posixaio: fix for HP-UX

[fio.git] / engines / posixaio.c

diff --git a/engines/posixaio.c b/engines/posixaio.c
index 9ff05c4914151387dae188d4c8798c7c3d0985c0..e315e972b73df4d085f220fb54ed4bc276083de9 100644 (file)
--- a/engines/posixaio.c
+++ b/engines/posixaio.c
@@ -8,7 +8,7 @@
 #include <stdlib.h>
 #include <unistd.h>
 #include <errno.h>
-#include <assert.h>
+#include <fcntl.h>
 
 #include "../fio.h"
 
@@ -16,6 +16,7 @@
 
 struct posixaio_data {
        struct io_u **aio_events;
+       unsigned int queued;
 };
 
 static int fill_timespec(struct timespec *ts)
@@ -55,7 +56,7 @@ static int fio_posixaio_cancel(struct thread_data fio_unused *td,
        struct fio_file *f = io_u->file;
        int r = aio_cancel(f->fd, &io_u->aiocb);
 
-       if (r == 1 || r == AIO_CANCELED)
+       if (r == AIO_ALLDONE || r == AIO_CANCELED)
                return 0;
 
        return 1;
@@ -71,36 +72,49 @@ static int fio_posixaio_prep(struct thread_data fio_unused *td,
        aiocb->aio_buf = io_u->xfer_buf;
        aiocb->aio_nbytes = io_u->xfer_buflen;
        aiocb->aio_offset = io_u->offset;
+       aiocb->aio_sigevent.sigev_notify = SIGEV_NONE;
 
        io_u->seen = 0;
        return 0;
 }
 
-static int fio_posixaio_getevents(struct thread_data *td, int min, int max,
-                                 struct timespec *t)
+#define SUSPEND_ENTRIES        8
+
+static int fio_posixaio_getevents(struct thread_data *td, unsigned int min,
+                                 unsigned int max, struct timespec *t)
 {
        struct posixaio_data *pd = td->io_ops->data;
-       struct list_head *entry;
+       struct aiocb *suspend_list[SUSPEND_ENTRIES];
+       struct flist_head *entry;
        struct timespec start;
-       int r, have_timeout = 0;
+       int have_timeout = 0;
+       int suspend_entries = 0;
+       unsigned int r;
 
        if (t && !fill_timespec(&start))
                have_timeout = 1;
 
        r = 0;
+       memset(suspend_list, 0, sizeof(*suspend_list));
 restart:
-       list_for_each(entry, &td->io_u_busylist) {
-               struct io_u *io_u = list_entry(entry, struct io_u, list);
+       flist_for_each(entry, &td->io_u_busylist) {
+               struct io_u *io_u = flist_entry(entry, struct io_u, list);
                int err;
 
                if (io_u->seen)
                        continue;
 
                err = aio_error(&io_u->aiocb);
-               if (err == EINPROGRESS)
+               if (err == EINPROGRESS) {
+                       if (suspend_entries < SUSPEND_ENTRIES) {
+                               suspend_list[suspend_entries] = &io_u->aiocb;
+                               suspend_entries++;
+                       }
                        continue;
+               }
 
                io_u->seen = 1;
+               pd->queued--;
                pd->aio_events[r++] = io_u;
 
                if (err == ECANCELED)
@@ -111,9 +125,6 @@ restart:
                        io_u->resid = io_u->xfer_buflen - retval;
                } else
                        io_u->error = err;
-
-               if (r >= max)
-                       break;
        }
 
        if (r >= min)
@@ -128,10 +139,10 @@ restart:
        }
 
        /*
-        * hrmpf, we need to wait for more. we should use aio_suspend, for
-        * now just sleep a little and recheck status of busy-and-not-seen
+        * must have some in-flight, wait for at least one
         */
-       usleep(1000);
+       aio_suspend((const struct aiocb * const *)suspend_list,
+                                                       suspend_entries, t);
        goto restart;
 }
 
@@ -142,9 +153,10 @@ static struct io_u *fio_posixaio_event(struct thread_data *td, int event)
        return pd->aio_events[event];
 }
 
-static int fio_posixaio_queue(struct thread_data fio_unused *td,
+static int fio_posixaio_queue(struct thread_data *td,
                              struct io_u *io_u)
 {
+       struct posixaio_data *pd = td->io_ops->data;
        struct aiocb *aiocb = &io_u->aiocb;
        int ret;
 
@@ -154,15 +166,40 @@ static int fio_posixaio_queue(struct thread_data fio_unused *td,
                ret = aio_read(aiocb);
        else if (io_u->ddir == DDIR_WRITE)
                ret = aio_write(aiocb);
-       else
+       else if (io_u->ddir == DDIR_TRIM) {
+               if (pd->queued)
+                       return FIO_Q_BUSY;
+
+               do_io_u_trim(td, io_u);
+               return FIO_Q_COMPLETED;
+       } else {
+#ifdef FIO_HAVE_POSIXAIO_FSYNC
                ret = aio_fsync(O_SYNC, aiocb);
+#else
+               if (pd->queued)
+                       return FIO_Q_BUSY;
 
+               do_io_u_sync(td, io_u);
+               return FIO_Q_COMPLETED;
+#endif
+       }
+               
        if (ret) {
+               /*
+                * At least OSX has a very low limit on the number of pending
+                * IOs, so if it returns EAGAIN, we are out of resources
+                * to queue more. Just return FIO_Q_BUSY to naturally
+                * drop off at this depth.
+                */
+               if (errno == EAGAIN)
+                       return FIO_Q_BUSY;
+
                io_u->error = errno;
                td_verror(td, io_u->error, "xfer");
                return FIO_Q_COMPLETED;
        }
 
+       pd->queued++;
        return FIO_Q_QUEUED;
 }
 
@@ -173,7 +210,6 @@ static void fio_posixaio_cleanup(struct thread_data *td)
        if (pd) {
                free(pd->aio_events);
                free(pd);
-               td->io_ops->data = NULL;
        }
 }
 
@@ -201,6 +237,7 @@ static struct ioengine_ops ioengine = {
        .cleanup        = fio_posixaio_cleanup,
        .open_file      = generic_open_file,
        .close_file     = generic_close_file,
+       .get_file_size  = generic_get_file_size,
 };
 
 #else /* FIO_HAVE_POSIXAIO */
@@ -212,7 +249,7 @@ static struct ioengine_ops ioengine = {
  */
 static int fio_posixaio_init(struct thread_data fio_unused *td)
 {
-       fprintf(stderr, "fio: posixaio not available\n");
+       log_err("fio: posixaio not available\n");
        return 1;
 }