filesetup: assign fileno at setup time
[fio.git] / filesetup.c
index abea1e60b726a8d82f503b054c83d76a66470210..3fc14641ed7d34380d5f735006eb129ab0b7faf0 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");
@@ -59,20 +59,32 @@ 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;
                }
        }
 
-       flags = O_WRONLY | O_CREAT;
+       flags = O_WRONLY;
+       if (td->o.allow_create)
+               flags |= O_CREAT;
        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) {
-               td_verror(td, errno, "open");
+               int err = errno;
+
+               if (err == ENOENT && !td->o.allow_create)
+                       log_err("fio: file creation disallowed by "
+                                       "allow_file_create=0\n");
+               else
+                       td_verror(td, err, "open");
                return 1;
        }
 
@@ -168,7 +180,7 @@ 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");
@@ -189,6 +201,8 @@ done:
 err:
        close(f->fd);
        f->fd = -1;
+       if (b)
+               free(b);
        return 1;
 }
 
@@ -250,21 +264,16 @@ error:
        return ret;
 }
 
-static unsigned long long get_rand_file_size(struct thread_data *td)
+unsigned long long get_rand_file_size(struct thread_data *td)
 {
        unsigned long long ret, sized;
+       uint64_t frand_max;
        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)));
-       }
-
+       frand_max = rand_max(&td->file_size_state);
+       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;
@@ -382,7 +391,11 @@ static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
                                   unsigned long long off,
                                   unsigned long long len)
 {
-       int ret = 0;
+       int errval = 0, ret = 0;
+
+#ifdef CONFIG_ESX
+       return 0;
+#endif
 
        if (len == -1ULL)
                len = f->io_size;
@@ -395,19 +408,28 @@ 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
+       if (td->io_ops->invalidate) {
+               ret = td->io_ops->invalidate(td, f);
+               if (ret < 0)
+                       errval = ret;
        } else if (f->filetype == FIO_TYPE_FILE) {
                ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
+               if (ret)
+                       errval = ret;
        } else if (f->filetype == FIO_TYPE_BD) {
+               int retry_count = 0;
+
                ret = blockdev_invalidate_cache(f);
+               while (ret < 0 && errno == EAGAIN && retry_count++ < 25) {
+                       /*
+                        * Linux multipath devices reject ioctl while
+                        * the maps are being updated. That window can
+                        * last tens of milliseconds; we'll try up to
+                        * a quarter of a second.
+                        */
+                       usleep(10000);
+                       ret = blockdev_invalidate_cache(f);
+               }
                if (ret < 0 && errno == EACCES && geteuid()) {
                        if (!root_warn) {
                                log_err("fio: only root may flush block "
@@ -416,18 +438,21 @@ static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
                        }
                        ret = 0;
                }
+               if (ret < 0)
+                       errval = errno;
        } 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 (errval)
+               log_info("fio: cache invalidation of %s failed: %s\n", f->file_name, strerror(errval));
 
-       return ret;
+       return 0;
 
 }
 
@@ -479,6 +504,10 @@ 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;
 }
@@ -509,11 +538,6 @@ int generic_open_file(struct thread_data *td, struct fio_file *f)
 
        dprint(FD_FILE, "fd open %s\n", f->file_name);
 
-       if (td_trim(td) && f->filetype != FIO_TYPE_BD) {
-               log_err("fio: trim only applies to block device\n");
-               return 1;
-       }
-
        if (!strcmp(f->file_name, "-")) {
                if (td_rw(td)) {
                        log_err("fio: can't read/write to stdin/out\n");
@@ -541,7 +565,7 @@ int generic_open_file(struct thread_data *td, struct fio_file *f)
        }
        if (td->o.sync_io)
                flags |= O_SYNC;
-       if (td->o.create_on_open)
+       if (td->o.create_on_open && td->o.allow_create)
                flags |= O_CREAT;
 skip_flags:
        if (f->filetype != FIO_TYPE_FILE)
@@ -552,7 +576,7 @@ open_again:
                if (!read_only)
                        flags |= O_RDWR;
 
-               if (f->filetype == FIO_TYPE_FILE)
+               if (f->filetype == FIO_TYPE_FILE && td->o.allow_create)
                        flags |= O_CREAT;
 
                if (is_std)
@@ -593,6 +617,7 @@ open_again:
                }
 
                td_verror(td, __e, buf);
+               return 1;
        }
 
        if (!from_hash && f->fd != -1) {
@@ -645,6 +670,7 @@ 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);
                }
@@ -686,7 +712,8 @@ 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)
@@ -708,8 +735,8 @@ 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);
@@ -721,7 +748,7 @@ static unsigned long long get_fs_free_counts(struct thread_data *td)
                fm = flist_entry(n, struct fio_mount, list);
                flist_del(&fm->list);
 
-               sz = get_fs_size(fm->base);
+               sz = get_fs_free_size(fm->base);
                if (sz && sz != -1ULL)
                        ret += sz;
 
@@ -739,7 +766,7 @@ 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;
 }
 
 /*
@@ -782,6 +809,7 @@ int setup_files(struct thread_data *td)
         */
        total_size = 0;
        for_each_file(td, f, i) {
+               f->fileno = i;
                if (f->real_file_size == -1ULL)
                        total_size = -1ULL;
                else
@@ -879,15 +907,41 @@ int setup_files(struct thread_data *td)
                }
        }
 
-       if (!o->size || o->size > total_size)
+       if (td->o.block_error_hist) {
+               int len;
+
+               assert(td->o.nr_files == 1);    /* checked in fixup_options */
+               f = td->files[0];
+               len = f->io_size / td->o.bs[DDIR_TRIM];
+               if (len > MAX_NR_BLOCK_INFOS || len <= 0) {
+                       log_err("fio: cannot calculate block histogram with "
+                               "%d trim blocks, maximum %d\n",
+                               len, MAX_NR_BLOCK_INFOS);
+                       td_verror(td, EINVAL, "block_error_hist");
+                       goto err_out;
+               }
+
+               td->ts.nr_block_infos = len;
+               for (i = 0; i < len; i++)
+                       td->ts.block_infos[i] =
+                               BLOCK_INFO(0, BLOCK_STATE_UNINIT);
+       } else
+               td->ts.nr_block_infos = 0;
+
+       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
         */
        if (need_extend) {
                temp_stall_ts = 1;
-               if (output_format == FIO_OUTPUT_NORMAL)
+               if (output_format & FIO_OUTPUT_NORMAL)
                        log_info("%s: Laying out IO file(s) (%u file(s) /"
                                 " %lluMB)\n", o->name, need_extend,
                                        extend_size >> 20);
@@ -912,7 +966,13 @@ 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;
@@ -930,8 +990,12 @@ 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)
@@ -964,12 +1028,12 @@ 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)
@@ -977,8 +1041,10 @@ 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.u.f, seed);
 
        return 1;
 }
@@ -1002,6 +1068,43 @@ static int init_rand_distribution(struct thread_data *td)
        return 1;
 }
 
+/*
+ * Check if the number of blocks exceeds the randomness capability of
+ * the selected generator. Tausworthe is 32-bit, the others are fullly
+ * 64-bit capable.
+ */
+static int check_rand_gen_limits(struct thread_data *td, struct fio_file *f,
+                                uint64_t blocks)
+{
+       if (blocks <= FRAND32_MAX)
+               return 0;
+       if (td->o.random_generator != FIO_RAND_GEN_TAUSWORTHE)
+               return 0;
+
+       /*
+        * If the user hasn't specified a random generator, switch
+        * to tausworthe64 with informational warning. If the user did
+        * specify one, just warn.
+        */
+       log_info("fio: file %s exceeds 32-bit tausworthe random generator.\n",
+                       f->file_name);
+
+       if (!fio_option_is_set(&td->o, random_generator)) {
+               log_info("fio: Switching to tausworthe64. Use the "
+                        "random_generator= option to get rid of this "
+                        " warning.\n");
+               td->o.random_generator = FIO_RAND_GEN_TAUSWORTHE64;
+               return 0;
+       }
+
+       /*
+        * Just make this information to avoid breaking scripts.
+        */
+       log_info("fio: Use the random_generator= option to switch to lfsr or "
+                        "tausworthe64.\n");
+       return 0;
+}
+
 int init_random_map(struct thread_data *td)
 {
        unsigned long long blocks;
@@ -1014,21 +1117,28 @@ 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 = fsize / (unsigned long long) td->o.rw_min_bs;
 
-               blocks = file_size / (unsigned long long) td->o.rw_min_bs;
+               if (check_rand_gen_limits(td, f, blocks))
+                       return 1;
 
                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;
 
@@ -1066,6 +1176,11 @@ 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);
 
@@ -1073,13 +1188,15 @@ 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);
        }
 
@@ -1212,7 +1329,7 @@ int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
        dprint(FD_FILE, "add file %s\n", fname);
 
        if (td->o.directory)
-               len = set_name_idx(file_name, td->o.directory, numjob);
+               len = set_name_idx(file_name, PATH_MAX, td->o.directory, numjob);
 
        sprintf(file_name + len, "%s", fname);
 
@@ -1323,8 +1440,11 @@ 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);
@@ -1402,7 +1522,8 @@ 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;
                        }
                }
 
@@ -1492,17 +1613,24 @@ 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]);
 }