cmdprio: Add support for per I/O priority hint
[fio.git] / engines / io_uring.c
index a2533c88e6a2c0bfa819dd782e37888f6f0bc960..e1abf6888fd50c886b87ecc29cf25519c8bbba9d 100644 (file)
 #include "../lib/types.h"
 #include "../os/linux/io_uring.h"
 #include "cmdprio.h"
+#include "zbd.h"
+#include "nvme.h"
+
+#include <sys/stat.h>
+
+enum uring_cmd_type {
+       FIO_URING_CMD_NVME = 1,
+};
 
 struct io_sq_ring {
        unsigned *head;
@@ -70,6 +78,8 @@ struct ioring_data {
        struct ioring_mmap mmap[3];
 
        struct cmdprio cmdprio;
+
+       struct nvme_dsm_range *dsm;
 };
 
 struct ioring_options {
@@ -85,6 +95,7 @@ struct ioring_options {
        unsigned int uncached;
        unsigned int nowait;
        unsigned int force_async;
+       enum uring_cmd_type cmd_type;
 };
 
 static const int ddir_to_op[2][2] = {
@@ -116,87 +127,6 @@ 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_options.percentage[DDIR_READ]),
-               .off2   = offsetof(struct ioring_options,
-                                  cmdprio_options.percentage[DDIR_WRITE]),
-               .minval = 0,
-               .maxval = 100,
-               .help   = "Send high priority I/O this percentage of the time",
-               .category = FIO_OPT_C_ENGINE,
-               .group  = FIO_OPT_G_IOURING,
-       },
-       {
-               .name   = "cmdprio_class",
-               .lname  = "Asynchronous I/O priority class",
-               .type   = FIO_OPT_INT,
-               .off1   = offsetof(struct ioring_options,
-                                  cmdprio_options.class[DDIR_READ]),
-               .off2   = offsetof(struct ioring_options,
-                                  cmdprio_options.class[DDIR_WRITE]),
-               .help   = "Set asynchronous IO priority class",
-               .minval = IOPRIO_MIN_PRIO_CLASS + 1,
-               .maxval = IOPRIO_MAX_PRIO_CLASS,
-               .interval = 1,
-               .category = FIO_OPT_C_ENGINE,
-               .group  = FIO_OPT_G_IOURING,
-       },
-       {
-               .name   = "cmdprio",
-               .lname  = "Asynchronous I/O priority level",
-               .type   = FIO_OPT_INT,
-               .off1   = offsetof(struct ioring_options,
-                                  cmdprio_options.level[DDIR_READ]),
-               .off2   = offsetof(struct ioring_options,
-                                  cmdprio_options.level[DDIR_WRITE]),
-               .help   = "Set asynchronous IO priority level",
-               .minval = IOPRIO_MIN_PRIO,
-               .maxval = IOPRIO_MAX_PRIO,
-               .interval = 1,
-               .category = FIO_OPT_C_ENGINE,
-               .group  = FIO_OPT_G_IOURING,
-       },
-       {
-               .name   = "cmdprio_bssplit",
-               .lname  = "Priority percentage block size split",
-               .type   = FIO_OPT_STR_STORE,
-               .off1   = offsetof(struct ioring_options,
-                                  cmdprio_options.bssplit_str),
-               .help   = "Set priority percentages for different block sizes",
-               .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",
-       },
-       {
-               .name   = "cmdprio_class",
-               .lname  = "Asynchronous I/O priority class",
-               .type   = FIO_OPT_UNSUPPORTED,
-               .help   = "Your platform does not support I/O priority classes",
-       },
-       {
-               .name   = "cmdprio",
-               .lname  = "Asynchronous I/O priority level",
-               .type   = FIO_OPT_UNSUPPORTED,
-               .help   = "Your platform does not support I/O priority classes",
-       },
-       {
-               .name   = "cmdprio_bssplit",
-               .lname  = "Priority percentage block size split",
-               .type   = FIO_OPT_UNSUPPORTED,
-               .help   = "Your platform does not support I/O priority classes",
-       },
-#endif
        {
                .name   = "fixedbufs",
                .lname  = "Fixed (pre-mapped) IO buffers",
@@ -218,7 +148,7 @@ static struct fio_option options[] = {
        {
                .name   = "sqthread_poll",
                .lname  = "Kernel SQ thread polling",
-               .type   = FIO_OPT_INT,
+               .type   = FIO_OPT_STR_SET,
                .off1   = offsetof(struct ioring_options, sqpoll_thread),
                .help   = "Offload submission/completion to kernel thread",
                .category = FIO_OPT_C_ENGINE,
@@ -270,6 +200,23 @@ static struct fio_option options[] = {
                .category = FIO_OPT_C_ENGINE,
                .group  = FIO_OPT_G_IOURING,
        },
+       {
+               .name   = "cmd_type",
+               .lname  = "Uring cmd type",
+               .type   = FIO_OPT_STR,
+               .off1   = offsetof(struct ioring_options, cmd_type),
+               .help   = "Specify uring-cmd type",
+               .def    = "nvme",
+               .posval = {
+                         { .ival = "nvme",
+                           .oval = FIO_URING_CMD_NVME,
+                           .help = "Issue nvme-uring-cmd",
+                         },
+               },
+               .category = FIO_OPT_C_ENGINE,
+               .group  = FIO_OPT_G_IOURING,
+       },
+       CMDPRIO_OPTIONS(struct ioring_options, FIO_OPT_G_IOURING),
        {
                .name   = NULL,
        },
@@ -278,8 +225,13 @@ 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)
 {
+#ifdef FIO_ARCH_HAS_SYSCALL
+       return __do_syscall6(__NR_io_uring_enter, ld->ring_fd, to_submit,
+                               min_complete, flags, NULL, 0);
+#else
        return syscall(__NR_io_uring_enter, ld->ring_fd, to_submit,
                        min_complete, flags, NULL, 0);
+#endif
 }
 
 static int fio_ioring_prep(struct thread_data *td, struct io_u *io_u)
@@ -333,8 +285,8 @@ static int fio_ioring_prep(struct thread_data *td, struct io_u *io_u)
                /*
                 * Since io_uring can have a submission context (sqthread_poll)
                 * that is different from the process context, we cannot rely on
-                * the IO priority set by ioprio_set() (option prio/prioclass)
-                * to be inherited.
+                * the IO priority set by ioprio_set() (options prio, prioclass,
+                * and priohint) to be inherited.
                 * td->ioprio will have the value of the "default prio", so set
                 * this unconditionally. This value might get overridden by
                 * fio_ioring_cmdprio_prep() if the option cmdprio_percentage or
@@ -368,6 +320,56 @@ static int fio_ioring_prep(struct thread_data *td, struct io_u *io_u)
        return 0;
 }
 
+static int fio_ioring_cmd_prep(struct thread_data *td, struct io_u *io_u)
+{
+       struct ioring_data *ld = td->io_ops_data;
+       struct ioring_options *o = td->eo;
+       struct fio_file *f = io_u->file;
+       struct nvme_uring_cmd *cmd;
+       struct io_uring_sqe *sqe;
+
+       /* only supports nvme_uring_cmd */
+       if (o->cmd_type != FIO_URING_CMD_NVME)
+               return -EINVAL;
+
+       if (io_u->ddir == DDIR_TRIM && td->io_ops->flags & FIO_ASYNCIO_SYNC_TRIM)
+               return 0;
+
+       sqe = &ld->sqes[(io_u->index) << 1];
+
+       if (o->registerfiles) {
+               sqe->fd = f->engine_pos;
+               sqe->flags = IOSQE_FIXED_FILE;
+       } else {
+               sqe->fd = f->fd;
+       }
+       sqe->rw_flags = 0;
+       if (!td->o.odirect && o->uncached)
+               sqe->rw_flags |= RWF_UNCACHED;
+       if (o->nowait)
+               sqe->rw_flags |= RWF_NOWAIT;
+
+       sqe->opcode = IORING_OP_URING_CMD;
+       sqe->user_data = (unsigned long) io_u;
+       if (o->nonvectored)
+               sqe->cmd_op = NVME_URING_CMD_IO;
+       else
+               sqe->cmd_op = NVME_URING_CMD_IO_VEC;
+       if (o->force_async && ++ld->prepped == o->force_async) {
+               ld->prepped = 0;
+               sqe->flags |= IOSQE_ASYNC;
+       }
+       if (o->fixedbufs) {
+               sqe->uring_cmd_flags = IORING_URING_CMD_FIXED;
+               sqe->buf_index = io_u->index;
+       }
+
+       cmd = (struct nvme_uring_cmd *)sqe->cmd;
+       return fio_nvme_uring_cmd_prep(cmd, io_u,
+                       o->nonvectored ? NULL : &ld->iovecs[io_u->index],
+                       &ld->dsm[io_u->index]);
+}
+
 static struct io_u *fio_ioring_event(struct thread_data *td, int event)
 {
        struct ioring_data *ld = td->io_ops_data;
@@ -391,6 +393,29 @@ static struct io_u *fio_ioring_event(struct thread_data *td, int event)
        return io_u;
 }
 
+static struct io_u *fio_ioring_cmd_event(struct thread_data *td, int event)
+{
+       struct ioring_data *ld = td->io_ops_data;
+       struct ioring_options *o = td->eo;
+       struct io_uring_cqe *cqe;
+       struct io_u *io_u;
+       unsigned index;
+
+       index = (event + ld->cq_ring_off) & ld->cq_ring_mask;
+       if (o->cmd_type == FIO_URING_CMD_NVME)
+               index <<= 1;
+
+       cqe = &ld->cq_ring.cqes[index];
+       io_u = (struct io_u *) (uintptr_t) cqe->user_data;
+
+       if (cqe->res != 0)
+               io_u->error = -cqe->res;
+       else
+               io_u->error = 0;
+
+       return io_u;
+}
+
 static int fio_ioring_cqring_reap(struct thread_data *td, unsigned int events,
                                   unsigned int max)
 {
@@ -427,6 +452,7 @@ static int fio_ioring_getevents(struct thread_data *td, unsigned int min,
                r = fio_ioring_cqring_reap(td, events, max);
                if (r) {
                        events += r;
+                       max -= r;
                        if (actual_min != 0)
                                actual_min -= r;
                        continue;
@@ -438,6 +464,7 @@ static int fio_ioring_getevents(struct thread_data *td, unsigned int min,
                        if (r < 0) {
                                if (errno == EAGAIN || errno == EINTR)
                                        continue;
+                               r = -errno;
                                td_verror(td, errno, "io_uring_enter");
                                break;
                        }
@@ -469,11 +496,12 @@ static enum fio_q_status fio_ioring_queue(struct thread_data *td,
        if (ld->queued == ld->iodepth)
                return FIO_Q_BUSY;
 
-       if (io_u->ddir == DDIR_TRIM) {
+       if (io_u->ddir == DDIR_TRIM && td->io_ops->flags & FIO_ASYNCIO_SYNC_TRIM) {
                if (ld->queued)
                        return FIO_Q_BUSY;
 
                do_io_u_trim(td, io_u);
+
                io_u_mark_submit(td, 1);
                io_u_mark_complete(td, 1);
                return FIO_Q_COMPLETED;
@@ -514,6 +542,12 @@ static void fio_ioring_queued(struct thread_data *td, int start, int nr)
 
                start++;
        }
+
+       /*
+        * only used for iolog
+        */
+       if (td->o.read_iolog_file)
+               memcpy(&td->last_issue, &now, sizeof(now));
 }
 
 static int fio_ioring_commit(struct thread_data *td)
@@ -532,12 +566,16 @@ static int fio_ioring_commit(struct thread_data *td)
         */
        if (o->sqpoll_thread) {
                struct io_sq_ring *ring = &ld->sq_ring;
+               unsigned start = *ld->sq_ring.tail - ld->queued;
                unsigned flags;
 
                flags = atomic_load_acquire(ring->flags);
                if (flags & IORING_SQ_NEED_WAKEUP)
                        io_uring_enter(ld, ld->queued, 0,
                                        IORING_ENTER_SQ_WAKEUP);
+               fio_ioring_queued(td, start, ld->queued);
+               io_u_mark_submit(td, ld->queued);
+
                ld->queued = 0;
                return 0;
        }
@@ -565,6 +603,7 @@ static int fio_ioring_commit(struct thread_data *td)
                                usleep(1);
                                continue;
                        }
+                       ret = -errno;
                        td_verror(td, errno, "io_uring_enter submit");
                        break;
                }
@@ -594,6 +633,7 @@ static void fio_ioring_cleanup(struct thread_data *td)
                free(ld->io_u_index);
                free(ld->iovecs);
                free(ld->fds);
+               free(ld->dsm);
                free(ld);
        }
 }
@@ -617,14 +657,22 @@ static int fio_ioring_mmap(struct ioring_data *ld, struct io_uring_params *p)
        sring->array = ptr + p->sq_off.array;
        ld->sq_ring_mask = *sring->ring_mask;
 
-       ld->mmap[1].len = p->sq_entries * sizeof(struct io_uring_sqe);
+       if (p->flags & IORING_SETUP_SQE128)
+               ld->mmap[1].len = 2 * p->sq_entries * sizeof(struct io_uring_sqe);
+       else
+               ld->mmap[1].len = p->sq_entries * sizeof(struct io_uring_sqe);
        ld->sqes = mmap(0, ld->mmap[1].len, PROT_READ | PROT_WRITE,
                                MAP_SHARED | MAP_POPULATE, ld->ring_fd,
                                IORING_OFF_SQES);
        ld->mmap[1].ptr = ld->sqes;
 
-       ld->mmap[2].len = p->cq_off.cqes +
-                               p->cq_entries * sizeof(struct io_uring_cqe);
+       if (p->flags & IORING_SETUP_CQE32) {
+               ld->mmap[2].len = p->cq_off.cqes +
+                                       2 * p->cq_entries * sizeof(struct io_uring_cqe);
+       } else {
+               ld->mmap[2].len = p->cq_off.cqes +
+                                       p->cq_entries * sizeof(struct io_uring_cqe);
+       }
        ptr = mmap(0, ld->mmap[2].len, PROT_READ | PROT_WRITE,
                        MAP_SHARED | MAP_POPULATE, ld->ring_fd,
                        IORING_OFF_CQ_RING);
@@ -652,11 +700,10 @@ static void fio_ioring_probe(struct thread_data *td)
        /* default to off, as that's always safe */
        o->nonvectored = 0;
 
-       p = malloc(sizeof(*p) + 256 * sizeof(struct io_uring_probe_op));
+       p = calloc(1, sizeof(*p) + 256 * sizeof(struct io_uring_probe_op));
        if (!p)
                return;
 
-       memset(p, 0, sizeof(*p) + 256 * sizeof(struct io_uring_probe_op));
        ret = syscall(__NR_io_uring_register, ld->ring_fd,
                        IORING_REGISTER_PROBE, p, 256);
        if (ret < 0)
@@ -690,6 +737,98 @@ static int fio_ioring_queue_init(struct thread_data *td)
                        p.flags |= IORING_SETUP_SQ_AFF;
                        p.sq_thread_cpu = o->sqpoll_cpu;
                }
+
+               /*
+                * Submission latency for sqpoll_thread is just the time it
+                * takes to fill in the SQ ring entries, and any syscall if
+                * IORING_SQ_NEED_WAKEUP is set, we don't need to log that time
+                * separately.
+                */
+               td->o.disable_slat = 1;
+       }
+
+       /*
+        * Clamp CQ ring size at our SQ ring size, we don't need more entries
+        * than that.
+        */
+       p.flags |= IORING_SETUP_CQSIZE;
+       p.cq_entries = depth;
+
+       /*
+        * Setup COOP_TASKRUN as we don't need to get IPI interrupted for
+        * completing IO operations.
+        */
+       p.flags |= IORING_SETUP_COOP_TASKRUN;
+
+       /*
+        * io_uring is always a single issuer, and we can defer task_work
+        * runs until we reap events.
+        */
+       p.flags |= IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN;
+
+retry:
+       ret = syscall(__NR_io_uring_setup, depth, &p);
+       if (ret < 0) {
+               if (errno == EINVAL && p.flags & IORING_SETUP_DEFER_TASKRUN) {
+                       p.flags &= ~IORING_SETUP_DEFER_TASKRUN;
+                       p.flags &= ~IORING_SETUP_SINGLE_ISSUER;
+                       goto retry;
+               }
+               if (errno == EINVAL && p.flags & IORING_SETUP_COOP_TASKRUN) {
+                       p.flags &= ~IORING_SETUP_COOP_TASKRUN;
+                       goto retry;
+               }
+               if (errno == EINVAL && p.flags & IORING_SETUP_CQSIZE) {
+                       p.flags &= ~IORING_SETUP_CQSIZE;
+                       goto retry;
+               }
+               return ret;
+       }
+
+       ld->ring_fd = ret;
+
+       fio_ioring_probe(td);
+
+       if (o->fixedbufs) {
+               ret = syscall(__NR_io_uring_register, ld->ring_fd,
+                               IORING_REGISTER_BUFFERS, ld->iovecs, depth);
+               if (ret < 0)
+                       return ret;
+       }
+
+       return fio_ioring_mmap(ld, &p);
+}
+
+static int fio_ioring_cmd_queue_init(struct thread_data *td)
+{
+       struct ioring_data *ld = td->io_ops_data;
+       struct ioring_options *o = td->eo;
+       int depth = td->o.iodepth;
+       struct io_uring_params p;
+       int ret;
+
+       memset(&p, 0, sizeof(p));
+
+       if (o->hipri)
+               p.flags |= IORING_SETUP_IOPOLL;
+       if (o->sqpoll_thread) {
+               p.flags |= IORING_SETUP_SQPOLL;
+               if (o->sqpoll_set) {
+                       p.flags |= IORING_SETUP_SQ_AFF;
+                       p.sq_thread_cpu = o->sqpoll_cpu;
+               }
+
+               /*
+                * Submission latency for sqpoll_thread is just the time it
+                * takes to fill in the SQ ring entries, and any syscall if
+                * IORING_SQ_NEED_WAKEUP is set, we don't need to log that time
+                * separately.
+                */
+               td->o.disable_slat = 1;
+       }
+       if (o->cmd_type == FIO_URING_CMD_NVME) {
+               p.flags |= IORING_SETUP_SQE128;
+               p.flags |= IORING_SETUP_CQE32;
        }
 
        /*
@@ -699,9 +838,30 @@ static int fio_ioring_queue_init(struct thread_data *td)
        p.flags |= IORING_SETUP_CQSIZE;
        p.cq_entries = depth;
 
+       /*
+        * Setup COOP_TASKRUN as we don't need to get IPI interrupted for
+        * completing IO operations.
+        */
+       p.flags |= IORING_SETUP_COOP_TASKRUN;
+
+       /*
+        * io_uring is always a single issuer, and we can defer task_work
+        * runs until we reap events.
+        */
+       p.flags |= IORING_SETUP_SINGLE_ISSUER | IORING_SETUP_DEFER_TASKRUN;
+
 retry:
        ret = syscall(__NR_io_uring_setup, depth, &p);
        if (ret < 0) {
+               if (errno == EINVAL && p.flags & IORING_SETUP_DEFER_TASKRUN) {
+                       p.flags &= ~IORING_SETUP_DEFER_TASKRUN;
+                       p.flags &= ~IORING_SETUP_SINGLE_ISSUER;
+                       goto retry;
+               }
+               if (errno == EINVAL && p.flags & IORING_SETUP_COOP_TASKRUN) {
+                       p.flags &= ~IORING_SETUP_COOP_TASKRUN;
+                       goto retry;
+               }
                if (errno == EINVAL && p.flags & IORING_SETUP_CQSIZE) {
                        p.flags &= ~IORING_SETUP_CQSIZE;
                        goto retry;
@@ -806,6 +966,52 @@ static int fio_ioring_post_init(struct thread_data *td)
        return 0;
 }
 
+static int fio_ioring_cmd_post_init(struct thread_data *td)
+{
+       struct ioring_data *ld = td->io_ops_data;
+       struct ioring_options *o = td->eo;
+       struct io_u *io_u;
+       int err, i;
+
+       for (i = 0; i < td->o.iodepth; i++) {
+               struct iovec *iov = &ld->iovecs[i];
+
+               io_u = ld->io_u_index[i];
+               iov->iov_base = io_u->buf;
+               iov->iov_len = td_max_bs(td);
+       }
+
+       err = fio_ioring_cmd_queue_init(td);
+       if (err) {
+               int init_err = errno;
+
+               td_verror(td, init_err, "io_queue_init");
+               return 1;
+       }
+
+       for (i = 0; i < td->o.iodepth; i++) {
+               struct io_uring_sqe *sqe;
+
+               if (o->cmd_type == FIO_URING_CMD_NVME) {
+                       sqe = &ld->sqes[i << 1];
+                       memset(sqe, 0, 2 * sizeof(*sqe));
+               } else {
+                       sqe = &ld->sqes[i];
+                       memset(sqe, 0, sizeof(*sqe));
+               }
+       }
+
+       if (o->registerfiles) {
+               err = fio_ioring_register_files(td);
+               if (err) {
+                       td_verror(td, errno, "ioring_register_files");
+                       return 1;
+               }
+       }
+
+       return 0;
+}
+
 static int fio_ioring_init(struct thread_data *td)
 {
        struct ioring_options *o = td->eo;
@@ -840,6 +1046,16 @@ static int fio_ioring_init(struct thread_data *td)
                return 1;
        }
 
+       /*
+        * For io_uring_cmd, trims are async operations unless we are operating
+        * in zbd mode where trim means zone reset.
+        */
+       if (!strcmp(td->io_ops->name, "io_uring_cmd") && td_trim(td) &&
+           td->o.zone_mode == ZONE_MODE_ZBD)
+               td->io_ops->flags |= FIO_ASYNCIO_SYNC_TRIM;
+       else
+               ld->dsm = calloc(ld->iodepth, sizeof(*ld->dsm));
+
        return 0;
 }
 
@@ -863,6 +1079,57 @@ static int fio_ioring_open_file(struct thread_data *td, struct fio_file *f)
        return 0;
 }
 
+static int fio_ioring_cmd_open_file(struct thread_data *td, struct fio_file *f)
+{
+       struct ioring_data *ld = td->io_ops_data;
+       struct ioring_options *o = td->eo;
+
+       if (o->cmd_type == FIO_URING_CMD_NVME) {
+               struct nvme_data *data = NULL;
+               unsigned int nsid, lba_size = 0;
+               __u32 ms = 0;
+               __u64 nlba = 0;
+               int ret;
+
+               /* Store the namespace-id and lba size. */
+               data = FILE_ENG_DATA(f);
+               if (data == NULL) {
+                       ret = fio_nvme_get_info(f, &nsid, &lba_size, &ms, &nlba);
+                       if (ret)
+                               return ret;
+
+                       data = calloc(1, sizeof(struct nvme_data));
+                       data->nsid = nsid;
+                       if (ms)
+                               data->lba_ext = lba_size + ms;
+                       else
+                               data->lba_shift = ilog2(lba_size);
+
+                       FILE_SET_ENG_DATA(f, data);
+               }
+
+               assert(data->lba_shift < 32);
+               lba_size = data->lba_ext ? data->lba_ext : (1U << data->lba_shift);
+
+               for_each_rw_ddir(ddir) {
+                       if (td->o.min_bs[ddir] % lba_size ||
+                               td->o.max_bs[ddir] % lba_size) {
+                               if (data->lba_ext)
+                                       log_err("block size must be a multiple of "
+                                               "(LBA data size + Metadata size)\n");
+                               else
+                                       log_err("block size must be a multiple of LBA data size\n");
+                               return 1;
+                       }
+                }
+       }
+       if (!ld || !o->registerfiles)
+               return generic_open_file(td, f);
+
+       f->fd = ld->fds[f->engine_pos];
+       return 0;
+}
+
 static int fio_ioring_close_file(struct thread_data *td, struct fio_file *f)
 {
        struct ioring_data *ld = td->io_ops_data;
@@ -875,10 +1142,116 @@ static int fio_ioring_close_file(struct thread_data *td, struct fio_file *f)
        return 0;
 }
 
-static struct ioengine_ops ioengine = {
+static int fio_ioring_cmd_close_file(struct thread_data *td,
+                                    struct fio_file *f)
+{
+       struct ioring_data *ld = td->io_ops_data;
+       struct ioring_options *o = td->eo;
+
+       if (o->cmd_type == FIO_URING_CMD_NVME) {
+               struct nvme_data *data = FILE_ENG_DATA(f);
+
+               FILE_SET_ENG_DATA(f, NULL);
+               free(data);
+       }
+       if (!ld || !o->registerfiles)
+               return generic_close_file(td, f);
+
+       f->fd = -1;
+       return 0;
+}
+
+static int fio_ioring_cmd_get_file_size(struct thread_data *td,
+                                       struct fio_file *f)
+{
+       struct ioring_options *o = td->eo;
+
+       if (fio_file_size_known(f))
+               return 0;
+
+       if (o->cmd_type == FIO_URING_CMD_NVME) {
+               struct nvme_data *data = NULL;
+               unsigned int nsid, lba_size = 0;
+               __u32 ms = 0;
+               __u64 nlba = 0;
+               int ret;
+
+               ret = fio_nvme_get_info(f, &nsid, &lba_size, &ms, &nlba);
+               if (ret)
+                       return ret;
+
+               data = calloc(1, sizeof(struct nvme_data));
+               data->nsid = nsid;
+               if (ms)
+                       data->lba_ext = lba_size + ms;
+               else
+                       data->lba_shift = ilog2(lba_size);
+
+               f->real_file_size = lba_size * nlba;
+               fio_file_set_size_known(f);
+
+               FILE_SET_ENG_DATA(f, data);
+               return 0;
+       }
+       return generic_get_file_size(td, f);
+}
+
+static int fio_ioring_cmd_get_zoned_model(struct thread_data *td,
+                                         struct fio_file *f,
+                                         enum zbd_zoned_model *model)
+{
+       return fio_nvme_get_zoned_model(td, f, model);
+}
+
+static int fio_ioring_cmd_report_zones(struct thread_data *td,
+                                      struct fio_file *f, uint64_t offset,
+                                      struct zbd_zone *zbdz,
+                                      unsigned int nr_zones)
+{
+       return fio_nvme_report_zones(td, f, offset, zbdz, nr_zones);
+}
+
+static int fio_ioring_cmd_reset_wp(struct thread_data *td, struct fio_file *f,
+                                  uint64_t offset, uint64_t length)
+{
+       return fio_nvme_reset_wp(td, f, offset, length);
+}
+
+static int fio_ioring_cmd_get_max_open_zones(struct thread_data *td,
+                                            struct fio_file *f,
+                                            unsigned int *max_open_zones)
+{
+       return fio_nvme_get_max_open_zones(td, f, max_open_zones);
+}
+
+static int fio_ioring_cmd_fetch_ruhs(struct thread_data *td, struct fio_file *f,
+                                    struct fio_ruhs_info *fruhs_info)
+{
+       struct nvme_fdp_ruh_status *ruhs;
+       int bytes, ret, i;
+
+       bytes = sizeof(*ruhs) + FDP_MAX_RUHS * sizeof(struct nvme_fdp_ruh_status_desc);
+       ruhs = scalloc(1, bytes);
+       if (!ruhs)
+               return -ENOMEM;
+
+       ret = fio_nvme_iomgmt_ruhs(td, f, ruhs, bytes);
+       if (ret)
+               goto free;
+
+       fruhs_info->nr_ruhs = le16_to_cpu(ruhs->nruhsd);
+       for (i = 0; i < fruhs_info->nr_ruhs; i++)
+               fruhs_info->plis[i] = le16_to_cpu(ruhs->ruhss[i].pid);
+free:
+       sfree(ruhs);
+       return ret;
+}
+
+static struct ioengine_ops ioengine_uring = {
        .name                   = "io_uring",
        .version                = FIO_IOOPS_VERSION,
-       .flags                  = FIO_ASYNCIO_SYNC_TRIM | FIO_NO_OFFLOAD,
+       .flags                  = FIO_ASYNCIO_SYNC_TRIM | FIO_NO_OFFLOAD |
+                                       FIO_ASYNCIO_SETS_ISSUE_TIME,
        .init                   = fio_ioring_init,
        .post_init              = fio_ioring_post_init,
        .io_u_init              = fio_ioring_io_u_init,
@@ -895,13 +1268,41 @@ static struct ioengine_ops ioengine = {
        .option_struct_size     = sizeof(struct ioring_options),
 };
 
+static struct ioengine_ops ioengine_uring_cmd = {
+       .name                   = "io_uring_cmd",
+       .version                = FIO_IOOPS_VERSION,
+       .flags                  = FIO_NO_OFFLOAD | FIO_MEMALIGN | FIO_RAWIO |
+                                       FIO_ASYNCIO_SETS_ISSUE_TIME,
+       .init                   = fio_ioring_init,
+       .post_init              = fio_ioring_cmd_post_init,
+       .io_u_init              = fio_ioring_io_u_init,
+       .prep                   = fio_ioring_cmd_prep,
+       .queue                  = fio_ioring_queue,
+       .commit                 = fio_ioring_commit,
+       .getevents              = fio_ioring_getevents,
+       .event                  = fio_ioring_cmd_event,
+       .cleanup                = fio_ioring_cleanup,
+       .open_file              = fio_ioring_cmd_open_file,
+       .close_file             = fio_ioring_cmd_close_file,
+       .get_file_size          = fio_ioring_cmd_get_file_size,
+       .get_zoned_model        = fio_ioring_cmd_get_zoned_model,
+       .report_zones           = fio_ioring_cmd_report_zones,
+       .reset_wp               = fio_ioring_cmd_reset_wp,
+       .get_max_open_zones     = fio_ioring_cmd_get_max_open_zones,
+       .options                = options,
+       .option_struct_size     = sizeof(struct ioring_options),
+       .fdp_fetch_ruhs         = fio_ioring_cmd_fetch_ruhs,
+};
+
 static void fio_init fio_ioring_register(void)
 {
-       register_ioengine(&ioengine);
+       register_ioengine(&ioengine_uring);
+       register_ioengine(&ioengine_uring_cmd);
 }
 
 static void fio_exit fio_ioring_unregister(void)
 {
-       unregister_ioengine(&ioengine);
+       unregister_ioengine(&ioengine_uring);
+       unregister_ioengine(&ioengine_uring_cmd);
 }
 #endif