Revamp file locking
[fio.git] / filesetup.c
index b8b22b5a54da51504b369b67975c46ef10f95084..4d2017dbed798d045f3c2727debc64ef5d4238eb 100644 (file)
@@ -8,6 +8,8 @@
 #include <sys/types.h>
 
 #include "fio.h"
+#include "smalloc.h"
+#include "filehash.h"
 
 static int root_warn;
 
@@ -18,6 +20,11 @@ static int extend_file(struct thread_data *td, struct fio_file *f)
        unsigned int bs;
        char *b;
 
+       if (read_only) {
+               log_err("fio: refusing extend of file due to read-only\n");
+               return 0;
+       }
+
        /*
         * check if we need to lay the file out complete again. fio
         * does that for operations involving reads, or for writes
@@ -28,8 +35,8 @@ static int extend_file(struct thread_data *td, struct fio_file *f)
        if (td_write(td) && !td->o.overwrite)
                unlink_file = 1;
 
-       if ((unlink_file || new_layout) && (f->flags & FIO_FILE_EXISTS)) {
-               if (unlink(f->file_name) < 0) {
+       if (unlink_file || new_layout) {
+               if ((unlink(f->file_name) < 0) && (errno != ENOENT)) {
                        td_verror(td, errno, "unlink");
                        return 1;
                }
@@ -39,20 +46,25 @@ static int extend_file(struct thread_data *td, struct fio_file *f)
        if (new_layout)
                flags |= O_TRUNC;
 
+       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");
                return 1;
        }
 
+       if (!new_layout)
+               goto done;
+
+       dprint(FD_FILE, "truncate file %s, size %llu\n", f->file_name,
+                                                       f->real_file_size);
        if (ftruncate(f->fd, f->real_file_size) == -1) {
                td_verror(td, errno, "ftruncate");
                goto err;
        }
 
-       if (!new_layout)
-               goto done;
-
+       dprint(FD_FILE, "fallocate file %s, size %llu\n", f->file_name,
+                                                       f->real_file_size);
        if (posix_fallocate(f->fd, 0, f->real_file_size) < 0) {
                td_verror(td, errno, "posix_fallocate");
                goto err;
@@ -104,7 +116,7 @@ static unsigned long long get_rand_file_size(struct thread_data *td)
        long r;
 
        r = os_random_long(&td->file_size_state);
-       ret = td->o.file_size_low + (unsigned long long) ((double) td->o.file_size_high * (r / (RAND_MAX + 1.0)));
+       ret = td->o.file_size_low + (unsigned long long) ((double) (td->o.file_size_high - td->o.file_size_low) * (r / (RAND_MAX + 1.0)));
        ret -= (ret % td->o.rw_min_bs);
        return ret;
 }
@@ -167,6 +179,8 @@ int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
 {
        int ret = 0;
 
+       dprint(FD_IO, "invalidate cache (%d)\n", td->o.odirect);
+
        if (td->o.odirect)
                return 0;
 
@@ -197,16 +211,53 @@ int file_invalidate_cache(struct thread_data *td, struct fio_file *f)
        return ret;
 }
 
-void generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
+int generic_close_file(struct thread_data fio_unused *td, struct fio_file *f)
 {
-       close(f->fd);
+       int ret = 0;
+
+       dprint(FD_FILE, "fd close %s\n", f->file_name);
+
+       remove_file_hash(f);
+
+       if (close(f->fd) < 0)
+               ret = errno;
+
        f->fd = -1;
+       return ret;
+}
+
+static int file_lookup_open(struct fio_file *f, int flags)
+{
+       struct fio_file *__f;
+       int from_hash;
+
+       __f = lookup_file_hash(f->file_name);
+       if (__f) {
+               /*
+                * racy, need the __f->lock locked
+                */
+               f->lock = __f->lock;
+               f->lock_owner = __f->lock_owner;
+               f->lock_batch = __f->lock_batch;
+               f->lock_ddir = __f->lock_ddir;
+               f->fd = dup(__f->fd);
+               f->references++;
+               from_hash = 1;
+       } else {
+               f->fd = open(f->file_name, flags, 0600);
+               from_hash = 0;
+       }
+
+       return from_hash;
 }
 
 int generic_open_file(struct thread_data *td, struct fio_file *f)
 {
        int is_std = 0;
        int flags = 0;
+       int from_hash = 0;
+
+       dprint(FD_FILE, "fd open %s\n", f->file_name);
 
        if (!strcmp(f->file_name, "-")) {
                if (td_rw(td)) {
@@ -226,8 +277,13 @@ int generic_open_file(struct thread_data *td, struct fio_file *f)
                flags |= OS_O_DIRECT;
        if (td->o.sync_io)
                flags |= O_SYNC;
+       if (f->filetype != FIO_TYPE_FILE)
+               flags |= O_NOATIME;
 
+open_again:
        if (td_write(td)) {
+               assert(!read_only);
+
                flags |= O_RDWR;
 
                if (f->filetype == FIO_TYPE_FILE)
@@ -236,9 +292,9 @@ int generic_open_file(struct thread_data *td, struct fio_file *f)
                if (is_std)
                        f->fd = dup(STDOUT_FILENO);
                else
-                       f->fd = open(f->file_name, flags, 0600);
+                       from_hash = file_lookup_open(f, flags);
        } else {
-               if (f->filetype == FIO_TYPE_CHAR)
+               if (f->filetype == FIO_TYPE_CHAR && !read_only)
                        flags |= O_RDWR;
                else
                        flags |= O_RDONLY;
@@ -246,13 +302,18 @@ int generic_open_file(struct thread_data *td, struct fio_file *f)
                if (is_std)
                        f->fd = dup(STDIN_FILENO);
                else
-                       f->fd = open(f->file_name, flags);
+                       from_hash = file_lookup_open(f, flags);
        }
 
        if (f->fd == -1) {
                char buf[FIO_VERROR_SIZE];
                int __e = errno;
 
+               if (errno == EPERM && (flags & O_NOATIME)) {
+                       flags &= ~O_NOATIME;
+                       goto open_again;
+               }
+
                snprintf(buf, sizeof(buf) - 1, "open(%s)", f->file_name);
 
                td_verror(td, __e, buf);
@@ -261,6 +322,18 @@ int generic_open_file(struct thread_data *td, struct fio_file *f)
        if (get_file_size(td, f))
                goto err;
 
+       if (!from_hash && f->fd != -1) {
+               if (add_file_hash(f)) {
+                       int ret;
+
+                       /*
+                        * OK to ignore, we haven't done anything with it
+                        */
+                       ret = generic_close_file(td, f);
+                       goto open_again;
+               }
+       }
+
        return 0;
 err:
        close(f->fd);
@@ -273,6 +346,8 @@ int open_files(struct thread_data *td)
        unsigned int i;
        int err = 0;
 
+       dprint(FD_FILE, "open files\n");
+
        for_each_file(td, f, i) {
                err = td_io_open_file(td, f);
                if (err) {
@@ -336,6 +411,8 @@ int setup_files(struct thread_data *td)
        unsigned int i;
        int err = 0, need_extend;
 
+       dprint(FD_FILE, "setup files\n");
+
        /*
         * if ioengine defines a setup() method, it's responsible for
         * opening the files and setting f->real_file_size to indicate
@@ -364,7 +441,8 @@ int setup_files(struct thread_data *td)
        /*
         * device/file sizes are zero and no size given, punt
         */
-       if ((!total_size || total_size == -1ULL) && !td->o.size) {
+       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);
                td_verror(td, EINVAL, "total_file_size");
                return 1;
@@ -387,11 +465,8 @@ int setup_files(struct thread_data *td)
                         * zero, set it to the real file size.
                         */
                        f->io_size = td->o.size / td->o.nr_files;
-                       if (!f->io_size) {
-                               if (f->file_offset > f->real_file_size)
-                                       goto err_offset;
+                       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) 
@@ -404,9 +479,7 @@ int setup_files(struct thread_data *td)
                                f->io_size = td->o.file_size_low - f->file_offset;
                        else
                                f->io_size = get_rand_file_size(td) - f->file_offset;
-               } else if (f->file_offset > f->real_file_size)
-                       goto err_offset;
-               else
+               } else
                        f->io_size = f->real_file_size - f->file_offset;
 
                if (f->io_size == -1ULL)
@@ -431,7 +504,7 @@ int setup_files(struct thread_data *td)
         */
        if (need_extend) {
                temp_stall_ts = 1;
-               log_info("%s: Laying out IO file(s) (%u files / %LuMiB)\n",
+               log_info("%s: Laying out IO file(s) (%u file(s) / %LuMiB)\n",
                        td->o.name, need_extend, extend_size >> 20);
 
                for_each_file(td, f, i) {
@@ -454,7 +527,12 @@ int setup_files(struct thread_data *td)
        if (!td->o.zone_size)
                td->o.zone_size = td->o.size;
 
-       td->total_io_size = td->o.size * td->o.loops;
+       /*
+        * 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;
        return 0;
 err_offset:
        log_err("%s: you need to specify valid offset=\n", td->o.name);
@@ -463,23 +541,22 @@ err_offset:
 
 int init_random_map(struct thread_data *td)
 {
-       int num_maps, blocks;
+       unsigned long long blocks, num_maps;
        struct fio_file *f;
        unsigned int i;
 
-       if (td->o.norandommap)
+       if (td->o.norandommap || !td_random(td))
                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));
+               blocks = (f->real_file_size + td->o.rw_min_bs - 1) / (unsigned long long) td->o.rw_min_bs;
+               num_maps = (blocks + BLOCKS_PER_MAP-1)/ (unsigned long long) BLOCKS_PER_MAP;
+               f->file_map = smalloc(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;
@@ -490,19 +567,31 @@ void close_files(struct thread_data *td)
        struct fio_file *f;
        unsigned int i;
 
+       for_each_file(td, f, i)
+               td_io_close_file(td, f);
+}
+
+void close_and_free_files(struct thread_data *td)
+{
+       struct fio_file *f;
+       unsigned int i;
+
+       dprint(FD_FILE, "close files\n");
+
        for_each_file(td, f, i) {
                if (td->o.unlink && f->filetype == FIO_TYPE_FILE)
                        unlink(f->file_name);
 
                td_io_close_file(td, f);
 
-               free(f->file_name);
+               sfree(f->file_name);
                f->file_name = NULL;
 
                if (f->file_map) {
-                       free(f->file_map);
+                       sfree(f->file_map);
                        f->file_map = NULL;
                }
+               sfree(f);
        }
 
        td->o.filename = NULL;
@@ -537,12 +626,15 @@ int add_file(struct thread_data *td, const char *fname)
        struct fio_file *f;
        int len = 0;
 
-       td->files = realloc(td->files, (cur_files + 1) * sizeof(*f));
+       dprint(FD_FILE, "add file %s\n", fname);
 
-       f = &td->files[cur_files];
-       memset(f, 0, sizeof(*f));
+       f = smalloc(sizeof(*f));
        f->fd = -1;
 
+       td->files = realloc(td->files, (cur_files + 1) * sizeof(f));
+
+       td->files[cur_files] = f;
+
        /*
         * init function, io engine may not be loaded yet
         */
@@ -553,10 +645,24 @@ int add_file(struct thread_data *td, const char *fname)
                len = sprintf(file_name, "%s/", td->o.directory);
 
        sprintf(file_name + len, "%s", fname);
-       f->file_name = strdup(file_name);
+       f->file_name = smalloc_strdup(file_name);
 
        get_file_type(f);
 
+       switch (td->o.file_lock_mode) {
+       case FILE_LOCK_NONE:
+               break;
+       case FILE_LOCK_READWRITE:
+               f->lock = fio_mutex_rw_init();
+               break;
+       case FILE_LOCK_EXCLUSIVE:
+               f->lock = fio_mutex_init(1);
+               break;
+       default:
+               log_err("fio: unknown lock mode: %d\n", td->o.file_lock_mode);
+               assert(0);
+       }
+
        td->files_index++;
        if (f->filetype == FIO_TYPE_FILE)
                td->nr_normal_files++;
@@ -566,27 +672,91 @@ int add_file(struct thread_data *td, const char *fname)
 
 void get_file(struct fio_file *f)
 {
+       dprint(FD_FILE, "get file %s, ref=%d\n", f->file_name, f->references);
        assert(f->flags & FIO_FILE_OPEN);
        f->references++;
 }
 
-void put_file(struct thread_data *td, struct fio_file *f)
+int put_file(struct thread_data *td, struct fio_file *f)
 {
+       int ret = 0;
+
+       dprint(FD_FILE, "put file %s, ref=%d\n", f->file_name, f->references);
+
        if (!(f->flags & FIO_FILE_OPEN))
-               return;
+               return 0;
 
        assert(f->references);
        if (--f->references)
-               return;
+               return 0;
 
        if (should_fsync(td) && td->o.fsync_on_close)
                fsync(f->fd);
 
        if (td->io_ops->close_file)
-               td->io_ops->close_file(td, f);
+               ret = td->io_ops->close_file(td, f);
 
        td->nr_open_files--;
        f->flags &= ~FIO_FILE_OPEN;
+       return ret;
+}
+
+void lock_file(struct thread_data *td, struct fio_file *f, enum fio_ddir ddir)
+{
+       if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
+               return;
+
+       if (f->lock_owner == td && f->lock_batch--)
+               return;
+
+       if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
+               if (ddir == DDIR_READ)
+                       fio_mutex_down_read(f->lock);
+               else
+                       fio_mutex_down_write(f->lock);
+       } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE)
+               fio_mutex_down(f->lock);
+
+       f->lock_owner = td;
+       f->lock_batch = td->o.lockfile_batch;
+       f->lock_ddir = ddir;
+}
+
+void unlock_file(struct thread_data *td, struct fio_file *f)
+{
+       if (!f->lock || td->o.file_lock_mode == FILE_LOCK_NONE)
+               return;
+       if (f->lock_batch)
+               return;
+
+       if (td->o.file_lock_mode == FILE_LOCK_READWRITE) {
+               const int is_read = f->lock_ddir == DDIR_READ;
+               int val = fio_mutex_getval(f->lock);
+
+               if ((is_read && val == 1) || (!is_read && val == -1))
+                       f->lock_owner = NULL;
+
+               if (is_read)
+                       fio_mutex_up_read(f->lock);
+               else
+                       fio_mutex_up_write(f->lock);
+       } else if (td->o.file_lock_mode == FILE_LOCK_EXCLUSIVE) {
+               int val = fio_mutex_getval(f->lock);
+
+               if (val == 0)
+                       f->lock_owner = NULL;
+
+               fio_mutex_up(f->lock);
+       }
+}
+
+void unlock_file_all(struct thread_data *td, struct fio_file *f)
+{
+       if (f->lock_owner != td)
+               return;
+
+       f->lock_batch = 0;
+       unlock_file(td, f);
 }
 
 static int recurse_dir(struct thread_data *td, const char *dirname)
@@ -655,13 +825,19 @@ void dup_files(struct thread_data *td, struct thread_data *org)
        if (!org->files)
                return;
 
-       bytes = org->files_index * sizeof(*f);
+       bytes = org->files_index * sizeof(f);
        td->files = malloc(bytes);
        memcpy(td->files, org->files, bytes);
 
        for_each_file(td, f, i) {
+               struct fio_file *__f;
+
+               __f = smalloc(sizeof(*__f));
+
                if (f->file_name)
-                       f->file_name = strdup(f->file_name);
+                       __f->file_name = smalloc_strdup(f->file_name);
+
+               td->files[i] = __f;
        }
 }