iolog: remove 'td' from trim_io_piece()
[fio.git] / iolog.c
diff --git a/iolog.c b/iolog.c
index feda9ed68fbcc0414556d91c307378300c00d966..d7474724f0d140cd4e66b5136f47aedb2002ddcd 100644 (file)
--- a/iolog.c
+++ b/iolog.c
@@ -4,7 +4,6 @@
  */
 #include <stdio.h>
 #include <stdlib.h>
-#include <libgen.h>
 #include <assert.h>
 #include <sys/types.h>
 #include <sys/stat.h>
 
 #include "flist.h"
 #include "fio.h"
-#include "verify.h"
 #include "trim.h"
 #include "filelock.h"
+#include "smalloc.h"
+#include "blktrace.h"
+#include "pshared.h"
+
+static int iolog_flush(struct io_log *log);
 
 static const char iolog_ver2[] = "fio version 2 iolog";
 
@@ -60,8 +63,9 @@ void log_file(struct thread_data *td, struct fio_file *f,
 static void iolog_delay(struct thread_data *td, unsigned long delay)
 {
        uint64_t usec = utime_since_now(&td->last_issue);
+       unsigned long orig_delay = delay;
        uint64_t this_delay;
-       struct timeval tv;
+       struct timespec ts;
 
        if (delay < td->time_offset) {
                td->time_offset = 0;
@@ -74,7 +78,7 @@ static void iolog_delay(struct thread_data *td, unsigned long delay)
 
        delay -= usec;
 
-       fio_gettime(&tv, NULL);
+       fio_gettime(&ts, NULL);
        while (delay && !td->terminate) {
                this_delay = delay;
                if (this_delay > 500000)
@@ -84,9 +88,9 @@ static void iolog_delay(struct thread_data *td, unsigned long delay)
                delay -= this_delay;
        }
 
-       usec = utime_since_now(&tv);
-       if (usec > delay)
-               td->time_offset = usec - delay;
+       usec = utime_since_now(&ts);
+       if (usec > orig_delay)
+               td->time_offset = usec - orig_delay;
        else
                td->time_offset = 0;
 }
@@ -106,6 +110,11 @@ static int ipo_special(struct thread_data *td, struct io_piece *ipo)
 
        switch (ipo->file_action) {
        case FIO_LOG_OPEN_FILE:
+               if (td->o.replay_redirect && fio_file_open(f)) {
+                       dprint(FD_FILE, "iolog: ignoring re-open of file %s\n",
+                                       f->file_name);
+                       break;
+               }
                ret = td_io_open_file(td, f);
                if (!ret)
                        break;
@@ -175,7 +184,7 @@ int read_iolog_get(struct thread_data *td, struct io_u *io_u)
 void prune_io_piece_log(struct thread_data *td)
 {
        struct io_piece *ipo;
-       struct rb_node *n;
+       struct fio_rb_node *n;
 
        while ((n = rb_first(&td->io_hist_tree)) != NULL) {
                ipo = rb_entry(n, struct io_piece, rb_node);
@@ -199,10 +208,10 @@ void prune_io_piece_log(struct thread_data *td)
  */
 void log_io_piece(struct thread_data *td, struct io_u *io_u)
 {
-       struct rb_node **p, *parent;
+       struct fio_rb_node **p, *parent;
        struct io_piece *ipo, *__ipo;
 
-       ipo = malloc(sizeof(struct io_piece));
+       ipo = calloc(1, sizeof(struct io_piece));
        init_ipo(ipo);
        ipo->file = io_u->file;
        ipo->offset = io_u->offset;
@@ -218,21 +227,11 @@ void log_io_piece(struct thread_data *td, struct io_u *io_u)
        }
 
        /*
-        * We don't need to sort the entries, if:
-        *
-        *      Sequential writes, or
-        *      Random writes that lay out the file as it goes along
-        *
-        * For both these cases, just reading back data in the order we
-        * wrote it out is the fastest.
-        *
-        * One exception is if we don't have a random map AND we are doing
-        * verifies, in that case we need to check for duplicate blocks and
-        * drop the old one, which we rely on the rb insert/lookup for
-        * handling.
+        * Only sort writes if we don't have a random map in which case we need
+        * to check for duplicate blocks and drop the old one, which we rely on
+        * the rb insert/lookup for handling.
         */
-       if (((!td->o.verifysort) || !td_random(td) || !td->o.overwrite) &&
-             (file_randommap(td, ipo->file) || td->o.verify == VERIFY_NONE)) {
+       if (file_randommap(td, ipo->file)) {
                INIT_FLIST_HEAD(&ipo->list);
                flist_add_tail(&ipo->list, &td->io_hist_list);
                ipo->flags |= IP_F_ONLIST;
@@ -269,13 +268,14 @@ restart:
                        overlap = 1;
 
                if (overlap) {
-                       dprint(FD_IO, "iolog: overlap %llu/%lu, %llu/%lu",
+                       dprint(FD_IO, "iolog: overlap %llu/%lu, %llu/%lu\n",
                                __ipo->offset, __ipo->len,
                                ipo->offset, ipo->len);
                        td->io_hist_len--;
                        rb_erase(parent, &td->io_hist_tree);
                        remove_trim_entry(td, __ipo);
-                       free(__ipo);
+                       if (!(__ipo->flags & IP_F_IN_FLIGHT))
+                               free(__ipo);
                        goto restart;
                }
        }
@@ -315,7 +315,7 @@ void unlog_io_piece(struct thread_data *td, struct io_u *io_u)
        td->io_hist_len--;
 }
 
-void trim_io_piece(struct thread_data *td, const struct io_u *io_u)
+void trim_io_piece(const struct io_u *io_u)
 {
        struct io_piece *ipo = io_u->ipo;
 
@@ -338,12 +338,12 @@ void write_iolog_close(struct thread_data *td)
  * Read version 2 iolog data. It is enhanced to include per-file logging,
  * syncs, etc.
  */
-static int read_iolog2(struct thread_data *td, FILE *f)
+static bool 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;
 
@@ -354,7 +354,7 @@ 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;
@@ -362,8 +362,12 @@ 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
@@ -389,8 +393,14 @@ static int read_iolog2(struct thread_data *td, FILE *f)
                } else if (r == 2) {
                        rw = DDIR_INVAL;
                        if (!strcmp(act, "add")) {
-                               fileno = add_file(td, fname, 0, 1);
-                               file_action = FIO_LOG_ADD_FILE;
+                               if (td->o.replay_redirect &&
+                                   get_fileno(td, fname) != -1) {
+                                       dprint(FD_FILE, "iolog: ignoring"
+                                               " re-add of file %s\n", fname);
+                               } else {
+                                       fileno = add_file(td, fname, 0, 1);
+                                       file_action = FIO_LOG_ADD_FILE;
+                               }
                                continue;
                        } else if (!strcmp(act, "open")) {
                                fileno = get_fileno(td, fname);
@@ -404,7 +414,7 @@ static int read_iolog2(struct thread_data *td, FILE *f)
                                continue;
                        }
                } else {
-                       log_err("bad iolog2: %s", p);
+                       log_err("bad iolog2: %s\n", p);
                        continue;
                }
 
@@ -418,6 +428,8 @@ static int read_iolog2(struct thread_data *td, FILE *f)
                                continue;
                        writes++;
                } else if (rw == DDIR_WAIT) {
+                       if (td->o.no_stall)
+                               continue;
                        waits++;
                } else if (rw == DDIR_INVAL) {
                } else if (!ddir_sync(rw)) {
@@ -428,13 +440,18 @@ static int read_iolog2(struct thread_data *td, FILE *f)
                /*
                 * Make note of file
                 */
-               ipo = malloc(sizeof(*ipo));
+               ipo = calloc(1, sizeof(*ipo));
                init_ipo(ipo);
                ipo->ddir = rw;
                if (rw == DDIR_WAIT) {
                        ipo->delay = offset;
                } else {
-                       ipo->offset = offset;
+                       if (td->o.replay_scale)
+                               ipo->offset = offset / td->o.replay_scale;
+                       else
+                               ipo->offset = offset;
+                       ipo_bytes_align(td->o.replay_align, ipo);
+
                        ipo->len = bytes;
                        if (rw != DDIR_INVAL && bytes > td->o.max_bs[rw])
                                td->o.max_bs[rw] = bytes;
@@ -448,7 +465,7 @@ 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"
@@ -457,7 +474,7 @@ static int read_iolog2(struct thread_data *td, FILE *f)
        }
 
        if (!reads && !writes && !waits)
-               return 1;
+               return false;
        else if (reads && !writes)
                td->o.td_ddir = TD_DDIR_READ;
        else if (!reads && writes)
@@ -465,22 +482,22 @@ static int read_iolog2(struct thread_data *td, FILE *f)
        else
                td->o.td_ddir = TD_DDIR_RW;
 
-       return 0;
+       return true;
 }
 
 /*
  * open iolog, check version, and call appropriate parser
  */
-static int init_iolog_read(struct thread_data *td)
+static bool init_iolog_read(struct thread_data *td)
 {
        char buffer[256], *p;
        FILE *f;
-       int ret;
+       bool ret;
 
        f = fopen(td->o.read_iolog_file, "r");
        if (!f) {
                perror("fopen read iolog");
-               return 1;
+               return false;
        }
 
        p = fgets(buffer, sizeof(buffer), f);
@@ -488,7 +505,7 @@ static int init_iolog_read(struct thread_data *td)
                td_verror(td, errno, "iolog read");
                log_err("fio: unable to read iolog\n");
                fclose(f);
-               return 1;
+               return false;
        }
 
        /*
@@ -499,7 +516,7 @@ static int init_iolog_read(struct thread_data *td)
                ret = read_iolog2(td, f);
        else {
                log_err("fio: iolog version 1 is no longer supported\n");
-               ret = 1;
+               ret = false;
        }
 
        fclose(f);
@@ -509,7 +526,7 @@ static int init_iolog_read(struct thread_data *td)
 /*
  * Set up a log for storing io patterns.
  */
-static int init_iolog_write(struct thread_data *td)
+static bool init_iolog_write(struct thread_data *td)
 {
        struct fio_file *ff;
        FILE *f;
@@ -518,7 +535,7 @@ static int init_iolog_write(struct thread_data *td)
        f = fopen(td->o.write_iolog_file, "a");
        if (!f) {
                perror("fopen write iolog");
-               return 1;
+               return false;
        }
 
        /*
@@ -533,7 +550,7 @@ static int init_iolog_write(struct thread_data *td)
         */
        if (fprintf(f, "%s\n", iolog_ver2) < 0) {
                perror("iolog init\n");
-               return 1;
+               return false;
        }
 
        /*
@@ -542,12 +559,12 @@ static int init_iolog_write(struct thread_data *td)
        for_each_file(td, ff, i)
                log_file(td, ff, FIO_LOG_ADD_FILE);
 
-       return 0;
+       return true;
 }
 
-int init_iolog(struct thread_data *td)
+bool init_iolog(struct thread_data *td)
 {
-       int ret = 0;
+       bool ret;
 
        if (td->o.read_iolog_file) {
                int need_swap;
@@ -562,8 +579,10 @@ int init_iolog(struct thread_data *td)
                        ret = init_iolog_read(td);
        } else if (td->o.write_iolog_file)
                ret = init_iolog_write(td);
+       else
+               ret = true;
 
-       if (ret)
+       if (!ret)
                td_verror(td, EINVAL, "failed initializing iolog");
 
        return ret;
@@ -573,19 +592,41 @@ 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 = calloc(1, sizeof(*l));
-       l->nr_samples = 0;
-       l->max_samples = 1024;
+       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->log = malloc(l->max_samples * log_entry_sz(l));
        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;
 
@@ -594,7 +635,8 @@ 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);
+               mutex_init_pshared(&l->deferred_free_lock);
                p->td->flags |= TD_F_COMPRESS_LOG;
        }
 
@@ -629,9 +671,85 @@ 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);
-       free(log);
+       sfree(log);
+}
+
+uint64_t hist_sum(int j, int stride, uint64_t *io_u_plat,
+               uint64_t *io_u_plat_last)
+{
+       uint64_t 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;
+       uint64_t *io_u_plat;
+       uint64_t *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 = s->data.plat_entry;
+               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, "%llu, ", (unsigned long long)
+                               hist_sum(j, stride, io_u_plat, io_u_plat_before));
+               }
+               fprintf(f, "%llu\n", (unsigned long 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)
@@ -652,16 +770,16 @@ void flush_samples(FILE *f, void *samples, uint64_t sample_size)
                s = __get_sample(samples, log_offset, i);
 
                if (!log_offset) {
-                       fprintf(f, "%lu, %lu, %u, %u\n",
+                       fprintf(f, "%lu, %" PRId64 ", %u, %u\n",
                                        (unsigned long) s->time,
-                                       (unsigned long) s->val,
+                                       s->data.val,
                                        io_sample_ddir(s), s->bs);
                } else {
                        struct io_sample_offset *so = (void *) s;
 
-                       fprintf(f, "%lu, %lu, %u, %u, %llu\n",
+                       fprintf(f, "%lu, %" PRId64 ", %u, %u, %llu\n",
                                        (unsigned long) s->time,
-                                       (unsigned long) s->val,
+                                       s->data.val,
                                        io_sample_ddir(s), s->bs,
                                        (unsigned long long) so->offset);
                }
@@ -672,14 +790,10 @@ void flush_samples(FILE *f, void *samples, uint64_t sample_size)
 
 struct iolog_flush_data {
        struct workqueue_work work;
-       pthread_mutex_t lock;
-       pthread_cond_t cv;
-       int wait;
-       volatile int done;
-       volatile int refs;
        struct io_log *log;
        void *samples;
-       uint64_t nr_samples;
+       uint32_t nr_samples;
+       bool free;
 };
 
 #define GZ_CHUNK       131072
@@ -706,6 +820,7 @@ 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;
@@ -740,7 +855,8 @@ 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);
@@ -757,7 +873,7 @@ 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) {
@@ -804,7 +920,7 @@ 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;
 }
@@ -860,7 +976,7 @@ int iolog_file_inflate(const char *file)
        struct iolog_compress ic;
        z_stream stream;
        struct stat sb;
-       ssize_t ret;
+       size_t ret;
        size_t total;
        void *buf;
        FILE *f;
@@ -882,12 +998,12 @@ int iolog_file_inflate(const char *file)
        ic.seq = 1;
 
        ret = fread(ic.buf, ic.len, 1, f);
-       if (ret < 0) {
+       if (ret == 0 && ferror(f)) {
                perror("fread");
                fclose(f);
                free(buf);
                return 1;
-       } else if (ret != 1) {
+       } else if (ferror(f) || (!feof(f) && ret != 1)) {
                log_err("fio: short read on reading log\n");
                fclose(f);
                free(buf);
@@ -942,7 +1058,7 @@ 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;
@@ -960,7 +1076,20 @@ 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);
@@ -969,7 +1098,7 @@ 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))
@@ -1009,29 +1138,41 @@ size_t log_chunk_sizes(struct io_log *log)
 
 #ifdef CONFIG_ZLIB
 
-static void drop_data_unlock(struct iolog_flush_data *data)
+static void iolog_put_deferred(struct io_log *log, void *ptr)
 {
-       int refs;
+       if (!ptr)
+               return;
 
-       refs = --data->refs;
-       pthread_mutex_unlock(&data->lock);
+       pthread_mutex_lock(&log->deferred_free_lock);
+       if (log->deferred < IOLOG_MAX_DEFER) {
+               log->deferred_items[log->deferred] = ptr;
+               log->deferred++;
+       } else if (!fio_did_warn(FIO_WARN_IOLOG_DROP))
+               log_err("fio: had to drop log entry free\n");
+       pthread_mutex_unlock(&log->deferred_free_lock);
+}
 
-       if (!refs) {
-               free(data);
-               pthread_mutex_destroy(&data->lock);
-               pthread_cond_destroy(&data->cv);
+static void iolog_free_deferred(struct io_log *log)
+{
+       int i;
+
+       if (!log->deferred)
+               return;
+
+       pthread_mutex_lock(&log->deferred_free_lock);
+
+       for (i = 0; i < log->deferred; i++) {
+               free(log->deferred_items[i]);
+               log->deferred_items[i] = NULL;
        }
+
+       log->deferred = 0;
+       pthread_mutex_unlock(&log->deferred_free_lock);
 }
 
-/*
- * Invoked from our compress helper thread, when logging would have exceeded
- * the specified memory limitation. Compresses the previously stored
- * entries.
- */
-static int gz_work(struct submit_worker *sw, struct workqueue_work *work)
+static int gz_work(struct iolog_flush_data *data)
 {
-       struct iolog_flush_data *data;
-       struct iolog_compress *c;
+       struct iolog_compress *c = NULL;
        struct flist_head list;
        unsigned int seq;
        z_stream stream;
@@ -1040,8 +1181,7 @@ static int gz_work(struct submit_worker *sw, struct workqueue_work *work)
 
        INIT_FLIST_HEAD(&list);
 
-       data = container_of(work, struct iolog_flush_data, work);
-
+       memset(&stream, 0, sizeof(stream));
        stream.zalloc = Z_NULL;
        stream.zfree = Z_NULL;
        stream.opaque = Z_NULL;
@@ -1049,7 +1189,7 @@ static int gz_work(struct submit_worker *sw, struct workqueue_work *work)
        ret = deflateInit(&stream, Z_DEFAULT_COMPRESSION);
        if (ret != Z_OK) {
                log_err("fio: failed to init gz stream\n");
-               return 0;
+               goto err;
        }
 
        seq = ++data->log->chunk_seq;
@@ -1057,9 +1197,13 @@ static int gz_work(struct submit_worker *sw, struct workqueue_work *work)
        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;
@@ -1079,9 +1223,26 @@ static int gz_work(struct submit_worker *sw, struct workqueue_work *work)
        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;
@@ -1090,6 +1251,8 @@ static int gz_work(struct submit_worker *sw, struct workqueue_work *work)
                        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);
        }
 
@@ -1099,7 +1262,7 @@ static int gz_work(struct submit_worker *sw, struct workqueue_work *work)
        if (ret != Z_OK)
                log_err("fio: deflateEnd %d\n", ret);
 
-       free(data->samples);
+       iolog_put_deferred(data->log, data->samples);
 
        if (!flist_empty(&list)) {
                pthread_mutex_lock(&data->log->chunk_lock);
@@ -1109,14 +1272,8 @@ static int gz_work(struct submit_worker *sw, struct workqueue_work *work)
 
        ret = 0;
 done:
-       if (data->wait) {
-               pthread_mutex_lock(&data->lock);
-               data->done = 1;
-               pthread_cond_signal(&data->cv);
-
-               drop_data_unlock(data);
-       } else
-               free(data);
+       if (data->free)
+               sfree(data);
        return ret;
 err:
        while (!flist_empty(&list)) {
@@ -1128,6 +1285,16 @@ err:
        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;
@@ -1144,7 +1311,7 @@ static int gz_init_worker(struct submit_worker *sw)
 }
 
 static struct workqueue_ops log_compress_wq_ops = {
-       .fn             = gz_work,
+       .fn             = gz_work_async,
        .init_worker_fn = gz_init_worker,
        .nice           = 1,
 };
@@ -1170,52 +1337,70 @@ 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 = 128;
-       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->wait = wait;
-       if (data->wait) {
-               pthread_mutex_init(&data->lock, NULL);
-               pthread_cond_init(&data->cv, NULL);
-               data->done = 0;
-               data->refs = 2;
-       }
+               data->samples = cur_log->log;
+               data->nr_samples = cur_log->nr_samples;
 
-       workqueue_enqueue(&log->td->log_compress_wq, &data->work);
+               sfree(cur_log);
 
-       if (wait) {
-               pthread_mutex_lock(&data->lock);
-               while (!data->done)
-                       pthread_cond_wait(&data->cv, &data->lock);
-
-               drop_data_unlock(data);
+               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 = smalloc(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);
+
+       iolog_free_deferred(log);
+
+       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;
 }
@@ -1231,6 +1416,29 @@ 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)
@@ -1239,29 +1447,88 @@ static int __write_log(struct thread_data *td, struct io_log *log, int try)
        return 0;
 }
 
-static int write_iops_log(struct thread_data *td, int try)
+static int write_iops_log(struct thread_data *td, int try, bool unit_log)
 {
-       return __write_log(td, td->iops_log, try);
+       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)
+static int write_slat_log(struct thread_data *td, int try, bool unit_log)
 {
-       return __write_log(td, td->slat_log, try);
+       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)
+static int write_clat_log(struct thread_data *td, int try, bool unit_log)
 {
-       return __write_log(td, td->clat_log, try);
+       int ret;
+
+       if (!unit_log)
+               return 0;
+
+       ret = __write_log(td, td->clat_log, try);
+       if (!ret)
+               td->clat_log = NULL;
+
+       return ret;
 }
 
-static int write_lat_log(struct thread_data *td, int try)
+static int write_clat_hist_log(struct thread_data *td, int try, bool unit_log)
 {
-       return __write_log(td, td->lat_log, try);
+       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_bandw_log(struct thread_data *td, int try)
+static int write_lat_log(struct thread_data *td, int try, bool unit_log)
 {
-       return __write_log(td, td->bw_log, try);
+       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 {
@@ -1270,13 +1537,14 @@ 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 {
        unsigned int mask;
-       int (*fn)(struct thread_data *, int);
+       int (*fn)(struct thread_data *, int, bool);
 };
 
 static struct log_type log_types[] = {
@@ -1300,9 +1568,13 @@ 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 fio_writeout_logs(struct thread_data *td)
+void td_writeout_logs(struct thread_data *td, bool unit_logs)
 {
        unsigned int log_mask = 0;
        unsigned int log_left = ALL_LOG_NR;
@@ -1310,7 +1582,7 @@ void fio_writeout_logs(struct thread_data *td)
 
        old_state = td_bump_runstate(td, TD_FINISHING);
 
-       finalize_logs(td);
+       finalize_logs(td, unit_logs);
 
        while (log_left) {
                int prev_log_left = log_left;
@@ -1320,7 +1592,7 @@ void fio_writeout_logs(struct thread_data *td)
                        int ret;
 
                        if (!(log_mask & lt->mask)) {
-                               ret = lt->fn(td, log_left != 1);
+                               ret = lt->fn(td, log_left != 1, unit_logs);
                                if (!ret) {
                                        log_left--;
                                        log_mask |= lt->mask;
@@ -1334,3 +1606,12 @@ void fio_writeout_logs(struct thread_data *td)
 
        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);
+}