X-Git-Url: https://git.kernel.dk/?a=blobdiff_plain;f=engines%2Flibaio.c;h=dd655355ba6995725c82f27ec3264690d92c3709;hb=a71ad043a3f4ad405911841e1fe5bb09823e3be2;hp=6c5d73ed27dd1751ac98c1177fcdbf2bdf2fe487;hpb=b220c6d6f37fc55760d66728f6f9f521ea2f9b46;p=fio.git

diff --git a/engines/libaio.c b/engines/libaio.c
index 6c5d73ed..dd655355 100644
--- a/engines/libaio.c
+++ b/engines/libaio.c
@@ -4,14 +4,31 @@
  * IO engine using the Linux native aio interface.
  *
  */
-#include <stdio.h>
 #include <stdlib.h>
 #include <unistd.h>
 #include <errno.h>
-#include <assert.h>
 #include <libaio.h>
+#include <sys/time.h>
+#include <sys/resource.h>
 
 #include "../fio.h"
+#include "../lib/pow2.h"
+#include "../optgroup.h"
+#include "../lib/memalign.h"
+#include "cmdprio.h"
+
+/* Should be defined in newest aio_abi.h */
+#ifndef IOCB_FLAG_IOPRIO
+#define IOCB_FLAG_IOPRIO    (1 << 1)
+#endif
+
+/* Hack for libaio < 0.3.111 */
+#ifndef CONFIG_LIBAIO_RW_FLAGS
+#define aio_rw_flags __pad2
+#endif
+
+static int fio_libaio_commit(struct thread_data *td);
+static int fio_libaio_init(struct thread_data *td);
 
 struct libaio_data {
 	io_context_t aio_ctx;
@@ -19,6 +36,8 @@ struct libaio_data {
 	struct iocb **iocbs;
 	struct io_u **io_us;
 
+	struct io_u **io_u_index;
+
 	/*
 	 * Basic ring buffer. 'head' is incremented in _queue(), and
 	 * 'tail' is incremented in _commit(). We keep 'queued' so
@@ -32,13 +51,26 @@ struct libaio_data {
 	unsigned int queued;
 	unsigned int head;
 	unsigned int tail;
+
+	bool use_cmdprio;
 };
 
 struct libaio_options {
 	struct thread_data *td;
 	unsigned int userspace_reap;
+	struct cmdprio cmdprio;
+	unsigned int nowait;
 };
 
+static int str_cmdprio_bssplit_cb(void *data, const char *input)
+{
+	struct libaio_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	= "userspace_reap",
@@ -49,6 +81,96 @@ static struct fio_option options[] = {
 		.category = FIO_OPT_C_ENGINE,
 		.group	= FIO_OPT_G_LIBAIO,
 	},
+#ifdef FIO_HAVE_IOPRIO_CLASS
+	{
+		.name	= "cmdprio_percentage",
+		.lname	= "high priority percentage",
+		.type	= FIO_OPT_INT,
+		.off1	= offsetof(struct libaio_options,
+				   cmdprio.percentage[DDIR_READ]),
+		.off2	= offsetof(struct libaio_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_LIBAIO,
+	},
+	{
+		.name	= "cmdprio_class",
+		.lname	= "Asynchronous I/O priority class",
+		.type	= FIO_OPT_INT,
+		.off1	= offsetof(struct libaio_options,
+				   cmdprio.class[DDIR_READ]),
+		.off2	= offsetof(struct libaio_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_LIBAIO,
+	},
+	{
+		.name	= "cmdprio",
+		.lname	= "Asynchronous I/O priority level",
+		.type	= FIO_OPT_INT,
+		.off1	= offsetof(struct libaio_options,
+				   cmdprio.level[DDIR_READ]),
+		.off2	= offsetof(struct libaio_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_LIBAIO,
+	},
+	{
+		.name   = "cmdprio_bssplit",
+		.lname  = "Priority percentage block size split",
+		.type   = FIO_OPT_STR_ULL,
+		.cb     = str_cmdprio_bssplit_cb,
+		.off1   = offsetof(struct libaio_options, cmdprio.bssplit),
+		.help   = "Set priority percentages for different block sizes",
+		.category = FIO_OPT_C_ENGINE,
+		.group	= FIO_OPT_G_LIBAIO,
+	},
+#else
+	{
+		.name	= "cmdprio_percentage",
+		.lname	= "high priority percentage",
+		.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	= "nowait",
+		.lname	= "RWF_NOWAIT",
+		.type	= FIO_OPT_BOOL,
+		.off1	= offsetof(struct libaio_options, nowait),
+		.help	= "Set RWF_NOWAIT for reads/writes",
+		.category = FIO_OPT_C_ENGINE,
+		.group	= FIO_OPT_G_LIBAIO,
+	},
 	{
 		.name	= NULL,
 	},
@@ -63,23 +185,59 @@ static inline void ring_inc(struct libaio_data *ld, unsigned int *val,
 		*val = (*val + add) % ld->entries;
 }
 
-static int fio_libaio_prep(struct thread_data fio_unused *td, struct io_u *io_u)
+static int fio_libaio_prep(struct thread_data *td, struct io_u *io_u)
 {
+	struct libaio_options *o = td->eo;
 	struct fio_file *f = io_u->file;
-
-	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);
+	struct iocb *iocb = &io_u->iocb;
+
+	if (io_u->ddir == DDIR_READ) {
+		io_prep_pread(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
+		if (o->nowait)
+			iocb->aio_rw_flags |= RWF_NOWAIT;
+	} else if (io_u->ddir == DDIR_WRITE) {
+		io_prep_pwrite(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
+		if (o->nowait)
+			iocb->aio_rw_flags |= RWF_NOWAIT;
+	} else if (ddir_sync(io_u->ddir))
+		io_prep_fsync(iocb, f->fd);
 
 	return 0;
 }
 
+static void fio_libaio_prio_prep(struct thread_data *td, struct io_u *io_u)
+{
+	struct libaio_options *o = td->eo;
+	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) {
+		io_u->ioprio = cmdprio_value;
+		io_u->iocb.aio_reqprio = cmdprio_value;
+		io_u->iocb.u.c.flags |= IOCB_FLAG_IOPRIO;
+		if (!td->ioprio || cmdprio_value < td->ioprio) {
+			/*
+			 * The async IO priority is higher (has a lower value)
+			 * than the default context priority.
+			 */
+			io_u->flags |= IO_U_F_HIGH_PRIO;
+		}
+	} else if (td->ioprio && td->ioprio < cmdprio_value) {
+		/*
+		 * The IO will be executed with the default context priority,
+		 * and this priority is higher (has a lower value) than the
+		 * async IO priority.
+		 */
+		io_u->flags |= IO_U_F_HIGH_PRIO;
+	}
+}
+
 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;
 
@@ -129,8 +287,8 @@ static int user_io_getevents(io_context_t aio_ctx, unsigned int max,
 		} else {
 			/* There is another completion to reap */
 			events[i] = ring->events[head];
-			read_barrier();
-			ring->head = (head + 1) % ring->nr;
+			atomic_store_release(&ring->head,
+					     (head + 1) % ring->nr);
 			i++;
 		}
 	}
@@ -141,9 +299,9 @@ 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, 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;
 
@@ -163,20 +321,23 @@ static int fio_libaio_getevents(struct thread_data *td, unsigned int min,
 			r = io_getevents(ld->aio_ctx, actual_min,
 				max, ld->aio_events + events, lt);
 		}
-		if (r >= 0)
+		if (r > 0)
 			events += r;
-		else if (r == -EAGAIN)
-			usleep(100);
-		else
+		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);
 
@@ -202,9 +363,14 @@ static int fio_libaio_queue(struct thread_data *td, struct io_u *io_u)
 			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;
 	}
 
+	if (ld->use_cmdprio)
+		fio_libaio_prio_prep(td, io_u);
+
 	ld->iocbs[ld->head] = &io_u->iocb;
 	ld->io_us[ld->head] = io_u;
 	ring_inc(ld, &ld->head, 1);
@@ -215,7 +381,7 @@ static int fio_libaio_queue(struct thread_data *td, struct io_u *io_u)
 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))
@@ -233,10 +399,10 @@ 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;
-	struct timeval tv;
+	struct timespec ts;
 	int ret, wait_start = 0;
 
 	if (!ld->queued)
@@ -277,9 +443,9 @@ static int fio_libaio_commit(struct thread_data *td)
 				break;
 			}
 			if (!wait_start) {
-				fio_gettime(&tv, NULL);
-				wait_start = 0;
-			} else if (mtime_since_now(&tv) > 30000) {
+				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;
 			}
@@ -303,17 +469,24 @@ static int fio_libaio_commit(struct thread_data *td)
 
 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);
@@ -321,44 +494,52 @@ 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_options *o = td->eo;
-	struct libaio_data *ld;
-	int err = 0;
-
-	ld = calloc(1, sizeof(*ld));
+	struct libaio_data *ld = td->io_ops_data;
+	int err;
 
-	/*
-	 * First try passing in 0 for queue depth, since we don't
-	 * care about the user ring. If that fails, the kernel is too old
-	 * and we need the right depth.
-	 */
-	if (!o->userspace_reap)
-		err = io_queue_init(INT_MAX, &ld->aio_ctx);
-	if (o->userspace_reap || err == -EINVAL)
-		err = io_queue_init(td->o.iodepth, &ld->aio_ctx);
+	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;
 	}
 
+	return 0;
+}
+
+static int fio_libaio_init(struct thread_data *td)
+{
+	struct libaio_data *ld;
+	struct libaio_options *o = td->eo;
+	struct cmdprio *cmdprio = &o->cmdprio;
+	int ret;
+
+	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;
+	td->io_ops_data = ld;
+
+	ret = fio_cmdprio_init(td, cmdprio, &ld->use_cmdprio);
+	if (ret) {
+		td_verror(td, EINVAL, "fio_libaio_init");
+		return 1;
+	}
+
 	return 0;
 }
 
-static struct ioengine_ops ioengine = {
+FIO_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,