store) {
+                       size_t ret;
+
+                       dprint(FD_COMPRESS, "log write chunk size=%lu, "
+                               "seq=%u\n", (unsigned long) ic->len, ic->seq);
+
+                       ret = fwrite(ic->buf, ic->len, 1, f);
+                       if (ret != 1 || ferror(f)) {
+                               iter.err = errno;
+                               log_err("fio: error writing compressed log\n");
+                       }
+               } else
+                       inflate_chunk(ic, log->log_gz_store, f, &stream, &iter);
+
+               free_chunk(ic);
+       }
+
+       if (iter.seq) {
+               finish_chunk(&stream, f, &iter);
+               free(iter.buf);
+       }
+
+       return iter.err;
+}
+
+/*
+ * Open compressed log file and decompress the stored chunks and
+ * write them to stdout. The chunks are stored sequentially in the
+ * file, so we iterate over them and do them one-by-one.
+ */
+int iolog_file_inflate(const char *file)
+{
+       struct inflate_chunk_iter iter = { .chunk_sz = 64 * 1024 * 1024, };
+       struct iolog_compress ic;
+       z_stream stream;
+       struct stat sb;
+       ssize_t ret;
+       size_t total;
+       void *buf;
        FILE *f;
 
-       f = fopen(name, "a");
+       f = fopen(file, "r");
        if (!f) {
-               perror("fopen log");
-               return;
+               perror("fopen");
+               return 1;
        }
 
-       for (i = 0; i < log->nr_samples; i++) {
-               fprintf(f, "%lu, %lu, %u, %u\n",
-                               (unsigned long) log->log[i].time,
-                               (unsigned long) log->log[i].val,
-                               log->log[i].ddir, log->log[i].bs);
+       if (stat(file, &sb) < 0) {
+               fclose(f);
+               perror("stat");
+               return 1;
+       }
+
+       ic.buf = buf = malloc(sb.st_size);
+       ic.len = sb.st_size;
+       ic.seq = 1;
+
+       ret = fread(ic.buf, ic.len, 1, f);
+       if (ret < 0) {
+               perror("fread");
+               fclose(f);
+               free(buf);
+               return 1;
+       } else if (ret != 1) {
+               log_err("fio: short read on reading log\n");
+               fclose(f);
+               free(buf);
+               return 1;
        }
 
        fclose(f);
-       free(log->log);
-       free(log);
+
+       /*
+        * Each chunk will return Z_STREAM_END. We don't know how many
+        * chunks are in the file, so we just keep looping and incrementing
+        * the sequence number until we have consumed the whole compressed
+        * file.
+        */
+       total = ic.len;
+       do {
+               size_t iret;
+
+               iret = inflate_chunk(&ic,  1, stdout, &stream, &iter);
+               total -= iret;
+               if (!total)
+                       break;
+               if (iter.err)
+                       break;
+
+               ic.seq++;
+               ic.len -= iret;
+               ic.buf += iret;
+       } while (1);
+
+       if (iter.seq) {
+               finish_chunk(&stream, stdout, &iter);
+               free(iter.buf);
+       }
+
+       free(buf);
+       return iter.err;
 }
 
-void finish_log_named(struct thread_data *td, struct io_log *log,
-                      const char *prefix, const char *postfix)
+#else
+
+static int inflate_gz_chunks(struct io_log *log, FILE *f)
 {
-       char file_name[256], *p;
+       return 0;
+}
+
+int iolog_file_inflate(const char *file)
+{
+       log_err("fio: log inflation not possible without zlib\n");
+       return 1;
+}
 
-       snprintf(file_name, sizeof(file_name), "%s_%s.log", prefix, postfix);
-       p = basename(file_name);
+#endif
 
-       if (td->client_type == FIO_CLIENT_TYPE_GUI) {
-               fio_send_iolog(td, log, p);
-               free(log->log);
-               free(log);
+void flush_log(struct io_log *log, int do_append)
+{
+       void *buf;
+       FILE *f;
+
+       if (!do_append)
+               f = fopen(log->filename, "w");
+       else
+               f = fopen(log->filename, "a");
+       if (!f) {
+               perror("fopen log");
+               return;
+       }
+
+       buf = set_file_buffer(f);
+
+       inflate_gz_chunks(log, f);
+
+       flush_samples(f, log->log, log->nr_samples * log_entry_sz(log));
+
+       fclose(f);
+       clear_file_buffer(buf);
+}
+
+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);
+
+       if (trylock) {
+               if (fio_trylock_file(log->filename))
+                       return 1;
        } else
-               __finish_log(log, p);
+               fio_lock_file(log->filename);
+
+       if (td->client_type == FIO_CLIENT_TYPE_GUI || is_backend)
+               fio_send_iolog(td, log, log->filename);
+       else
+               flush_log(log, !td->o.per_job_logs);
+
+       fio_unlock_file(log->filename);
+       free_log(log);
+       return 0;
+}
+
+size_t log_chunk_sizes(struct io_log *log)
+{
+       struct flist_head *entry;
+       size_t ret;
+
+       if (flist_empty(&log->chunk_list))
+               return 0;
+
+       ret = 0;
+       pthread_mutex_lock(&log->chunk_lock);
+       flist_for_each(entry, &log->chunk_list) {
+               struct iolog_compress *c;
+
+               c = flist_entry(entry, struct iolog_compress, list);
+               ret += c->len;
+       }
+       pthread_mutex_unlock(&log->chunk_lock);
+       return ret;
+}
+
+#ifdef CONFIG_ZLIB
+
+static int gz_work(struct iolog_flush_data *data)
+{
+       struct iolog_compress *c;
+       struct flist_head list;
+       unsigned int seq;
+       z_stream stream;
+       size_t total = 0;
+       int ret;
+
+       INIT_FLIST_HEAD(&list);
+
+       stream.zalloc = Z_NULL;
+       stream.zfree = Z_NULL;
+       stream.opaque = Z_NULL;
+
+       ret = deflateInit(&stream, Z_DEFAULT_COMPRESSION);
+       if (ret != Z_OK) {
+               log_err("fio: failed to init gz stream\n");
+               goto err;
+       }
+
+       seq = ++data->log->chunk_seq;
+
+       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);
+       do {
+               c = get_new_chunk(seq);
+               stream.avail_out = GZ_CHUNK;
+               stream.next_out = c->buf;
+               ret = deflate(&stream, Z_NO_FLUSH);
+               if (ret < 0) {
+                       log_err("fio: deflate log (%d)\n", ret);
+                       free_chunk(c);
+                       goto err;
+               }
+
+               c->len = GZ_CHUNK - stream.avail_out;
+               flist_add_tail(&c->list, &list);
+               total += c->len;
+       } while (stream.avail_in);
+
+       stream.next_out = c->buf + c->len;
+       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 {
+               do {
+                       c = get_new_chunk(seq);
+                       stream.avail_out = GZ_CHUNK;
+                       stream.next_out = c->buf;
+                       ret = deflate(&stream, Z_FINISH);
+                       c->len = GZ_CHUNK - stream.avail_out;
+                       total += c->len;
+                       flist_add_tail(&c->list, &list);
+               } while (ret != Z_STREAM_END);
+       }
+
+       dprint(FD_COMPRESS, "deflated to size=%lu\n", (unsigned long) total);
+
+       ret = deflateEnd(&stream);
+       if (ret != Z_OK)
+               log_err("fio: deflateEnd %d\n", ret);
+
+       free(data->samples);
+
+       if (!flist_empty(&list)) {
+               pthread_mutex_lock(&data->log->chunk_lock);
+               flist_splice_tail(&list, &data->log->chunk_list);
+               pthread_mutex_unlock(&data->log->chunk_lock);
+       }
+
+       ret = 0;
+done:
+       free(data);
+       return ret;
+err:
+       while (!flist_empty(&list)) {
+               c = flist_first_entry(list.next, struct iolog_compress, list);
+               flist_del(&c->list);
+               free_chunk(c);
+       }
+       ret = 1;
+       goto done;
+}
+
+/*
+ * Invoked from our compress helper thread, when logging would have exceeded
+ * the specified memory limitation. Compresses the previously stored
+ * entries.
+ */
+static int gz_work_async(struct submit_worker *sw, struct workqueue_work *work)
+{
+       return gz_work(container_of(work, struct iolog_flush_data, work));
+}
+
+static int gz_init_worker(struct submit_worker *sw)
+{
+       struct thread_data *td = sw->wq->td;
+
+       if (!fio_option_is_set(&td->o, log_gz_cpumask))
+               return 0;
+
+       if (fio_setaffinity(gettid(), td->o.log_gz_cpumask) == -1) {
+               log_err("gz: failed to set CPU affinity\n");
+               return 1;
+       }
+
+       return 0;
+}
+
+static struct workqueue_ops log_compress_wq_ops = {
+       .fn             = gz_work_async,
+       .init_worker_fn = gz_init_worker,
+       .nice           = 1,
+};
+
+int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
+{
+       if (!(td->flags & TD_F_COMPRESS_LOG))
+               return 0;
+
+       workqueue_init(td, &td->log_compress_wq, &log_compress_wq_ops, 1, sk_out);
+       return 0;
+}
+
+void iolog_compress_exit(struct thread_data *td)
+{
+       if (!(td->flags & TD_F_COMPRESS_LOG))
+               return;
+
+       workqueue_exit(&td->log_compress_wq);
+}
+
+/*
+ * 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.
+ */
+int iolog_flush(struct io_log *log, int wait)
+{
+       struct iolog_flush_data *data;
+
+       io_u_quiesce(log->td);
+
+       data = malloc(sizeof(*data));
+       if (!data)
+               return 1;
+
+       data->log = log;
+
+       data->samples = log->log;
+       data->nr_samples = log->nr_samples;
+
+       log->nr_samples = 0;
+       log->max_samples = DEF_LOG_ENTRIES;
+       log->log = malloc(log->max_samples * log_entry_sz(log));
+
+       if (!wait)
+               workqueue_enqueue(&log->td->log_compress_wq, &data->work);
+       else
+               gz_work(data);
+
+       return 0;
+}
+
+#else
+
+int iolog_flush(struct io_log *log, int wait)
+{
+       return 1;
+}
+
+int iolog_compress_init(struct thread_data *td, struct sk_out *sk_out)
+{
+       return 0;
+}
+
+void iolog_compress_exit(struct thread_data *td)
+{
+}
+
+#endif
+
+static int __write_log(struct thread_data *td, struct io_log *log, int try)
+{
+       if (log)
+               return finish_log(td, log, try);
+
+       return 0;
+}
+
+static int write_iops_log(struct thread_data *td, int try, bool unit_log)
+{
+       int ret;
+
+       if (per_unit_log(td->iops_log) != unit_log)
+               return 0;
+
+       ret = __write_log(td, td->iops_log, try);
+       if (!ret)
+               td->iops_log = NULL;
+
+       return ret;
+}
+
+static int write_slat_log(struct thread_data *td, int try, bool unit_log)
+{
+       int ret;
+
+       if (!unit_log)
+               return 0;
+
+       ret = __write_log(td, td->slat_log, try);
+       if (!ret)
+               td->slat_log = NULL;
+
+       return ret;
+}
+
+static int write_clat_log(struct thread_data *td, int try, bool unit_log)
+{
+       int ret;
+
+       if (!unit_log)
+               return 0;
+
+       ret = __write_log(td, td->clat_log, try);
+       if (!ret)
+               td->clat_log = NULL;
+
+       return ret;
 }
 
-void finish_log(struct thread_data *td, struct io_log *log, const char *name)
+static int write_lat_log(struct thread_data *td, int try, bool unit_log)
 {
-       finish_log_named(td, log, td->o.name, name);
+       int ret;
+
+       if (!unit_log)
+               return 0;
+
+       ret = __write_log(td, td->lat_log, try);
+       if (!ret)
+               td->lat_log = NULL;
+
+       return ret;
+}
+
+static int write_bandw_log(struct thread_data *td, int try, bool unit_log)
+{
+       int ret;
+
+       if (per_unit_log(td->bw_log) != unit_log)
+               return 0;
+
+       ret = __write_log(td, td->bw_log, try);
+       if (!ret)
+               td->bw_log = NULL;
+
+       return ret;
+}
+
+enum {
+       BW_LOG_MASK     = 1,
+       LAT_LOG_MASK    = 2,
+       SLAT_LOG_MASK   = 4,
+       CLAT_LOG_MASK   = 8,
+       IOPS_LOG_MASK   = 16,
+
+       ALL_LOG_NR      = 5,
+};
+
+struct log_type {
+       unsigned int mask;
+       int (*fn)(struct thread_data *, int, bool);
+};
+
+static struct log_type log_types[] = {
+       {
+               .mask   = BW_LOG_MASK,
+               .fn     = write_bandw_log,
+       },
+       {
+               .mask   = LAT_LOG_MASK,
+               .fn     = write_lat_log,
+       },
+       {
+               .mask   = SLAT_LOG_MASK,
+               .fn     = write_slat_log,
+       },
+       {
+               .mask   = CLAT_LOG_MASK,
+               .fn     = write_clat_log,
+       },
+       {
+               .mask   = IOPS_LOG_MASK,
+               .fn     = write_iops_log,
+       },
+};
+
+void td_writeout_logs(struct thread_data *td, bool unit_logs)
+{
+       unsigned int log_mask = 0;
+       unsigned int log_left = ALL_LOG_NR;
+       int old_state, i;
+
+       old_state = td_bump_runstate(td, TD_FINISHING);
+
+       finalize_logs(td, unit_logs);
+
+       while (log_left) {
+               int prev_log_left = log_left;
+
+               for (i = 0; i < ALL_LOG_NR && log_left; i++) {
+                       struct log_type *lt = &log_types[i];
+                       int ret;
+
+                       if (!(log_mask & lt->mask)) {
+                               ret = lt->fn(td, log_left != 1, unit_logs);
+                               if (!ret) {
+                                       log_left--;
+                                       log_mask |= lt->mask;
+                               }
+                       }
+               }
+
+               if (prev_log_left == log_left)
+                       usleep(5000);
+       }
+
+       td_restore_runstate(td, old_state);
+}
+
+void fio_writeout_logs(bool unit_logs)
+{
+       struct thread_data *td;
+       int i;
+
+       for_each_td(td, i)
+               td_writeout_logs(td, unit_logs);
 }