#include <libaio.h>
#include "../fio.h"
+#include "../lib/pow2.h"
+
+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 {
- struct thread_data *td;
+ void *pad;
unsigned int userspace_reap;
};
.type = FIO_OPT_STR_SET,
.off1 = offsetof(struct libaio_options, userspace_reap),
.help = "Use alternative user-space reap implementation",
- .category = FIO_OPT_C_IO,
+ .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;
}
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;
struct libaio_options *o = td->eo;
unsigned actual_min = td->o.iodepth_batch_complete == 0 ? 0 : min;
+ struct timespec __lt, *lt = NULL;
int r, events = 0;
+ if (t) {
+ __lt = *t;
+ lt = &__lt;
+ }
+
do {
if (o->userspace_reap == 1
&& actual_min == 0
ld->aio_events + events);
} else {
r = io_getevents(ld->aio_ctx, actual_min,
- max, ld->aio_events + events, t);
+ max, ld->aio_events + events, lt);
}
- if (r >= 0)
+ if (r > 0)
events += r;
- else if (r == -EAGAIN)
+ else if ((min && r == 0) || r == -EAGAIN) {
+ fio_libaio_commit(td);
usleep(100);
+ } else if (r != -EINTR)
+ break;
} while (events < min);
return r < 0 ? r : events;
fio_ro_check(td, io_u);
- if (ld->iocbs_nr == (int) td->o.iodepth)
+ if (ld->queued == td->o.iodepth)
return FIO_Q_BUSY;
/*
* have pending io, to let fio complete those first.
*/
if (ddir_sync(io_u->ddir)) {
- if (ld->iocbs_nr)
+ if (ld->queued)
return FIO_Q_BUSY;
do_io_u_sync(td, io_u);
}
if (io_u->ddir == DDIR_TRIM) {
- if (ld->iocbs_nr)
+ if (ld->queued)
return FIO_Q_BUSY;
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 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);
io_u_mark_submit(td, ret);
- ld->iocbs_nr -= ret;
- io_us += ret;
- iocbs += 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 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);
static int fio_libaio_init(struct thread_data *td)
{
- struct libaio_data *ld = malloc(sizeof(*ld));
struct libaio_options *o = td->eo;
+ struct libaio_data *ld;
int err = 0;
- memset(ld, 0, sizeof(*ld));
+ ld = calloc(1, sizeof(*ld));
/*
* First try passing in 0 for queue depth, since we don't
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;
projects / fio.git / blobdiff