Add cross-stripe intel sample verify job
[fio.git] / iolog.c
diff --git a/iolog.c b/iolog.c
index 3723e0a86025afe460c11a7585e14a365f37022b..b72dcf9791cfa242ceace8b4080a03934e9fa554 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"
+
+#include <netinet/in.h>
+#include <netinet/tcp.h>
+#include <arpa/inet.h>
+#include <sys/stat.h>
+#include <sys/socket.h>
+#include <sys/un.h>
 
 static int iolog_flush(struct io_log *log);
 
@@ -35,7 +42,7 @@ void log_io_u(const struct thread_data *td, const struct io_u *io_u)
        if (!td->o.write_iolog_file)
                return;
 
-       fprintf(td->iolog_f, "%s %s %llu %lu\n", io_u->file->file_name,
+       fprintf(td->iolog_f, "%s %s %llu %llu\n", io_u->file->file_name,
                                                io_ddir_name(io_u->ddir),
                                                io_u->offset, io_u->buflen);
 }
@@ -63,8 +70,9 @@ 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;
@@ -77,7 +85,7 @@ 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)
@@ -87,9 +95,9 @@ 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;
 }
@@ -109,6 +117,11 @@ 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;
@@ -128,6 +141,8 @@ static int ipo_special(struct thread_data *td, struct io_piece *ipo)
        return 1;
 }
 
+static bool read_iolog2(struct thread_data *td);
+
 int read_iolog_get(struct thread_data *td, struct io_u *io_u)
 {
        struct io_piece *ipo;
@@ -135,7 +150,13 @@ int read_iolog_get(struct thread_data *td, struct io_u *io_u)
 
        while (!flist_empty(&td->io_log_list)) {
                int ret;
-
+               if (td->o.read_iolog_chunked) {
+                       if (td->io_log_checkmark == td->io_log_current) {
+                               if (!read_iolog2(td))
+                                       return 1;
+                       }
+                       td->io_log_current--;
+               }
                ipo = flist_first_entry(&td->io_log_list, struct io_piece, list);
                flist_del(&ipo->list);
                remove_trim_entry(td, ipo);
@@ -155,7 +176,7 @@ int read_iolog_get(struct thread_data *td, struct io_u *io_u)
                        io_u->buflen = ipo->len;
                        io_u->file = td->files[ipo->fileno];
                        get_file(io_u->file);
-                       dprint(FD_IO, "iolog: get %llu/%lu/%s\n", io_u->offset,
+                       dprint(FD_IO, "iolog: get %llu/%llu/%s\n", io_u->offset,
                                                io_u->buflen, io_u->file->file_name);
                        if (ipo->delay)
                                iolog_delay(td, ipo->delay);
@@ -178,7 +199,7 @@ 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);
@@ -202,10 +223,10 @@ 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;
@@ -221,21 +242,11 @@ 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;
@@ -272,13 +283,14 @@ 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;
                }
        }
@@ -318,7 +330,7 @@ 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;
 
@@ -337,36 +349,74 @@ void write_iolog_close(struct thread_data *td)
        td->iolog_buf = NULL;
 }
 
+static int64_t iolog_items_to_fetch(struct thread_data *td)
+{
+       struct timespec now;
+       uint64_t elapsed;
+       uint64_t for_1s;
+       int64_t items_to_fetch;
+
+       if (!td->io_log_highmark)
+               return 10;
+
+
+       fio_gettime(&now, NULL);
+       elapsed = ntime_since(&td->io_log_highmark_time, &now);
+       if (elapsed) {
+               for_1s = (td->io_log_highmark - td->io_log_current) * 1000000000 / elapsed;
+               items_to_fetch = for_1s - td->io_log_current;
+               if (items_to_fetch < 0)
+                       items_to_fetch = 0;
+       } else
+               items_to_fetch = 0;
+
+       td->io_log_highmark = td->io_log_current + items_to_fetch;
+       td->io_log_checkmark = (td->io_log_highmark + 1) / 2;
+       fio_gettime(&td->io_log_highmark_time, NULL);
+
+       return items_to_fetch;
+}
+
 /*
  * 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)
 {
        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;
+       bool realloc = false;
+       int64_t items_to_fetch = 0;
 
-       free_release_files(td);
+       if (td->o.read_iolog_chunked) {
+               items_to_fetch = iolog_items_to_fetch(td);
+               if (!items_to_fetch)
+                       return true;
+       }
 
        /*
         * Read in the read iolog and store it, reuse the infrastructure
         * for doing verifications.
         */
        str = malloc(4096);
-       fname = malloc(256+16);
+       rfname = fname = malloc(256+16);
        act = malloc(256+16);
 
        reads = writes = waits = 0;
-       while ((p = fgets(str, 4096, f)) != NULL) {
+       while ((p = fgets(str, 4096, td->io_log_rfile)) != NULL) {
                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
@@ -392,8 +442,14 @@ 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, td->subjob_number, 1);
+                                       file_action = FIO_LOG_ADD_FILE;
+                               }
                                continue;
                        } else if (!strcmp(act, "open")) {
                                fileno = get_fileno(td, fname);
@@ -407,7 +463,7 @@ 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;
                }
 
@@ -421,6 +477,8 @@ 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)) {
@@ -431,27 +489,47 @@ 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])
+                       if (rw != DDIR_INVAL && bytes > td->o.max_bs[rw]) {
+                               realloc = true;
                                td->o.max_bs[rw] = bytes;
+                       }
                        ipo->fileno = fileno;
                        ipo->file_action = file_action;
                        td->o.size += bytes;
                }
 
                queue_io_piece(td, ipo);
+
+               if (td->o.read_iolog_chunked) {
+                       td->io_log_current++;
+                       items_to_fetch--;
+                       if (items_to_fetch == 0)
+                               break;
+               }
        }
 
        free(str);
        free(act);
-       free(fname);
+       free(rfname);
+
+       if (td->o.read_iolog_chunked) {
+               td->io_log_highmark = td->io_log_current;
+               td->io_log_checkmark = (td->io_log_highmark + 1) / 2;
+               fio_gettime(&td->io_log_highmark_time, NULL);
+       }
 
        if (writes && read_only) {
                log_err("fio: <%s> skips replay of %d writes due to"
@@ -459,8 +537,22 @@ static int read_iolog2(struct thread_data *td, FILE *f)
                writes = 0;
        }
 
+       if (td->o.read_iolog_chunked) {
+               if (td->io_log_current == 0) {
+                       return false;
+               }
+               td->o.td_ddir = TD_DDIR_RW;
+               if (realloc && td->orig_buffer)
+               {
+                       io_u_quiesce(td);
+                       free_io_mem(td);
+                       init_io_u_buffers(td);
+               }
+               return true;
+       }
+
        if (!reads && !writes && !waits)
-               return 1;
+               return false;
        else if (reads && !writes)
                td->o.td_ddir = TD_DDIR_READ;
        else if (!reads && writes)
@@ -468,22 +560,70 @@ static int read_iolog2(struct thread_data *td, FILE *f)
        else
                td->o.td_ddir = TD_DDIR_RW;
 
-       return 0;
+       return true;
+}
+
+static bool is_socket(const char *path)
+{
+       struct stat buf;
+       int r;
+
+       r = stat(path, &buf);
+       if (r == -1)
+               return false;
+
+       return S_ISSOCK(buf.st_mode);
+}
+
+static int open_socket(const char *path)
+{
+       struct sockaddr_un addr;
+       int ret, fd;
+
+       fd = socket(AF_UNIX, SOCK_STREAM, 0);
+       if (fd < 0)
+               return fd;
+
+       addr.sun_family = AF_UNIX;
+       if (snprintf(addr.sun_path, sizeof(addr.sun_path), "%s", path) >=
+           sizeof(addr.sun_path)) {
+               log_err("%s: path name %s is too long for a Unix socket\n",
+                       __func__, path);
+       }
+
+       ret = connect(fd, (const struct sockaddr *)&addr, strlen(path) + sizeof(addr.sun_family));
+       if (!ret)
+               return fd;
+
+       close(fd);
+       return -1;
 }
 
 /*
  * 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;
+       char buffer[256], *p, *fname;
+       FILE *f = NULL;
+
+       fname = get_name_by_idx(td->o.read_iolog_file, td->subjob_number);
+       dprint(FD_IO, "iolog: name=%s\n", fname);
+
+       if (is_socket(fname)) {
+               int fd;
+
+               fd = open_socket(fname);
+               if (fd >= 0)
+                       f = fdopen(fd, "r");
+       } else
+               f = fopen(fname, "r");
+
+       free(fname);
 
-       f = fopen(td->o.read_iolog_file, "r");
        if (!f) {
                perror("fopen read iolog");
-               return 1;
+               return false;
        }
 
        p = fgets(buffer, sizeof(buffer), f);
@@ -491,28 +631,28 @@ 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;
        }
 
        /*
         * version 2 of the iolog stores a specific string as the
         * first line, check for that
         */
-       if (!strncmp(iolog_ver2, buffer, strlen(iolog_ver2)))
-               ret = read_iolog2(td, f);
-       else {
-               log_err("fio: iolog version 1 is no longer supported\n");
-               ret = 1;
+       if (!strncmp(iolog_ver2, buffer, strlen(iolog_ver2))) {
+               free_release_files(td);
+               td->io_log_rfile = f;
+               return read_iolog2(td);
        }
 
+       log_err("fio: iolog version 1 is no longer supported\n");
        fclose(f);
-       return ret;
+       return false;
 }
 
 /*
  * 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;
@@ -521,7 +661,7 @@ 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;
        }
 
        /*
@@ -536,7 +676,7 @@ 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;
        }
 
        /*
@@ -545,12 +685,12 @@ 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;
@@ -565,8 +705,10 @@ 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;
@@ -576,6 +718,9 @@ 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 = scalloc(1, sizeof(*l));
        INIT_FLIST_HEAD(&l->io_logs);
@@ -584,16 +729,28 @@ void setup_log(struct io_log **log, struct log_params *p,
        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;
+               struct io_logs *__p;
 
-               p = calloc(1, sizeof(*l->pending));
-               p->max_samples = l->td->o.iodepth;
-               p->log = calloc(p->max_samples, log_entry_sz(l));
-               l->pending = 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)
@@ -604,7 +761,8 @@ 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;
        }
 
@@ -645,6 +803,7 @@ void free_log(struct io_log *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) {
@@ -658,6 +817,67 @@ void free_log(struct io_log *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, %llu, ", (unsigned long) s->time,
+                                               io_sample_ddir(s), (unsigned long long) 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)
 {
        struct io_sample *s;
@@ -676,17 +896,17 @@ 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, %llu\n",
                                        (unsigned long) s->time,
-                                       (unsigned long) s->val,
-                                       io_sample_ddir(s), s->bs);
+                                       s->data.val,
+                                       io_sample_ddir(s), (unsigned long long) s->bs);
                } else {
                        struct io_sample_offset *so = (void *) s;
 
-                       fprintf(f, "%lu, %lu, %u, %u, %llu\n",
+                       fprintf(f, "%lu, %" PRId64 ", %u, %llu, %llu\n",
                                        (unsigned long) s->time,
-                                       (unsigned long) s->val,
-                                       io_sample_ddir(s), s->bs,
+                                       s->data.val,
+                                       io_sample_ddir(s), (unsigned long long) s->bs,
                                        (unsigned long long) so->offset);
                }
        }
@@ -882,7 +1102,7 @@ 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;
@@ -904,12 +1124,12 @@ 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);
@@ -964,7 +1184,7 @@ 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;
@@ -987,7 +1207,14 @@ void flush_log(struct io_log *log, int do_append)
 
                cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
                flist_del_init(&cur_log->list);
-               flush_samples(f, cur_log->log, cur_log->nr_samples * log_entry_sz(log));
+               
+               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);
@@ -1037,6 +1264,38 @@ size_t log_chunk_sizes(struct io_log *log)
 
 #ifdef CONFIG_ZLIB
 
+static void iolog_put_deferred(struct io_log *log, void *ptr)
+{
+       if (!ptr)
+               return;
+
+       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);
+}
+
+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);
+}
+
 static int gz_work(struct iolog_flush_data *data)
 {
        struct iolog_compress *c = NULL;
@@ -1069,7 +1328,8 @@ static int gz_work(struct iolog_flush_data *data)
                                data->log->filename);
        do {
                if (c)
-                       dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq, c->len);
+                       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;
@@ -1106,7 +1366,7 @@ static int gz_work(struct iolog_flush_data *data)
        total -= c->len;
        c->len = GZ_CHUNK - stream.avail_out;
        total += c->len;
-       dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq, c->len);
+       dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq, (unsigned long) c->len);
 
        if (ret != Z_STREAM_END) {
                do {
@@ -1117,7 +1377,8 @@ 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, c->len);
+                       dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq,
+                               (unsigned long) c->len);
                } while (ret != Z_STREAM_END);
        }
 
@@ -1127,7 +1388,7 @@ static int gz_work(struct iolog_flush_data *data)
        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);
@@ -1138,7 +1399,7 @@ static int gz_work(struct iolog_flush_data *data)
        ret = 0;
 done:
        if (data->free)
-               free(data);
+               sfree(data);
        return ret;
 err:
        while (!flist_empty(&list)) {
@@ -1226,9 +1487,7 @@ static int iolog_flush(struct io_log *log)
                data->samples = cur_log->log;
                data->nr_samples = cur_log->nr_samples;
 
-               cur_log->nr_samples = 0;
-               cur_log->max_samples = 0;
-               cur_log->log = NULL;
+               sfree(cur_log);
 
                gz_work(data);
        }
@@ -1241,7 +1500,7 @@ int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
 {
        struct iolog_flush_data *data;
 
-       data = malloc(sizeof(*data));
+       data = smalloc(sizeof(*data));
        if (!data)
                return 1;
 
@@ -1255,6 +1514,9 @@ int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
        cur_log->log = NULL;
 
        workqueue_enqueue(&log->td->log_compress_wq, &data->work);
+
+       iolog_free_deferred(log);
+
        return 0;
 }
 #else
@@ -1353,6 +1615,20 @@ 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;
@@ -1387,8 +1663,9 @@ 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 {
@@ -1417,6 +1694,10 @@ 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)