Add roundup_pow2() as a generic helper
[fio.git] / engines / io_uring.c
index 329f2f071dcd5c1c61d89de2c298d7c651c33f21..0ccd23184fab3b17a9c68c93a9c30375b5673986 100644 (file)
@@ -17,6 +17,7 @@
 #include "../optgroup.h"
 #include "../lib/memalign.h"
 #include "../lib/fls.h"
+#include "../lib/roundup.h"
 
 #ifdef ARCH_HAVE_IOURING
 
@@ -63,6 +64,8 @@ struct ioring_data {
        int queued;
        int cq_ring_off;
        unsigned iodepth;
+       bool ioprio_class_set;
+       bool ioprio_set;
 
        struct ioring_mmap mmap[3];
 };
@@ -70,6 +73,7 @@ struct ioring_data {
 struct ioring_options {
        void *pad;
        unsigned int hipri;
+       unsigned int cmdprio_percentage;
        unsigned int fixedbufs;
        unsigned int registerfiles;
        unsigned int sqpoll_thread;
@@ -77,6 +81,7 @@ struct ioring_options {
        unsigned int sqpoll_cpu;
        unsigned int nonvectored;
        unsigned int uncached;
+       unsigned int nowait;
 };
 
 static const int ddir_to_op[2][2] = {
@@ -108,6 +113,26 @@ static struct fio_option options[] = {
                .category = FIO_OPT_C_ENGINE,
                .group  = FIO_OPT_G_IOURING,
        },
+#ifdef FIO_HAVE_IOPRIO_CLASS
+       {
+               .name   = "cmdprio_percentage",
+               .lname  = "high priority percentage",
+               .type   = FIO_OPT_INT,
+               .off1   = offsetof(struct ioring_options, cmdprio_percentage),
+               .minval = 1,
+               .maxval = 100,
+               .help   = "Send high priority I/O this percentage of the time",
+               .category = FIO_OPT_C_ENGINE,
+               .group  = FIO_OPT_G_IOURING,
+       },
+#else
+       {
+               .name   = "cmdprio_percentage",
+               .lname  = "high priority percentage",
+               .type   = FIO_OPT_UNSUPPORTED,
+               .help   = "Your platform does not support I/O priority classes",
+       },
+#endif
        {
                .name   = "fixedbufs",
                .lname  = "Fixed (pre-mapped) IO buffers",
@@ -162,6 +187,15 @@ static struct fio_option options[] = {
                .category = FIO_OPT_C_ENGINE,
                .group  = FIO_OPT_G_IOURING,
        },
+       {
+               .name   = "nowait",
+               .lname  = "RWF_NOWAIT",
+               .type   = FIO_OPT_BOOL,
+               .off1   = offsetof(struct ioring_options, nowait),
+               .help   = "Use RWF_NOWAIT for reads/writes",
+               .category = FIO_OPT_C_ENGINE,
+               .group  = FIO_OPT_G_IOURING,
+       },
        {
                .name   = NULL,
        },
@@ -170,7 +204,7 @@ static struct fio_option options[] = {
 static int io_uring_enter(struct ioring_data *ld, unsigned int to_submit,
                         unsigned int min_complete, unsigned int flags)
 {
-       return syscall(__NR_sys_io_uring_enter, ld->ring_fd, to_submit,
+       return syscall(__NR_io_uring_enter, ld->ring_fd, to_submit,
                        min_complete, flags, NULL, 0);
 }
 
@@ -212,9 +246,11 @@ static int fio_ioring_prep(struct thread_data *td, struct io_u *io_u)
                }
                if (!td->o.odirect && o->uncached)
                        sqe->rw_flags = RWF_UNCACHED;
-               if (fio_option_is_set(&td->o, ioprio_class))
+               if (o->nowait)
+                       sqe->rw_flags |= RWF_NOWAIT;
+               if (ld->ioprio_class_set)
                        sqe->ioprio = td->o.ioprio_class << 13;
-               if (fio_option_is_set(&td->o, ioprio))
+               if (ld->ioprio_set)
                        sqe->ioprio |= td->o.ioprio;
                sqe->off = io_u->offset;
        } else if (ddir_sync(io_u->ddir)) {
@@ -266,15 +302,15 @@ static int fio_ioring_cqring_reap(struct thread_data *td, unsigned int events,
 
        head = *ring->head;
        do {
-               read_barrier();
-               if (head == *ring->tail)
+               if (head == atomic_load_acquire(ring->tail))
                        break;
                reaped++;
                head++;
        } while (reaped + events < max);
 
-       *ring->head = head;
-       write_barrier();
+       if (reaped)
+               atomic_store_release(ring->head, head);
+
        return reaped;
 }
 
@@ -313,11 +349,23 @@ static int fio_ioring_getevents(struct thread_data *td, unsigned int min,
        return r < 0 ? r : events;
 }
 
+static void fio_ioring_prio_prep(struct thread_data *td, struct io_u *io_u)
+{
+       struct ioring_options *o = td->eo;
+       struct ioring_data *ld = td->io_ops_data;
+       if (rand_between(&td->prio_state, 0, 99) < o->cmdprio_percentage) {
+               ld->sqes[io_u->index].ioprio = IOPRIO_CLASS_RT << IOPRIO_CLASS_SHIFT;
+               io_u->flags |= IO_U_F_PRIORITY;
+       }
+       return;
+}
+
 static enum fio_q_status fio_ioring_queue(struct thread_data *td,
                                          struct io_u *io_u)
 {
        struct ioring_data *ld = td->io_ops_data;
        struct io_sq_ring *ring = &ld->sq_ring;
+       struct ioring_options *o = td->eo;
        unsigned tail, next_tail;
 
        fio_ro_check(td, io_u);
@@ -337,15 +385,13 @@ static enum fio_q_status fio_ioring_queue(struct thread_data *td,
 
        tail = *ring->tail;
        next_tail = tail + 1;
-       read_barrier();
-       if (next_tail == *ring->head)
+       if (next_tail == atomic_load_acquire(ring->head))
                return FIO_Q_BUSY;
 
-       /* ensure sqe stores are ordered with tail update */
-       write_barrier();
+       if (o->cmdprio_percentage)
+               fio_ioring_prio_prep(td, io_u);
        ring->array[tail & ld->sq_ring_mask] = io_u->index;
-       *ring->tail = next_tail;
-       write_barrier();
+       atomic_store_release(ring->tail, next_tail);
 
        ld->queued++;
        return FIO_Q_QUEUED;
@@ -513,7 +559,7 @@ static int fio_ioring_queue_init(struct thread_data *td)
                }
        }
 
-       ret = syscall(__NR_sys_io_uring_setup, depth, &p);
+       ret = syscall(__NR_io_uring_setup, depth, &p);
        if (ret < 0)
                return ret;
 
@@ -528,7 +574,7 @@ static int fio_ioring_queue_init(struct thread_data *td)
                if (setrlimit(RLIMIT_MEMLOCK, &rlim) < 0)
                        return -1;
 
-               ret = syscall(__NR_sys_io_uring_register, ld->ring_fd,
+               ret = syscall(__NR_io_uring_register, ld->ring_fd,
                                IORING_REGISTER_BUFFERS, ld->iovecs, depth);
                if (ret < 0)
                        return ret;
@@ -554,7 +600,7 @@ static int fio_ioring_register_files(struct thread_data *td)
                f->engine_pos = i;
        }
 
-       ret = syscall(__NR_sys_io_uring_register, ld->ring_fd,
+       ret = syscall(__NR_io_uring_register, ld->ring_fd,
                        IORING_REGISTER_FILES, ld->fds, td->o.nr_files);
        if (ret) {
 err:
@@ -609,15 +655,11 @@ static int fio_ioring_post_init(struct thread_data *td)
        return 0;
 }
 
-static unsigned roundup_pow2(unsigned depth)
-{
-       return 1UL << __fls(depth - 1);
-}
-
 static int fio_ioring_init(struct thread_data *td)
 {
        struct ioring_options *o = td->eo;
        struct ioring_data *ld;
+       struct thread_options *to = &td->o;
 
        /* sqthread submission requires registered files */
        if (o->sqpoll_thread)
@@ -640,6 +682,23 @@ static int fio_ioring_init(struct thread_data *td)
        ld->iovecs = calloc(td->o.iodepth, sizeof(struct iovec));
 
        td->io_ops_data = ld;
+
+       /*
+        * Check for option conflicts
+        */
+       if ((fio_option_is_set(to, ioprio) || fio_option_is_set(to, ioprio_class)) &&
+                       o->cmdprio_percentage != 0) {
+               log_err("%s: cmdprio_percentage option and mutually exclusive "
+                               "prio or prioclass option is set, exiting\n", to->name);
+               td_verror(td, EINVAL, "fio_io_uring_init");
+               return 1;
+       }
+
+       if (fio_option_is_set(&td->o, ioprio_class))
+               ld->ioprio_class_set = true;
+       if (fio_option_is_set(&td->o, ioprio))
+               ld->ioprio_set = true;
+
        return 0;
 }