X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=engines%2Fio_uring.c;h=5279b1d0ffb7d38ba9ceb144b99ba26dd690574d;hp=15a4d4753639ce3bf6ef73b1d752a491fc520129;hb=b532dd6d476679b08e4a56a60e8a7dd958779df9;hpb=e2239016d4839aeb54b5da61f94baf0d518233fb diff --git a/engines/io_uring.c b/engines/io_uring.c index 15a4d475..5279b1d0 100644 --- a/engines/io_uring.c +++ b/engines/io_uring.c @@ -20,7 +20,8 @@ #ifdef ARCH_HAVE_IOURING -#include "../os/io_uring.h" +#include "../lib/types.h" +#include "../os/linux/io_uring.h" struct io_sq_ring { unsigned *head; @@ -36,7 +37,7 @@ struct io_cq_ring { unsigned *tail; unsigned *ring_mask; unsigned *ring_entries; - struct io_uring_event *events; + struct io_uring_cqe *cqes; }; struct ioring_mmap { @@ -47,11 +48,10 @@ struct ioring_mmap { struct ioring_data { int ring_fd; - struct io_u **io_us; struct io_u **io_u_index; struct io_sq_ring sq_ring; - struct io_uring_iocb *iocbs; + struct io_uring_sqe *sqes; struct iovec *iovecs; unsigned sq_ring_mask; @@ -72,18 +72,17 @@ struct ioring_options { void *pad; unsigned int hipri; unsigned int fixedbufs; - unsigned int sqthread; - unsigned int sqthread_set; - unsigned int sqthread_poll; - unsigned int sqwq; + unsigned int sqpoll_thread; + unsigned int sqpoll_set; + unsigned int sqpoll_cpu; }; -static int fio_ioring_sqthread_cb(void *data, unsigned long long *val) +static int fio_ioring_sqpoll_cb(void *data, unsigned long long *val) { struct ioring_options *o = data; - o->sqthread = *val; - o->sqthread_set = 1; + o->sqpoll_cpu = *val; + o->sqpoll_set = 1; return 0; } @@ -106,30 +105,21 @@ static struct fio_option options[] = { .category = FIO_OPT_C_ENGINE, .group = FIO_OPT_G_LIBAIO, }, - { - .name = "sqthread", - .lname = "Use kernel SQ thread on this CPU", - .type = FIO_OPT_INT, - .cb = fio_ioring_sqthread_cb, - .help = "Offload submission to kernel thread", - .category = FIO_OPT_C_ENGINE, - .group = FIO_OPT_G_LIBAIO, - }, { .name = "sqthread_poll", - .lname = "Kernel SQ thread should poll", - .type = FIO_OPT_STR_SET, - .off1 = offsetof(struct ioring_options, sqthread_poll), - .help = "Used with sqthread, enables kernel side polling", + .lname = "Kernel SQ thread polling", + .type = FIO_OPT_INT, + .off1 = offsetof(struct ioring_options, sqpoll_thread), + .help = "Offload submission/completion to kernel thread", .category = FIO_OPT_C_ENGINE, .group = FIO_OPT_G_LIBAIO, }, { - .name = "sqwq", - .lname = "Offload submission to kernel workqueue", - .type = FIO_OPT_STR_SET, - .off1 = offsetof(struct ioring_options, sqwq), - .help = "Offload submission to kernel workqueue", + .name = "sqthread_poll_cpu", + .lname = "SQ Thread Poll CPU", + .type = FIO_OPT_INT, + .cb = fio_ioring_sqpoll_cb, + .help = "What CPU to run SQ thread polling on", .category = FIO_OPT_C_ENGINE, .group = FIO_OPT_G_LIBAIO, }, @@ -148,50 +138,67 @@ static int io_uring_enter(struct ioring_data *ld, unsigned int to_submit, static int fio_ioring_prep(struct thread_data *td, struct io_u *io_u) { struct ioring_data *ld = td->io_ops_data; + struct ioring_options *o = td->eo; struct fio_file *f = io_u->file; - struct io_uring_iocb *iocb; + struct io_uring_sqe *sqe; - iocb = &ld->iocbs[io_u->index]; - iocb->fd = f->fd; - iocb->flags = 0; - iocb->ioprio = 0; + sqe = &ld->sqes[io_u->index]; + sqe->fd = f->fd; + sqe->flags = 0; + sqe->ioprio = 0; + sqe->buf_index = 0; if (io_u->ddir == DDIR_READ || io_u->ddir == DDIR_WRITE) { - if (io_u->ddir == DDIR_READ) - iocb->opcode = IORING_OP_READ; - else - iocb->opcode = IORING_OP_WRITE; - iocb->off = io_u->offset; - iocb->addr = io_u->xfer_buf; - iocb->len = io_u->xfer_buflen; - } else if (ddir_sync(io_u->ddir)) - iocb->opcode = IORING_OP_FSYNC; + if (o->fixedbufs) { + if (io_u->ddir == DDIR_READ) + sqe->opcode = IORING_OP_READ_FIXED; + else + sqe->opcode = IORING_OP_WRITE_FIXED; + sqe->addr = (unsigned long) io_u->xfer_buf; + sqe->len = io_u->xfer_buflen; + sqe->buf_index = io_u->index; + } else { + if (io_u->ddir == DDIR_READ) + sqe->opcode = IORING_OP_READV; + else + sqe->opcode = IORING_OP_WRITEV; + sqe->addr = (unsigned long) &ld->iovecs[io_u->index]; + sqe->len = 1; + } + sqe->off = io_u->offset; + } else if (ddir_sync(io_u->ddir)) { + sqe->fsync_flags = 0; + if (io_u->ddir == DDIR_DATASYNC) + sqe->fsync_flags |= IORING_FSYNC_DATASYNC; + sqe->opcode = IORING_OP_FSYNC; + } + sqe->user_data = (unsigned long) io_u; return 0; } static struct io_u *fio_ioring_event(struct thread_data *td, int event) { struct ioring_data *ld = td->io_ops_data; - struct io_uring_event *ev; + struct io_uring_cqe *cqe; struct io_u *io_u; unsigned index; index = (event + ld->cq_ring_off) & ld->cq_ring_mask; - ev = &ld->cq_ring.events[index]; - io_u = ld->io_u_index[ev->index]; + cqe = &ld->cq_ring.cqes[index]; + io_u = (struct io_u *) (uintptr_t) cqe->user_data; - if (ev->res != io_u->xfer_buflen) { - if (ev->res > io_u->xfer_buflen) - io_u->error = -ev->res; + if (cqe->res != io_u->xfer_buflen) { + if (cqe->res > io_u->xfer_buflen) + io_u->error = -cqe->res; else - io_u->resid = io_u->xfer_buflen - ev->res; + io_u->resid = io_u->xfer_buflen - cqe->res; } else io_u->error = 0; if (io_u->ddir == DDIR_READ) { - if (ev->flags & IOEV_FLAG_CACHEHIT) + if (cqe->flags & IOCQE_FLAG_CACHEHIT) ld->cachehit++; else ld->cachemiss++; @@ -239,7 +246,7 @@ static int fio_ioring_getevents(struct thread_data *td, unsigned int min, continue; } - if (!o->sqthread_poll) { + if (!o->sqpoll_thread) { r = io_uring_enter(ld, 0, actual_min, IORING_ENTER_GETEVENTS); if (r < 0) { @@ -282,6 +289,8 @@ static enum fio_q_status fio_ioring_queue(struct thread_data *td, if (next_tail == *ring->head) return FIO_Q_BUSY; + /* ensure sqe stores are ordered with tail update */ + write_barrier(); ring->array[tail & ld->sq_ring_mask] = io_u->index; *ring->tail = next_tail; write_barrier(); @@ -321,12 +330,18 @@ static int fio_ioring_commit(struct thread_data *td) if (!ld->queued) return 0; - /* Nothing to do */ - if (o->sqthread_poll) { + /* + * Kernel side does submission. just need to check if the ring is + * flagged as needing a kick, if so, call io_uring_enter(). This + * only happens if we've been idle too long. + */ + if (o->sqpoll_thread) { struct io_sq_ring *ring = &ld->sq_ring; + read_barrier(); if (*ring->flags & IORING_SQ_NEED_WAKEUP) - io_uring_enter(ld, ld->queued, 0, 0); + io_uring_enter(ld, ld->queued, 0, + IORING_ENTER_SQ_WAKEUP); ld->queued = 0; return 0; } @@ -383,7 +398,6 @@ static void fio_ioring_cleanup(struct thread_data *td) fio_ioring_unmap(ld); free(ld->io_u_index); - free(ld->io_us); free(ld->iovecs); free(ld); } @@ -408,14 +422,14 @@ static int fio_ioring_mmap(struct ioring_data *ld, struct io_uring_params *p) sring->array = ptr + p->sq_off.array; ld->sq_ring_mask = *sring->ring_mask; - ld->mmap[1].len = p->sq_entries * sizeof(struct io_uring_iocb); - ld->iocbs = mmap(0, ld->mmap[1].len, PROT_READ | PROT_WRITE, + ld->mmap[1].len = p->sq_entries * sizeof(struct io_uring_sqe); + ld->sqes = mmap(0, ld->mmap[1].len, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, ld->ring_fd, - IORING_OFF_IOCB); - ld->mmap[1].ptr = ld->iocbs; + IORING_OFF_SQES); + ld->mmap[1].ptr = ld->sqes; - ld->mmap[2].len = p->cq_off.events + - p->cq_entries * sizeof(struct io_uring_event); + ld->mmap[2].len = p->cq_off.cqes + + p->cq_entries * sizeof(struct io_uring_cqe); ptr = mmap(0, ld->mmap[2].len, PROT_READ | PROT_WRITE, MAP_SHARED | MAP_POPULATE, ld->ring_fd, IORING_OFF_CQ_RING); @@ -424,7 +438,7 @@ static int fio_ioring_mmap(struct ioring_data *ld, struct io_uring_params *p) cring->tail = ptr + p->cq_off.tail; cring->ring_mask = ptr + p->cq_off.ring_mask; cring->ring_entries = ptr + p->cq_off.ring_entries; - cring->events = ptr + p->cq_off.events; + cring->cqes = ptr + p->cq_off.cqes; ld->cq_ring_mask = *cring->ring_mask; return 0; } @@ -441,14 +455,19 @@ static int fio_ioring_queue_init(struct thread_data *td) if (o->hipri) p.flags |= IORING_SETUP_IOPOLL; - if (o->sqthread_set) { - p.sq_thread_cpu = o->sqthread; - p.flags |= IORING_SETUP_SQTHREAD; - if (o->sqthread_poll) - p.flags |= IORING_SETUP_SQPOLL; + if (o->sqpoll_thread) { + p.flags |= IORING_SETUP_SQPOLL; + if (o->sqpoll_set) { + p.flags |= IORING_SETUP_SQ_AFF; + p.sq_thread_cpu = o->sqpoll_cpu; + } } - if (o->sqwq) - p.flags |= IORING_SETUP_SQWQ; + + ret = syscall(__NR_sys_io_uring_setup, depth, &p); + if (ret < 0) + return ret; + + ld->ring_fd = ret; if (o->fixedbufs) { struct rlimit rlim = { @@ -456,35 +475,30 @@ static int fio_ioring_queue_init(struct thread_data *td) .rlim_max = RLIM_INFINITY, }; - setrlimit(RLIMIT_MEMLOCK, &rlim); - p.flags |= IORING_SETUP_FIXEDBUFS; - } + if (setrlimit(RLIMIT_MEMLOCK, &rlim) < 0) + return -1; - ret = syscall(__NR_sys_io_uring_setup, depth, ld->iovecs, &p); - if (ret < 0) - return ret; + ret = syscall(__NR_sys_io_uring_register, ld->ring_fd, + IORING_REGISTER_BUFFERS, ld->iovecs, depth); + if (ret < 0) + return ret; + } - ld->ring_fd = ret; return fio_ioring_mmap(ld, &p); } static int fio_ioring_post_init(struct thread_data *td) { struct ioring_data *ld = td->io_ops_data; - struct ioring_options *o = td->eo; struct io_u *io_u; - int err; + int err, i; - if (o->fixedbufs) { - int i; - - for (i = 0; i < td->o.iodepth; i++) { - struct iovec *iov = &ld->iovecs[i]; + for (i = 0; i < td->o.iodepth; i++) { + struct iovec *iov = &ld->iovecs[i]; - io_u = ld->io_u_index[i]; - iov->iov_base = io_u->buf; - iov->iov_len = td_max_bs(td); - } + io_u = ld->io_u_index[i]; + iov->iov_base = io_u->buf; + iov->iov_len = td_max_bs(td); } err = fio_ioring_queue_init(td); @@ -513,8 +527,6 @@ static int fio_ioring_init(struct thread_data *td) /* io_u index */ ld->io_u_index = calloc(td->o.iodepth, sizeof(struct io_u *)); - ld->io_us = calloc(td->o.iodepth, sizeof(struct io_u *)); - ld->iovecs = calloc(td->o.iodepth, sizeof(struct iovec)); td->io_ops_data = ld;