The following changes since commit d175907a7bd0b78cdd18bf47882428368a448057:
Fio 1.44 (2010-10-07 10:31:16 +0200)
are available in the git repository at:
git://git.kernel.dk/fio.git master
Jens Axboe (4):
binject: setup and teardown mappings internally
binject updates
Add missing file.h f->file_data change
Initial support for explicit write barriers
engines/binject.c | 212 ++++++++++++++++++++++++++++++++++++++++++----------
file.h | 9 +--
filesetup.c | 1 +
fio.h | 1 +
io_u.c | 15 +++-
ioengine.h | 2 +
options.c | 7 ++
os/binject.h | 2 +
8 files changed, 199 insertions(+), 50 deletions(-)
---
Diff of recent changes:
diff --git a/engines/binject.c b/engines/binject.c
index 68b6353..44a3796 100644
--- a/engines/binject.c
+++ b/engines/binject.c
@@ -24,6 +24,12 @@ struct binject_data {
int *fd_flags;
};
+struct binject_file {
+ unsigned int bs;
+ int minor;
+ int fd;
+};
+
static void binject_buc_init(struct binject_data *bd, struct io_u *io_u)
{
struct b_user_cmd *buc = &io_u->buc;
@@ -51,6 +57,37 @@ static int pollin_events(struct pollfd *pfds, int fds)
return 0;
}
+static unsigned int binject_read_commands(struct thread_data *td, void *p,
+ int left, int *err)
+{
+ struct binject_file *bf;
+ struct fio_file *f;
+ int i, ret, events;
+
+one_more:
+ events = 0;
+ for_each_file(td, f, i) {
+ bf = f->file_data;
+ ret = read(bf->fd, p, left * sizeof(struct b_user_cmd));
+ if (ret < 0) {
+ if (errno == EAGAIN)
+ continue;
+ *err = -errno;
+ td_verror(td, errno, "read");
+ break;
+ } else if (ret) {
+ p += ret;
+ events += ret / sizeof(struct b_user_cmd);
+ }
+ }
+
+ if (*err || events)
+ return events;
+
+ usleep(1000);
+ goto one_more;
+}
+
static int fio_binject_getevents(struct thread_data *td, unsigned int min,
unsigned int max, struct timespec fio_unused *t)
{
@@ -59,29 +96,27 @@ static int fio_binject_getevents(struct thread_data *td, unsigned int min,
void *buf = bd->cmds;
unsigned int i, events;
struct fio_file *f;
+ struct binject_file *bf;
/*
* Fill in the file descriptors
*/
for_each_file(td, f, i) {
+ bf = f->file_data;
+
/*
* don't block for min events == 0
*/
if (!min) {
- bd->fd_flags[i] = fcntl(f->fd, F_GETFL);
- fcntl(f->fd, F_SETFL, bd->fd_flags[i] | O_NONBLOCK);
+ bd->fd_flags[i] = fcntl(bf->fd, F_GETFL);
+ fcntl(bf->fd, F_SETFL, bd->fd_flags[i] | O_NONBLOCK);
}
- bd->pfds[i].fd = f->fd;
+ bd->pfds[i].fd = bf->fd;
bd->pfds[i].events = POLLIN;
}
while (left) {
- void *p;
-
- do {
- if (!min)
- break;
-
+ while (!min) {
ret = poll(bd->pfds, td->o.nr_files, -1);
if (ret < 0) {
if (!r)
@@ -93,34 +128,15 @@ static int fio_binject_getevents(struct thread_data *td, unsigned int min,
if (pollin_events(bd->pfds, td->o.nr_files))
break;
- } while (1);
+ }
if (r < 0)
break;
-re_read:
- p = buf;
- events = 0;
- for_each_file(td, f, i) {
- ret = read(f->fd, p, left * sizeof(struct b_user_cmd));
- if (ret < 0) {
- if (errno == EAGAIN)
- continue;
- r = -errno;
- td_verror(td, errno, "read");
- break;
- } else if (ret) {
- p += ret;
- events += ret / sizeof(struct b_user_cmd);
- }
- }
+ events = binject_read_commands(td, buf, left, &r);
if (r < 0)
break;
- if (!events) {
- usleep(1000);
- goto re_read;
- }
left -= events;
r += events;
@@ -134,8 +150,10 @@ re_read:
}
if (!min) {
- for_each_file(td, f, i)
- fcntl(f->fd, F_SETFL, bd->fd_flags[i]);
+ for_each_file(td, f, i) {
+ bf = f->file_data;
+ fcntl(bf->fd, F_SETFL, bd->fd_flags[i]);
+ }
}
if (r > 0)
@@ -147,10 +165,10 @@ re_read:
static int fio_binject_doio(struct thread_data *td, struct io_u *io_u)
{
struct b_user_cmd *buc = &io_u->buc;
- struct fio_file *f = io_u->file;
+ struct binject_file *bf = io_u->file->file_data;
int ret;
- ret = write(f->fd, buc, sizeof(*buc));
+ ret = write(bf->fd, buc, sizeof(*buc));
if (ret < 0)
return ret;
@@ -160,10 +178,10 @@ static int fio_binject_doio(struct thread_data *td, struct io_u *io_u)
static int fio_binject_prep(struct thread_data *td, struct io_u *io_u)
{
struct binject_data *bd = td->io_ops->data;
- unsigned int bs = io_u->file->file_data;
struct b_user_cmd *buc = &io_u->buc;
+ struct binject_file *bf = io_u->file->file_data;
- if (io_u->xfer_buflen & (bs - 1)) {
+ if (io_u->xfer_buflen & (bf->bs - 1)) {
log_err("read/write not sector aligned\n");
return EINVAL;
}
@@ -173,7 +191,10 @@ static int fio_binject_prep(struct thread_data *td, struct io_u *io_u)
buc->type = B_TYPE_READ;
} else if (io_u->ddir == DDIR_WRITE) {
binject_buc_init(bd, io_u);
- buc->type = B_TYPE_WRITE;
+ if (io_u->flags & IO_U_F_BARRIER)
+ buc->type = B_TYPE_WRITEBARRIER;
+ else
+ buc->type = B_TYPE_WRITE;
} else if (io_u->ddir == DDIR_TRIM) {
binject_buc_init(bd, io_u);
buc->type = B_TYPE_DISCARD;
@@ -210,8 +231,107 @@ static struct io_u *fio_binject_event(struct thread_data *td, int event)
return bd->events[event];
}
+static void binject_unmap_dev(struct thread_data *td, struct binject_file *bf)
+{
+ struct b_ioctl_cmd bic;
+ int fdb;
+
+ if (bf->fd >= 0) {
+ close(bf->fd);
+ bf->fd = -1;
+ }
+
+ fdb = open("/dev/binject-ctl", O_RDWR);
+ if (fdb < 0) {
+ td_verror(td, errno, "open binject-ctl");
+ return;
+ }
+
+ bic.minor = bf->minor;
+
+ if (ioctl(fdb, 1, &bic) < 0) {
+ td_verror(td, errno, "binject dev unmap");
+ close(fdb);
+ return;
+ }
+
+ close(fdb);
+}
+
+static int binject_map_dev(struct thread_data *td, struct binject_file *bf,
+ int fd)
+{
+ struct b_ioctl_cmd bic;
+ char name[80];
+ struct stat sb;
+ int fdb, dev_there, loops;
+
+ fdb = open("/dev/binject-ctl", O_RDWR);
+ if (fdb < 0) {
+ td_verror(td, errno, "binject ctl open");
+ return 1;
+ }
+
+ bic.fd = fd;
+
+ if (ioctl(fdb, 0, &bic) < 0) {
+ td_verror(td, errno, "binject dev map");
+ close(fdb);
+ return 1;
+ }
+
+ bf->minor = bic.minor;
+
+ sprintf(name, "/dev/binject%u", bf->minor);
+
+ /*
+ * Wait for udev to create the node...
+ */
+ dev_there = loops = 0;
+ do {
+ if (!stat(name, &sb)) {
+ dev_there = 1;
+ break;
+ }
+
+ usleep(10000);
+ } while (++loops < 100);
+
+ close(fdb);
+
+ if (!dev_there) {
+ log_err("fio: timed out waiting for binject dev\n");
+ goto err_unmap;
+ }
+
+ bf->fd = open(name, O_RDWR);
+ if (bf->fd < 0) {
+ td_verror(td, errno, "binject dev open");
+err_unmap:
+ binject_unmap_dev(td, bf);
+ return 1;
+ }
+
+ return 0;
+}
+
+static int fio_binject_close_file(struct thread_data *td, struct fio_file *f)
+{
+ struct binject_file *bf = f->file_data;
+
+ if (bf) {
+ binject_unmap_dev(td, bf);
+ free(bf);
+ f->file_data = NULL;
+ return generic_close_file(td, f);
+ }
+
+ return 0;
+}
+
static int fio_binject_open_file(struct thread_data *td, struct fio_file *f)
{
+ struct binject_file *bf;
unsigned int bs;
int ret;
@@ -221,14 +341,24 @@ static int fio_binject_open_file(struct thread_data *td, struct fio_file *f)
if (f->filetype != FIO_TYPE_BD) {
log_err("fio: binject only works with block devices\n");
- return 1;
+ goto err_close;
}
if (ioctl(f->fd, BLKSSZGET, &bs) < 0) {
td_verror(td, errno, "BLKSSZGET");
+ goto err_close;
+ }
+
+ bf = malloc(sizeof(*bf));
+ bf->bs = bs;
+ bf->minor = bf->fd = -1;
+ f->file_data = bf;
+
+ if (binject_map_dev(td, bf, f->fd)) {
+err_close:
+ ret = generic_close_file(td, f);
return 1;
}
- f->file_data = bs;
return 0;
}
@@ -278,9 +408,9 @@ static struct ioengine_ops ioengine = {
.event = fio_binject_event,
.cleanup = fio_binject_cleanup,
.open_file = fio_binject_open_file,
- .close_file = generic_close_file,
+ .close_file = fio_binject_close_file,
.get_file_size = generic_get_file_size,
- .flags = FIO_RAWIO,
+ .flags = FIO_RAWIO | FIO_BARRIER,
};
#else /* FIO_HAVE_BINJECT */
diff --git a/file.h b/file.h
index 6349465..fa8c1d2 100644
--- a/file.h
+++ b/file.h
@@ -47,13 +47,8 @@ struct fio_file {
struct flist_head hash_list;
enum fio_filetype filetype;
- /*
- * A file may not be a file descriptor, let the io engine decide
- */
- union {
- unsigned long file_data;
- int fd;
- };
+ void *file_data;
+ int fd;
/*
* filename and possible memory mapping
diff --git a/filesetup.c b/filesetup.c
index 5a8105a..ae893c7 100644
--- a/filesetup.c
+++ b/filesetup.c
@@ -244,6 +244,7 @@ static int bdev_size(struct thread_data *td, struct fio_file *f)
r = blockdev_size(f->fd, &bytes);
if (r) {
td_verror(td, r, "blockdev_size");
+ printf("fd is %d\n", f->fd);
goto err;
}
diff --git a/fio.h b/fio.h
index c2a0d4d..4ed8cb1 100644
--- a/fio.h
+++ b/fio.h
@@ -217,6 +217,7 @@ struct thread_options {
unsigned int thinktime_blocks;
unsigned int fsync_blocks;
unsigned int fdatasync_blocks;
+ unsigned int barrier_blocks;
unsigned long start_delay;
unsigned long long timeout;
unsigned long long ramp_time;
diff --git a/io_u.c b/io_u.c
index baa961b..6d539a0 100644
--- a/io_u.c
+++ b/io_u.c
@@ -501,6 +501,17 @@ static enum fio_ddir get_rw_ddir(struct thread_data *td)
return td->rwmix_ddir;
}
+static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
+{
+ io_u->ddir = get_rw_ddir(td);
+
+ if (io_u->ddir == DDIR_WRITE && (td->io_ops->flags & FIO_BARRIER) &&
+ td->o.barrier_blocks &&
+ !(td->io_issues[DDIR_WRITE] % td->o.barrier_blocks) &&
+ td->io_issues[DDIR_WRITE])
+ io_u->flags |= IO_U_F_BARRIER;
+}
+
void put_file_log(struct thread_data *td, struct fio_file *f)
{
int ret = put_file(td, f);
@@ -560,7 +571,7 @@ static int fill_io_u(struct thread_data *td, struct io_u *io_u)
if (td->io_ops->flags & FIO_NOIO)
goto out;
- io_u->ddir = get_rw_ddir(td);
+ set_rw_ddir(td, io_u);
/*
* fsync() or fdatasync() or trim etc, we are done
@@ -963,7 +974,7 @@ again:
if (io_u) {
assert(io_u->flags & IO_U_F_FREE);
io_u->flags &= ~(IO_U_F_FREE | IO_U_F_FREE_DEF);
- io_u->flags &= ~IO_U_F_TRIMMED;
+ io_u->flags &= ~(IO_U_F_TRIMMED | IO_U_F_BARRIER);
io_u->error = 0;
flist_del(&io_u->list);
diff --git a/ioengine.h b/ioengine.h
index 344cdbf..7a3c08f 100644
--- a/ioengine.h
+++ b/ioengine.h
@@ -10,6 +10,7 @@ enum {
IO_U_F_IN_CUR_DEPTH = 1 << 3,
IO_U_F_BUSY_OK = 1 << 4,
IO_U_F_TRIMMED = 1 << 5,
+ IO_U_F_BARRIER = 1 << 6,
};
/*
@@ -126,6 +127,7 @@ enum fio_ioengine_flags {
FIO_NOIO = 1 << 6, /* thread does only pseudo IO */
FIO_SIGQUIT = 1 << 7, /* needs SIGQUIT to exit */
FIO_PIPEIO = 1 << 8, /* input/output no seekable */
+ FIO_BARRIER = 1 << 9, /* engine supports barriers */
};
/*
diff --git a/options.c b/options.c
index bdf3582..d111018 100644
--- a/options.c
+++ b/options.c
@@ -1185,6 +1185,13 @@ static struct fio_option options[FIO_MAX_OPTS] = {
.help = "Issue fdatasync for writes every given number of blocks",
.def = "0",
},
+ {
+ .name = "write_barrier",
+ .type = FIO_OPT_INT,
+ .off1 = td_var_offset(barrier_blocks),
+ .help = "Make every Nth write a barrier write",
+ .def = "0",
+ },
#ifdef FIO_HAVE_SYNC_FILE_RANGE
{
.name = "sync_file_range",
diff --git a/os/binject.h b/os/binject.h
index 1f358f3..24eee0d 100644
--- a/os/binject.h
+++ b/os/binject.h
@@ -31,6 +31,8 @@ enum {
B_TYPE_DISCARD,
B_TYPE_READVOID,
B_TYPE_WRITEZERO,
+ B_TYPE_READBARRIER,
+ B_TYPE_WRITEBARRIER,
B_TYPE_NR
};
--
To unsubscribe from this list: send the line "unsubscribe fio" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
