First cut at doing sequences of atoms in one commit
authorJens Axboe <jens.axboe@oracle.com>
Fri, 23 Feb 2007 12:21:45 +0000 (13:21 +0100)
committerJens Axboe <jens.axboe@oracle.com>
Fri, 23 Feb 2007 12:21:45 +0000 (13:21 +0100)
Needs error handling and stop conditions, so it's in an experimental
branch for now.

Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
engines/syslet-rw.c

index d439810..74a6b31 100644 (file)
@@ -19,8 +19,67 @@ struct syslet_data {
        
        struct async_head_user ahu;
        struct syslet_uatom **ring;
+
+       struct syslet_uatom *head, *tail;
+       struct syslet_uatom **event_map;
+       unsigned int event_map_idx;
 };
 
+static void fio_syslet_complete_atom(struct thread_data *td,
+                                    struct syslet_uatom *atom)
+{
+       struct syslet_data *sd = td->io_ops->data;
+       struct io_u *io_u;
+       int i, end;
+
+       if (!sd->event_map_idx)
+               return;
+
+       /*
+        * Find the start of the string of atoms for this sequence
+        */
+       for (end = sd->event_map_idx - 1; end >= 0; end--)
+               if (atom == sd->event_map[end])
+                       break;
+
+       if (end < 0 || atom != sd->event_map[end]) {
+               printf("didn't find atom\n");
+               return;
+       }
+
+       //printf("end=%d, total %d\n", end, sd->event_map_idx);
+
+       /*
+        * now complete in right order
+        */
+       for (i = 0; i <= end; i++) {
+               long ret;
+
+               atom = sd->event_map[i];
+               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;
+       }
+
+       /*
+        * Move later completions to the front, if we didn't complete all
+        */
+       if (end == (int) sd->event_map_idx - 1)
+               sd->event_map_idx = 0;
+       else {
+               int nr = sd->event_map_idx - end - 1;
+
+               memmove(sd->event_map, &sd->event_map[end + 1], nr * sizeof(struct syslet_uatom *));
+               sd->event_map_idx = nr;
+       }
+}
+
 /*
  * Inspect the ring to see if we have completed events
  */
@@ -30,8 +89,6 @@ 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 +98,7 @@ 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 +107,6 @@ 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 +121,7 @@ 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);
@@ -156,11 +204,13 @@ static unsigned long thread_stack_alloc()
        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 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;
 
        if (!sd->ahu.new_thread_stack)
                sd->ahu.new_thread_stack = thread_stack_alloc();
@@ -169,30 +219,28 @@ static int fio_syslet_queue(struct thread_data *td, struct io_u *io_u)
         * 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);
 
-       assert(sd->nr_events < td->iodepth);
+       return 0;
+}
+
+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, "xfer");
+       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;
+       sd->event_map[sd->event_map_idx++] = sd->tail;
+       return FIO_Q_QUEUED;
 }
 
 static int async_head_init(struct syslet_data *sd, unsigned int depth)
@@ -227,6 +275,7 @@ static void fio_syslet_cleanup(struct thread_data *td)
        if (sd) {
                async_head_exit(sd);
                free(sd->events);
+               free(sd->event_map);
                free(sd);
                td->io_ops->data = NULL;
        }
@@ -241,6 +290,8 @@ static int fio_syslet_init(struct thread_data *td)
        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->event_map = malloc(sizeof(struct syslet_uatom *) * td->iodepth);
+       memset(sd->event_map, 0, sizeof(struct syslet_uatom *) * td->iodepth);
 
        /*
         * This will handily fail for kernels where syslet isn't available
@@ -261,6 +312,7 @@ 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,