X-Git-Url: https://git.kernel.dk/?a=blobdiff_plain;f=engines%2Fio_uring.c;h=a5e77d8f52fcfe29e1985824f953b0b594e64f85;hb=01fe773df4bc4a35450ce3ef50c8075b3bf55cd0;hp=ebca08c874d446c11ab36c3dc8a94831742269e3;hpb=bffad86f1e15793b6acda42ab9ca73e148f7032e;p=fio.git

diff --git a/engines/io_uring.c b/engines/io_uring.c
index ebca08c8..a5e77d8f 100644
--- a/engines/io_uring.c
+++ b/engines/io_uring.c
@@ -20,29 +20,24 @@
 
 #ifdef ARCH_HAVE_IOURING
 
-typedef uint64_t u64;
-typedef uint32_t u32;
-typedef int32_t s32;
-typedef uint16_t u16;
-typedef uint8_t u8;
-
-#include "../os/io_uring.h"
+#include "../lib/types.h"
+#include "../os/linux/io_uring.h"
 
 struct io_sq_ring {
-	u32 *head;
-	u32 *tail;
-	u32 *ring_mask;
-	u32 *ring_entries;
-	u32 *flags;
-	u32 *array;
+	unsigned *head;
+	unsigned *tail;
+	unsigned *ring_mask;
+	unsigned *ring_entries;
+	unsigned *flags;
+	unsigned *array;
 };
 
 struct io_cq_ring {
-	u32 *head;
-	u32 *tail;
-	u32 *ring_mask;
-	u32 *ring_entries;
-	struct io_uring_event *events;
+	unsigned *head;
+	unsigned *tail;
+	unsigned *ring_mask;
+	unsigned *ring_entries;
+	struct io_uring_cqe *cqes;
 };
 
 struct ioring_mmap {
@@ -53,11 +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;
 
@@ -68,9 +62,6 @@ struct ioring_data {
 	int cq_ring_off;
 	unsigned iodepth;
 
-	uint64_t cachehit;
-	uint64_t cachemiss;
-
 	struct ioring_mmap mmap[3];
 };
 
@@ -78,18 +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;
 }
 
@@ -112,30 +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,
 	},
@@ -148,61 +129,71 @@ 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)
 {
 	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)
-			ld->cachehit++;
-		else
-			ld->cachemiss++;
-	}
-
 	return io_u;
 }
 
@@ -211,7 +202,7 @@ static int fio_ioring_cqring_reap(struct thread_data *td, unsigned int events,
 {
 	struct ioring_data *ld = td->io_ops_data;
 	struct io_cq_ring *ring = &ld->cq_ring;
-	u32 head, reaped = 0;
+	unsigned head, reaped = 0;
 
 	head = *ring->head;
 	do {
@@ -242,10 +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) {
@@ -288,6 +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();
@@ -327,12 +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;
 	}
@@ -382,14 +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);
 	}
@@ -401,7 +398,7 @@ static int fio_ioring_mmap(struct ioring_data *ld, struct io_uring_params *p)
 	struct io_cq_ring *cring = &ld->cq_ring;
 	void *ptr;
 
-	ld->mmap[0].len = p->sq_off.array + p->sq_entries * sizeof(u32);
+	ld->mmap[0].len = p->sq_off.array + p->sq_entries * sizeof(__u32);
 	ptr = mmap(0, ld->mmap[0].len, PROT_READ | PROT_WRITE,
 			MAP_SHARED | MAP_POPULATE, ld->ring_fd,
 			IORING_OFF_SQ_RING);
@@ -414,14 +411,14 @@ 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);
@@ -430,7 +427,7 @@ 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;
 }
@@ -447,14 +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 = {
@@ -462,35 +464,30 @@ 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);
@@ -519,8 +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;