Add aioring engine
authorJens Axboe <axboe@kernel.dk>
Thu, 13 Dec 2018 02:48:15 +0000 (19:48 -0700)
committerJens Axboe <axboe@kernel.dk>
Thu, 13 Dec 2018 02:48:15 +0000 (19:48 -0700)
This is a new Linux aio engine, built on top of the new aio
interfaces. It supports polled aio, regular aio, and buffered
aio.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
Makefile
arch/arch-x86_64.h
engines/aioring.c [new file with mode: 0644]
options.c

index 284621d..f111ae6 100644 (file)
--- a/Makefile
+++ b/Makefile
@@ -68,6 +68,9 @@ endif
 ifdef CONFIG_LIBAIO
   SOURCE += engines/libaio.c
 endif
+ifdef CONFIG_LIBAIO
+  SOURCE += engines/aioring.c
+endif
 ifdef CONFIG_RDMA
   SOURCE += engines/rdma.c
 endif
index ac670d0..c69cbef 100644 (file)
@@ -4,6 +4,9 @@
 #ifndef __NR_sys_io_setup2
 #define __NR_sys_io_setup2     335
 #endif
+#ifndef __NR_sys_io_ring_enter
+#define __NR_sys_io_ring_enter 336
+#endif
 
 static inline void do_cpuid(unsigned int *eax, unsigned int *ebx,
                            unsigned int *ecx, unsigned int *edx)
diff --git a/engines/aioring.c b/engines/aioring.c
new file mode 100644 (file)
index 0000000..c15f4a4
--- /dev/null
@@ -0,0 +1,554 @@
+/*
+ * aioring engine
+ *
+ * IO engine using the new native Linux libaio ring interface
+ *
+ */
+#include <stdlib.h>
+#include <unistd.h>
+#include <errno.h>
+#include <libaio.h>
+#include <sys/time.h>
+#include <sys/resource.h>
+
+#include "../fio.h"
+#include "../lib/pow2.h"
+#include "../optgroup.h"
+#include "../lib/memalign.h"
+
+#ifndef IOCB_FLAG_HIPRI
+#define IOCB_FLAG_HIPRI        (1 << 2)
+#endif
+
+/*
+ * io_setup2(2) flags
+ */
+#ifndef IOCTX_FLAG_IOPOLL
+#define IOCTX_FLAG_IOPOLL      (1 << 0)
+#endif
+#ifndef IOCTX_FLAG_SCQRING
+#define IOCTX_FLAG_SCQRING     (1 << 1)
+#endif
+#ifndef IOCTX_FLAG_FIXEDBUFS
+#define IOCTX_FLAG_FIXEDBUFS   (1 << 2)
+#endif
+#ifndef IOCTX_FLAG_SQTHREAD
+#define IOCTX_FLAG_SQTHREAD    (1 << 3)
+#endif
+#ifndef IOCTX_FLAG_SQWQ
+#define IOCTX_FLAG_SQWQ                (1 << 4)
+#endif
+
+/*
+ * io_ring_enter(2) flags
+ */
+#ifndef IORING_FLAG_SUBMIT
+#define IORING_FLAG_SUBMIT     (1 << 0)
+#endif
+#ifndef IORING_FLAG_GETEVENTS
+#define IORING_FLAG_GETEVENTS  (1 << 1)
+#endif
+
+typedef uint64_t u64;
+typedef uint32_t u32;
+typedef uint16_t u16;
+
+struct aio_sq_ring {
+       union {
+               struct {
+                       u32 head;
+                       u32 tail;
+                       u32 nr_events;
+                       u16 sq_thread_cpu;
+                       u64 iocbs;
+               };
+               u32 pad[16];
+       };
+       u32 array[0];
+};
+
+struct aio_cq_ring {
+       union {
+               struct {
+                       u32 head;
+                       u32 tail;
+                       u32 nr_events;
+               };
+               struct io_event pad;
+       };
+       struct io_event events[0];
+};
+
+struct aioring_data {
+       io_context_t aio_ctx;
+       struct io_u **io_us;
+       struct io_u **io_u_index;
+
+       struct aio_sq_ring *sq_ring;
+       struct iocb *iocbs;
+
+       struct aio_cq_ring *cq_ring;
+       struct io_event *events;
+
+       int queued;
+       int cq_ring_off;
+};
+
+struct aioring_options {
+       void *pad;
+       unsigned int hipri;
+       unsigned int fixedbufs;
+};
+
+static struct fio_option options[] = {
+       {
+               .name   = "hipri",
+               .lname  = "High Priority",
+               .type   = FIO_OPT_STR_SET,
+               .off1   = offsetof(struct aioring_options, hipri),
+               .help   = "Use polled IO completions",
+               .category = FIO_OPT_C_ENGINE,
+               .group  = FIO_OPT_G_LIBAIO,
+       },
+       {
+               .name   = "fixedbufs",
+               .lname  = "Fixed (pre-mapped) IO buffers",
+               .type   = FIO_OPT_STR_SET,
+               .off1   = offsetof(struct aioring_options, fixedbufs),
+               .help   = "Pre map IO buffers",
+               .category = FIO_OPT_C_ENGINE,
+               .group  = FIO_OPT_G_LIBAIO,
+       },
+       {
+               .name   = NULL,
+       },
+};
+
+static int fio_aioring_commit(struct thread_data *td);
+
+static int io_ring_enter(io_context_t ctx, unsigned int to_submit,
+                        unsigned int min_complete, unsigned int flags)
+{
+#ifdef __NR_sys_io_ring_enter
+       return syscall(__NR_sys_io_ring_enter, ctx, to_submit, min_complete,
+                       flags);
+#else
+       return -1;
+#endif
+}
+
+static int fio_aioring_prep(struct thread_data *td, struct io_u *io_u)
+{
+       struct aioring_data *ld = td->io_ops_data;
+       struct fio_file *f = io_u->file;
+       struct aioring_options *o = td->eo;
+       struct iocb *iocb;
+
+       iocb = &ld->iocbs[io_u->index];
+
+       if (io_u->ddir == DDIR_READ) {
+               if (o->fixedbufs) {
+                       iocb->aio_fildes = f->fd;
+                       iocb->aio_lio_opcode = IO_CMD_PREAD;
+                       iocb->u.c.offset = io_u->offset;
+               } else {
+                       io_prep_pread(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
+                       if (o->hipri)
+                               iocb->u.c.flags |= IOCB_FLAG_HIPRI;
+               }
+       } else if (io_u->ddir == DDIR_WRITE) {
+               if (o->fixedbufs) {
+                       iocb->aio_fildes = f->fd;
+                       iocb->aio_lio_opcode = IO_CMD_PWRITE;
+                       iocb->u.c.offset = io_u->offset;
+               } else {
+                       io_prep_pwrite(iocb, f->fd, io_u->xfer_buf, io_u->xfer_buflen, io_u->offset);
+                       if (o->hipri)
+                               iocb->u.c.flags |= IOCB_FLAG_HIPRI;
+               }
+       } else if (ddir_sync(io_u->ddir))
+               io_prep_fsync(iocb, f->fd);
+
+       iocb->data = io_u;
+       return 0;
+}
+
+static struct io_u *fio_aioring_event(struct thread_data *td, int event)
+{
+       struct aioring_data *ld = td->io_ops_data;
+       struct io_event *ev;
+       struct io_u *io_u;
+       int index;
+
+       index = event + ld->cq_ring_off;
+       if (index >= ld->cq_ring->nr_events)
+               index -= ld->cq_ring->nr_events;
+
+       ev = &ld->cq_ring->events[index];
+       io_u = ev->data;
+
+       if (ev->res != io_u->xfer_buflen) {
+               if (ev->res > io_u->xfer_buflen)
+                       io_u->error = -ev->res;
+               else
+                       io_u->resid = io_u->xfer_buflen - ev->res;
+       } else
+               io_u->error = 0;
+
+       return io_u;
+}
+
+static int fio_aioring_cqring_reap(struct thread_data *td, unsigned int events,
+                                  unsigned int max)
+{
+       struct aioring_data *ld = td->io_ops_data;
+       struct aio_cq_ring *ring = ld->cq_ring;
+       u32 head, reaped = 0;
+
+       head = ring->head;
+       do {
+               read_barrier();
+               if (head == ring->tail)
+                       break;
+               reaped++;
+               head++;
+               if (head == ring->nr_events)
+                       head = 0;
+       } while (reaped + events < max);
+
+       ring->head = head;
+       write_barrier();
+       return reaped;
+}
+
+static int fio_aioring_getevents(struct thread_data *td, unsigned int min,
+                                unsigned int max, const struct timespec *t)
+{
+       struct aioring_data *ld = td->io_ops_data;
+       unsigned actual_min = td->o.iodepth_batch_complete_min == 0 ? 0 : min;
+       struct aio_cq_ring *ring = ld->cq_ring;
+       int r, events = 0;
+
+       ld->cq_ring_off = ring->head;
+       do {
+               r = fio_aioring_cqring_reap(td, events, max);
+               if (r) {
+                       events += r;
+                       continue;
+               }
+
+               r = io_ring_enter(ld->aio_ctx, 0, actual_min,
+                                       IORING_FLAG_GETEVENTS);
+               if (r < 0) {
+                       if (errno == EAGAIN)
+                               continue;
+                       perror("ring enter");
+                       break;
+               }
+       } while (events < min);
+
+       return r < 0 ? r : events;
+}
+
+static enum fio_q_status fio_aioring_queue(struct thread_data *td,
+                                          struct io_u *io_u)
+{
+       struct aioring_data *ld = td->io_ops_data;
+       struct aio_sq_ring *ring = ld->sq_ring;
+       unsigned tail, next_tail;
+
+       fio_ro_check(td, io_u);
+
+       if (ld->queued == td->o.iodepth)
+               return FIO_Q_BUSY;
+
+       /*
+        * fsync is tricky, since it can fail and we need to do it
+        * serialized with other io. the reason is that linux doesn't
+        * support aio fsync yet. So return busy for the case where we
+        * have pending io, to let fio complete those first.
+        */
+       if (ddir_sync(io_u->ddir)) {
+               if (ld->queued)
+                       return FIO_Q_BUSY;
+
+               do_io_u_sync(td, io_u);
+               return FIO_Q_COMPLETED;
+       }
+
+       if (io_u->ddir == DDIR_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;
+       }
+
+       tail = ring->tail;
+       next_tail = tail + 1;
+       if (next_tail == ring->nr_events)
+               next_tail = 0;
+       read_barrier();
+       if (next_tail == ring->head)
+               return FIO_Q_BUSY;
+
+       ring->array[tail] = io_u->index;
+       ring->tail = next_tail;
+       write_barrier();
+
+       ld->queued++;
+       return FIO_Q_QUEUED;
+}
+
+static void fio_aioring_queued(struct thread_data *td, int start, int nr)
+{
+       struct aioring_data *ld = td->io_ops_data;
+       struct timespec now;
+
+       if (!fio_fill_issue_time(td))
+               return;
+
+       fio_gettime(&now, NULL);
+
+       while (nr--) {
+               int index = ld->sq_ring->array[start];
+               struct io_u *io_u = io_u = ld->io_u_index[index];
+
+               memcpy(&io_u->issue_time, &now, sizeof(now));
+               io_u_queued(td, io_u);
+
+               start++;
+               if (start == ld->sq_ring->nr_events)
+                       start = 0;
+       }
+}
+
+static int fio_aioring_commit(struct thread_data *td)
+{
+       struct aioring_data *ld = td->io_ops_data;
+       int ret;
+
+       if (!ld->queued)
+               return 0;
+
+       do {
+               int start = ld->sq_ring->head;
+               long nr = ld->queued;
+
+               ret = io_ring_enter(ld->aio_ctx, nr, 0, IORING_FLAG_SUBMIT |
+                                               IORING_FLAG_GETEVENTS);
+               if (ret == -1)
+                       perror("io_ring_enter");
+               if (ret > 0) {
+                       fio_aioring_queued(td, start, ret);
+                       io_u_mark_submit(td, ret);
+
+                       ld->queued -= ret;
+                       ret = 0;
+               } else if (ret == -EINTR || !ret) {
+                       if (!ret)
+                               io_u_mark_submit(td, ret);
+                       continue;
+               } else if (ret == -EAGAIN) {
+                       /*
+                        * If we get EAGAIN, we should break out without
+                        * error and let the upper layer reap some
+                        * events for us. If we have no queued IO, we
+                        * must loop here. If we loop for more than 30s,
+                        * just error out, something must be buggy in the
+                        * IO path.
+                        */
+                       if (ld->queued) {
+                               ret = 0;
+                               break;
+                       }
+                       usleep(1);
+                       continue;
+               } else if (ret == -ENOMEM) {
+                       /*
+                        * If we get -ENOMEM, reap events if we can. If
+                        * we cannot, treat it as a fatal event since there's
+                        * nothing we can do about it.
+                        */
+                       if (ld->queued)
+                               ret = 0;
+                       break;
+               } else
+                       break;
+       } while (ld->queued);
+
+       return ret;
+}
+
+static size_t aioring_cq_size(struct thread_data *td)
+{
+       return sizeof(struct aio_cq_ring) + 2 * td->o.iodepth * sizeof(struct io_event);
+}
+
+static size_t aioring_sq_iocb(struct thread_data *td)
+{
+       return sizeof(struct iocb) * td->o.iodepth;
+}
+
+static size_t aioring_sq_size(struct thread_data *td)
+{
+       return sizeof(struct aio_sq_ring) + td->o.iodepth * sizeof(u32);
+}
+
+static void fio_aioring_cleanup(struct thread_data *td)
+{
+       struct aioring_data *ld = td->io_ops_data;
+
+       if (ld) {
+               /*
+                * Work-around to avoid huge RCU stalls at exit time. If we
+                * don't do this here, then it'll be torn down by exit_aio().
+                * But for that case we can parallellize the freeing, thus
+                * speeding it up a lot.
+                */
+               if (!(td->flags & TD_F_CHILD))
+                       io_destroy(ld->aio_ctx);
+               free(ld->io_u_index);
+               free(ld->io_us);
+               fio_memfree(ld->sq_ring, aioring_sq_size(td), false);
+               fio_memfree(ld->iocbs, aioring_sq_iocb(td), false);
+               fio_memfree(ld->cq_ring, aioring_cq_size(td), false);
+               free(ld);
+       }
+}
+
+static int fio_aioring_queue_init(struct thread_data *td)
+{
+#ifdef __NR_sys_io_setup2
+       struct aioring_data *ld = td->io_ops_data;
+       struct aioring_options *o = td->eo;
+       int flags = IOCTX_FLAG_SCQRING;
+       int depth = td->o.iodepth;
+
+       if (o->hipri)
+               flags |= IOCTX_FLAG_IOPOLL;
+       if (o->fixedbufs) {
+               struct rlimit rlim = {
+                       .rlim_cur = RLIM_INFINITY,
+                       .rlim_max = RLIM_INFINITY,
+               };
+
+               setrlimit(RLIMIT_MEMLOCK, &rlim);
+               flags |= IOCTX_FLAG_FIXEDBUFS;
+       }
+
+       return syscall(__NR_sys_io_setup2, depth, flags,
+                       ld->sq_ring, ld->cq_ring, &ld->aio_ctx);
+#else
+       return -1;
+#endif
+}
+
+static int fio_aioring_post_init(struct thread_data *td)
+{
+       struct aioring_data *ld = td->io_ops_data;
+       struct aioring_options *o = td->eo;
+       struct io_u *io_u;
+       struct iocb *iocb;
+       int err = 0;
+
+       if (o->fixedbufs) {
+               int i;
+
+               for (i = 0; i < td->o.iodepth; i++) {
+                       io_u = ld->io_u_index[i];
+                       iocb = &ld->iocbs[i];
+                       iocb->u.c.buf = io_u->buf;
+                       iocb->u.c.nbytes = td_max_bs(td);
+
+                       if (o->hipri)
+                               iocb->u.c.flags |= IOCB_FLAG_HIPRI;
+               }
+       }
+
+       err = fio_aioring_queue_init(td);
+       if (err) {
+               td_verror(td, -err, "io_queue_init");
+               return 1;
+       }
+
+       return 0;
+}
+
+static int fio_aioring_init(struct thread_data *td)
+{
+       struct aioring_data *ld;
+
+       if (td->o.iodepth <= 1) {
+               printf("aio-ring: needs a minimum QD of 2\n");
+               return 1;
+       }
+
+       ld = calloc(1, sizeof(*ld));
+
+       /* io_u index */
+       ld->io_u_index = calloc(td->o.iodepth, sizeof(struct io_u *));
+       ld->io_us = calloc(td->o.iodepth, sizeof(struct io_u *));
+
+       ld->iocbs = fio_memalign(page_size, aioring_sq_iocb(td), false);
+       memset(ld->iocbs, 0, aioring_sq_iocb(td));
+
+       ld->sq_ring = fio_memalign(page_size, aioring_sq_size(td), false);
+       memset(ld->sq_ring, 0, aioring_sq_size(td));
+       ld->sq_ring->nr_events = td->o.iodepth;
+       ld->sq_ring->iocbs = (u64) ld->iocbs;
+
+       ld->cq_ring = fio_memalign(page_size, aioring_cq_size(td), false);
+       memset(ld->cq_ring, 0, aioring_cq_size(td));
+       ld->cq_ring->nr_events = td->o.iodepth * 2;
+
+       td->io_ops_data = ld;
+       return 0;
+}
+
+static int fio_aioring_io_u_init(struct thread_data *td, struct io_u *io_u)
+{
+       struct aioring_data *ld = td->io_ops_data;
+
+       ld->io_u_index[io_u->index] = io_u;
+       return 0;
+}
+
+static struct ioengine_ops ioengine = {
+       .name                   = "aio-ring",
+       .version                = FIO_IOOPS_VERSION,
+       .init                   = fio_aioring_init,
+       .post_init              = fio_aioring_post_init,
+       .io_u_init              = fio_aioring_io_u_init,
+       .prep                   = fio_aioring_prep,
+       .queue                  = fio_aioring_queue,
+       .commit                 = fio_aioring_commit,
+       .getevents              = fio_aioring_getevents,
+       .event                  = fio_aioring_event,
+       .cleanup                = fio_aioring_cleanup,
+       .open_file              = generic_open_file,
+       .close_file             = generic_close_file,
+       .get_file_size          = generic_get_file_size,
+       .options                = options,
+       .option_struct_size     = sizeof(struct aioring_options),
+};
+
+static void fio_init fio_aioring_register(void)
+{
+#ifdef __NR_sys_io_setup2
+#ifdef __NR_sys_io_ring_enter
+       register_ioengine(&ioengine);
+#endif
+#endif
+}
+
+static void fio_exit fio_aioring_unregister(void)
+{
+#ifdef __NR_sys_io_setup2
+#ifdef __NR_sys_io_ring_enter
+       unregister_ioengine(&ioengine);
+#endif
+#endif
+}
index 7a7006c..81fe9ff 100644 (file)
--- a/options.c
+++ b/options.c
@@ -1773,6 +1773,11 @@ struct fio_option fio_options[FIO_MAX_OPTS] = {
                            .help = "Linux native asynchronous IO",
                          },
 #endif
+#ifdef CONFIG_LIBAIO
+                         { .ival = "aio-ring",
+                           .help = "Linux native asynchronous IO",
+                         },
+#endif
 #ifdef CONFIG_POSIXAIO
                          { .ival = "posixaio",
                            .help = "POSIX asynchronous IO",