Add the file sharing bits

When you use the same filename for several jobs now, they will share
the same file structure. Enable locking through two new options:

- lockfile. If set, a semaphore is associated with the file and it is
  held from ->prep() to ->queue() has done its work.
- lockfile_batch. This controls how many IOs the job gets to do per
  semaphore acqusition.

Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
diff --git a/HOWTO b/HOWTO
index d8c6b94..3e0a31b 100644
--- a/HOWTO
+++ b/HOWTO
@@ -219,6 +219,14 @@
 opendir=str	Tell fio to recursively add any file it can find in this
 		directory and down the file system tree.
 
+lockfile=bool	If set, fio will lock a file internally before doing IO to it.
+		This makes it safe to share file descriptors across fio
+		jobs that run at the same time.
+
+lockfile_batch=int	Acquiring a semaphore can be quite expensive, so
+		allow a process to complete this number of IOs before releasing
+		the semaphore again. Defaults to 1.
+
 readwrite=str
 rw=str		Type of io pattern. Accepted values are:
 
diff --git a/filesetup.c b/filesetup.c
index e2c71cd..4e2a36c 100644
--- a/filesetup.c
+++ b/filesetup.c
@@ -228,8 +228,10 @@
 
 int generic_open_file(struct thread_data *td, struct fio_file *f)
 {
+	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);
 
@@ -265,8 +267,16 @@
 
 		if (is_std)
 			f->fd = dup(STDOUT_FILENO);
-		else
-			f->fd = open(f->file_name, flags, 0600);
+		else {
+			__f = lookup_file_hash(f->file_name);
+			if (__f) {
+				f->sem = __f->sem;
+				f->fd = dup(__f->fd);
+				f->references++;
+				from_hash = 1;
+			} else
+				f->fd = open(f->file_name, flags, 0600);
+		}
 	} else {
 		if (f->filetype == FIO_TYPE_CHAR && !read_only)
 			flags |= O_RDWR;
@@ -275,8 +285,16 @@
 
 		if (is_std)
 			f->fd = dup(STDIN_FILENO);
-		else
-			f->fd = open(f->file_name, flags);
+		else {
+			__f = lookup_file_hash(f->file_name);
+			if (__f) {
+				f->sem = __f->sem;
+				f->fd = dup(__f->fd);
+				f->references++;
+				from_hash = 1;
+			} else
+				f->fd = open(f->file_name, flags);
+		}
 	}
 
 	if (f->fd == -1) {
@@ -296,7 +314,17 @@
 	if (get_file_size(td, f))
 		goto err;
 
-	add_file_hash(f);
+	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:
@@ -613,6 +641,9 @@
 
 	get_file_type(f);
 
+	if (td->o.lockfile)
+		f->sem = fio_sem_init(1);
+
 	td->files_index++;
 	if (f->filetype == FIO_TYPE_FILE)
 		td->nr_normal_files++;
@@ -653,10 +684,29 @@
 
 void lock_file(struct thread_data *td, struct fio_file *f)
 {
+	if (f && f->sem) {
+		if (f->sem_owner == td && f->sem_batch--)
+			return;
+
+		fio_sem_down(f->sem);
+		f->sem_owner = td;
+		f->sem_batch = td->o.lockfile_batch;
+	}
 }
 
 void unlock_file(struct fio_file *f)
 {
+	if (f && f->sem) {
+		int sem_val;
+
+		if (f->sem_batch)
+			return;
+
+		sem_getvalue(&f->sem->sem, &sem_val);
+		if (!sem_val)
+			f->sem_owner = NULL;
+		fio_sem_up(f->sem);
+	}
 }
 
 static int recurse_dir(struct thread_data *td, const char *dirname)
diff --git a/fio.h b/fio.h
index fe1a227..ffd3d7d 100644
--- a/fio.h
+++ b/fio.h
@@ -22,6 +22,7 @@
 #include "arch/arch.h"
 #include "os/os.h"
 #include "mutex.h"
+#include "sem.h"
 #include "log.h"
 #include "debug.h"
 
@@ -310,6 +311,13 @@
 	unsigned long long last_pos;
 
 	/*
+	 * if io is protected by a semaphore, this is set
+	 */
+	struct fio_sem *sem;
+	void *sem_owner;
+	unsigned int sem_batch;
+
+	/*
 	 * block map for random io
 	 */
 	unsigned long *file_map;
@@ -407,6 +415,8 @@
 
 	unsigned int nr_files;
 	unsigned int open_files;
+	unsigned int lockfile;
+	unsigned int lockfile_batch;
 
 	unsigned int odirect;
 	unsigned int invalidate_cache;
diff --git a/ioengines.c b/ioengines.c
index 5a2d6b9..bd2eb4a 100644
--- a/ioengines.c
+++ b/ioengines.c
@@ -359,5 +359,10 @@
 	 */
 	f->flags |= FIO_FILE_CLOSING;
 
+	if (f->sem_owner == td && f->sem_batch) {
+		f->sem_batch = 0;
+		unlock_file(f);
+	}
+
 	return put_file(td, f);
 }
diff --git a/options.c b/options.c
index 1119a98..27b4fb6 100644
--- a/options.c
+++ b/options.c
@@ -385,6 +385,22 @@
 		.help	= "File(s) to use for the workload",
 	},
 	{
+		.name	= "lockfile",
+		.type	= FIO_OPT_BOOL,
+		.off1	= td_var_offset(lockfile),
+		.help	= "Lock file when doing IO to it",
+		.parent	= "filename",
+		.def	= "0",
+	},
+	{
+		.name	= "lockfile_batch",
+		.type	= FIO_OPT_INT,
+		.off1	= td_var_offset(lockfile_batch),
+		.help	= "Number of IOs to allow per file lock",
+		.parent	= "lockfile",
+		.def	= "1",
+	},
+	{
 		.name	= "opendir",
 		.type	= FIO_OPT_STR_STORE,
 		.off1	= td_var_offset(opendir),