Add support for normal/gaussian random distributions
[fio.git] / filesetup.c
index 5ce0f6803ae7d952c262d25728b1a60ec7bff86f..7dceddbbf4dc21077d312dace3fa2321ba04f037 100644 (file)
@@ -38,7 +38,7 @@ static int extend_file(struct thread_data *td, struct fio_file *f)
        int r, new_layout = 0, unlink_file = 0, flags;
        unsigned long long left;
        unsigned int bs;
-       char *b;
+       char *b = NULL;
 
        if (read_only) {
                log_err("fio: refusing extend of file due to read-only\n");
@@ -59,7 +59,7 @@ static int extend_file(struct thread_data *td, struct fio_file *f)
 
        if (unlink_file || new_layout) {
                dprint(FD_FILE, "layout unlink %s\n", f->file_name);
-               if ((unlink(f->file_name) < 0) && (errno != ENOENT)) {
+               if ((td_io_unlink_file(td, f) < 0) && (errno != ENOENT)) {
                        td_verror(td, errno, "unlink");
                        return 1;
                }
@@ -69,6 +69,10 @@ static int extend_file(struct thread_data *td, struct fio_file *f)
        if (new_layout)
                flags |= O_TRUNC;
 
+#ifdef WIN32
+       flags |= _O_BINARY;
+#endif
+
        dprint(FD_FILE, "open file %s, flags %x\n", f->file_name, flags);
        f->fd = open(f->file_name, flags, 0644);
        if (f->fd < 0) {
@@ -168,7 +172,7 @@ static int extend_file(struct thread_data *td, struct fio_file *f)
 
        if (td->terminate) {
                dprint(FD_FILE, "terminate unlink %s\n", f->file_name);
-               unlink(f->file_name);
+               td_io_unlink_file(td, f);
        } else if (td->o.create_fsync) {
                if (fsync(f->fd) < 0) {
                        td_verror(td, errno, "fsync");
@@ -189,12 +193,14 @@ done:
 err:
        close(f->fd);
        f->fd = -1;
+       if (b)
+               free(b);
        return 1;
 }
 
 static int pre_read_file(struct thread_data *td, struct fio_file *f)
 {
-       int r, did_open = 0, old_runstate;
+       int ret = 0, r, did_open = 0, old_runstate;
        unsigned long long left;
        unsigned int bs;
        char *b;
@@ -216,7 +222,13 @@ static int pre_read_file(struct thread_data *td, struct fio_file *f)
        b = malloc(bs);
        memset(b, 0, bs);
 
-       lseek(f->fd, f->file_offset, SEEK_SET);
+       if (lseek(f->fd, f->file_offset, SEEK_SET) < 0) {
+               td_verror(td, errno, "lseek");
+               log_err("fio: failed to lseek pre-read file\n");
+               ret = 1;
+               goto error;
+       }
+
        left = f->io_size;
 
        while (left && !td->terminate) {
@@ -234,12 +246,14 @@ static int pre_read_file(struct thread_data *td, struct fio_file *f)
                }
        }
 
+error:
        td_restore_runstate(td, old_runstate);
 
        if (did_open)
                td->io_ops->close_file(td, f);
+
        free(b);
-       return 0;
+       return ret;
 }
 
 static unsigned long long get_rand_file_size(struct thread_data *td)
@@ -247,16 +261,9 @@ static unsigned long long get_rand_file_size(struct thread_data *td)
        unsigned long long ret, sized;
        unsigned long r;
 
-       if (td->o.use_os_rand) {
-               r = os_random_long(&td->file_size_state);
-               sized = td->o.file_size_high - td->o.file_size_low;
-               ret = (unsigned long long) ((double) sized * (r / (OS_RAND_MAX + 1.0)));
-       } else {
-               r = __rand(&td->__file_size_state);
-               sized = td->o.file_size_high - td->o.file_size_low;
-               ret = (unsigned long long) ((double) sized * (r / (FRAND_MAX + 1.0)));
-       }
-
+       r = __rand(&td->file_size_state);
+       sized = td->o.file_size_high - td->o.file_size_low;
+       ret = (unsigned long long) ((double) sized * (r / (FRAND_MAX + 1.0)));
        ret += td->o.file_size_low;
        ret -= (ret % td->o.rw_min_bs);
        return ret;
@@ -376,6 +383,10 @@ static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
 {
        int ret = 0;
 
+#ifdef CONFIG_ESX
+       return 0;
+#endif
+
        if (len == -1ULL)
                len = f->io_size;
        if (off == -1ULL)
@@ -387,18 +398,11 @@ static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
        dprint(FD_IO, "invalidate cache %s: %llu/%llu\n", f->file_name, off,
                                                                len);
 
-       /*
-        * FIXME: add blockdev flushing too
-        */
-       if (f->mmap_ptr) {
-               ret = posix_madvise(f->mmap_ptr, f->mmap_sz, POSIX_MADV_DONTNEED);
-#ifdef FIO_MADV_FREE
-               if (f->filetype == FIO_TYPE_BD)
-                       (void) posix_madvise(f->mmap_ptr, f->mmap_sz, FIO_MADV_FREE);
-#endif
-       } else if (f->filetype == FIO_TYPE_FILE) {
+       if (td->io_ops->invalidate)
+               ret = td->io_ops->invalidate(td, f);
+       else if (f->filetype == FIO_TYPE_FILE)
                ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
-       else if (f->filetype == FIO_TYPE_BD) {
+       else if (f->filetype == FIO_TYPE_BD) {
                ret = blockdev_invalidate_cache(f);
                if (ret < 0 && errno == EACCES && geteuid()) {
                        if (!root_warn) {
@@ -411,15 +415,18 @@ static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
        } else if (f->filetype == FIO_TYPE_CHAR || f->filetype == FIO_TYPE_PIPE)
                ret = 0;
 
-       if (ret < 0) {
-               td_verror(td, errno, "invalidate_cache");
-               return 1;
-       } else if (ret > 0) {
-               td_verror(td, ret, "invalidate_cache");
-               return 1;
+       /*
+        * Cache flushing isn't a fatal condition, and we know it will
+        * happen on some platforms where we don't have the proper
+        * function to flush eg block device caches. So just warn and
+        * continue on our way.
+        */
+       if (ret) {
+               log_info("fio: cache invalidation of %s failed: %s\n", f->file_name, strerror(errno));
+               ret = 0;
        }
 
-       return ret;
+       return 0;
 
 }
 
@@ -471,6 +478,10 @@ int file_lookup_open(struct fio_file *f, int flags)
                from_hash = 0;
        }
 
+#ifdef WIN32
+       flags |= _O_BINARY;
+#endif
+
        f->fd = open(f->file_name, flags, 0600);
        return from_hash;
 }
@@ -585,6 +596,7 @@ open_again:
                }
 
                td_verror(td, __e, buf);
+               return 1;
        }
 
        if (!from_hash && f->fd != -1) {
@@ -637,6 +649,7 @@ static int get_file_sizes(struct thread_data *td)
                        if (td->error != ENOENT) {
                                log_err("%s\n", td->verror);
                                err = 1;
+                               break;
                        }
                        clear_error(td);
                }
@@ -678,7 +691,8 @@ static unsigned long long get_fs_free_counts(struct thread_data *td)
                } else if (f->filetype != FIO_TYPE_FILE)
                        continue;
 
-               strcpy(buf, f->file_name);
+               buf[255] = '\0';
+               strncpy(buf, f->file_name, 255);
 
                if (stat(buf, &sb) < 0) {
                        if (errno != ENOENT)
@@ -700,8 +714,8 @@ static unsigned long long get_fs_free_counts(struct thread_data *td)
                if (fm)
                        continue;
 
-               fm = malloc(sizeof(*fm));
-               strcpy(fm->__base, buf);
+               fm = calloc(1, sizeof(*fm));
+               strncpy(fm->__base, buf, sizeof(fm->__base) - 1);
                fm->base = basename(fm->__base);
                fm->key = sb.st_dev;
                flist_add(&fm->list, &list);
@@ -731,7 +745,7 @@ uint64_t get_start_offset(struct thread_data *td, struct fio_file *f)
                return f->real_file_size;
 
        return td->o.start_offset +
-               (td->thread_number - 1) * td->o.offset_increment;
+               td->subjob_number * td->o.offset_increment;
 }
 
 /*
@@ -798,7 +812,7 @@ int setup_files(struct thread_data *td)
         * Calculate per-file size and potential extra size for the
         * first files, if needed.
         */
-       if (!o->file_size_low) {
+       if (!o->file_size_low && o->nr_files) {
                uint64_t all_fs;
 
                fs = o->size / o->nr_files;
@@ -871,9 +885,14 @@ int setup_files(struct thread_data *td)
                }
        }
 
-       if (!o->size || o->size > total_size)
+       if (!o->size || (total_size && o->size > total_size))
                o->size = total_size;
 
+       if (o->size < td_min_bs(td)) {
+               log_err("fio: blocksize too large for data set\n");
+               goto err_out;
+       }
+
        /*
         * See if we need to extend some files
         */
@@ -904,7 +923,13 @@ int setup_files(struct thread_data *td)
 
                        err = __file_invalidate_cache(td, f, old_len,
                                                                extend_len);
-                       close(f->fd);
+
+                       /*
+                        * Shut up static checker
+                        */
+                       if (f->fd != -1)
+                               close(f->fd);
+
                        f->fd = -1;
                        if (err)
                                break;
@@ -922,8 +947,12 @@ int setup_files(struct thread_data *td)
         * iolog already set the total io size, if we read back
         * stored entries.
         */
-       if (!o->read_iolog_file)
-               td->total_io_size = o->size * o->loops;
+       if (!o->read_iolog_file) {
+               if (o->io_limit)
+                       td->total_io_size = o->io_limit * o->loops;
+               else
+                       td->total_io_size = o->size * o->loops;
+       }
 
 done:
        if (o->create_only)
@@ -956,12 +985,12 @@ static int __init_rand_distribution(struct thread_data *td, struct fio_file *f)
 {
        unsigned int range_size, seed;
        unsigned long nranges;
-       uint64_t file_size;
+       uint64_t fsize;
 
        range_size = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
-       file_size = min(f->real_file_size, f->io_size);
+       fsize = min(f->real_file_size, f->io_size);
 
-       nranges = (file_size + range_size - 1) / range_size;
+       nranges = (fsize + range_size - 1) / range_size;
 
        seed = jhash(f->file_name, strlen(f->file_name), 0) * td->thread_number;
        if (!td->o.rand_repeatable)
@@ -969,8 +998,10 @@ static int __init_rand_distribution(struct thread_data *td, struct fio_file *f)
 
        if (td->o.random_distribution == FIO_RAND_DIST_ZIPF)
                zipf_init(&f->zipf, nranges, td->o.zipf_theta.u.f, seed);
-       else
+       else if (td->o.random_distribution == FIO_RAND_DIST_PARETO)
                pareto_init(&f->zipf, nranges, td->o.pareto_h.u.f, seed);
+       else if (td->o.random_distribution == FIO_RAND_DIST_GAUSS)
+               gauss_init(&f->gauss, nranges, td->o.gauss_dev, seed);
 
        return 1;
 }
@@ -1006,21 +1037,25 @@ int init_random_map(struct thread_data *td)
                return 0;
 
        for_each_file(td, f, i) {
-               uint64_t file_size = min(f->real_file_size, f->io_size);
+               uint64_t fsize = min(f->real_file_size, f->io_size);
 
-               blocks = file_size / (unsigned long long) td->o.rw_min_bs;
+               blocks = fsize / (unsigned long long) td->o.rw_min_bs;
 
                if (td->o.random_generator == FIO_RAND_GEN_LFSR) {
                        unsigned long seed;
 
                        seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
-                       
-                       if (!lfsr_init(&f->lfsr, blocks, seed, 0))
+
+                       if (!lfsr_init(&f->lfsr, blocks, seed, 0)) {
+                               fio_file_set_lfsr(f);
                                continue;
+                       }
                } else if (!td->o.norandommap) {
                        f->io_axmap = axmap_new(blocks);
-                       if (f->io_axmap)
+                       if (f->io_axmap) {
+                               fio_file_set_axmap(f);
                                continue;
+                       }
                } else if (td->o.norandommap)
                        continue;
 
@@ -1058,6 +1093,11 @@ void close_and_free_files(struct thread_data *td)
        dprint(FD_FILE, "close files\n");
 
        for_each_file(td, f, i) {
+               if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
+                       dprint(FD_FILE, "free unlink %s\n", f->file_name);
+                       td_io_unlink_file(td, f);
+               }
+
                if (fio_file_open(f))
                        td_io_close_file(td, f);
 
@@ -1065,13 +1105,15 @@ void close_and_free_files(struct thread_data *td)
 
                if (td->o.unlink && f->filetype == FIO_TYPE_FILE) {
                        dprint(FD_FILE, "free unlink %s\n", f->file_name);
-                       unlink(f->file_name);
+                       td_io_unlink_file(td, f);
                }
 
                sfree(f->file_name);
                f->file_name = NULL;
-               axmap_free(f->io_axmap);
-               f->io_axmap = NULL;
+               if (fio_file_axmap(f)) {
+                       axmap_free(f->io_axmap);
+                       f->io_axmap = NULL;
+               }
                sfree(f);
        }
 
@@ -1109,45 +1151,89 @@ static void get_file_type(struct fio_file *f)
        }
 }
 
-static void set_already_allocated(const char *fname) {
-       struct file_name *fn;
+static int __is_already_allocated(const char *fname)
+{
+       struct flist_head *entry;
+       char *filename;
 
-       fn = malloc(sizeof(struct file_name));
-       fn->filename = strdup(fname);
-       flist_add_tail(&fn->list, &filename_list);
+       if (flist_empty(&filename_list))
+               return 0;
+
+       flist_for_each(entry, &filename_list) {
+               filename = flist_entry(entry, struct file_name, list)->filename;
+
+               if (strcmp(filename, fname) == 0)
+                       return 1;
+       }
+
+       return 0;
 }
 
 static int is_already_allocated(const char *fname)
 {
-       struct flist_head *entry;
-       char *filename;
+       int ret;
 
-       if (!flist_empty(&filename_list))
-       {
-               flist_for_each(entry, &filename_list) {
-                       filename = flist_entry(entry, struct file_name, list)->filename;
+       fio_file_hash_lock();
+       ret = __is_already_allocated(fname);
+       fio_file_hash_unlock();
+       return ret;
+}
 
-                       if (strcmp(filename, fname) == 0)
-                               return 1;
-               }
+static void set_already_allocated(const char *fname)
+{
+       struct file_name *fn;
+
+       fn = malloc(sizeof(struct file_name));
+       fn->filename = strdup(fname);
+
+       fio_file_hash_lock();
+       if (!__is_already_allocated(fname)) {
+               flist_add_tail(&fn->list, &filename_list);
+               fn = NULL;
        }
+       fio_file_hash_unlock();
 
-       return 0;
+       if (fn) {
+               free(fn->filename);
+               free(fn);
+       }
 }
 
-static void free_already_allocated() {
+
+static void free_already_allocated(void)
+{
        struct flist_head *entry, *tmp;
        struct file_name *fn;
 
-       if (!flist_empty(&filename_list))
-       {
-               flist_for_each_safe(entry, tmp, &filename_list) {
-                       fn = flist_entry(entry, struct file_name, list);
-                       free(fn->filename);
-                       flist_del(&fn->list);
-                       free(fn);
-               }
+       if (flist_empty(&filename_list))
+               return;
+
+       fio_file_hash_lock();
+       flist_for_each_safe(entry, tmp, &filename_list) {
+               fn = flist_entry(entry, struct file_name, list);
+               free(fn->filename);
+               flist_del(&fn->list);
+               free(fn);
+       }
+
+       fio_file_hash_unlock();
+}
+
+static struct fio_file *alloc_new_file(struct thread_data *td)
+{
+       struct fio_file *f;
+
+       f = smalloc(sizeof(*f));
+       if (!f) {
+               log_err("fio: smalloc OOM\n");
+               assert(0);
+               return NULL;
        }
+
+       f->fd = -1;
+       f->shadow_fd = -1;
+       fio_file_reset(td, f);
+       return f;
 }
 
 int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
@@ -1168,15 +1254,7 @@ int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
        if (numjob && is_already_allocated(file_name))
                return 0;
 
-       f = smalloc(sizeof(*f));
-       if (!f) {
-               log_err("fio: smalloc OOM\n");
-               assert(0);
-       }
-
-       f->fd = -1;
-       f->shadow_fd = -1;
-       fio_file_reset(td, f);
+       f = alloc_new_file(td);
 
        if (td->files_size <= td->files_index) {
                unsigned int new_size = td->o.nr_files + 1;
@@ -1235,13 +1313,6 @@ int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
 
        set_already_allocated(file_name);
 
-       /*
-        * For adding files after the fact - if openfiles= isn't
-        * given as an option, ensure we allow at least one file open
-        */
-       if (!td->o.open_files)
-               td->o.open_files = 1;
-
        if (inc)
                td->o.nr_files++;
 
@@ -1286,8 +1357,11 @@ int put_file(struct thread_data *td, struct fio_file *f)
        if (--f->references)
                return 0;
 
-       if (should_fsync(td) && td->o.fsync_on_close)
+       if (should_fsync(td) && td->o.fsync_on_close) {
                f_ret = fsync(f->fd);
+               if (f_ret < 0)
+                       f_ret = errno;
+       }
 
        if (td->io_ops->close_file)
                ret = td->io_ops->close_file(td, f);
@@ -1365,7 +1439,8 @@ static int recurse_dir(struct thread_data *td, const char *dirname)
                if (lstat(full_path, &sb) == -1) {
                        if (errno != ENOENT) {
                                td_verror(td, errno, "stat");
-                               return 1;
+                               ret = 1;
+                               break;
                        }
                }
 
@@ -1413,13 +1488,7 @@ void dup_files(struct thread_data *td, struct thread_data *org)
        for_each_file(org, f, i) {
                struct fio_file *__f;
 
-               __f = smalloc(sizeof(*__f));
-               if (!__f) {
-                       log_err("fio: smalloc OOM\n");
-                       assert(0);
-               }
-               __f->fd = -1;
-               fio_file_reset(td, __f);
+               __f = alloc_new_file(td);
 
                if (f->file_name) {
                        __f->file_name = smalloc_strdup(f->file_name);
@@ -1461,17 +1530,24 @@ int get_fileno(struct thread_data *td, const char *fname)
 void free_release_files(struct thread_data *td)
 {
        close_files(td);
+       td->o.nr_files = 0;
+       td->o.open_files = 0;
        td->files_index = 0;
        td->nr_normal_files = 0;
 }
 
 void fio_file_reset(struct thread_data *td, struct fio_file *f)
 {
-       f->last_pos = f->file_offset;
-       f->last_start = -1ULL;
-       if (f->io_axmap)
+       int i;
+
+       for (i = 0; i < DDIR_RWDIR_CNT; i++) {
+               f->last_pos[i] = f->file_offset;
+               f->last_start[i] = -1ULL;
+       }
+
+       if (fio_file_axmap(f))
                axmap_reset(f->io_axmap);
-       if (td->o.random_generator == FIO_RAND_GEN_LFSR)
+       else if (fio_file_lfsr(f))
                lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
 }
 
@@ -1488,6 +1564,7 @@ int fio_files_done(struct thread_data *td)
 }
 
 /* free memory used in initialization phase only */
-void filesetup_mem_free() {
+void filesetup_mem_free(void)
+{
        free_already_allocated();
 }