Make sure the ->files array is job private
[fio.git] / filesetup.c
index d0e3ed212366c27c05d7e7618912f80508fdb907..d8b7401b35e0230d9f40afe53dcbfcd5e84b83e1 100644 (file)
@@ -2,8 +2,10 @@
 #include <fcntl.h>
 #include <string.h>
 #include <assert.h>
+#include <dirent.h>
 #include <sys/stat.h>
 #include <sys/mman.h>
+#include <sys/types.h>
 
 #include "fio.h"
 #include "os.h"
@@ -15,12 +17,19 @@ static int file_ok(struct thread_data *td, struct fio_file *f)
 {
        struct stat st;
 
-       if (td->filetype != FIO_TYPE_FILE)
+       if (f->filetype != FIO_TYPE_FILE ||
+           (td->io_ops->flags & FIO_DISKLESSIO))
                return 0;
 
-       if (stat(f->file_name, &st) == -1)
+       if (lstat(f->file_name, &st) == -1)
                return 1;
-       else if (st.st_size < (off_t) f->file_size)
+
+       /*
+        * if it's a special file, size is always ok for now
+        */
+       if (S_ISCHR(st.st_mode) || S_ISBLK(st.st_mode))
+               return 0;
+       if (st.st_size < (off_t) f->file_size)
                return 1;
 
        return 0;
@@ -35,21 +44,26 @@ static int create_file(struct thread_data *td, struct fio_file *f)
 
        f->fd = open(f->file_name, O_WRONLY | O_CREAT | O_TRUNC, 0644);
        if (f->fd < 0) {
-               td_verror(td, errno);
+               td_verror(td, errno, "open");
                return 1;
        }
 
        if (ftruncate(f->fd, f->file_size) == -1) {
-               td_verror(td, errno);
+               td_verror(td, errno, "ftruncate");
+               goto err;
+       }
+
+       if (posix_fallocate(f->fd, 0, f->file_size) < 0) {
+               td_verror(td, errno, "posix_fallocate");
                goto err;
        }
 
-       b = malloc(td->max_bs);
-       memset(b, 0, td->max_bs);
+       b = malloc(td->o.max_bs[DDIR_WRITE]);
+       memset(b, 0, td->o.max_bs[DDIR_WRITE]);
 
        left = f->file_size;
        while (left && !td->terminate) {
-               bs = td->max_bs;
+               bs = td->o.max_bs[DDIR_WRITE];
                if (bs > left)
                        bs = left;
 
@@ -60,9 +74,9 @@ static int create_file(struct thread_data *td, struct fio_file *f)
                        continue;
                } else {
                        if (r < 0)
-                               td_verror(td, errno);
+                               td_verror(td, errno, "write");
                        else
-                               td_verror(td, EIO);
+                               td_verror(td, EIO, "write");
 
                        break;
                }
@@ -70,7 +84,7 @@ static int create_file(struct thread_data *td, struct fio_file *f)
 
        if (td->terminate)
                unlink(f->file_name);
-       else if (td->create_fsync)
+       else if (td->o.create_fsync)
                fsync(f->fd);
 
        free(b);
@@ -83,47 +97,158 @@ err:
        return 1;
 }
 
+static unsigned long long set_rand_file_size(struct thread_data *td,
+                                            unsigned long long total_size)
+{
+       unsigned long long upper = total_size;
+       unsigned long long ret;
+       long r;
+
+       if (upper > td->o.file_size_high)
+               upper = td->o.file_size_high;
+       else if (upper < td->o.file_size_low)
+               return 0;
+       else if (!upper)
+               return 0;
+
+       r = os_random_long(&td->file_size_state);
+       ret = td->o.file_size_low + (unsigned long long) ((double) upper * (r / (RAND_MAX + 1.0)));
+       ret -= (ret % td->o.rw_min_bs);
+       if (ret > upper)
+               ret = upper;
+       return ret;
+}
+
+static int fill_file_size(struct thread_data *td, struct fio_file *f,
+                         unsigned long long *file_size, int new_files)
+{
+       if (!td->o.file_size_low) {
+               f->file_size = *file_size / new_files;
+               f->real_file_size = f->file_size;
+       } else {
+               /*
+                * If we don't have enough space left for a file
+                * of the minimum size, bail.
+                */
+               if (*file_size < td->o.file_size_low)
+                       return 1;
+
+               f->file_size = set_rand_file_size(td, *file_size);
+               f->real_file_size = f->file_size;
+               *file_size -= f->file_size;
+       }
+
+       return 0;
+}
+
 static int create_files(struct thread_data *td)
 {
        struct fio_file *f;
-       int i, err, need_create;
+       int err, need_create, can_extend;
+       unsigned long long total_file_size, local_file_size, create_size;
+       unsigned int i, new_files;
+
+       new_files = 0;
+       total_file_size = td->o.size;
+       for_each_file(td, f, i) {
+               unsigned long long s;
+
+               f->file_offset = td->o.start_offset;
+
+               if (f->filetype != FIO_TYPE_FILE)
+                       continue;
+               if (!total_file_size)
+                       continue;
+
+               if (f->flags & FIO_FILE_EXISTS) {
+                       if ((f->file_size > td->o.size / td->o.nr_files) ||
+                           !f->file_size)
+                               f->file_size = td->o.size / td->o.nr_files;
+
+                       s = f->file_size;
+                       if (s > total_file_size)
+                               s = total_file_size;
+
+                       total_file_size -= s;
+               } else
+                       new_files++;
+       }
 
        /*
         * unless specifically asked for overwrite, let normal io extend it
         */
-       if (!td->overwrite) {
-               for_each_file(td, f, i)
-                       f->file_size = td->total_file_size / td->nr_files;
+       can_extend = !td->o.overwrite && !(td->io_ops->flags & FIO_NOEXTEND);
+       if (can_extend && new_files) {
+               for_each_file(td, f, i) {
+                       if (fill_file_size(td, f, &total_file_size, new_files)) {
+                               log_info("fio: limited to %d files\n", i);
+                               td->o.nr_files = i;
+                               break;
+                       }
+               }
 
                return 0;
        }
 
-       if (!td->total_file_size) {
-               log_err("Need size for create\n");
-               td_verror(td, EINVAL);
-               return 1;
-       }
+       local_file_size = total_file_size;
+       if (!local_file_size)
+               local_file_size = -1;
 
+       total_file_size = 0;
        need_create = 0;
+       create_size = 0;
        for_each_file(td, f, i) {
-               f->file_size = td->total_file_size / td->nr_files;
-               need_create += file_ok(td, f);
-       }
+               int file_there;
 
-       td->io_size = td->total_file_size;
+               if (f->filetype != FIO_TYPE_FILE)
+                       continue;
+               if (f->flags & FIO_FILE_EXISTS) {
+                       total_file_size += f->file_size;
+                       continue;
+               }
+
+               if (fill_file_size(td, f, &local_file_size, new_files)) {
+                       log_info("fio: limited to %d files\n", i);
+                       new_files -= (td->o.nr_files - i);
+                       td->o.nr_files = i;
+                       break;
+               }
+
+               total_file_size += f->file_size;
+               create_size += f->file_size;
+               file_there = !file_ok(td, f);
+
+               if (file_there && td_write(td) && !td->o.overwrite) {
+                       unlink(f->file_name);
+                       file_there = 0;
+               }
+
+               need_create += !file_there;
+       }
 
        if (!need_create)
                return 0;
 
+       if (!td->o.size && !total_file_size) {
+               log_err("Need size for create\n");
+               td_verror(td, EINVAL, "file_size");
+               return 1;
+       }
+
        temp_stall_ts = 1;
-       fprintf(f_out, "%s: Laying out IO file(s) (%d x %LuMiB == %LuMiB)\n",
-                               td->name, td->nr_files,
-                               (td->total_file_size >> 20) / td->nr_files,
-                               td->total_file_size >> 20);
+       log_info("%s: Laying out IO file(s) (%u files / %LuMiB)\n",
+                               td->o.name, new_files, create_size >> 20);
 
        err = 0;
        for_each_file(td, f, i) {
+               /*
+                * Only unlink files that we created.
+                */
+               f->flags &= ~FIO_FILE_UNLINK;
                if (file_ok(td, f)) {
+                       if (td->o.unlink)
+                               f->flags |= FIO_FILE_UNLINK;
+
                        err = create_file(td, f);
                        if (err)
                                break;
@@ -138,9 +263,9 @@ static int file_size(struct thread_data *td, struct fio_file *f)
 {
        struct stat st;
 
-       if (td->overwrite) {
+       if (td->o.overwrite) {
                if (fstat(f->fd, &st) == -1) {
-                       td_verror(td, errno);
+                       td_verror(td, errno, "fstat");
                        return 1;
                }
 
@@ -148,9 +273,9 @@ static int file_size(struct thread_data *td, struct fio_file *f)
 
                if (!f->file_size || f->file_size > f->real_file_size)
                        f->file_size = f->real_file_size;
-       }
+       } else
+               f->real_file_size = f->file_size;
 
-       f->file_size -= f->file_offset;
        return 0;
 }
 
@@ -161,7 +286,7 @@ static int bdev_size(struct thread_data *td, struct fio_file *f)
 
        r = blockdev_size(f->fd, &bytes);
        if (r) {
-               td_verror(td, r);
+               td_verror(td, r, "blockdev_size");
                return 1;
        }
 
@@ -181,9 +306,10 @@ static int get_file_size(struct thread_data *td, struct fio_file *f)
 {
        int ret = 0;
 
-       if (td->filetype == FIO_TYPE_FILE)
-               ret = file_size(td, f);
-       else if (td->filetype == FIO_TYPE_BD)
+       if (f->filetype == FIO_TYPE_FILE) {
+               if (!(f->flags & FIO_FILE_EXISTS))
+                       ret = file_size(td, f);
+       } else if (f->filetype == FIO_TYPE_BD)
                ret = bdev_size(td, f);
        else
                f->real_file_size = -1;
@@ -192,7 +318,7 @@ static int get_file_size(struct thread_data *td, struct fio_file *f)
                return ret;
 
        if (f->file_offset > f->real_file_size) {
-               log_err("%s: offset extends end (%Lu > %Lu)\n", td->name, f->file_offset, f->real_file_size);
+               log_err("%s: offset extends end (%Lu > %Lu)\n", td->o.name, f->file_offset, f->real_file_size);
                return 1;
        }
 
@@ -203,206 +329,355 @@ int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
 {
        int ret = 0;
 
+       if (td->o.odirect)
+               return 0;
+
        /*
         * FIXME: add blockdev flushing too
         */
-       if (td->io_ops->flags & FIO_MMAPIO)
+       if (f->mmap)
                ret = madvise(f->mmap, f->file_size, MADV_DONTNEED);
-       else if (td->filetype == FIO_TYPE_FILE)
+       else if (f->filetype == FIO_TYPE_FILE)
                ret = fadvise(f->fd, f->file_offset, f->file_size, POSIX_FADV_DONTNEED);
-       else if (td->filetype == FIO_TYPE_BD)
+       else if (f->filetype == FIO_TYPE_BD)
                ret = blockdev_invalidate_cache(f->fd);
-       else if (td->filetype == FIO_TYPE_CHAR)
+       else if (f->filetype == FIO_TYPE_CHAR)
                ret = 0;
 
        if (ret < 0) {
-               td_verror(td, errno);
+               td_verror(td, errno, "invalidate_cache");
                return 1;
        }
 
-       return 0;
+       return ret;
 }
 
-static int __setup_file_mmap(struct thread_data *td, struct fio_file *f)
+void generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
 {
-       int flags;
+       close(f->fd);
+       f->fd = -1;
+}
 
-       if (td_rw(td))
-               flags = PROT_READ | PROT_WRITE;
-       else if (td_write(td)) {
-               flags = PROT_WRITE;
+int generic_open_file(struct thread_data *td, struct fio_file *f)
+{
+       int flags = 0;
 
-               if (td->verify != VERIFY_NONE)
-                       flags |= PROT_READ;
-       } else
-               flags = PROT_READ;
+       if (td->o.odirect)
+               flags |= OS_O_DIRECT;
+       if (td->o.sync_io)
+               flags |= O_SYNC;
 
-       f->mmap = mmap(NULL, f->file_size, flags, MAP_SHARED, f->fd, f->file_offset);
-       if (f->mmap == MAP_FAILED) {
-               f->mmap = NULL;
-               td_verror(td, errno);
-               return 1;
+       if (td_write(td) || td_rw(td)) {
+               flags |= O_RDWR;
+
+               if (f->filetype == FIO_TYPE_FILE)
+                       flags |= O_CREAT;
+
+               f->fd = open(f->file_name, flags, 0600);
+       } else {
+               if (f->filetype == FIO_TYPE_CHAR)
+                       flags |= O_RDWR;
+               else
+                       flags |= O_RDONLY;
+
+               f->fd = open(f->file_name, flags);
        }
 
-       if (td->invalidate_cache && file_invalidate_cache(td, f))
+       if (f->fd == -1) {
+               char buf[FIO_VERROR_SIZE];
+               int __e = errno;
+
+               snprintf(buf, sizeof(buf) - 1, "open(%s)", f->file_name);
+
+               td_verror(td, __e, buf);
+               if (__e == EINVAL && td->o.odirect)
+                       log_err("fio: destination does not support O_DIRECT\n");
+               if (__e == EMFILE)
+                       log_err("fio: try reducing/setting openfiles (failed at %u of %u)\n", td->nr_open_files, td->o.nr_files);
                return 1;
+       }
 
-       if (td->sequential) {
-               if (madvise(f->mmap, f->file_size, MADV_SEQUENTIAL) < 0) {
-                       td_verror(td, errno);
-                       return 1;
+       if (get_file_size(td, f))
+               goto err;
+
+       if (td->o.invalidate_cache && file_invalidate_cache(td, f))
+               goto err;
+
+       if (!td->o.fadvise_hint)
+               return 0;
+
+       if (!td_random(td)) {
+               if (fadvise(f->fd, f->file_offset, f->file_size, POSIX_FADV_SEQUENTIAL) < 0) {
+                       td_verror(td, errno, "fadvise");
+                       goto err;
                }
        } else {
-               if (madvise(f->mmap, f->file_size, MADV_RANDOM) < 0) {
-                       td_verror(td, errno);
-                       return 1;
+               if (fadvise(f->fd, f->file_offset, f->file_size, POSIX_FADV_RANDOM) < 0) {
+                       td_verror(td, errno, "fadvise");
+                       goto err;
                }
        }
 
        return 0;
+err:
+       close(f->fd);
+       return 1;
 }
 
-static int setup_files_mmap(struct thread_data *td)
+int open_files(struct thread_data *td)
 {
        struct fio_file *f;
-       int i, err = 0;
+       unsigned int i;
+       int err = 0;
 
        for_each_file(td, f, i) {
-               err = __setup_file_mmap(td, f);
+               err = td_io_open_file(td, f);
                if (err)
                        break;
+
+               if (td->o.open_files == td->nr_open_files)
+                       break;
        }
 
+       if (!err)
+               return 0;
+
+       for_each_file(td, f, i)
+               td_io_close_file(td, f);
+
        return err;
 }
 
-static int __setup_file_plain(struct thread_data *td, struct fio_file *f)
+int setup_files(struct thread_data *td)
 {
-       if (td->invalidate_cache && file_invalidate_cache(td, f))
+       struct fio_file *f;
+       unsigned int i;
+       int err;
+
+       /*
+        * if ioengine defines a setup() method, it's responsible for
+        * setting up everything in the td->files[] area.
+        */
+       if (td->io_ops->setup)
+               return td->io_ops->setup(td);
+
+       if (create_files(td))
                return 1;
 
-       if (td->sequential) {
-               if (fadvise(f->fd, f->file_offset, f->file_size, POSIX_FADV_SEQUENTIAL) < 0) {
-                       td_verror(td, errno);
-                       return 1;
-               }
-       } else {
-               if (fadvise(f->fd, f->file_offset, f->file_size, POSIX_FADV_RANDOM) < 0) {
-                       td_verror(td, errno);
+       err = open_files(td);
+       if (err)
+               return err;
+
+       /*
+        * Recalculate the total file size now that files are set up.
+        */
+       td->o.size = 0;
+       for_each_file(td, f, i)
+               td->o.size += f->file_size;
+
+       td->io_size = td->o.size;
+       if (td->io_size == 0) {
+               log_err("%s: no io blocks\n", td->o.name);
+               td_verror(td, EINVAL, "total_file_size");
+               return 1;
+       }
+
+       if (!td->o.zone_size)
+               td->o.zone_size = td->io_size;
+
+       td->total_io_size = td->io_size * td->o.loops;
+
+       for_each_file(td, f, i)
+               td_io_close_file(td, f);
+
+       return err;
+}
+
+int init_random_map(struct thread_data *td)
+{
+       int num_maps, blocks;
+       struct fio_file *f;
+       unsigned int i;
+
+       if (td->o.norandommap)
+               return 0;
+
+       for_each_file(td, f, i) {
+               blocks = (f->real_file_size + td->o.rw_min_bs - 1) / td->o.rw_min_bs;
+               num_maps = (blocks + BLOCKS_PER_MAP-1)/ BLOCKS_PER_MAP;
+               f->file_map = malloc(num_maps * sizeof(long));
+               if (!f->file_map) {
+                       log_err("fio: failed allocating random map. If running a large number of jobs, try the 'norandommap' option\n");
                        return 1;
                }
+               f->num_maps = num_maps;
+               memset(f->file_map, 0, num_maps * sizeof(long));
        }
 
        return 0;
 }
 
-static int setup_files_plain(struct thread_data *td)
+void close_files(struct thread_data *td)
 {
        struct fio_file *f;
-       int i, err = 0;
+       unsigned int i;
 
        for_each_file(td, f, i) {
-               err = __setup_file_plain(td, f);
-               if (err)
-                       break;
+               if ((f->flags & FIO_FILE_UNLINK) &&
+                   f->filetype == FIO_TYPE_FILE)
+                       unlink(f->file_name);
+
+               td_io_close_file(td, f);
+
+               free(f->file_name);
+               f->file_name = NULL;
+
+               if (f->file_map) {
+                       free(f->file_map);
+                       f->file_map = NULL;
+               }
        }
 
-       return err;
+       td->o.filename = NULL;
+       free(td->files);
+       td->files = NULL;
+       td->o.nr_files = 0;
 }
 
-static int setup_file(struct thread_data *td, struct fio_file *f)
+static void get_file_type(struct fio_file *f)
 {
-       int flags = 0;
-
-       if (td->odirect)
-               flags |= OS_O_DIRECT;
+       struct stat sb;
+
+       f->filetype = FIO_TYPE_FILE;
+
+       if (!lstat(f->file_name, &sb)) {
+               f->flags |= FIO_FILE_EXISTS;
+
+               if (S_ISBLK(sb.st_mode))
+                       f->filetype = FIO_TYPE_BD;
+               else if (S_ISCHR(sb.st_mode))
+                       f->filetype = FIO_TYPE_CHAR;
+               else {
+                       /*
+                        * might as well do this here, and save a stat later on
+                        */
+                       f->real_file_size = sb.st_size;
+                       f->file_size = f->real_file_size;
+               }
+       }
+}
 
-       if (td_write(td) || td_rw(td)) {
-               if (td->filetype == FIO_TYPE_FILE) {
-                       if (!td->overwrite)
-                               flags |= O_TRUNC;
+void add_file(struct thread_data *td, const char *fname)
+{
+       int cur_files = td->files_index;
+       char file_name[PATH_MAX];
+       struct fio_file *f;
+       int len = 0;
 
-                       flags |= O_CREAT;
-               }
-               if (td->sync_io)
-                       flags |= O_SYNC;
+       td->files = realloc(td->files, (cur_files + 1) * sizeof(*f));
 
-               flags |= O_RDWR;
+       f = &td->files[cur_files];
+       memset(f, 0, sizeof(*f));
+       f->fd = -1;
 
-               f->fd = open(f->file_name, flags, 0600);
-       } else {
-               if (td->filetype == FIO_TYPE_CHAR)
-                       flags |= O_RDWR;
-               else
-                       flags |= O_RDONLY;
+       if (td->o.directory)
+               len = sprintf(file_name, "%s/", td->o.directory);
 
-               f->fd = open(f->file_name, flags);
-       }
+       sprintf(file_name + len, "%s", fname);
+       f->file_name = strdup(file_name);
 
-       if (f->fd == -1) {
-               td_verror(td, errno);
-               return 1;
-       }
+       get_file_type(f);
 
-       if (get_file_size(td, f))
-               return 1;
+       td->files_index++;
+       if (f->filetype == FIO_TYPE_FILE)
+               td->nr_normal_files++;
+}
 
-       return 0;
+void get_file(struct fio_file *f)
+{
+       f->references++;
 }
 
-int setup_files(struct thread_data *td)
+void put_file(struct thread_data *td, struct fio_file *f)
 {
-       struct fio_file *f;
-       int i, err;
+       if (!(f->flags & FIO_FILE_OPEN))
+               return;
 
-       /*
-        * if ioengine defines a setup() method, it's responsible for
-        * setting up everything in the td->files[] area.
-        */
-       if (td->io_ops->setup)
-               return td->io_ops->setup(td);
+       assert(f->references);
+       if (--f->references)
+               return;
 
-       if (create_files(td))
-               return 1;
+       if (should_fsync(td) && td->o.fsync_on_close)
+               fsync(f->fd);
 
-       for_each_file(td, f, i) {
-               err = setup_file(td, f);
-               if (err)
-                       break;
-       }
+       if (td->io_ops->close_file)
+               td->io_ops->close_file(td, f);
+       td->nr_open_files--;
+       f->flags &= ~FIO_FILE_OPEN;
+}
 
-       if (td->io_size == 0) {
-               log_err("%s: no io blocks\n", td->name);
-               td_verror(td, EINVAL);
+static int recurse_dir(struct thread_data *td, const char *dirname)
+{
+       struct dirent *dir;
+       int ret = 0;
+       DIR *D;
+
+       D = opendir(dirname);
+       if (!D) {
+               td_verror(td, errno, "opendir");
                return 1;
        }
 
-       if (!td->zone_size)
-               td->zone_size = td->io_size;
+       while ((dir = readdir(D)) != NULL) {
+               char full_path[PATH_MAX];
+               struct stat sb;
+
+               if (!strcmp(dir->d_name, ".") || !strcmp(dir->d_name, ".."))
+                       continue;
 
-       td->total_io_size = td->io_size * td->loops;
+               sprintf(full_path, "%s/%s", dirname, dir->d_name);
 
-       if (td->io_ops->flags & FIO_MMAPIO)
-               return setup_files_mmap(td);
-       else
-               return setup_files_plain(td);
+               if (lstat(full_path, &sb) == -1) {
+                       if (errno != ENOENT) {
+                               td_verror(td, errno, "stat");
+                               return 1;
+                       }
+               }
+
+               if (S_ISREG(sb.st_mode)) {
+                       add_file(td, full_path);
+                       td->o.nr_files++;
+                       continue;
+               }
+
+               if ((ret = recurse_dir(td, full_path)) != 0)
+                       break;
+       }
+
+       closedir(D);
+       return ret;
 }
 
-void close_files(struct thread_data *td)
+int add_dir_files(struct thread_data *td, const char *path)
+{
+       return recurse_dir(td, path);
+}
+
+void dup_files(struct thread_data *td, struct thread_data *org)
 {
        struct fio_file *f;
-       int i;
+       unsigned int i;
+       size_t bytes;
+
+       if (!org->files)
+               return;
+
+       bytes = org->files_index * sizeof(*f);
+       td->files = malloc(bytes);
+       memcpy(td->files, org->files, bytes);
 
        for_each_file(td, f, i) {
-               if (f->fd != -1) {
-                       if (td->unlink && td->filetype == FIO_TYPE_FILE)
-                               unlink(f->file_name);
-                       close(f->fd);
-                       f->fd = -1;
-               }
-               if (f->mmap) {
-                       munmap(f->mmap, f->file_size);
-                       f->mmap = NULL;
-               }
+               if (f->file_name)
+                       f->file_name = strdup(f->file_name);
        }
 }