X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=engines%2Fsyslet-rw.c;h=2eab207cd8c7cb4c0ab175a321b3d9842fe72634;hp=32697a88c25b64a281a794b77b1674d54ac20410;hb=12d9d841526ad75a67bb43a90edeefd05f85f11e;hpb=a2e1b08afa641078832d69871ae5e404db3e4fc1

diff --git a/engines/syslet-rw.c b/engines/syslet-rw.c
index 32697a88..2eab207c 100644
--- a/engines/syslet-rw.c
+++ b/engines/syslet-rw.c
@@ -1,5 +1,8 @@
 /*
- * read/write() engine that uses syslet to be async
+ * syslet engine
+ *
+ * IO engine that does regular pread(2)/pwrite(2) to transfer data, but
+ * with syslets to make the execution async.
  *
  */
 #include <stdio.h>
@@ -7,76 +10,108 @@
 #include <unistd.h>
 #include <errno.h>
 #include <assert.h>
+#include <malloc.h>
+#include <asm/unistd.h>
 
 #include "../fio.h"
-#include "../os.h"
+#include "../lib/fls.h"
 
 #ifdef FIO_HAVE_SYSLET
 
+#ifdef __NR_pread64
+#define __NR_fio_pread	__NR_pread64
+#define __NR_fio_pwrite	__NR_pwrite64
+#else
+#define __NR_fio_pread	__NR_pread
+#define __NR_fio_pwrite	__NR_pwrite
+#endif
+
 struct syslet_data {
 	struct io_u **events;
 	unsigned int nr_events;
 	
-	struct syslet_uatom **ring;
-	unsigned int ring_index;
+	struct syslet_ring *ring;
+	unsigned int ring_mask;
+	void *stack;
 };
 
-/*
- * Inspect the ring to see if we have completed events
- */
-static void fio_syslet_complete(struct thread_data *td)
+static void fio_syslet_add_event(struct thread_data *td, struct io_u *io_u)
 {
 	struct syslet_data *sd = td->io_ops->data;
 
-	do {
-		struct syslet_uatom *atom;
-		struct io_u *io_u;
-		long ret;
+	assert(sd->nr_events < td->o.iodepth);
+	sd->events[sd->nr_events++] = io_u;
+}
 
-		atom = sd->ring[sd->ring_index];
-		if (!atom)
-			break;
+static void fio_syslet_add_events(struct thread_data *td, unsigned int nr)
+{
+	struct syslet_data *sd = td->io_ops->data;
+	unsigned int i, uidx;
+
+	uidx = sd->ring->user_tail;
+	read_barrier();
 
-		sd->ring[sd->ring_index] = NULL;
-		if (++sd->ring_index == td->iodepth)
-			sd->ring_index = 0;
+	for (i = 0; i < nr; i++) {
+		unsigned int idx = (i + uidx) & sd->ring_mask;
+		struct syslet_completion *comp = &sd->ring->comp[idx];
+		struct io_u *io_u = (struct io_u *) (long) comp->caller_data;
+		long ret;
 
-		io_u = atom->private;
-		ret = *atom->ret_ptr;
-		if (ret > 0)
+		ret = comp->status;
+		if (ret <= 0) {
+			io_u->resid = io_u->xfer_buflen;
+			io_u->error = -ret;
+		} else {
 			io_u->resid = io_u->xfer_buflen - ret;
-		else if (ret < 0)
-			io_u->error = ret;
+			io_u->error = 0;
+		}
 
-		sd->events[sd->nr_events++] = io_u;
-	} while (1);
+		fio_syslet_add_event(td, io_u);
+	}
 }
 
-static int fio_syslet_getevents(struct thread_data *td, int min,
-				int fio_unused max,
-				struct timespec fio_unused *t)
+static void fio_syslet_wait_for_events(struct thread_data *td)
 {
 	struct syslet_data *sd = td->io_ops->data;
-	int get_events;
-	long ret;
+	struct syslet_ring *ring = sd->ring;
 
 	do {
-		fio_syslet_complete(td);
+		unsigned int kh = ring->kernel_head;
+		int ret;
 
 		/*
-		 * do we have enough immediate completions?
+		 * first reap events that are already completed
 		 */
-		if (sd->nr_events >= (unsigned int) min)
+		if (ring->user_tail != kh) {
+			unsigned int nr = kh - ring->user_tail;
+
+			fio_syslet_add_events(td, nr);
+			ring->user_tail = kh;
 			break;
+		}
 
 		/*
-		 * OK, we need to wait for some events...
+		 * block waiting for at least one event
 		 */
-		get_events = min - sd->nr_events;
-		ret = async_wait(get_events);
-		if (ret < 0)
-			return errno;
+		ret = syscall(__NR_syslet_ring_wait, ring, ring->user_tail);
+		assert(!ret);
 	} while (1);
+}
+
+static int fio_syslet_getevents(struct thread_data *td, unsigned int min,
+				unsigned int fio_unused max,
+				struct timespec fio_unused *t)
+{
+	struct syslet_data *sd = td->io_ops->data;
+	long ret;
+
+	/*
+	 * While we have less events than requested, block waiting for them
+	 * (if we have to, there may already be more completed events ready
+	 * for us - see fio_syslet_wait_for_events()
+	 */
+	while (sd->nr_events < min)
+		fio_syslet_wait_for_events(td);
 
 	ret = sd->nr_events;
 	sd->nr_events = 0;
@@ -90,125 +125,93 @@ static struct io_u *fio_syslet_event(struct thread_data *td, int event)
 	return sd->events[event];
 }
 
-static void init_atom(struct syslet_uatom *atom, int nr, void *arg0,
-		      void *arg1, void *arg2, void *arg3, void *ret_ptr,
-		      unsigned long flags, void *priv)
+static void fio_syslet_prep_sync(struct fio_file *f,
+				 struct indirect_registers *regs)
 {
-	atom->flags = flags;
-	atom->nr = nr;
-	atom->ret_ptr = ret_ptr;
-	atom->next = NULL;
-	atom->arg_ptr[0] = arg0;
-	atom->arg_ptr[1] = arg1;
-	atom->arg_ptr[2] = arg2;
-	atom->arg_ptr[3] = arg3;
-	atom->arg_ptr[4] = atom->arg_ptr[5] = NULL;
-	atom->private = priv;
+	FILL_IN(*regs, __NR_fsync, (long) f->fd);
 }
 
-/*
- * Use seek atom for sync
- */
-static void fio_syslet_prep_sync(struct io_u *io_u, struct fio_file *f)
-{
-	init_atom(&io_u->req.atom, __NR_fsync, &f->fd, NULL, NULL, NULL,
-		  &io_u->req.ret, SYSLET_STOP_ON_NEGATIVE, io_u);
-}
-
-static void fio_syslet_prep_rw(struct io_u *io_u, struct fio_file *f)
+static void fio_syslet_prep_rw(struct io_u *io_u, struct fio_file *f,
+			       struct indirect_registers *regs)
 {
-	int nr;
+	long nr;
 
 	/*
 	 * prepare rw
 	 */
 	if (io_u->ddir == DDIR_READ)
-		nr = __NR_pread64;
+		nr = __NR_fio_pread;
 	else
-		nr = __NR_pwrite64;
+		nr = __NR_fio_pwrite;
 
-	init_atom(&io_u->req.atom, nr, &f->fd, &io_u->xfer_buf,
-		  &io_u->xfer_buflen, &io_u->offset, &io_u->req.ret,
-		  SYSLET_STOP_ON_NEGATIVE, io_u);
+	FILL_IN(*regs, nr, (long) f->fd, (long) io_u->xfer_buf,
+		(long) io_u->xfer_buflen, (long) io_u->offset);
 }
 
-static int fio_syslet_prep(struct thread_data fio_unused *td, struct io_u *io_u)
+static void fio_syslet_prep(struct io_u *io_u, struct indirect_registers *regs)
 {
 	struct fio_file *f = io_u->file;
 
 	if (io_u->ddir == DDIR_SYNC)
-		fio_syslet_prep_sync(io_u, f);
+		fio_syslet_prep_sync(f, regs);
 	else
-		fio_syslet_prep_rw(io_u, f);
+		fio_syslet_prep_rw(io_u, f, regs);
+}
 
-	return 0;
+static void ret_func(void)
+{
+	syscall(__NR_exit);
 }
 
 static int fio_syslet_queue(struct thread_data *td, struct io_u *io_u)
 {
 	struct syslet_data *sd = td->io_ops->data;
-	struct syslet_uatom *done;
-	long ret;
+	union indirect_params params;
+	struct indirect_registers regs;
+	int ret;
 
-	done = async_exec(&io_u->req.atom);
-	if (!done)
-		return 0;
-
-	/*
-	 * completed sync
-	 */
-	ret = io_u->req.ret;
-	if (ret != (long) io_u->xfer_buflen) {
-		if (ret > 0) {
-			io_u->resid = io_u->xfer_buflen - ret;
-			io_u->error = 0;
-			return ret;
-		} else
-			io_u->error = errno;
-	}
-
-	if (!io_u->error)
-		sd->events[sd->nr_events++] = io_u;
-	else
-		td_verror(td, io_u->error);
+	fio_ro_check(td, io_u);
 
-	return io_u->error;
-}
-
-static int async_head_init(struct syslet_data *sd, unsigned int depth)
-{
-	struct async_head_user ahu;
-	unsigned long ring_size;
-
-	ring_size = sizeof(struct syslet_uatom *) * depth;
-	sd->ring = malloc(ring_size);
-	memset(sd->ring, 0, ring_size);
+	memset(&params, 0, sizeof(params));
+	fill_syslet_args(&params.syslet, sd->ring, (long)io_u, ret_func, sd->stack);
 
-	memset(&ahu, 0, sizeof(ahu));
-	ahu.completion_ring = sd->ring;
-	ahu.ring_size_bytes = ring_size;
-	ahu.max_nr_threads = -1;
+	fio_syslet_prep(io_u, &regs);
 
-	if (async_register(&ahu, sizeof(ahu)) < 0) {
-		perror("async_register");
-		fprintf(stderr, "fio: syslet likely not supported\n");
-		free(sd->ring);
-		return 1;
+	ret = syscall(__NR_indirect, &regs, &params, sizeof(params), 0);
+	if (ret == (int) io_u->xfer_buflen) {
+		/*
+		 * completed sync, account. this also catches fsync().
+		 */
+		return FIO_Q_COMPLETED;
+	} else if (ret < 0) {
+		/*
+		 * queued for async execution
+		 */
+		if (errno == ESYSLETPENDING)
+			return FIO_Q_QUEUED;
 	}
 
-	return 0;
+	io_u->error = errno;
+	td_verror(td, io_u->error, "xfer");
+	return FIO_Q_COMPLETED;
 }
 
-static void async_head_exit(struct syslet_data *sd, unsigned int depth)
+static int check_syslet_support(struct syslet_data *sd)
 {
-	struct async_head_user ahu;
+	union indirect_params params;
+	struct indirect_registers regs;
+	pid_t pid, my_pid = getpid();
+
+	memset(&params, 0, sizeof(params));
+	fill_syslet_args(&params.syslet, sd->ring, 0, ret_func, sd->stack);
 
-	memset(&ahu, 0, sizeof(ahu));
-	ahu.completion_ring = sd->ring;
-	ahu.ring_size_bytes = sizeof(struct syslet_uatom *) * depth;
+	FILL_IN(regs, __NR_getpid);
 
-	if (async_unregister(&ahu, sizeof(ahu)) < 0)
-		perror("async_register");
+	pid = syscall(__NR_indirect, &regs, &params, sizeof(params), 0);
+	if (pid == my_pid)
+		return 0;
+
+	return 1;
 }
 
 static void fio_syslet_cleanup(struct thread_data *td)
@@ -216,45 +219,71 @@ static void fio_syslet_cleanup(struct thread_data *td)
 	struct syslet_data *sd = td->io_ops->data;
 
 	if (sd) {
-		async_head_exit(sd, td->iodepth);
 		free(sd->events);
+		free(sd->ring);
 		free(sd);
-		td->io_ops->data = NULL;
 	}
 }
 
 static int fio_syslet_init(struct thread_data *td)
 {
 	struct syslet_data *sd;
-
+	void *ring = NULL, *stack = NULL;
+	unsigned int ring_size, ring_nr;
 
 	sd = malloc(sizeof(*sd));
 	memset(sd, 0, sizeof(*sd));
-	sd->events = malloc(sizeof(struct io_u *) * td->iodepth);
-	memset(sd->events, 0, sizeof(struct io_u *) * td->iodepth);
+
+	sd->events = malloc(sizeof(struct io_u *) * td->o.iodepth);
+	memset(sd->events, 0, sizeof(struct io_u *) * td->o.iodepth);
 
 	/*
-	 * This will handily fail for kernels where syslet isn't available
+	 * The ring needs to be a power-of-2, so round it up if we have to
 	 */
-	if (async_head_init(sd, td->iodepth)) {
-		free(sd->events);
-		free(sd);
-		return 1;
+	ring_nr = td->o.iodepth;
+	if (ring_nr & (ring_nr - 1))
+		ring_nr = 1 << __fls(ring_nr);
+
+	ring_size = sizeof(struct syslet_ring) +
+			ring_nr * sizeof(struct syslet_completion);
+	if (posix_memalign(&ring, sizeof(uint64_t), ring_size))
+		goto err_mem;
+	if (posix_memalign(&stack, page_size, page_size))
+		goto err_mem;
+
+	sd->ring = ring;
+	sd->ring_mask = ring_nr - 1;
+	sd->stack = stack;
+
+	memset(sd->ring, 0, ring_size);
+	sd->ring->elements = ring_nr;
+
+	if (!check_syslet_support(sd)) {
+		td->io_ops->data = sd;
+		return 0;
 	}
 
-	td->io_ops->data = sd;
-	return 0;
+	log_err("fio: syslets do not appear to work\n");
+err_mem:
+	free(sd->events);
+	if (ring)
+		free(ring);
+	if (stack)
+		free(stack);
+	free(sd);
+	return 1;
 }
 
 static struct ioengine_ops ioengine = {
 	.name		= "syslet-rw",
 	.version	= FIO_IOOPS_VERSION,
 	.init		= fio_syslet_init,
-	.prep		= fio_syslet_prep,
 	.queue		= fio_syslet_queue,
 	.getevents	= fio_syslet_getevents,
 	.event		= fio_syslet_event,
 	.cleanup	= fio_syslet_cleanup,
+	.open_file	= generic_open_file,
+	.close_file	= generic_close_file,
 };
 
 #else /* FIO_HAVE_SYSLET */