syslet engine: submission latency numbers

[fio.git] / engines / syslet-rw.c

diff --git a/engines/syslet-rw.c b/engines/syslet-rw.c
index f188d1c665fac1a7240504c3957364e036a2a451..8ab217f40b407e0c628016d8794bd90ed5f1e6aa 100644 (file)
--- a/engines/syslet-rw.c
+++ b/engines/syslet-rw.c
@@ -19,8 +19,50 @@ struct syslet_data {
        
        struct async_head_user ahu;
        struct syslet_uatom **ring;
+
+       struct syslet_uatom *head, *tail;
 };
 
+static void fio_syslet_complete_atom(struct thread_data *td,
+                                    struct syslet_uatom *atom)
+{
+       struct syslet_data *sd = td->io_ops->data;
+       struct syslet_uatom *last;
+       struct io_u *io_u;
+
+       /*
+        * complete from the beginning of the sequence up to (and
+        * including) this atom
+        */
+       last = atom;
+       io_u = atom->private;
+       atom = io_u->req.head;
+
+       /*
+        * now complete in right order
+        */
+       do {
+               long ret;
+
+               io_u = atom->private;
+               ret = *atom->ret_ptr;
+               if (ret > 0)
+                       io_u->resid = io_u->xfer_buflen - ret;
+               else if (ret < 0)
+                       io_u->error = ret;
+
+               assert(sd->nr_events < td->iodepth);
+               sd->events[sd->nr_events++] = io_u;
+
+               if (atom == last)
+                       break;
+
+               atom = atom->next;
+       } while (1);
+
+       assert(!last->next);
+}
+
 /*
  * Inspect the ring to see if we have completed events
  */
@@ -30,8 +72,6 @@ static void fio_syslet_complete(struct thread_data *td)
 
        do {
                struct syslet_uatom *atom;
-               struct io_u *io_u;
-               long ret;
 
                atom = sd->ring[sd->ahu.user_ring_idx];
                if (!atom)
@@ -41,14 +81,7 @@ static void fio_syslet_complete(struct thread_data *td)
                if (++sd->ahu.user_ring_idx == td->iodepth)
                        sd->ahu.user_ring_idx = 0;
 
-               io_u = atom->private;
-               ret = *atom->ret_ptr;
-               if (ret > 0)
-                       io_u->resid = io_u->xfer_buflen - ret;
-               else if (ret < 0)
-                       io_u->error = ret;
-
-               sd->events[sd->nr_events++] = io_u;
+               fio_syslet_complete_atom(td, atom);
        } while (1);
 }
 
@@ -57,7 +90,6 @@ static int fio_syslet_getevents(struct thread_data *td, int min,
                                struct timespec fio_unused *t)
 {
        struct syslet_data *sd = td->io_ops->data;
-       int get_events;
        long ret;
 
        do {
@@ -72,8 +104,7 @@ static int fio_syslet_getevents(struct thread_data *td, int min,
                /*
                 * OK, we need to wait for some events...
                 */
-               get_events = min - sd->nr_events;
-               ret = async_wait(get_events, sd->ahu.user_ring_idx, &sd->ahu);
+               ret = async_wait(1, sd->ahu.user_ring_idx, &sd->ahu);
                if (ret < 0)
                        return -errno;
        } while (1);
@@ -146,53 +177,74 @@ static int fio_syslet_prep(struct thread_data fio_unused *td, struct io_u *io_u)
 static void cachemiss_thread_start(void)
 {
        while (1)
-               async_thread();
+               async_thread(NULL, NULL);
 }
 
 #define THREAD_STACK_SIZE (16384)
 
 static unsigned long thread_stack_alloc()
 {
-       return (unsigned long)malloc(THREAD_STACK_SIZE) + THREAD_STACK_SIZE;
+       return (unsigned long) malloc(THREAD_STACK_SIZE) + THREAD_STACK_SIZE;
 }
 
-static int fio_syslet_queue(struct thread_data *td, struct io_u *io_u)
+static void fio_syslet_queued(struct thread_data *td, struct syslet_data *sd)
+{
+       struct syslet_uatom *atom;
+       struct timeval now;
+
+       fio_gettime(&now, NULL);
+
+       atom = sd->head;
+       while (atom) {
+               struct io_u *io_u = atom->private;
+
+               memcpy(&io_u->issue_time, &now, sizeof(now));
+               io_u_queued(td, io_u);
+               atom = atom->next;
+       }
+}
+
+static int fio_syslet_commit(struct thread_data *td)
 {
        struct syslet_data *sd = td->io_ops->data;
        struct syslet_uatom *done;
-       long ret;
+
+       if (!sd->head)
+               return 0;
+
+       assert(!sd->tail->next);
 
        if (!sd->ahu.new_thread_stack)
                sd->ahu.new_thread_stack = thread_stack_alloc();
 
+       fio_syslet_queued(td, sd);
+
        /*
         * On sync completion, the atom is returned. So on NULL return
         * it's queued asynchronously.
         */
-       done = async_exec(&io_u->req.atom, &sd->ahu);
+       done = async_exec(sd->head, &sd->ahu);
 
-       if (!done)
-               return FIO_Q_QUEUED;
+       sd->head = sd->tail = NULL;
 
-       /*
-        * 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 FIO_Q_COMPLETED;
-               } else
-                       io_u->error = errno;
-       }
+       if (done)
+               fio_syslet_complete_atom(td, done);
+
+       return 0;
+}
 
-       assert(sd->nr_events < td->iodepth);
+static int fio_syslet_queue(struct thread_data *td, struct io_u *io_u)
+{
+       struct syslet_data *sd = td->io_ops->data;
 
-       if (io_u->error)
-               td_verror(td, io_u->error);
+       if (sd->tail) {
+               sd->tail->next = &io_u->req.atom;
+               sd->tail = &io_u->req.atom;
+       } else
+               sd->head = sd->tail = &io_u->req.atom;
 
-       return FIO_Q_COMPLETED;
+       io_u->req.head = sd->head;
+       return FIO_Q_QUEUED;
 }
 
 static int async_head_init(struct syslet_data *sd, unsigned int depth)
@@ -209,8 +261,8 @@ static int async_head_init(struct syslet_data *sd, unsigned int depth)
        sd->ahu.completion_ring = sd->ring;
        sd->ahu.ring_size_bytes = ring_size;
        sd->ahu.head_stack = thread_stack_alloc();
-       sd->ahu.head_eip = (unsigned long)cachemiss_thread_start;
-       sd->ahu.new_thread_eip = (unsigned long)cachemiss_thread_start;
+       sd->ahu.head_eip = (unsigned long) cachemiss_thread_start;
+       sd->ahu.new_thread_eip = (unsigned long) cachemiss_thread_start;
 
        return 0;
 }
@@ -261,6 +313,7 @@ static struct ioengine_ops ioengine = {
        .init           = fio_syslet_init,
        .prep           = fio_syslet_prep,
        .queue          = fio_syslet_queue,
+       .commit         = fio_syslet_commit,
        .getevents      = fio_syslet_getevents,
        .event          = fio_syslet_event,
        .cleanup        = fio_syslet_cleanup,