#include <unistd.h>
#include <errno.h>
#include <assert.h>
+#include <libaio.h>
#include "../fio.h"
+#include "../lib/pow2.h"
-#ifdef FIO_HAVE_LIBAIO
-
-#define ev_to_iou(ev) (struct io_u *) ((unsigned long) (ev)->obj)
+static int fio_libaio_commit(struct thread_data *td);
struct libaio_data {
io_context_t aio_ctx;
struct io_event *aio_events;
struct iocb **iocbs;
struct io_u **io_us;
- int iocbs_nr;
+
+ /*
+ * Basic ring buffer. 'head' is incremented in _queue(), and
+ * 'tail' is incremented in _commit(). We keep 'queued' so
+ * that we know if the ring is full or empty, when
+ * 'head' == 'tail'. 'entries' is the ring size, and
+ * 'is_pow2' is just an optimization to use AND instead of
+ * modulus to get the remainder on ring increment.
+ */
+ int is_pow2;
+ unsigned int entries;
+ unsigned int queued;
+ unsigned int head;
+ unsigned int tail;
+};
+
+struct libaio_options {
+ void *pad;
+ unsigned int userspace_reap;
+};
+
+static struct fio_option options[] = {
+ {
+ .name = "userspace_reap",
+ .lname = "Libaio userspace reaping",
+ .type = FIO_OPT_STR_SET,
+ .off1 = offsetof(struct libaio_options, userspace_reap),
+ .help = "Use alternative user-space reap implementation",
+ .category = FIO_OPT_C_ENGINE,
+ .group = FIO_OPT_G_LIBAIO,
+ },
+ {
+ .name = NULL,
+ },
};
+static inline void ring_inc(struct libaio_data *ld, unsigned int *val,
+ unsigned int add)
+{
+ if (ld->is_pow2)
+ *val = (*val + add) & (ld->entries - 1);
+ else
+ *val = (*val + add) % ld->entries;
+}
+
static int fio_libaio_prep(struct thread_data fio_unused *td, struct io_u *io_u)
{
struct fio_file *f = io_u->file;
io_prep_pread(&io_u->iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
else if (io_u->ddir == DDIR_WRITE)
io_prep_pwrite(&io_u->iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
- else if (io_u->ddir == DDIR_SYNC)
+ else if (ddir_sync(io_u->ddir))
io_prep_fsync(&io_u->iocb, f->fd);
- else
- return 1;
return 0;
}
struct io_u *io_u;
ev = ld->aio_events + event;
- io_u = ev_to_iou(ev);
+ io_u = container_of(ev->obj, struct io_u, iocb);
if (ev->res != io_u->xfer_buflen) {
if (ev->res > io_u->xfer_buflen)
return io_u;
}
+struct aio_ring {
+ unsigned id; /** kernel internal index number */
+ unsigned nr; /** number of io_events */
+ unsigned head;
+ unsigned tail;
+
+ unsigned magic;
+ unsigned compat_features;
+ unsigned incompat_features;
+ unsigned header_length; /** size of aio_ring */
+
+ struct io_event events[0];
+};
+
+#define AIO_RING_MAGIC 0xa10a10a1
+
+static int user_io_getevents(io_context_t aio_ctx, unsigned int max,
+ struct io_event *events)
+{
+ long i = 0;
+ unsigned head;
+ struct aio_ring *ring = (struct aio_ring*) aio_ctx;
+
+ while (i < max) {
+ head = ring->head;
+
+ if (head == ring->tail) {
+ /* There are no more completions */
+ break;
+ } else {
+ /* There is another completion to reap */
+ events[i] = ring->events[head];
+ read_barrier();
+ ring->head = (head + 1) % ring->nr;
+ i++;
+ }
+ }
+
+ return i;
+}
+
static int fio_libaio_getevents(struct thread_data *td, unsigned int min,
- unsigned int max, struct timespec *t)
+ unsigned int max, const struct timespec *t)
{
struct libaio_data *ld = td->io_ops->data;
- long r;
+ struct libaio_options *o = td->eo;
+ unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
+ struct timespec __lt, *lt = NULL;
+ int r, events = 0;
+
+ if (t) {
+ __lt = *t;
+ lt = &__lt;
+ }
do {
- r = io_getevents(ld->aio_ctx, min, max, ld->aio_events, t);
- if (r >= min)
- break;
- else if (r == -EAGAIN) {
+ if (o->userspace_reap == 1
+ && actual_min == 0
+ && ((struct aio_ring *)(ld->aio_ctx))->magic
+ == AIO_RING_MAGIC) {
+ r = user_io_getevents(ld->aio_ctx, max,
+ ld->aio_events + events);
+ } else {
+ r = io_getevents(ld->aio_ctx, actual_min,
+ max, ld->aio_events + events, lt);
+ }
+ if (r > 0)
+ events += r;
+ else if ((min && r == 0) || r == -EAGAIN) {
+ fio_libaio_commit(td);
usleep(100);
- continue;
- } else if (r == -EINTR)
- continue;
- else if (r != 0)
+ } else if (r != -EINTR)
break;
- } while (1);
+ } while (events < min);
- return r;
+ return r < 0 ? r : events;
}
static int fio_libaio_queue(struct thread_data *td, struct io_u *io_u)
fio_ro_check(td, io_u);
- if (ld->iocbs_nr == (int) td->o.iodepth)
+ if (ld->queued == td->o.iodepth)
return FIO_Q_BUSY;
/*
* support aio fsync yet. So return busy for the case where we
* have pending io, to let fio complete those first.
*/
- if (io_u->ddir == DDIR_SYNC) {
- if (ld->iocbs_nr)
+ if (ddir_sync(io_u->ddir)) {
+ if (ld->queued)
+ return FIO_Q_BUSY;
+
+ do_io_u_sync(td, io_u);
+ return FIO_Q_COMPLETED;
+ }
+
+ if (io_u->ddir == DDIR_TRIM) {
+ if (ld->queued)
return FIO_Q_BUSY;
- if (fsync(io_u->file->fd) < 0)
- io_u->error = errno;
+ do_io_u_trim(td, io_u);
return FIO_Q_COMPLETED;
}
- ld->iocbs[ld->iocbs_nr] = &io_u->iocb;
- ld->io_us[ld->iocbs_nr] = io_u;
- ld->iocbs_nr++;
+ ld->iocbs[ld->head] = &io_u->iocb;
+ ld->io_us[ld->head] = io_u;
+ ring_inc(ld, &ld->head, 1);
+ ld->queued++;
return FIO_Q_QUEUED;
}
struct timeval now;
unsigned int i;
+ if (!fio_fill_issue_time(td))
+ return;
+
fio_gettime(&now, NULL);
for (i = 0; i < nr; i++) {
struct libaio_data *ld = td->io_ops->data;
struct iocb **iocbs;
struct io_u **io_us;
- int ret;
+ struct timeval tv;
+ int ret, wait_start = 0;
- if (!ld->iocbs_nr)
+ if (!ld->queued)
return 0;
- io_us = ld->io_us;
- iocbs = ld->iocbs;
do {
- ret = io_submit(ld->aio_ctx, ld->iocbs_nr, iocbs);
+ long nr = ld->queued;
+
+ nr = min((unsigned int) nr, ld->entries - ld->tail);
+ io_us = ld->io_us + ld->tail;
+ iocbs = ld->iocbs + ld->tail;
+
+ ret = io_submit(ld->aio_ctx, nr, iocbs);
if (ret > 0) {
fio_libaio_queued(td, io_us, ret);
- ld->iocbs_nr -= ret;
- io_us += ret;
- iocbs += ret;
+ io_u_mark_submit(td, ret);
+
+ ld->queued -= ret;
+ ring_inc(ld, &ld->tail, ret);
ret = 0;
- } else if (!ret || ret == -EAGAIN || ret == -EINTR)
+ wait_start = 0;
+ } else if (ret == -EINTR || !ret) {
+ if (!ret)
+ io_u_mark_submit(td, ret);
+ wait_start = 0;
continue;
- else
+ } else if (ret == -EAGAIN) {
+ /*
+ * If we get EAGAIN, we should break out without
+ * error and let the upper layer reap some
+ * events for us. If we have no queued IO, we
+ * must loop here. If we loop for more than 30s,
+ * just error out, something must be buggy in the
+ * IO path.
+ */
+ if (ld->queued) {
+ ret = 0;
+ break;
+ }
+ if (!wait_start) {
+ fio_gettime(&tv, NULL);
+ wait_start = 1;
+ } else if (mtime_since_now(&tv) > 30000) {
+ log_err("fio: aio appears to be stalled, giving up\n");
+ break;
+ }
+ usleep(1);
+ continue;
+ } else if (ret == -ENOMEM) {
+ /*
+ * If we get -ENOMEM, reap events if we can. If
+ * we cannot, treat it as a fatal event since there's
+ * nothing we can do about it.
+ */
+ if (ld->queued)
+ ret = 0;
+ break;
+ } else
break;
- } while (ld->iocbs_nr);
+ } while (ld->queued);
return ret;
}
struct libaio_data *ld = td->io_ops->data;
if (ld) {
- io_destroy(ld->aio_ctx);
+ /*
+ * Work-around to avoid huge RCU stalls at exit time. If we
+ * don't do this here, then it'll be torn down by exit_aio().
+ * But for that case we can parallellize the freeing, thus
+ * speeding it up a lot.
+ */
+ if (!(td->flags & TD_F_CHILD))
+ io_destroy(ld->aio_ctx);
free(ld->aio_events);
free(ld->iocbs);
free(ld->io_us);
free(ld);
- td->io_ops->data = NULL;
}
}
static int fio_libaio_init(struct thread_data *td)
{
- struct libaio_data *ld = malloc(sizeof(*ld));
- static int warn_print;
- int err;
+ struct libaio_options *o = td->eo;
+ struct libaio_data *ld;
+ int err = 0;
- if (td->o.iodepth > 1 && !td->o.odirect && !warn_print) {
- log_info("fio: libaio engine is only async for non-buffered IO\n");
- warn_print = 1;
- }
-
- memset(ld, 0, sizeof(*ld));
+ ld = calloc(1, sizeof(*ld));
- err = io_queue_init(td->o.iodepth, &ld->aio_ctx);
+ /*
+ * First try passing in 0 for queue depth, since we don't
+ * care about the user ring. If that fails, the kernel is too old
+ * and we need the right depth.
+ */
+ if (!o->userspace_reap)
+ err = io_queue_init(INT_MAX, &ld->aio_ctx);
+ if (o->userspace_reap || err == -EINVAL)
+ err = io_queue_init(td->o.iodepth, &ld->aio_ctx);
if (err) {
td_verror(td, -err, "io_queue_init");
log_err("fio: check /proc/sys/fs/aio-max-nr\n");
return 1;
}
- ld->aio_events = malloc(td->o.iodepth * sizeof(struct io_event));
- memset(ld->aio_events, 0, td->o.iodepth * sizeof(struct io_event));
- ld->iocbs = malloc(td->o.iodepth * sizeof(struct iocb *));
- memset(ld->iocbs, 0, sizeof(struct iocb *));
- ld->io_us = malloc(td->o.iodepth * sizeof(struct io_u *));
- memset(ld->io_us, 0, td->o.iodepth * sizeof(struct io_u *));
- ld->iocbs_nr = 0;
+ ld->entries = td->o.iodepth;
+ ld->is_pow2 = is_power_of_2(ld->entries);
+ ld->aio_events = calloc(ld->entries, sizeof(struct io_event));
+ ld->iocbs = calloc(ld->entries, sizeof(struct iocb *));
+ ld->io_us = calloc(ld->entries, sizeof(struct io_u *));
td->io_ops->data = ld;
return 0;
}
static struct ioengine_ops ioengine = {
- .name = "libaio",
- .version = FIO_IOOPS_VERSION,
- .init = fio_libaio_init,
- .prep = fio_libaio_prep,
- .queue = fio_libaio_queue,
- .commit = fio_libaio_commit,
- .cancel = fio_libaio_cancel,
- .getevents = fio_libaio_getevents,
- .event = fio_libaio_event,
- .cleanup = fio_libaio_cleanup,
- .open_file = generic_open_file,
- .close_file = generic_close_file,
-};
-
-#else /* FIO_HAVE_LIBAIO */
-
-/*
- * When we have a proper configure system in place, we simply wont build
- * and install this io engine. For now install a crippled version that
- * just complains and fails to load.
- */
-static int fio_libaio_init(struct thread_data fio_unused *td)
-{
- fprintf(stderr, "fio: libaio not available\n");
- return 1;
-}
-
-static struct ioengine_ops ioengine = {
- .name = "libaio",
- .version = FIO_IOOPS_VERSION,
- .init = fio_libaio_init,
+ .name = "libaio",
+ .version = FIO_IOOPS_VERSION,
+ .init = fio_libaio_init,
+ .prep = fio_libaio_prep,
+ .queue = fio_libaio_queue,
+ .commit = fio_libaio_commit,
+ .cancel = fio_libaio_cancel,
+ .getevents = fio_libaio_getevents,
+ .event = fio_libaio_event,
+ .cleanup = fio_libaio_cleanup,
+ .open_file = generic_open_file,
+ .close_file = generic_close_file,
+ .get_file_size = generic_get_file_size,
+ .options = options,
+ .option_struct_size = sizeof(struct libaio_options),
};
-#endif
-
static void fio_init fio_libaio_register(void)
{
register_ioengine(&ioengine);
projects / fio.git / blobdiff