Fix ETA display for iolog replay
[fio.git] / log.c
diff --git a/log.c b/log.c
index a705e5021415ad7069e169ba06906435724d1f37..759771ae6a5d3167b7fcb36f92990164d3faf351 100644 (file)
--- a/log.c
+++ b/log.c
+/*
+ * Code related to writing an iolog of what a thread is doing, and to
+ * later read that back and replay
+ */
 #include <stdio.h>
 #include <stdlib.h>
+#include <assert.h>
 #include "list.h"
 #include "fio.h"
 
-void write_iolog_put(struct thread_data *td, struct io_u *io_u)
+static const char iolog_ver2[] = "fio version 2 iolog";
+
+void log_io_u(struct thread_data *td, struct io_u *io_u)
 {
-       fprintf(td->iolog_f, "%u,%llu,%u\n", io_u->ddir, io_u->offset, io_u->buflen);
+       const char *act[] = { "read", "write", "sync" };
+
+       assert(io_u->ddir < 3);
+
+       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);
+}
+
+void log_file(struct thread_data *td, struct fio_file *f,
+             enum file_log_act what)
+{
+       const char *act[] = { "add", "open", "close" };
+
+       assert(what < 3);
+
+       if (!td->o.write_iolog_file)
+               return;
+
+       fprintf(td->iolog_f, "%s %s\n", f->file_name, act[what]);
+}
+
+static void iolog_delay(struct thread_data *td, unsigned long delay)
+{
+       unsigned long usec = utime_since_now(&td->last_issue);
+
+       if (delay < usec)
+               return;
+
+       delay -= usec;
+
+       /*
+        * less than 100 usec delay, just regard it as noise
+        */
+       if (delay < 100)
+               return;
+
+       usec_sleep(td, delay);
 }
 
 int read_iolog_get(struct thread_data *td, struct io_u *io_u)
 {
        struct io_piece *ipo;
 
-       if (!list_empty(&td->io_log_list)) {
+       while (!list_empty(&td->io_log_list)) {
                ipo = list_entry(td->io_log_list.next, struct io_piece, list);
                list_del(&ipo->list);
+
+               /*
+                * invalid ddir, this is a file action
+                */
+               if (ipo->ddir == DDIR_INVAL) {
+                       struct fio_file *f = &td->files[ipo->fileno];
+
+                       if (ipo->file_action == FIO_LOG_OPEN_FILE) {
+                               assert(!td_io_open_file(td, f));
+                               free(ipo);
+                               continue;
+                       } else if (ipo->file_action == FIO_LOG_CLOSE_FILE) {
+                               td_io_close_file(td, f);
+                               free(ipo);
+                               continue;
+                       }
+               }
+
                io_u->offset = ipo->offset;
                io_u->buflen = ipo->len;
                io_u->ddir = ipo->ddir;
-               io_u->file = ipo->file;
+               io_u->file = &td->files[ipo->fileno];
+               get_file(io_u->file);
+
+               if (ipo->delay)
+                       iolog_delay(td, ipo->delay);
+
                free(ipo);
                return 0;
        }
 
+       td->done = 1;
        return 1;
 }
 
 void prune_io_piece_log(struct thread_data *td)
 {
        struct io_piece *ipo;
+       struct rb_node *n;
 
-       while (!list_empty(&td->io_hist_list)) {
-               ipo = list_entry(td->io_hist_list.next, struct io_piece, list);
-
-               list_del(&ipo->list);
+       while ((n = rb_first(&td->io_hist_tree)) != NULL) {
+               ipo = rb_entry(n, struct io_piece, rb_node);
+               rb_erase(n, &td->io_hist_tree);
                free(ipo);
        }
 }
 
 /*
- * log a succesful write, so we can unwind the log for verify
+ * log a successful write, so we can unwind the log for verify
  */
 void log_io_piece(struct thread_data *td, struct io_u *io_u)
 {
-       struct io_piece *ipo = malloc(sizeof(struct io_piece));
-       struct list_head *entry;
+       struct rb_node **p, *parent;
+       struct io_piece *ipo, *__ipo;
 
-       INIT_LIST_HEAD(&ipo->list);
+       ipo = malloc(sizeof(struct io_piece));
        ipo->file = io_u->file;
        ipo->offset = io_u->offset;
        ipo->len = io_u->buflen;
 
        /*
-        * for random io where the writes extend the file, it will typically
-        * be laid out with the block scattered as written. it's faster to
-        * read them in in that order again, so don't sort
+        * 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.
         */
-       if (td->sequential || !td->overwrite) {
+       if (!td_random(td) || !td->o.overwrite ||
+            (io_u->file->flags & FIO_FILE_NOSORT)) {
+               INIT_LIST_HEAD(&ipo->list);
                list_add_tail(&ipo->list, &td->io_hist_list);
                return;
        }
 
+       RB_CLEAR_NODE(&ipo->rb_node);
+       p = &td->io_hist_tree.rb_node;
+       parent = NULL;
+
        /*
-        * for random io, sort the list so verify will run faster
+        * Sort the entry into the verification list
         */
-       entry = &td->io_hist_list;
-       while ((entry = entry->prev) != &td->io_hist_list) {
-               struct io_piece *__ipo = list_entry(entry, struct io_piece, list);
-
-               if (__ipo->offset < ipo->offset)
-                       break;
+       while (*p) {
+               parent = *p;
+
+               __ipo = rb_entry(parent, struct io_piece, rb_node);
+               if (ipo->offset <= __ipo->offset)
+                       p = &(*p)->rb_left;
+               else
+                       p = &(*p)->rb_right;
        }
 
-       list_add(&ipo->list, entry);
+       rb_link_node(&ipo->rb_node, parent, p);
+       rb_insert_color(&ipo->rb_node, &td->io_hist_tree);
 }
 
 void write_iolog_close(struct thread_data *td)
@@ -80,25 +162,130 @@ void write_iolog_close(struct thread_data *td)
        fflush(td->iolog_f);
        fclose(td->iolog_f);
        free(td->iolog_buf);
+       td->iolog_f = NULL;
+       td->iolog_buf = NULL;
 }
 
 /*
- * Open a stored log and read in the entries.
+ * Read version 2 iolog data. It is enhanced to include per-file logging,
+ * syncs, etc.
  */
-static int init_iolog_read(struct thread_data *td)
+static int read_iolog2(struct thread_data *td, FILE *f)
 {
        unsigned long long offset;
        unsigned int bytes;
+       int reads, writes, fileno = 0, file_action = 0; /* stupid gcc */
+       char *fname, *act;
        char *str, *p;
-       FILE *f;
-       int rw, reads, writes;
+       enum fio_ddir rw;
 
-       f = fopen(td->read_iolog_file, "r");
-       if (!f) {
-               perror("fopen read iolog");
-               return 1;
+       free_release_files(td);
+
+       /*
+        * Read in the read iolog and store it, reuse the infrastructure
+        * for doing verifications.
+        */
+       str = malloc(4096);
+       fname = malloc(256+16);
+       act = malloc(256+16);
+
+       reads = writes = 0;
+       while ((p = fgets(str, 4096, f)) != NULL) {
+               struct io_piece *ipo;
+               int r;
+
+               r = sscanf(p, "%256s %256s %llu %u", fname, act, &offset, &bytes);
+               if (r == 4) {
+                       /*
+                        * Check action first
+                        */
+                       if (!strcmp(act, "read"))
+                               rw = DDIR_READ;
+                       else if (!strcmp(act, "write"))
+                               rw = DDIR_WRITE;
+                       else if (!strcmp(act, "sync"))
+                               rw = DDIR_SYNC;
+                       else {
+                               log_err("fio: bad iolog file action: %s\n",act);
+                               continue;
+                       }
+               } else if (r == 2) {
+                       rw = DDIR_INVAL;
+                       if (!strcmp(act, "add")) {
+                               td->o.nr_files++;
+                               fileno = add_file(td, fname);
+                               file_action = FIO_LOG_ADD_FILE;
+                               continue;
+                       } else if (!strcmp(act, "open")) {
+                               fileno = get_fileno(td, fname);
+                               file_action = FIO_LOG_OPEN_FILE;
+                       } else if (!strcmp(act, "close")) {
+                               fileno = get_fileno(td, fname);
+                               file_action = FIO_LOG_CLOSE_FILE;
+                       } else {
+                               log_err("fio: bad iolog file action: %s\n",act);
+                               continue;
+                       }
+               } else {
+                       log_err("bad iolog2: %s", p);
+                       continue;
+               }
+                       
+               if (rw == DDIR_READ)
+                       reads++;
+               else if (rw == DDIR_WRITE)
+                       writes++;
+               else if (rw != DDIR_SYNC && rw != DDIR_INVAL) {
+                       log_err("bad ddir: %d\n", rw);
+                       continue;
+               }
+
+               /*
+                * Make note of file
+                */
+               ipo = malloc(sizeof(*ipo));
+               memset(ipo, 0, sizeof(*ipo));
+               INIT_LIST_HEAD(&ipo->list);
+               ipo->offset = offset;
+               ipo->len = bytes;
+               ipo->ddir = rw;
+               if (bytes > td->o.max_bs[rw])
+                       td->o.max_bs[rw] = bytes;
+               if (rw == DDIR_INVAL) {
+                       ipo->fileno = fileno;
+                       ipo->file_action = file_action;
+               }
+               list_add_tail(&ipo->list, &td->io_log_list);
+               td->total_io_size += bytes;
        }
 
+       free(str);
+       free(act);
+       free(fname);
+
+       if (!reads && !writes)
+               return 1;
+       else if (reads && !writes)
+               td->o.td_ddir = TD_DDIR_READ;
+       else if (!reads && writes)
+               td->o.td_ddir = TD_DDIR_WRITE;
+       else
+               td->o.td_ddir = TD_DDIR_RW;
+
+       return 0;
+}
+
+/*
+ * Read version 1 iolog data.
+ */
+static int read_iolog(struct thread_data *td, FILE *f)
+{
+       unsigned long long offset;
+       unsigned int bytes;
+       char *str, *p;
+       int reads, writes;
+       enum fio_ddir rw;
+
        /*
         * Read in the read iolog and store it, reuse the infrastructure
         * for doing verifications.
@@ -116,44 +303,92 @@ static int init_iolog_read(struct thread_data *td)
                        reads++;
                else if (rw == DDIR_WRITE)
                        writes++;
-               else {
+               else if (rw != DDIR_SYNC) {
                        log_err("bad ddir: %d\n", rw);
                        continue;
                }
 
                ipo = malloc(sizeof(*ipo));
+               memset(ipo, 0, sizeof(*ipo));
                INIT_LIST_HEAD(&ipo->list);
                ipo->offset = offset;
                ipo->len = bytes;
-               ipo->ddir = (enum fio_ddir) rw;
-               if (bytes > td->max_bs[rw])
-                       td->max_bs[rw] = bytes;
+               ipo->ddir = rw;
+               if (bytes > td->o.max_bs[rw])
+                       td->o.max_bs[rw] = bytes;
                list_add_tail(&ipo->list, &td->io_log_list);
+               td->total_io_size += bytes;
        }
 
        free(str);
-       fclose(f);
 
        if (!reads && !writes)
                return 1;
        else if (reads && !writes)
-               td->ddir = DDIR_READ;
+               td->o.td_ddir = TD_DDIR_READ;
        else if (!reads && writes)
-               td->ddir = DDIR_READ;
+               td->o.td_ddir = TD_DDIR_WRITE;
        else
-               td->iomix = 1;
+               td->o.td_ddir = TD_DDIR_RW;
 
        return 0;
 }
 
+/*
+ * open iolog, check version, and call appropriate parser
+ */
+static int init_iolog_read(struct thread_data *td)
+{
+       char buffer[256], *p;
+       FILE *f;
+       int ret;
+
+       f = fopen(td->o.read_iolog_file, "r");
+       if (!f) {
+               perror("fopen read iolog");
+               return 1;
+       }
+
+       p = fgets(buffer, sizeof(buffer), f);
+       if (!p) {
+               td_verror(td, errno, "iolog read");
+               log_err("fio: unable to read iolog\n");
+               return 1;
+       }
+
+       /*
+        * 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 {
+               /*
+                * seek back to the beginning
+                */
+               if (fseek(f, 0, SEEK_SET) < 0) {
+                       td_verror(td, errno, "iolog read");
+                       log_err("fio: unable to read iolog\n");
+                       return 1;
+               }
+
+               ret = read_iolog(td, f);
+       }
+
+       fclose(f);
+       return ret;
+}
+
 /*
  * Setup a log for storing io patterns.
  */
 static int init_iolog_write(struct thread_data *td)
 {
+       struct fio_file *ff;
        FILE *f;
+       unsigned int i;
 
-       f = fopen(td->write_iolog_file, "w+");
+       f = fopen(td->o.write_iolog_file, "w+");
        if (!f) {
                perror("fopen write iolog");
                return 1;
@@ -165,6 +400,21 @@ static int init_iolog_write(struct thread_data *td)
        td->iolog_f = f;
        td->iolog_buf = malloc(8192);
        setvbuf(f, td->iolog_buf, _IOFBF, 8192);
+
+       /*
+        * write our version line
+        */
+       if (fprintf(f, "%s\n", iolog_ver2) < 0) {
+               perror("iolog init\n");
+               return 1;
+       }
+
+       /*
+        * add all known files
+        */
+       for_each_file(td, ff, i)
+               log_file(td, ff, FIO_LOG_ADD_FILE);
+
        return 0;
 }
 
@@ -172,42 +422,24 @@ int init_iolog(struct thread_data *td)
 {
        int ret = 0;
 
-       if (td->io_ops->flags & FIO_CPUIO)
+       if (td->io_ops->flags & FIO_DISKLESSIO)
                return 0;
 
-       if (td->read_iolog_file)
-               ret = init_iolog_read(td);
-       else if (td->write_iolog_file)
+       if (td->o.read_iolog_file) {
+               /*
+                * Check if it's a blktrace file and load that if possible.
+                * Otherwise assume it's a normal log file and load that.
+                */
+               if (is_blktrace(td->o.read_iolog_file))
+                       ret = load_blktrace(td, td->o.read_iolog_file);
+               else
+                       ret = init_iolog_read(td);
+       } else if (td->o.write_iolog_file)
                ret = init_iolog_write(td);
 
        return ret;
 }
 
-int setup_rate(struct thread_data *td)
-{
-       unsigned long long rate;
-       int nr_reads_per_msec;
-
-       if (!td->rate)
-               return 0;
-
-       if (td->rate < td->ratemin) {
-               log_err("min rate larger than nominal rate\n");
-               return -1;
-       }
-
-       rate = td->rate;
-       nr_reads_per_msec = (rate * 1024 * 1000) / td->min_bs[DDIR_READ];
-       if (!nr_reads_per_msec) {
-               log_err("rate lower than supported\n");
-               return -1;
-       }
-
-       td->rate_usec_cycle = 1000000000ULL / nr_reads_per_msec;
-       td->rate_pending_usleep = 0;
-       return 0;
-}
-
 void setup_log(struct io_log **log)
 {
        struct io_log *l = malloc(sizeof(*l));