X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=engines%2Flibaio.c;h=9cc910d73bc19fdbb690baa5295131fc0f7addbc;hp=bc9c42877fdaaadf98d1c5ab51520c4e85d3903b;hb=954cd73a9a93102c24afa869fbe67ac38af6e416;hpb=f423479dad8916c758792308863dcd833a1a36c8 diff --git a/engines/libaio.c b/engines/libaio.c index bc9c4287..9cc910d7 100644 --- a/engines/libaio.c +++ b/engines/libaio.c @@ -1,5 +1,7 @@ /* - * native linux aio io engine + * libaio engine + * + * IO engine using the Linux native aio interface. * */ #include @@ -7,13 +9,9 @@ #include #include #include +#include #include "../fio.h" -#include "../os.h" - -#ifdef FIO_HAVE_LIBAIO - -#define ev_to_iou(ev) (struct io_u *) ((unsigned long) (ev)->obj) struct libaio_data { io_context_t aio_ctx; @@ -23,6 +21,26 @@ struct libaio_data { int iocbs_nr; }; +struct libaio_options { + struct thread_data *td; + 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 int fio_libaio_prep(struct thread_data fio_unused *td, struct io_u *io_u) { struct fio_file *f = io_u->file; @@ -31,10 +49,8 @@ static int fio_libaio_prep(struct thread_data fio_unused *td, struct io_u *io_u) 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; } @@ -46,7 +62,7 @@ static struct io_u *fio_libaio_event(struct thread_data *td, int event) 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) @@ -59,33 +75,82 @@ static struct io_u *fio_libaio_event(struct thread_data *td, int event) return io_u; } -static int fio_libaio_getevents(struct thread_data *td, int min, int max, - struct timespec *t) +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) { struct libaio_data *ld = td->io_ops->data; - long r; + struct libaio_options *o = td->eo; + unsigned actual_min = td->o.iodepth_batch_complete == 0 ? 0 : min; + int r, events = 0; 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, t); + } + if (r >= 0) + events += r; + else if (r == -EAGAIN) usleep(100); - continue; - } else if (r == -EINTR) - continue; - else if (r != 0) - 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) { struct libaio_data *ld = td->io_ops->data; - if (ld->iocbs_nr == (int) td->iodepth) + fio_ro_check(td, io_u); + + if (ld->iocbs_nr == (int) td->o.iodepth) return FIO_Q_BUSY; /* @@ -94,12 +159,19 @@ static int fio_libaio_queue(struct thread_data *td, struct io_u *io_u) * 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 (ddir_sync(io_u->ddir)) { + if (ld->iocbs_nr) + return FIO_Q_BUSY; + + do_io_u_sync(td, io_u); + return FIO_Q_COMPLETED; + } + + if (io_u->ddir == DDIR_TRIM) { if (ld->iocbs_nr) 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; } @@ -115,6 +187,9 @@ static void fio_libaio_queued(struct thread_data *td, struct io_u **io_us, struct timeval now; unsigned int i; + if (!fio_fill_issue_time(td)) + return; + fio_gettime(&now, NULL); for (i = 0; i < nr; i++) { @@ -130,36 +205,29 @@ static int fio_libaio_commit(struct thread_data *td) struct libaio_data *ld = td->io_ops->data; struct iocb **iocbs; struct io_u **io_us; - int ret, iocbs_nr; + int ret; if (!ld->iocbs_nr) return 0; - iocbs_nr = ld->iocbs_nr; io_us = ld->io_us; iocbs = ld->iocbs; do { - ret = io_submit(ld->aio_ctx, iocbs_nr, iocbs); - if (ret == iocbs_nr) { - fio_libaio_queued(td, io_us, ret); - ret = 0; - break; - } else if (ret > 0) { + ret = io_submit(ld->aio_ctx, ld->iocbs_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; - iocbs_nr -= ret; - continue; - } else if (ret == -EAGAIN || !ret) - usleep(100); - else if (ret == -EINTR) + ret = 0; + } else if (!ret || ret == -EAGAIN || ret == -EINTR) { + if (!ret) + io_u_mark_submit(td, ret); continue; - else + } else break; - } while (1); - - if (!ret) - ld->iocbs_nr = 0; + } while (ld->iocbs_nr); return ret; } @@ -181,27 +249,39 @@ static void fio_libaio_cleanup(struct thread_data *td) 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)); + struct libaio_options *o = td->eo; + int err = 0; memset(ld, 0, sizeof(*ld)); - if (io_queue_init(td->iodepth, &ld->aio_ctx)) { - td_verror(td, errno, "io_queue_init"); + + /* + * 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"); free(ld); return 1; } - ld->aio_events = malloc(td->iodepth * sizeof(struct io_event)); - memset(ld->aio_events, 0, td->iodepth * sizeof(struct io_event)); - ld->iocbs = malloc(td->iodepth * sizeof(struct iocb *)); + 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->iodepth * sizeof(struct io_u *)); - memset(ld->io_us, 0, td->iodepth * sizeof(struct io_u *)); + 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; td->io_ops->data = ld; @@ -209,39 +289,23 @@ static int fio_libaio_init(struct thread_data *td) } 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, + .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), }; -#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, -}; - -#endif - static void fio_init fio_libaio_register(void) { register_ioengine(&ioengine);