Branch and cache miss speedups
[fio.git] / filesetup.c
index 96be3d1c4faf9b15b267ff6635de35e4226398c2..2744d4fce0f9597166a3a749445c530594ff33b7 100644 (file)
@@ -11,6 +11,7 @@
 #include "fio.h"
 #include "smalloc.h"
 #include "filehash.h"
+#include "options.h"
 #include "os/os.h"
 #include "hash.h"
 #include "lib/axmap.h"
@@ -21,6 +22,8 @@
 
 static int root_warn;
 
+static FLIST_HEAD(filename_list);
+
 static inline void clear_error(struct thread_data *td)
 {
        td->error = 0;
@@ -121,13 +124,14 @@ static int extend_file(struct thread_data *td, struct fio_file *f)
                dprint(FD_FILE, "truncate file %s, size %llu\n", f->file_name,
                                        (unsigned long long) f->real_file_size);
                if (ftruncate(f->fd, f->real_file_size) == -1) {
-                       td_verror(td, errno, "ftruncate");
-                       goto err;
+                       if (errno != EFBIG) {
+                               td_verror(td, errno, "ftruncate");
+                               goto err;
+                       }
                }
        }
 
        b = malloc(td->o.max_bs[DDIR_WRITE]);
-       memset(b, 0, td->o.max_bs[DDIR_WRITE]);
 
        left = f->real_file_size;
        while (left && !td->terminate) {
@@ -135,6 +139,8 @@ static int extend_file(struct thread_data *td, struct fio_file *f)
                if (bs > left)
                        bs = left;
 
+               fill_io_buffer(td, b, bs, bs);
+
                r = write(f->fd, b, bs);
 
                if (r > 0) {
@@ -355,7 +361,8 @@ static int get_file_size(struct thread_data *td, struct fio_file *f)
 
        if (f->file_offset > f->real_file_size) {
                log_err("%s: offset extends end (%llu > %llu)\n", td->o.name,
-                                       f->file_offset, f->real_file_size);
+                                       (unsigned long long) f->file_offset,
+                                       (unsigned long long) f->real_file_size);
                return 1;
        }
 
@@ -386,7 +393,8 @@ static int __file_invalidate_cache(struct thread_data *td, struct fio_file *f,
        if (f->mmap_ptr) {
                ret = posix_madvise(f->mmap_ptr, f->mmap_sz, POSIX_MADV_DONTNEED);
 #ifdef FIO_MADV_FREE
-               (void) posix_madvise(f->mmap_ptr, f->mmap_sz, 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) {
                ret = posix_fadvise(f->fd, off, len, POSIX_FADV_DONTNEED);
@@ -441,6 +449,7 @@ int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
                f->shadow_fd = -1;
        }
 
+       f->engine_data = 0;
        return ret;
 }
 
@@ -515,6 +524,13 @@ int generic_open_file(struct thread_data *td, struct fio_file *f)
                goto skip_flags;
        if (td->o.odirect)
                flags |= OS_O_DIRECT;
+       if (td->o.oatomic) {
+               if (!FIO_O_ATOMIC) {
+                       td_verror(td, EINVAL, "OS does not support atomic IO");
+                       return 1;
+               }
+               flags |= OS_O_DIRECT | FIO_O_ATOMIC;
+       }
        if (td->o.sync_io)
                flags |= O_SYNC;
        if (td->o.create_on_open)
@@ -580,7 +596,7 @@ open_again:
                         * work-around a "feature" on Linux, where a close of
                         * an fd that has been opened for write will trigger
                         * udev to call blkid to check partitions, fs id, etc.
-                        * That polutes the device cache, which can slow down
+                        * That pollutes the device cache, which can slow down
                         * unbuffered accesses.
                         */
                        if (f->shadow_fd == -1)
@@ -719,13 +735,20 @@ uint64_t get_start_offset(struct thread_data *td)
 int setup_files(struct thread_data *td)
 {
        unsigned long long total_size, extend_size;
+       struct thread_options *o = &td->o;
        struct fio_file *f;
-       unsigned int i;
+       unsigned int i, nr_fs_extra = 0;
        int err = 0, need_extend;
+       int old_state;
+       const unsigned int bs = td_min_bs(td);
+       uint64_t fs = 0;
 
        dprint(FD_FILE, "setup files\n");
 
-       if (td->o.read_iolog_file)
+       old_state = td->runstate;
+       td_set_runstate(td, TD_SETTING_UP);
+
+       if (o->read_iolog_file)
                goto done;
 
        /*
@@ -739,7 +762,7 @@ int setup_files(struct thread_data *td)
                err = get_file_sizes(td);
 
        if (err)
-               return err;
+               goto err_out;
 
        /*
         * check sizes. if the files/devices do not exist and the size
@@ -753,17 +776,32 @@ int setup_files(struct thread_data *td)
                        total_size += f->real_file_size;
        }
 
-       if (td->o.fill_device)
+       if (o->fill_device)
                td->fill_device_size = get_fs_free_counts(td);
 
        /*
         * device/file sizes are zero and no size given, punt
         */
-       if ((!total_size || total_size == -1ULL) && !td->o.size &&
-           !(td->io_ops->flags & FIO_NOIO) && !td->o.fill_device) {
-               log_err("%s: you need to specify size=\n", td->o.name);
+       if ((!total_size || total_size == -1ULL) && !o->size &&
+           !(td->io_ops->flags & FIO_NOIO) && !o->fill_device &&
+           !(o->nr_files && (o->file_size_low || o->file_size_high))) {
+               log_err("%s: you need to specify size=\n", o->name);
                td_verror(td, EINVAL, "total_file_size");
-               return 1;
+               goto err_out;
+       }
+
+       /*
+        * Calculate per-file size and potential extra size for the
+        * first files, if needed.
+        */
+       if (!o->file_size_low) {
+               uint64_t all_fs;
+
+               fs = o->size / o->nr_files;
+               all_fs = fs * o->nr_files;
+
+               if (all_fs < o->size)
+                       nr_fs_extra = (o->size - all_fs) / bs;
        }
 
        /*
@@ -776,27 +814,33 @@ int setup_files(struct thread_data *td)
        for_each_file(td, f, i) {
                f->file_offset = get_start_offset(td);
 
-               if (!td->o.file_size_low) {
+               if (!o->file_size_low) {
                        /*
                         * no file size range given, file size is equal to
-                        * total size divided by number of files. if that is
-                        * zero, set it to the real file size.
+                        * total size divided by number of files. If that is
+                        * zero, set it to the real file size. If the size
+                        * doesn't divide nicely with the min blocksize,
+                        * make the first files bigger.
                         */
-                       f->io_size = td->o.size / td->o.nr_files;
+                       f->io_size = fs;
+                       if (nr_fs_extra) {
+                               nr_fs_extra--;
+                               f->io_size += bs;
+                       }
+
                        if (!f->io_size)
                                f->io_size = f->real_file_size - f->file_offset;
-               } else if (f->real_file_size < td->o.file_size_low ||
-                          f->real_file_size > td->o.file_size_high) {
-                       if (f->file_offset > td->o.file_size_low)
+               } else if (f->real_file_size < o->file_size_low ||
+                          f->real_file_size > o->file_size_high) {
+                       if (f->file_offset > o->file_size_low)
                                goto err_offset;
                        /*
                         * file size given. if it's fixed, use that. if it's a
                         * range, generate a random size in-between.
                         */
-                       if (td->o.file_size_low == td->o.file_size_high) {
-                               f->io_size = td->o.file_size_low
-                                               - f->file_offset;
-                       } else {
+                       if (o->file_size_low == o->file_size_high)
+                               f->io_size = o->file_size_low - f->file_offset;
+                       else {
                                f->io_size = get_rand_file_size(td)
                                                - f->file_offset;
                        }
@@ -806,15 +850,15 @@ int setup_files(struct thread_data *td)
                if (f->io_size == -1ULL)
                        total_size = -1ULL;
                else {
-                        if (td->o.size_percent)
-                                f->io_size = (f->io_size * td->o.size_percent) / 100;
+                        if (o->size_percent)
+                                f->io_size = (f->io_size * o->size_percent) / 100;
                        total_size += f->io_size;
                }
 
                if (f->filetype == FIO_TYPE_FILE &&
                    (f->io_size + f->file_offset) > f->real_file_size &&
                    !(td->io_ops->flags & FIO_DISKLESSIO)) {
-                       if (!td->o.create_on_open) {
+                       if (!o->create_on_open) {
                                need_extend++;
                                extend_size += (f->io_size + f->file_offset);
                        } else
@@ -823,8 +867,8 @@ int setup_files(struct thread_data *td)
                }
        }
 
-       if (!td->o.size || td->o.size > total_size)
-               td->o.size = total_size;
+       if (!o->size || o->size > total_size)
+               o->size = total_size;
 
        /*
         * See if we need to extend some files
@@ -833,7 +877,7 @@ int setup_files(struct thread_data *td)
                temp_stall_ts = 1;
                if (output_format == FIO_OUTPUT_NORMAL)
                        log_info("%s: Laying out IO file(s) (%u file(s) /"
-                                " %lluMB)\n", td->o.name, need_extend,
+                                " %lluMB)\n", o->name, need_extend,
                                        extend_size >> 20);
 
                for_each_file(td, f, i) {
@@ -844,7 +888,7 @@ int setup_files(struct thread_data *td)
 
                        assert(f->filetype == FIO_TYPE_FILE);
                        fio_file_clear_extend(f);
-                       if (!td->o.fill_device) {
+                       if (!o->fill_device) {
                                old_len = f->real_file_size;
                                extend_len = f->io_size + f->file_offset -
                                                old_len;
@@ -865,25 +909,28 @@ int setup_files(struct thread_data *td)
        }
 
        if (err)
-               return err;
+               goto err_out;
 
-       if (!td->o.zone_size)
-               td->o.zone_size = td->o.size;
+       if (!o->zone_size)
+               o->zone_size = o->size;
 
        /*
         * iolog already set the total io size, if we read back
         * stored entries.
         */
-       if (!td->o.read_iolog_file)
-               td->total_io_size = td->o.size * td->o.loops;
+       if (!o->read_iolog_file)
+               td->total_io_size = o->size * o->loops;
 
 done:
-       if (td->o.create_only)
+       if (o->create_only)
                td->done = 1;
 
+       td_set_runstate(td, old_state);
        return 0;
 err_offset:
-       log_err("%s: you need to specify valid offset=\n", td->o.name);
+       log_err("%s: you need to specify valid offset=\n", o->name);
+err_out:
+       td_set_runstate(td, old_state);
        return 1;
 }
 
@@ -963,7 +1010,7 @@ int init_random_map(struct thread_data *td)
 
                        seed = td->rand_seeds[FIO_RAND_BLOCK_OFF];
                        
-                       if (!lfsr_init(&f->lfsr, blocks, seed, seed & 0xF))
+                       if (!lfsr_init(&f->lfsr, blocks, seed, 0))
                                continue;
                } else if (!td->o.norandommap) {
                        f->io_axmap = axmap_new(blocks);
@@ -1029,6 +1076,7 @@ void close_and_free_files(struct thread_data *td)
        td->files_index = 0;
        td->files = NULL;
        td->file_locks = NULL;
+       td->o.file_lock_mode = FILE_LOCK_NONE;
        td->o.nr_files = 0;
 }
 
@@ -1056,7 +1104,48 @@ static void get_file_type(struct fio_file *f)
        }
 }
 
-int add_file(struct thread_data *td, const char *fname)
+static void set_already_allocated(const char *fname) {
+       struct file_name *fn;
+
+       fn = malloc(sizeof(struct file_name));
+       fn->filename = strdup(fname);
+       flist_add_tail(&fn->list, &filename_list);
+}
+
+static int is_already_allocated(const char *fname)
+{
+       struct flist_head *entry;
+       char *filename;
+
+       if (!flist_empty(&filename_list))
+       {
+               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 void free_already_allocated() {
+       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);
+               }
+       }
+}
+
+int add_file(struct thread_data *td, const char *fname, int numjob, int inc)
 {
        int cur_files = td->files_index;
        char file_name[PATH_MAX];
@@ -1065,6 +1154,15 @@ int add_file(struct thread_data *td, const char *fname)
 
        dprint(FD_FILE, "add file %s\n", fname);
 
+       if (td->o.directory)
+               len = set_name_idx(file_name, td->o.directory, numjob);
+
+       sprintf(file_name + len, "%s", fname);
+
+       /* clean cloned siblings using existing files */
+       if (numjob && is_already_allocated(file_name))
+               return 0;
+
        f = smalloc(sizeof(*f));
        if (!f) {
                log_err("fio: smalloc OOM\n");
@@ -1104,10 +1202,6 @@ int add_file(struct thread_data *td, const char *fname)
        if (td->io_ops && (td->io_ops->flags & FIO_DISKLESSIO))
                f->real_file_size = -1ULL;
 
-       if (td->o.directory)
-               len = sprintf(file_name, "%s/", td->o.directory);
-
-       sprintf(file_name + len, "%s", fname);
        f->file_name = smalloc_strdup(file_name);
        if (!f->file_name) {
                log_err("fio: smalloc OOM\n");
@@ -1134,6 +1228,18 @@ int add_file(struct thread_data *td, const char *fname)
        if (f->filetype == FIO_TYPE_FILE)
                td->nr_normal_files++;
 
+       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++;
+
        dprint(FD_FILE, "file %p \"%s\" added at %d\n", f, f->file_name,
                                                        cur_files);
 
@@ -1150,7 +1256,7 @@ int add_file_exclusive(struct thread_data *td, const char *fname)
                        return i;
        }
 
-       return add_file(td, fname);
+       return add_file(td, fname, 0, 1);
 }
 
 void get_file(struct fio_file *f)
@@ -1221,6 +1327,8 @@ void unlock_file(struct thread_data *td, struct fio_file *f)
 
 void unlock_file_all(struct thread_data *td, struct fio_file *f)
 {
+       if (td->o.file_lock_mode == FILE_LOCK_NONE || !td->file_locks)
+               return;
        if (td->file_locks[f->fileno] != FILE_LOCK_NONE)
                unlock_file(td, f);
 }
@@ -1257,8 +1365,7 @@ static int recurse_dir(struct thread_data *td, const char *dirname)
                }
 
                if (S_ISREG(sb.st_mode)) {
-                       add_file(td, full_path);
-                       td->o.nr_files++;
+                       add_file(td, full_path, 0, 1);
                        continue;
                }
                if (!S_ISDIR(sb.st_mode))
@@ -1319,6 +1426,11 @@ void dup_files(struct thread_data *td, struct thread_data *org)
                        __f->filetype = f->filetype;
                }
 
+               if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
+                       __f->lock = f->lock;
+               else if (td->o.file_lock_mode == FILE_LOCK_READWRITE)
+                       __f->rwlock = f->rwlock;
+
                td->files[i] = __f;
        }
 }
@@ -1357,3 +1469,20 @@ void fio_file_reset(struct thread_data *td, struct fio_file *f)
        if (td->o.random_generator == FIO_RAND_GEN_LFSR)
                lfsr_reset(&f->lfsr, td->rand_seeds[FIO_RAND_BLOCK_OFF]);
 }
+
+int fio_files_done(struct thread_data *td)
+{
+       struct fio_file *f;
+       unsigned int i;
+
+       for_each_file(td, f, i)
+               if (!fio_file_done(f))
+                       return 0;
+
+       return 1;
+}
+
+/* free memory used in initialization phase only */
+void filesetup_mem_free() {
+       free_already_allocated();
+}