engines/io_uring: add new I/O engine for uring passthrough support
authorAnuj Gupta <anuj20.g@samsung.com>
Tue, 31 May 2022 13:31:51 +0000 (19:01 +0530)
committerJens Axboe <axboe@kernel.dk>
Thu, 2 Jun 2022 08:19:41 +0000 (02:19 -0600)
Add a new I/O engine (io_uring_cmd) for sending uring passthrough
commands. It will also use most of the existing helpers from the
I/O engine io_uring. The I/O preparation, completion, file open,
file close and post init paths are going to differ and hence
io_uring_cmd will have its own helper for them.

Add a new io_uring_cmd engine specific option to support nvme
passthrough commands. Filename name for this specific option
must specify nvme-ns generic character device (dev/ngXnY).
This provides io_uring_cmd I/O engine a bandwidth to support
various passthrough commands in future.

The engine_pos and engine_data fields in struct fio_file are
separated now. This will help I/O engine io_uring_cmd to store
specific data as well as keep track of register files.

Added a new option cmd_type. This specifies the type of uring
command to submit. Currently it only supports nvme uring command

Signed-off-by: Anuj Gupta <anuj20.g@samsung.com>
Co-authored-by: Ankit Kumar <ankit.kumar@samsung.com>
Link: https://lore.kernel.org/r/20220531133155.17493-6-ankit.kumar@samsung.com
Signed-off-by: Jens Axboe <axboe@kernel.dk>
engines/io_uring.c
file.h
os/linux/io_uring.h

index 1e15647ede45210a5090fca2eb10e923261f6dd5..a7b7b1663c70cd7e661c6f59040915dd589b05ac 100644 (file)
 #include "../lib/types.h"
 #include "../os/linux/io_uring.h"
 #include "cmdprio.h"
+#include "nvme.h"
+
+#include <sys/stat.h>
+
+enum uring_cmd_type {
+       FIO_URING_CMD_NVME = 1,
+};
 
 struct io_sq_ring {
        unsigned *head;
@@ -85,6 +92,7 @@ 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] = {
@@ -270,6 +278,22 @@ 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,
+       },
        {
                .name   = NULL,
        },
@@ -373,6 +397,52 @@ 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 io_uring_sqe *sqe;
+       int ret;
+
+       /* nvme_uring_cmd case */
+       if (o->cmd_type == FIO_URING_CMD_NVME) {
+               struct nvme_uring_cmd *cmd;
+
+               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;
+               }
+
+               cmd = (struct nvme_uring_cmd *)sqe->cmd;
+               ret = fio_nvme_uring_cmd_prep(cmd, io_u,
+                               o->nonvectored ? NULL : &ld->iovecs[io_u->index]);
+
+               return ret;
+       }
+       return -EINVAL;
+}
+
 static struct io_u *fio_ioring_event(struct thread_data *td, int event)
 {
        struct ioring_data *ld = td->io_ops_data;
@@ -396,6 +466,29 @@ 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)
 {
@@ -622,14 +715,22 @@ 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);
@@ -728,6 +829,61 @@ retry:
        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;
+               }
+       }
+       if (o->cmd_type == FIO_URING_CMD_NVME) {
+               p.flags |= IORING_SETUP_SQE128;
+               p.flags |= IORING_SETUP_CQE32;
+       }
+
+       /*
+        * 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;
+
+retry:
+       ret = syscall(__NR_io_uring_setup, depth, &p);
+       if (ret < 0) {
+               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_register_files(struct thread_data *td)
 {
        struct ioring_data *ld = td->io_ops_data;
@@ -811,6 +967,52 @@ 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;
@@ -868,6 +1070,38 @@ 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;
+               unsigned long long 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, &nlba);
+                       if (ret)
+                               return ret;
+
+                       data = calloc(1, sizeof(struct nvme_data));
+                       data->nsid = nsid;
+                       data->lba_shift = ilog2(lba_size);
+
+                       FILE_SET_ENG_DATA(f, data);
+               }
+       }
+       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;
@@ -880,7 +1114,57 @@ 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;
+               unsigned long long nlba = 0;
+               int ret;
+
+               ret = fio_nvme_get_info(f, &nsid, &lba_size, &nlba);
+               if (ret)
+                       return ret;
+
+               data = calloc(1, sizeof(struct nvme_data));
+               data->nsid = nsid;
+               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 struct ioengine_ops ioengine_uring = {
        .name                   = "io_uring",
        .version                = FIO_IOOPS_VERSION,
        .flags                  = FIO_ASYNCIO_SYNC_TRIM | FIO_NO_OFFLOAD,
@@ -900,13 +1184,35 @@ 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_ASYNCIO_SYNC_TRIM | FIO_NO_OFFLOAD | FIO_MEMALIGN | FIO_RAWIO,
+       .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,
+       .options                = options,
+       .option_struct_size     = sizeof(struct ioring_options),
+};
+
 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
diff --git a/file.h b/file.h
index faf65a2a013846fb716f0b80403d6549b90120d2..da1b894706a64227af9a484d760fadc79edcd0d5 100644 (file)
--- a/file.h
+++ b/file.h
@@ -126,12 +126,14 @@ struct fio_file {
        unsigned int last_write_idx;
 
        /*
-        * For use by the io engine for offset or private data storage
+        * For use by the io engine to store offset
         */
-       union {
-               uint64_t engine_pos;
-               void *engine_data;
-       };
+       uint64_t engine_pos;
+
+       /*
+        * For use by the io engine for private data storage
+        */
+       void *engine_data;
 
        /*
         * if io is protected by a semaphore, this is set
index 2fa661350b63ecabad3f91ce191ea616e398695a..929997f8277d9b09f8555d90785074e9da534bc7 100644 (file)
@@ -22,6 +22,7 @@ struct io_uring_sqe {
        union {
                __u64   off;    /* offset into file */
                __u64   addr2;
+               __u32   cmd_op;
        };
        union {
                __u64   addr;   /* pointer to buffer or iovecs */
@@ -171,6 +172,14 @@ enum {
        IORING_OP_MKDIRAT,
        IORING_OP_SYMLINKAT,
        IORING_OP_LINKAT,
+       IORING_OP_MSG_RING,
+       IORING_OP_FSETXATTR,
+       IORING_OP_SETXATTR,
+       IORING_OP_FGETXATTR,
+       IORING_OP_GETXATTR,
+       IORING_OP_SOCKET,
+       IORING_OP_URING_CMD,
+
 
        /* this goes last, obviously */
        IORING_OP_LAST,