X-Git-Url: https://git.kernel.dk/?a=blobdiff_plain;f=engines%2Fio_uring.c;h=a5e77d8f52fcfe29e1985824f953b0b594e64f85;hb=01fe773df4bc4a35450ce3ef50c8075b3bf55cd0;hp=06355e9c18263b896f053a75b13d342d249ec3a1;hpb=650346e17d045366b817814dd3e10dc94d0d990f;p=fio.git diff --git a/engines/io_uring.c b/engines/io_uring.c index 06355e9c..a5e77d8f 100644 --- a/engines/io_uring.c +++ b/engines/io_uring.c @@ -21,7 +21,7 @@ #ifdef ARCH_HAVE_IOURING #include "../lib/types.h" -#include "../os/io_uring.h" +#include "../os/linux/io_uring.h" struct io_sq_ring { unsigned *head; @@ -48,7 +48,6 @@ 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; @@ -63,9 +62,6 @@ struct ioring_data { int cq_ring_off; unsigned iodepth; - uint64_t cachehit; - uint64_t cachemiss; - struct ioring_mmap mmap[3]; }; @@ -73,18 +69,17 @@ 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; } @@ -107,30 +102,21 @@ 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, }, @@ -143,7 +129,7 @@ static int io_uring_enter(struct ioring_data *ld, unsigned int to_submit, unsigned int min_complete, unsigned int flags) { return syscall(__NR_sys_io_uring_enter, ld->ring_fd, to_submit, - min_complete, flags); + min_complete, flags, NULL, 0); } static int fio_ioring_prep(struct thread_data *td, struct io_u *io_u) @@ -157,6 +143,7 @@ static int fio_ioring_prep(struct thread_data *td, struct io_u *io_u) 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 (o->fixedbufs) { @@ -164,20 +151,26 @@ static int fio_ioring_prep(struct thread_data *td, struct io_u *io_u) sqe->opcode = IORING_OP_READ_FIXED; else sqe->opcode = IORING_OP_WRITE_FIXED; - sqe->addr = io_u->xfer_buf; + 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 = &ld->iovecs[io_u->index]; + sqe->addr = (unsigned long) &ld->iovecs[io_u->index]; sqe->len = 1; } sqe->off = io_u->offset; - } else if (ddir_sync(io_u->ddir)) + } 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; } @@ -191,7 +184,7 @@ static struct io_u *fio_ioring_event(struct thread_data *td, int event) index = (event + ld->cq_ring_off) & ld->cq_ring_mask; cqe = &ld->cq_ring.cqes[index]; - io_u = ld->io_u_index[cqe->index]; + io_u = (struct io_u *) (uintptr_t) cqe->user_data; if (cqe->res != io_u->xfer_buflen) { if (cqe->res > io_u->xfer_buflen) @@ -201,13 +194,6 @@ static struct io_u *fio_ioring_event(struct thread_data *td, int event) } else io_u->error = 0; - if (io_u->ddir == DDIR_READ) { - if (cqe->flags & IOCQE_FLAG_CACHEHIT) - ld->cachehit++; - else - ld->cachemiss++; - } - return io_u; } @@ -247,10 +233,12 @@ static int fio_ioring_getevents(struct thread_data *td, unsigned int min, r = fio_ioring_cqring_reap(td, events, max); if (r) { events += r; + if (actual_min != 0) + actual_min -= r; continue; } - if (!o->sqthread_poll) { + if (!o->sqpoll_thread) { r = io_uring_enter(ld, 0, actual_min, IORING_ENTER_GETEVENTS); if (r < 0) { @@ -293,6 +281,8 @@ 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(); @@ -332,12 +322,18 @@ 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; } @@ -387,14 +383,10 @@ static void fio_ioring_cleanup(struct thread_data *td) struct ioring_data *ld = td->io_ops_data; if (ld) { - td->ts.cachehit += ld->cachehit; - td->ts.cachemiss += ld->cachemiss; - if (!(td->flags & TD_F_CHILD)) fio_ioring_unmap(ld); free(ld->io_u_index); - free(ld->io_us); free(ld->iovecs); free(ld); } @@ -445,7 +437,6 @@ static int fio_ioring_queue_init(struct thread_data *td) struct ioring_data *ld = td->io_ops_data; struct ioring_options *o = td->eo; int depth = td->o.iodepth; - struct iovec *vecs = NULL; struct io_uring_params p; int ret; @@ -453,14 +444,19 @@ 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 = { @@ -468,15 +464,15 @@ static int fio_ioring_queue_init(struct thread_data *td) .rlim_max = RLIM_INFINITY, }; - setrlimit(RLIMIT_MEMLOCK, &rlim); - vecs = ld->iovecs; - } + if (setrlimit(RLIMIT_MEMLOCK, &rlim) < 0) + return -1; - ret = syscall(__NR_sys_io_uring_setup, depth, vecs, &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); } @@ -520,7 +516,6 @@ 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;