engines/libaio: add support for user mapped iocbs

For polled IO, we can support having the kernel map our iocbs,
instead of having to copy them for each IO submission.

Signed-off-by: Jens Axboe <axboe@kernel.dk>

diff --git a/engines/libaio.c b/engines/libaio.c
index 2a4d653467849e0d764c8a0c279e15264f13c690..bae141cabdd1bf57e974c1348fb5ab114e1da76d 100644 (file)
--- a/engines/libaio.c
+++ b/engines/libaio.c
@@ -19,6 +19,10 @@
 #ifndef IOCTX_FLAG_IOPOLL
 #define IOCTX_FLAG_IOPOLL      (1 << 0)
 #endif
+#ifndef IOCTX_FLAG_USERIOCB
+#define IOCTX_FLAG_USERIOCB    (1 << 1)
+#endif
+
 
 static int fio_libaio_commit(struct thread_data *td);
 
@@ -28,6 +32,9 @@ struct libaio_data {
        struct iocb **iocbs;
        struct io_u **io_us;
 
+       struct iocb *user_iocbs;
+       struct io_u **io_u_index;
+
        /*
         * Basic ring buffer. 'head' is incremented in _queue(), and
         * 'tail' is incremented in _commit(). We keep 'queued' so
@@ -47,6 +54,7 @@ struct libaio_options {
        void *pad;
        unsigned int userspace_reap;
        unsigned int hipri;
+       unsigned int useriocb;
 };
 
 static struct fio_option options[] = {
@@ -68,6 +76,15 @@ static struct fio_option options[] = {
                .category = FIO_OPT_C_ENGINE,
                .group  = FIO_OPT_G_LIBAIO,
        },
+       {
+               .name   = "useriocb",
+               .lname  = "User IOCBs",
+               .type   = FIO_OPT_STR_SET,
+               .off1   = offsetof(struct libaio_options, useriocb),
+               .help   = "Use user mapped IOCBs",
+               .category = FIO_OPT_C_ENGINE,
+               .group  = FIO_OPT_G_LIBAIO,
+       },
        {
                .name   = NULL,
        },
@@ -84,19 +101,26 @@ static inline void ring_inc(struct libaio_data *ld, unsigned int *val,
 
 static int fio_libaio_prep(struct thread_data fio_unused *td, struct io_u *io_u)
 {
+       struct libaio_data *ld = td->io_ops_data;
        struct fio_file *f = io_u->file;
        struct libaio_options *o = td->eo;
+       struct iocb *iocb;
+
+       if (o->useriocb)
+               iocb = &ld->user_iocbs[io_u->index];
+       else
+               iocb = &io_u->iocb;
 
        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);
+               io_prep_pread(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
                if (o->hipri)
-                       io_u->iocb.u.c.flags |= IOCB_FLAG_HIPRI;
+                       iocb->u.c.flags |= IOCB_FLAG_HIPRI;
        } 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);
+               io_prep_pwrite(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
                if (o->hipri)
-                       io_u->iocb.u.c.flags |= IOCB_FLAG_HIPRI;
+                       iocb->u.c.flags |= IOCB_FLAG_HIPRI;
        } else if (ddir_sync(io_u->ddir))
-               io_prep_fsync(&io_u->iocb, f->fd);
+               io_prep_fsync(iocb, f->fd);
 
        return 0;
 }
@@ -104,11 +128,16 @@ static int fio_libaio_prep(struct thread_data fio_unused *td, struct io_u *io_u)
 static struct io_u *fio_libaio_event(struct thread_data *td, int event)
 {
        struct libaio_data *ld = td->io_ops_data;
+       struct libaio_options *o = td->eo;
        struct io_event *ev;
        struct io_u *io_u;
 
        ev = ld->aio_events + event;
-       io_u = container_of(ev->obj, struct io_u, iocb);
+       if (o->useriocb) {
+               int index = (int) (uintptr_t) ev->obj;
+               io_u = ld->io_u_index[index];
+       } else
+               io_u = container_of(ev->obj, struct io_u, iocb);
 
        if (ev->res != io_u->xfer_buflen) {
                if (ev->res > io_u->xfer_buflen)
@@ -204,6 +233,7 @@ 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_options *o = td->eo;
 
        fio_ro_check(td, io_u);
 
@@ -234,7 +264,11 @@ static enum fio_q_status fio_libaio_queue(struct thread_data *td,
                return FIO_Q_COMPLETED;
        }
 
-       ld->iocbs[ld->head] = &io_u->iocb;
+       if (o->useriocb)
+               ld->iocbs[ld->head] = (struct iocb *) (uintptr_t) io_u->index;
+       else
+               ld->iocbs[ld->head] = &io_u->iocb;
+
        ld->io_us[ld->head] = io_u;
        ring_inc(ld, &ld->head, 1);
        ld->queued++;
@@ -353,25 +387,34 @@ static void fio_libaio_cleanup(struct thread_data *td)
                free(ld->aio_events);
                free(ld->iocbs);
                free(ld->io_us);
+               if (ld->user_iocbs)
+                       free(ld->user_iocbs);
                free(ld);
        }
 }
 
 static int fio_libaio_queue_init(struct libaio_data *ld, unsigned int depth,
-                                bool hipri)
+                                bool hipri, bool useriocb)
 {
 #ifdef __NR_sys_io_setup2
        int flags = 0;
 
        if (hipri)
-               flags = IOCTX_FLAG_IOPOLL;
+               flags |= IOCTX_FLAG_IOPOLL;
+       if (useriocb)
+               flags |= IOCTX_FLAG_USERIOCB;
 
-       return syscall(__NR_sys_io_setup2, depth, flags, &ld->aio_ctx);
+       return syscall(__NR_sys_io_setup2, depth, flags, ld->user_iocbs, &ld->aio_ctx);
 #else
        if (hipri) {
                log_err("fio: polled aio not available on your platform\n");
                return 1;
        }
+       if (useriocb) {
+               log_err("fio: user mapped iocbs not available on your platform\n");
+               return 1;
+       }
+
        return io_queue_init(depth, &ld->aio_ctx);
 #endif
 }
@@ -384,15 +427,31 @@ static int fio_libaio_init(struct thread_data *td)
 
        ld = calloc(1, sizeof(*ld));
 
+       if (o->useriocb) {
+               size_t size;
+               void *p;
+
+               ld->io_u_index = calloc(td->o.iodepth, sizeof(struct io_u *));
+               size = td->o.iodepth * sizeof(struct iocb);
+               if (posix_memalign(&p, page_size, size)) {
+                       log_err("fio: libaio iocb allocation failure\n");
+                       free(ld);
+                       return 1;
+               }
+               ld->user_iocbs = p;
+       }
+
        /*
         * 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.
         */
-       err = fio_libaio_queue_init(ld, td->o.iodepth, o->hipri);
+       err = fio_libaio_queue_init(ld, td->o.iodepth, o->hipri, o->useriocb);
        if (err) {
                td_verror(td, -err, "io_queue_init");
                log_err("fio: check /proc/sys/fs/aio-max-nr\n");
+               if (ld->user_iocbs)
+                       free(ld->user_iocbs);
                free(ld);
                return 1;
        }
@@ -407,10 +466,24 @@ static int fio_libaio_init(struct thread_data *td)
        return 0;
 }
 
+static int fio_libaio_io_u_init(struct thread_data *td, struct io_u *io_u)
+{
+       struct libaio_options *o = td->eo;
+
+       if (o->useriocb) {
+               struct libaio_data *ld = td->io_ops_data;
+
+               ld->io_u_index[io_u->index] = io_u;
+       }
+
+       return 0;
+}
+
 static struct ioengine_ops ioengine = {
        .name                   = "libaio",
        .version                = FIO_IOOPS_VERSION,
        .init                   = fio_libaio_init,
+       .io_u_init              = fio_libaio_io_u_init,
        .prep                   = fio_libaio_prep,
        .queue                  = fio_libaio_queue,
        .commit                 = fio_libaio_commit,