iolog: dprint() casts for 32-bit warnings
[fio.git] / iolog.c
diff --git a/iolog.c b/iolog.c
index 8bfd5bcf25ee5401ce7bbeaf29e1f1868cab3509..6576ca55237e58668bb892d1f4b1ec227eaef515 100644 (file)
--- a/iolog.c
+++ b/iolog.c
@@ -18,7 +18,9 @@
 #include "verify.h"
 #include "trim.h"
 #include "filelock.h"
-#include "lib/tp.h"
+#include "smalloc.h"
+
+static int iolog_flush(struct io_log *log);
 
 static const char iolog_ver2[] = "fio version 2 iolog";
 
@@ -28,19 +30,14 @@ 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,
@@ -65,20 +62,22 @@ 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);
-       unsigned long this_delay;
+       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)
@@ -87,6 +86,12 @@ static void iolog_delay(struct thread_data *td, unsigned long delay)
                usec_sleep(td, this_delay);
                delay -= this_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)
@@ -113,7 +118,7 @@ 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);
@@ -131,7 +136,7 @@ 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);
 
@@ -184,7 +189,7 @@ 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--;
@@ -247,6 +252,7 @@ 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);
@@ -254,11 +260,18 @@ 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);
@@ -280,6 +293,18 @@ 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;
 
@@ -293,7 +318,7 @@ 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, struct io_u *io_u)
+void trim_io_piece(struct thread_data *td, const struct io_u *io_u)
 {
        struct io_piece *ipo = io_u->ipo;
 
@@ -550,20 +575,42 @@ int init_iolog(struct thread_data *td)
 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;
+       int i;
+       struct io_u_plat_entry *entry;
+       struct flist_head *list;
 
-       memset(l, 0, 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;
 
@@ -571,8 +618,8 @@ 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) {
-               pthread_mutex_init(&l->chunk_lock, NULL);
+       else if (l->log_gz || l->log_gz_store) {
+               mutex_init_pshared(&l->chunk_lock);
                p->td->flags |= TD_F_COMPRESS_LOG;
        }
 
@@ -607,12 +654,88 @@ 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);
 }
 
-static void flush_samples(FILE *f, void *samples, uint64_t sample_size)
+inline unsigned long hist_sum(int j, int stride, unsigned int *io_u_plat,
+               unsigned int *io_u_plat_last)
+{
+       unsigned long sum;
+       int k;
+
+       if (io_u_plat_last) {
+               for (k = sum = 0; k < stride; k++)
+                       sum += io_u_plat[j + k] - io_u_plat_last[j + k];
+       } else {
+               for (k = sum = 0; k < stride; k++)
+                       sum += io_u_plat[j + k];
+       }
+
+       return sum;
+}
+
+static void flush_hist_samples(FILE *f, int hist_coarseness, void *samples,
+                              uint64_t sample_size)
+{
+       struct io_sample *s;
+       int log_offset;
+       uint64_t i, j, nr_samples;
+       struct io_u_plat_entry *entry, *entry_before;
+       unsigned int *io_u_plat;
+       unsigned int *io_u_plat_before;
+
+       int stride = 1 << hist_coarseness;
+       
+       if (!sample_size)
+               return;
+
+       s = __get_sample(samples, 0, 0);
+       log_offset = (s->__ddir & LOG_OFFSET_SAMPLE_BIT) != 0;
+
+       nr_samples = sample_size / __log_entry_sz(log_offset);
+
+       for (i = 0; i < nr_samples; i++) {
+               s = __get_sample(samples, log_offset, i);
+
+               entry = (struct io_u_plat_entry *) (uintptr_t) s->val;
+               io_u_plat = entry->io_u_plat;
+
+               entry_before = flist_first_entry(&entry->list, struct io_u_plat_entry, list);
+               io_u_plat_before = entry_before->io_u_plat;
+
+               fprintf(f, "%lu, %u, %u, ", (unsigned long) s->time,
+                                               io_sample_ddir(s), s->bs);
+               for (j = 0; j < FIO_IO_U_PLAT_NR - stride; j += stride) {
+                       fprintf(f, "%lu, ", hist_sum(j, stride, io_u_plat,
+                                               io_u_plat_before));
+               }
+               fprintf(f, "%lu\n", (unsigned long)
+                       hist_sum(FIO_IO_U_PLAT_NR - stride, stride, io_u_plat,
+                                       io_u_plat_before));
+
+               flist_del(&entry_before->list);
+               free(entry_before);
+       }
+}
+
+void flush_samples(FILE *f, void *samples, uint64_t sample_size)
 {
        struct io_sample *s;
        int log_offset;
@@ -649,17 +772,11 @@ static void flush_samples(FILE *f, void *samples, uint64_t sample_size)
 #ifdef CONFIG_ZLIB
 
 struct iolog_flush_data {
-       struct tp_work work;
+       struct workqueue_work work;
        struct io_log *log;
        void *samples;
-       uint64_t nr_samples;
-};
-
-struct iolog_compress {
-       struct flist_head list;
-       void *buf;
-       size_t len;
-       unsigned int seq;
+       uint32_t nr_samples;
+       bool free;
 };
 
 #define GZ_CHUNK       131072
@@ -686,11 +803,16 @@ 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;
        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;
 
@@ -700,8 +822,9 @@ static int z_stream_init(z_stream *stream, int gz_hdr)
        return 0;
 }
 
-struct flush_chunk_iter {
+struct inflate_chunk_iter {
        unsigned int seq;
+       int err;
        void *buf;
        size_t buf_size;
        size_t buf_used;
@@ -709,13 +832,14 @@ struct flush_chunk_iter {
 };
 
 static void finish_chunk(z_stream *stream, FILE *f,
-                        struct flush_chunk_iter *iter)
+                        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);
+               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);
@@ -723,9 +847,18 @@ static void finish_chunk(z_stream *stream, FILE *f,
        iter->buf_size = iter->buf_used = 0;
 }
 
-static int flush_chunk(struct iolog_compress *ic, int gz_hdr, FILE *f,
-                      z_stream *stream, struct flush_chunk_iter *iter)
+/*
+ * 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\n",
+                               (unsigned long) ic->len, ic->seq);
+
        if (ic->seq != iter->seq) {
                if (iter->seq)
                        finish_chunk(stream, f, iter);
@@ -752,6 +885,7 @@ static int flush_chunk(struct iolog_compress *ic, int gz_hdr, FILE *f,
                err = inflate(stream, Z_NO_FLUSH);
                if (err < 0) {
                        log_err("fio: failed inflating log: %d\n", err);
+                       iter->err = err;
                        break;
                }
 
@@ -767,26 +901,41 @@ static int flush_chunk(struct iolog_compress *ic, int gz_hdr, FILE *f,
                        break;
        }
 
-       return 0;
+       ret = (void *) stream->next_in - ic->buf;
+
+       dprint(FD_COMPRESS, "inflated to size=%lu\n", (unsigned long) iter->buf_size);
+
+       return ret;
 }
 
-static void flush_gz_chunks(struct io_log *log, FILE *f)
+/*
+ * 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 flush_chunk_iter iter = { .chunk_sz = log->log_gz, };
-       struct flist_head *node;
+       struct inflate_chunk_iter iter = { .chunk_sz = log->log_gz, };
        z_stream stream;
 
        while (!flist_empty(&log->chunk_list)) {
                struct iolog_compress *ic;
 
-               node = log->chunk_list.next;
-               ic = flist_entry(node, struct iolog_compress, list);
+               ic = flist_first_entry(&log->chunk_list, struct iolog_compress, list);
                flist_del(&ic->list);
 
-               if (log->log_gz_store)
-                       fwrite(ic->buf, ic->len, 1, f);
-               else
-                       flush_chunk(ic, log->log_gz_store, f, &stream, &iter);
+               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);
        }
@@ -795,15 +944,24 @@ static void flush_gz_chunks(struct io_log *log, FILE *f)
                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 flush_chunk_iter iter = { .chunk_sz = 64 * 1024 * 1024, };
+       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(file, "r");
@@ -818,7 +976,7 @@ int iolog_file_inflate(const char *file)
                return 1;
        }
 
-       ic.buf = malloc(sb.st_size);
+       ic.buf = buf = malloc(sb.st_size);
        ic.len = sb.st_size;
        ic.seq = 1;
 
@@ -826,40 +984,72 @@ int iolog_file_inflate(const char *file)
        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);
 
-       flush_chunk(&ic,  1, stdout, &stream, &iter);
+       /*
+        * 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(ic.buf);
-       return 0;
+       free(buf);
+       return iter.err;
 }
 
 #else
 
-static void flush_gz_chunks(struct io_log *log, FILE *f)
+static int inflate_gz_chunks(struct io_log *log, FILE *f)
 {
+       return 0;
+}
+
+int iolog_file_inflate(const char *file)
+{
+       log_err("fio: log inflation not possible without zlib\n");
+       return 1;
 }
 
 #endif
 
-void flush_log(struct io_log *log)
+void flush_log(struct io_log *log, bool do_append)
 {
        void *buf;
        FILE *f;
 
-       f = fopen(log->filename, "w");
+       if (!do_append)
+               f = fopen(log->filename, "w");
+       else
+               f = fopen(log->filename, "a");
        if (!f) {
                perror("fopen log");
                return;
@@ -867,9 +1057,22 @@ void flush_log(struct io_log *log)
 
        buf = set_file_buffer(f);
 
-       flush_gz_chunks(log, f);
+       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 == 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);
@@ -877,8 +1080,8 @@ void flush_log(struct io_log *log)
 
 static int finish_log(struct thread_data *td, struct io_log *log, int trylock)
 {
-       if (td->tp_data)
-               iolog_flush(log, 1);
+       if (td->flags & TD_F_COMPRESS_LOG)
+               iolog_flush(log);
 
        if (trylock) {
                if (fio_trylock_file(log->filename))
@@ -886,22 +1089,41 @@ static int finish_log(struct thread_data *td, struct io_log *log, int trylock)
        } else
                fio_lock_file(log->filename);
 
-       if (td->client_type == FIO_CLIENT_TYPE_GUI)
+       if (td->client_type == FIO_CLIENT_TYPE_GUI || is_backend)
                fio_send_iolog(td, log, log->filename);
        else
-               flush_log(log);
+               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 tp_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;
@@ -910,8 +1132,7 @@ static int gz_work(struct tp_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;
@@ -919,7 +1140,7 @@ static int gz_work(struct tp_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;
@@ -927,14 +1148,21 @@ static int gz_work(struct tp_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, 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;
                ret = deflate(&stream, Z_NO_FLUSH);
                if (ret < 0) {
                        log_err("fio: deflate log (%d)\n", ret);
-                       break;
+                       free_chunk(c);
+                       goto err;
                }
 
                c->len = GZ_CHUNK - stream.avail_out;
@@ -946,19 +1174,41 @@ static int gz_work(struct tp_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;
                        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);
+                       dprint(FD_COMPRESS, "seq=%d, chunk=%lu\n", seq,
+                               (unsigned long) c->len);
                } 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);
@@ -971,116 +1221,262 @@ static int gz_work(struct tp_work *work)
                pthread_mutex_unlock(&data->log->chunk_lock);
        }
 
-       if (work->wait) {
-               work->done = 1;
-               pthread_cond_signal(&work->cv);
-       } else
+       ret = 0;
+done:
+       if (data->free)
                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;
 }
 
-int iolog_flush(struct io_log *log, int wait)
+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 finish == true, will not return
+ * until the log compression has completed, and will flush all previous
+ * logs too
+ */
+static int iolog_flush(struct io_log *log)
 {
-       struct tp_data *tdat = log->td->tp_data;
        struct iolog_flush_data *data;
-       size_t sample_size;
 
        data = malloc(sizeof(*data));
        if (!data)
                return 1;
 
        data->log = log;
+       data->free = false;
 
-       sample_size = log->nr_samples * log_entry_sz(log);
-       data->samples = malloc(sample_size);
-       if (!data->samples) {
-               free(data);
-               return 1;
-       }
+       while (!flist_empty(&log->io_logs)) {
+               struct io_logs *cur_log;
 
-       memcpy(data->samples, log->log, sample_size);
-       data->nr_samples = log->nr_samples;
-       data->work.fn = gz_work;
-       log->nr_samples = 0;
+               cur_log = flist_first_entry(&log->io_logs, struct io_logs, list);
+               flist_del_init(&cur_log->list);
 
-       if (wait) {
-               pthread_mutex_init(&data->work.lock, NULL);
-               pthread_cond_init(&data->work.cv, NULL);
-               data->work.wait = 1;
-       } else
-               data->work.wait = 0;
+               data->samples = cur_log->log;
+               data->nr_samples = cur_log->nr_samples;
 
-       tp_queue_work(tdat, &data->work);
+               sfree(cur_log);
 
-       if (wait) {
-               pthread_mutex_lock(&data->work.lock);
-               while (!data->work.done)
-                       pthread_cond_wait(&data->work.cv, &data->work.lock);
-               pthread_mutex_unlock(&data->work.lock);
-               free(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 = malloc(sizeof(*data));
+       if (!data)
+               return 1;
+
+       data->log = log;
+
+       data->samples = cur_log->log;
+       data->nr_samples = cur_log->nr_samples;
+       data->free = true;
+
+       cur_log->nr_samples = cur_log->max_samples = 0;
+       cur_log->log = NULL;
+
+       workqueue_enqueue(&log->td->log_compress_wq, &data->work);
+       return 0;
+}
 #else
 
-int iolog_flush(struct io_log *log, int wait)
+static int iolog_flush(struct io_log *log)
 {
        return 1;
 }
 
+int iolog_cur_flush(struct io_log *log, struct io_logs *cur_log)
+{
+       return 1;
+}
+
+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_iops_log(struct thread_data *td, int try)
+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)
+               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)
 {
-       struct io_log *log = td->iops_log;
+       int ret;
 
-       if (!log)
+       if (!unit_log)
                return 0;
 
-       return finish_log(td, log, try);
+       ret = __write_log(td, td->slat_log, try);
+       if (!ret)
+               td->slat_log = NULL;
+
+       return ret;
 }
 
-static int write_slat_log(struct thread_data *td, int try)
+static int write_clat_log(struct thread_data *td, int try, bool unit_log)
 {
-       struct io_log *log = td->slat_log;
+       int ret;
 
-       if (!log)
+       if (!unit_log)
                return 0;
 
-       return finish_log(td, log, try);
+       ret = __write_log(td, td->clat_log, try);
+       if (!ret)
+               td->clat_log = NULL;
+
+       return ret;
 }
 
-static int write_clat_log(struct thread_data *td, int try)
+static int write_clat_hist_log(struct thread_data *td, int try, bool unit_log)
 {
-       struct io_log *log = td->clat_log;
+       int ret;
 
-       if (!log)
+       if (!unit_log)
                return 0;
 
-       return finish_log(td, log, try);
+       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)
+static int write_lat_log(struct thread_data *td, int try, bool unit_log)
 {
-       struct io_log *log = td->lat_log;
+       int ret;
 
-       if (!log)
+       if (!unit_log)
                return 0;
 
-       return finish_log(td, log, try);
+       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)
+static int write_bandw_log(struct thread_data *td, int try, bool unit_log)
 {
-       struct io_log *log = td->bw_log;
+       int ret;
 
-       if (!log)
+       if (per_unit_log(td->bw_log) != unit_log)
                return 0;
 
-       return finish_log(td, log, try);
+       ret = __write_log(td, td->bw_log, try);
+       if (!ret)
+               td->bw_log = NULL;
+
+       return ret;
 }
 
 enum {
@@ -1089,13 +1485,14 @@ 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[] = {
@@ -1119,9 +1516,13 @@ 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;
@@ -1129,7 +1530,7 @@ 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;
@@ -1139,7 +1540,7 @@ 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;
@@ -1153,3 +1554,12 @@ 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);
+}