X-Git-Url: https://git.kernel.dk/?a=blobdiff_plain;f=engines%2Fio_uring.c;h=df2d6c4c7e4a1e3903b4446974dd70abd8700025;hb=1437d6357429f4294c5b64eaa1380fc594079a54;hp=0ccd23184fab3b17a9c68c93a9c30375b5673986;hpb=f0ed01ed095cf1ca7c1945a5a0267e8f73b7b4a9;p=fio.git

diff --git a/engines/io_uring.c b/engines/io_uring.c
index 0ccd2318..df2d6c4c 100644
--- a/engines/io_uring.c
+++ b/engines/io_uring.c
@@ -23,6 +23,7 @@
 
 #include "../lib/types.h"
 #include "../os/linux/io_uring.h"
+#include "cmdprio.h"
 
 struct io_sq_ring {
 	unsigned *head;
@@ -64,16 +65,17 @@ struct ioring_data {
 	int queued;
 	int cq_ring_off;
 	unsigned iodepth;
-	bool ioprio_class_set;
-	bool ioprio_set;
+	int prepped;
 
 	struct ioring_mmap mmap[3];
+
+	bool use_cmdprio;
 };
 
 struct ioring_options {
-	void *pad;
+	struct thread_data *td;
 	unsigned int hipri;
-	unsigned int cmdprio_percentage;
+	struct cmdprio cmdprio;
 	unsigned int fixedbufs;
 	unsigned int registerfiles;
 	unsigned int sqpoll_thread;
@@ -82,6 +84,7 @@ struct ioring_options {
 	unsigned int nonvectored;
 	unsigned int uncached;
 	unsigned int nowait;
+	unsigned int force_async;
 };
 
 static const int ddir_to_op[2][2] = {
@@ -103,6 +106,15 @@ static int fio_ioring_sqpoll_cb(void *data, unsigned long long *val)
 	return 0;
 }
 
+static int str_cmdprio_bssplit_cb(void *data, const char *input)
+{
+	struct ioring_options *o = data;
+	struct thread_data *td = o->td;
+	struct cmdprio *cmdprio = &o->cmdprio;
+
+	return fio_cmdprio_bssplit_parse(td, input, cmdprio);
+}
+
 static struct fio_option options[] = {
 	{
 		.name	= "hipri",
@@ -118,13 +130,56 @@ static struct fio_option options[] = {
 		.name	= "cmdprio_percentage",
 		.lname	= "high priority percentage",
 		.type	= FIO_OPT_INT,
-		.off1	= offsetof(struct ioring_options, cmdprio_percentage),
-		.minval	= 1,
+		.off1	= offsetof(struct ioring_options,
+				   cmdprio.percentage[DDIR_READ]),
+		.off2	= offsetof(struct ioring_options,
+				   cmdprio.percentage[DDIR_WRITE]),
+		.minval	= 0,
 		.maxval	= 100,
 		.help	= "Send high priority I/O this percentage of the time",
 		.category = FIO_OPT_C_ENGINE,
 		.group	= FIO_OPT_G_IOURING,
 	},
+	{
+		.name	= "cmdprio_class",
+		.lname	= "Asynchronous I/O priority class",
+		.type	= FIO_OPT_INT,
+		.off1	= offsetof(struct ioring_options,
+				   cmdprio.class[DDIR_READ]),
+		.off2	= offsetof(struct ioring_options,
+				   cmdprio.class[DDIR_WRITE]),
+		.help	= "Set asynchronous IO priority class",
+		.minval	= IOPRIO_MIN_PRIO_CLASS + 1,
+		.maxval	= IOPRIO_MAX_PRIO_CLASS,
+		.interval = 1,
+		.category = FIO_OPT_C_ENGINE,
+		.group	= FIO_OPT_G_IOURING,
+	},
+	{
+		.name	= "cmdprio",
+		.lname	= "Asynchronous I/O priority level",
+		.type	= FIO_OPT_INT,
+		.off1	= offsetof(struct ioring_options,
+				   cmdprio.level[DDIR_READ]),
+		.off2	= offsetof(struct ioring_options,
+				   cmdprio.level[DDIR_WRITE]),
+		.help	= "Set asynchronous IO priority level",
+		.minval	= IOPRIO_MIN_PRIO,
+		.maxval	= IOPRIO_MAX_PRIO,
+		.interval = 1,
+		.category = FIO_OPT_C_ENGINE,
+		.group	= FIO_OPT_G_IOURING,
+	},
+	{
+		.name   = "cmdprio_bssplit",
+		.lname  = "Priority percentage block size split",
+		.type   = FIO_OPT_STR_ULL,
+		.cb     = str_cmdprio_bssplit_cb,
+		.off1   = offsetof(struct ioring_options, cmdprio.bssplit),
+		.help   = "Set priority percentages for different block sizes",
+		.category = FIO_OPT_C_ENGINE,
+		.group	= FIO_OPT_G_IOURING,
+	},
 #else
 	{
 		.name	= "cmdprio_percentage",
@@ -132,6 +187,24 @@ static struct fio_option options[] = {
 		.type	= FIO_OPT_UNSUPPORTED,
 		.help	= "Your platform does not support I/O priority classes",
 	},
+	{
+		.name	= "cmdprio_class",
+		.lname	= "Asynchronous I/O priority class",
+		.type	= FIO_OPT_UNSUPPORTED,
+		.help	= "Your platform does not support I/O priority classes",
+	},
+	{
+		.name	= "cmdprio",
+		.lname	= "Asynchronous I/O priority level",
+		.type	= FIO_OPT_UNSUPPORTED,
+		.help	= "Your platform does not support I/O priority classes",
+	},
+	{
+		.name   = "cmdprio_bssplit",
+		.lname  = "Priority percentage block size split",
+		.type	= FIO_OPT_UNSUPPORTED,
+		.help	= "Your platform does not support I/O priority classes",
+	},
 #endif
 	{
 		.name	= "fixedbufs",
@@ -174,6 +247,7 @@ static struct fio_option options[] = {
 		.lname	= "Non-vectored",
 		.type	= FIO_OPT_INT,
 		.off1	= offsetof(struct ioring_options, nonvectored),
+		.def	= "-1",
 		.help	= "Use non-vectored read/write commands",
 		.category = FIO_OPT_C_ENGINE,
 		.group	= FIO_OPT_G_IOURING,
@@ -196,6 +270,15 @@ static struct fio_option options[] = {
 		.category = FIO_OPT_C_ENGINE,
 		.group	= FIO_OPT_G_IOURING,
 	},
+	{
+		.name	= "force_async",
+		.lname	= "Force async",
+		.type	= FIO_OPT_INT,
+		.off1	= offsetof(struct ioring_options, force_async),
+		.help	= "Set IOSQE_ASYNC every N requests",
+		.category = FIO_OPT_C_ENGINE,
+		.group	= FIO_OPT_G_IOURING,
+	},
 	{
 		.name	= NULL,
 	},
@@ -217,14 +300,12 @@ static int fio_ioring_prep(struct thread_data *td, struct io_u *io_u)
 
 	sqe = &ld->sqes[io_u->index];
 
-	/* zero out fields not used in this submission */
-	memset(sqe, 0, sizeof(*sqe));
-
 	if (o->registerfiles) {
 		sqe->fd = f->engine_pos;
 		sqe->flags = IOSQE_FIXED_FILE;
 	} else {
 		sqe->fd = f->fd;
+		sqe->flags = 0;
 	}
 
 	if (io_u->ddir == DDIR_READ || io_u->ddir == DDIR_WRITE) {
@@ -234,38 +315,52 @@ static int fio_ioring_prep(struct thread_data *td, struct io_u *io_u)
 			sqe->len = io_u->xfer_buflen;
 			sqe->buf_index = io_u->index;
 		} else {
+			struct iovec *iov = &ld->iovecs[io_u->index];
+
+			/*
+			 * Update based on actual io_u, requeue could have
+			 * adjusted these
+			 */
+			iov->iov_base = io_u->xfer_buf;
+			iov->iov_len = io_u->xfer_buflen;
+
 			sqe->opcode = ddir_to_op[io_u->ddir][!!o->nonvectored];
 			if (o->nonvectored) {
-				sqe->addr = (unsigned long)
-						ld->iovecs[io_u->index].iov_base;
-				sqe->len = ld->iovecs[io_u->index].iov_len;
+				sqe->addr = (unsigned long) iov->iov_base;
+				sqe->len = iov->iov_len;
 			} else {
-				sqe->addr = (unsigned long) &ld->iovecs[io_u->index];
+				sqe->addr = (unsigned long) iov;
 				sqe->len = 1;
 			}
 		}
+		sqe->rw_flags = 0;
 		if (!td->o.odirect && o->uncached)
-			sqe->rw_flags = RWF_UNCACHED;
+			sqe->rw_flags |= RWF_UNCACHED;
 		if (o->nowait)
 			sqe->rw_flags |= RWF_NOWAIT;
-		if (ld->ioprio_class_set)
-			sqe->ioprio = td->o.ioprio_class << 13;
-		if (ld->ioprio_set)
-			sqe->ioprio |= td->o.ioprio;
 		sqe->off = io_u->offset;
 	} else if (ddir_sync(io_u->ddir)) {
+		sqe->ioprio = 0;
 		if (io_u->ddir == DDIR_SYNC_FILE_RANGE) {
 			sqe->off = f->first_write;
 			sqe->len = f->last_write - f->first_write;
 			sqe->sync_range_flags = td->o.sync_file_range;
 			sqe->opcode = IORING_OP_SYNC_FILE_RANGE;
 		} else {
+			sqe->off = 0;
+			sqe->addr = 0;
+			sqe->len = 0;
 			if (io_u->ddir == DDIR_DATASYNC)
 				sqe->fsync_flags |= IORING_FSYNC_DATASYNC;
 			sqe->opcode = IORING_OP_FSYNC;
 		}
 	}
 
+	if (o->force_async && ++ld->prepped == o->force_async) {
+		ld->prepped = 0;
+		sqe->flags |= IOSQE_ASYNC;
+	}
+
 	sqe->user_data = (unsigned long) io_u;
 	return 0;
 }
@@ -353,11 +448,35 @@ static void fio_ioring_prio_prep(struct thread_data *td, struct io_u *io_u)
 {
 	struct ioring_options *o = td->eo;
 	struct ioring_data *ld = td->io_ops_data;
-	if (rand_between(&td->prio_state, 0, 99) < o->cmdprio_percentage) {
-		ld->sqes[io_u->index].ioprio = IOPRIO_CLASS_RT << IOPRIO_CLASS_SHIFT;
-		io_u->flags |= IO_U_F_PRIORITY;
+	struct io_uring_sqe *sqe = &ld->sqes[io_u->index];
+	struct cmdprio *cmdprio = &o->cmdprio;
+	enum fio_ddir ddir = io_u->ddir;
+	unsigned int p = fio_cmdprio_percentage(cmdprio, io_u);
+	unsigned int cmdprio_value =
+		ioprio_value(cmdprio->class[ddir], cmdprio->level[ddir]);
+
+	if (p && rand_between(&td->prio_state, 0, 99) < p) {
+		sqe->ioprio = cmdprio_value;
+		if (!td->ioprio || cmdprio_value < td->ioprio) {
+			/*
+			 * The async IO priority is higher (has a lower value)
+			 * than the priority set by "prio" and "prioclass"
+			 * options.
+			 */
+			io_u->flags |= IO_U_F_PRIORITY;
+		}
+	} else {
+		sqe->ioprio = td->ioprio;
+		if (cmdprio_value && td->ioprio && td->ioprio < cmdprio_value) {
+			/*
+			 * The IO will be executed with the priority set by
+			 * "prio" and "prioclass" options, and this priority
+			 * is higher (has a lower value) than the async IO
+			 * priority.
+			 */
+			io_u->flags |= IO_U_F_PRIORITY;
+		}
 	}
-	return;
 }
 
 static enum fio_q_status fio_ioring_queue(struct thread_data *td,
@@ -365,7 +484,6 @@ static enum fio_q_status fio_ioring_queue(struct thread_data *td,
 {
 	struct ioring_data *ld = td->io_ops_data;
 	struct io_sq_ring *ring = &ld->sq_ring;
-	struct ioring_options *o = td->eo;
 	unsigned tail, next_tail;
 
 	fio_ro_check(td, io_u);
@@ -388,7 +506,7 @@ static enum fio_q_status fio_ioring_queue(struct thread_data *td,
 	if (next_tail == atomic_load_acquire(ring->head))
 		return FIO_Q_BUSY;
 
-	if (o->cmdprio_percentage)
+	if (ld->use_cmdprio)
 		fio_ioring_prio_prep(td, io_u);
 	ring->array[tail & ld->sq_ring_mask] = io_u->index;
 	atomic_store_release(ring->tail, next_tail);
@@ -435,9 +553,10 @@ static int fio_ioring_commit(struct thread_data *td)
 	 */
 	if (o->sqpoll_thread) {
 		struct io_sq_ring *ring = &ld->sq_ring;
+		unsigned flags;
 
-		read_barrier();
-		if (*ring->flags & IORING_SQ_NEED_WAKEUP)
+		flags = atomic_load_acquire(ring->flags);
+		if (flags & IORING_SQ_NEED_WAKEUP)
 			io_uring_enter(ld, ld->queued, 0,
 					IORING_ENTER_SQ_WAKEUP);
 		ld->queued = 0;
@@ -479,7 +598,7 @@ static void fio_ioring_unmap(struct ioring_data *ld)
 {
 	int i;
 
-	for (i = 0; i < ARRAY_SIZE(ld->mmap); i++)
+	for (i = 0; i < FIO_ARRAY_SIZE(ld->mmap); i++)
 		munmap(ld->mmap[i].ptr, ld->mmap[i].len);
 	close(ld->ring_fd);
 }
@@ -539,6 +658,40 @@ static int fio_ioring_mmap(struct ioring_data *ld, struct io_uring_params *p)
 	return 0;
 }
 
+static void fio_ioring_probe(struct thread_data *td)
+{
+	struct ioring_data *ld = td->io_ops_data;
+	struct ioring_options *o = td->eo;
+	struct io_uring_probe *p;
+	int ret;
+
+	/* already set by user, don't touch */
+	if (o->nonvectored != -1)
+		return;
+
+	/* default to off, as that's always safe */
+	o->nonvectored = 0;
+
+	p = malloc(sizeof(*p) + 256 * sizeof(struct io_uring_probe_op));
+	if (!p)
+		return;
+
+	memset(p, 0, sizeof(*p) + 256 * sizeof(struct io_uring_probe_op));
+	ret = syscall(__NR_io_uring_register, ld->ring_fd,
+			IORING_REGISTER_PROBE, p, 256);
+	if (ret < 0)
+		goto out;
+
+	if (IORING_OP_WRITE > p->ops_len)
+		goto out;
+
+	if ((p->ops[IORING_OP_READ].flags & IO_URING_OP_SUPPORTED) &&
+	    (p->ops[IORING_OP_WRITE].flags & IO_URING_OP_SUPPORTED))
+		o->nonvectored = 1;
+out:
+	free(p);
+}
+
 static int fio_ioring_queue_init(struct thread_data *td)
 {
 	struct ioring_data *ld = td->io_ops_data;
@@ -565,15 +718,9 @@ static int fio_ioring_queue_init(struct thread_data *td)
 
 	ld->ring_fd = ret;
 
-	if (o->fixedbufs) {
-		struct rlimit rlim = {
-			.rlim_cur = RLIM_INFINITY,
-			.rlim_max = RLIM_INFINITY,
-		};
-
-		if (setrlimit(RLIMIT_MEMLOCK, &rlim) < 0)
-			return -1;
+	fio_ioring_probe(td);
 
+	if (o->fixedbufs) {
 		ret = syscall(__NR_io_uring_register, ld->ring_fd,
 				IORING_REGISTER_BUFFERS, ld->iovecs, depth);
 		if (ret < 0)
@@ -640,10 +787,21 @@ static int fio_ioring_post_init(struct thread_data *td)
 
 	err = fio_ioring_queue_init(td);
 	if (err) {
-		td_verror(td, errno, "io_queue_init");
+		int init_err = errno;
+
+		if (init_err == ENOSYS)
+			log_err("fio: your kernel doesn't support io_uring\n");
+		td_verror(td, init_err, "io_queue_init");
 		return 1;
 	}
 
+	for (i = 0; i < td->o.iodepth; i++) {
+		struct io_uring_sqe *sqe;
+
+		sqe = &ld->sqes[i];
+		memset(sqe, 0, sizeof(*sqe));
+	}
+
 	if (o->registerfiles) {
 		err = fio_ioring_register_files(td);
 		if (err) {
@@ -659,7 +817,9 @@ static int fio_ioring_init(struct thread_data *td)
 {
 	struct ioring_options *o = td->eo;
 	struct ioring_data *ld;
-	struct thread_options *to = &td->o;
+	struct cmdprio *cmdprio = &o->cmdprio;
+	bool has_cmdprio = false;
+	int ret;
 
 	/* sqthread submission requires registered files */
 	if (o->sqpoll_thread)
@@ -683,21 +843,21 @@ static int fio_ioring_init(struct thread_data *td)
 
 	td->io_ops_data = ld;
 
-	/*
-	 * Check for option conflicts
-	 */
-	if ((fio_option_is_set(to, ioprio) || fio_option_is_set(to, ioprio_class)) &&
-			o->cmdprio_percentage != 0) {
-		log_err("%s: cmdprio_percentage option and mutually exclusive "
-				"prio or prioclass option is set, exiting\n", to->name);
-		td_verror(td, EINVAL, "fio_io_uring_init");
+	ret = fio_cmdprio_init(td, cmdprio, &has_cmdprio);
+	if (ret) {
+		td_verror(td, EINVAL, "fio_ioring_init");
 		return 1;
 	}
 
-	if (fio_option_is_set(&td->o, ioprio_class))
-		ld->ioprio_class_set = true;
-	if (fio_option_is_set(&td->o, ioprio))
-		ld->ioprio_set = true;
+	/*
+	 * Since io_uring can have a submission context (sqthread_poll) that is
+	 * different from the process context, we cannot rely on the the IO
+	 * priority set by ioprio_set() (option prio/prioclass) to be inherited.
+	 * Therefore, we set the sqe->ioprio field when prio/prioclass is used.
+	 */
+	ld->use_cmdprio = has_cmdprio ||
+		fio_option_is_set(&td->o, ioprio_class) ||
+		fio_option_is_set(&td->o, ioprio);
 
 	return 0;
 }
@@ -737,7 +897,7 @@ static int fio_ioring_close_file(struct thread_data *td, struct fio_file *f)
 static struct ioengine_ops ioengine = {
 	.name			= "io_uring",
 	.version		= FIO_IOOPS_VERSION,
-	.flags			= FIO_ASYNCIO_SYNC_TRIM,
+	.flags			= FIO_ASYNCIO_SYNC_TRIM | FIO_NO_OFFLOAD,
 	.init			= fio_ioring_init,
 	.post_init		= fio_ioring_post_init,
 	.io_u_init		= fio_ioring_io_u_init,