X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=engines%2Flibaio.c;h=cd5b89f98a36a500f0854a02e68dc228aa7ecbf1;hp=2b8c6dac533c8bf2c9f1cdd0cc02e7ba7a599b72;hb=ad4e92984ec2d92b18abede6cfbf8ef7f353d77a;hpb=b0df7176176935b4d39fd196ad35f046bbe10333 diff --git a/engines/libaio.c b/engines/libaio.c index 2b8c6dac..cd5b89f9 100644 --- a/engines/libaio.c +++ b/engines/libaio.c @@ -4,29 +4,47 @@ * IO engine using the Linux native aio interface. * */ -#include #include #include #include -#include +#include +#include +#include #include "../fio.h" +#include "../lib/pow2.h" +#include "../optgroup.h" +#include "../lib/memalign.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; + + struct io_u **io_u_index; + + /* + * 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; + unsigned int hipri; }; static struct fio_option options[] = { @@ -36,35 +54,46 @@ static struct fio_option options[] = { .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; + struct iocb *iocb = &io_u->iocb; - if (io_u->ddir == DDIR_READ) - 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 (ddir_sync(io_u->ddir)) - io_prep_fsync(&io_u->iocb, f->fd); + if (io_u->ddir == DDIR_READ) { + io_prep_pread(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset); + } else if (io_u->ddir == DDIR_WRITE) { + io_prep_pwrite(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset); + } else if (ddir_sync(io_u->ddir)) + io_prep_fsync(iocb, f->fd); return 0; } static struct io_u *fio_libaio_event(struct thread_data *td, int event) { - struct libaio_data *ld = td->io_ops->data; + struct libaio_data *ld = td->io_ops_data; struct io_event *ev; 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) @@ -119,13 +148,19 @@ static int user_io_getevents(io_context_t aio_ctx, unsigned int max, } 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_data *ld = td->io_ops_data; struct libaio_options *o = td->eo; - unsigned actual_min = td->o.iodepth_batch_complete == 0 ? 0 : min; + 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 { if (o->userspace_reap == 1 && actual_min == 0 @@ -135,24 +170,29 @@ static int fio_libaio_getevents(struct thread_data *td, unsigned int min, 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) - usleep(100); + else if ((min && r == 0) || r == -EAGAIN) { + fio_libaio_commit(td); + if (actual_min) + usleep(10); + } else if (r != -EINTR) + break; } while (events < min); return r < 0 ? r : events; } -static int fio_libaio_queue(struct thread_data *td, struct io_u *io_u) +static enum fio_q_status fio_libaio_queue(struct thread_data *td, + struct io_u *io_u) { - struct libaio_data *ld = td->io_ops->data; + struct libaio_data *ld = td->io_ops_data; fio_ro_check(td, io_u); - if (ld->iocbs_nr == (int) td->o.iodepth) + if (ld->queued == td->o.iodepth) return FIO_Q_BUSY; /* @@ -162,7 +202,7 @@ static int fio_libaio_queue(struct thread_data *td, struct io_u *io_u) * 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); @@ -170,23 +210,26 @@ static int fio_libaio_queue(struct thread_data *td, struct io_u *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); + io_u_mark_submit(td, 1); + io_u_mark_complete(td, 1); 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; } static void fio_libaio_queued(struct thread_data *td, struct io_u **io_us, unsigned int nr) { - struct timeval now; + struct timespec now; unsigned int i; if (!fio_fill_issue_time(td)) @@ -204,49 +247,94 @@ static void fio_libaio_queued(struct thread_data *td, struct io_u **io_us, static int fio_libaio_commit(struct thread_data *td) { - struct libaio_data *ld = td->io_ops->data; + struct libaio_data *ld = td->io_ops_data; struct iocb **iocbs; struct io_u **io_us; - int ret; + struct timespec ts; + 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(&ts, NULL); + wait_start = 1; + } else if (mtime_since_now(&ts) > 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; } static int fio_libaio_cancel(struct thread_data *td, struct io_u *io_u) { - struct libaio_data *ld = td->io_ops->data; + struct libaio_data *ld = td->io_ops_data; return io_cancel(ld->aio_ctx, &io_u->iocb, ld->aio_events); } static void fio_libaio_cleanup(struct thread_data *td) { - struct libaio_data *ld = td->io_ops->data; + 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); @@ -254,37 +342,42 @@ static void fio_libaio_cleanup(struct thread_data *td) } } -static int fio_libaio_init(struct thread_data *td) +static int fio_libaio_post_init(struct thread_data *td) { - struct libaio_data *ld = malloc(sizeof(*ld)); + struct libaio_data *ld = td->io_ops_data; int err; - memset(ld, 0, sizeof(*ld)); - 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->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; + return 0; +} - td->io_ops->data = ld; +static int fio_libaio_init(struct thread_data *td) +{ + struct libaio_data *ld; + + ld = calloc(1, sizeof(*ld)); + + 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, + .flags = FIO_ASYNCIO_SYNC_TRIM, .init = fio_libaio_init, + .post_init = fio_libaio_post_init, .prep = fio_libaio_prep, .queue = fio_libaio_queue, .commit = fio_libaio_commit, @@ -299,27 +392,6 @@ static struct ioengine_ops ioengine = { .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) -{ - log_err("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);