[v2,7/8] dataplane: add virtio-blk data plane code

virtio-blk-data-plane is a subset implementation of virtio-blk.  It only
handles read, write, and flush requests.  It does this using a dedicated
thread that executes an epoll(2)-based event loop and processes I/O
using Linux AIO.

This approach performs very well but can be used for raw image files
only.  The number of IOPS achieved has been reported to be several times
higher than the existing virtio-blk implementation.

Eventually it should be possible to unify virtio-blk-data-plane with the
main body of QEMU code once the block layer and hardware emulation is
able to run outside the global mutex.

Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
---
 hw/dataplane/Makefile.objs |   2 +-
 hw/dataplane/virtio-blk.c  | 418 +++++++++++++++++++++++++++++++++++++++++++++
 hw/dataplane/virtio-blk.h  |  41 +++++
 trace-events               |   6 +
 4 files changed, 466 insertions(+), 1 deletion(-)
 create mode 100644 hw/dataplane/virtio-blk.c
 create mode 100644 hw/dataplane/virtio-blk.h

On Tue, Nov 20, 2012 at 01:31:51PM +0100, Stefan Hajnoczi wrote:
> virtio-blk-data-plane is a subset implementation of virtio-blk.  It only
> handles read, write, and flush requests.  It does this using a dedicated
> thread that executes an epoll(2)-based event loop and processes I/O
> using Linux AIO.
> 
> This approach performs very well but can be used for raw image files
> only.  The number of IOPS achieved has been reported to be several times
> higher than the existing virtio-blk implementation.
> 
> Eventually it should be possible to unify virtio-blk-data-plane with the
> main body of QEMU code once the block layer and hardware emulation is
> able to run outside the global mutex.
> 
> Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
> ---
>  hw/dataplane/Makefile.objs |   2 +-
>  hw/dataplane/virtio-blk.c  | 418 +++++++++++++++++++++++++++++++++++++++++++++
>  hw/dataplane/virtio-blk.h  |  41 +++++
>  trace-events               |   6 +
>  4 files changed, 466 insertions(+), 1 deletion(-)
>  create mode 100644 hw/dataplane/virtio-blk.c
>  create mode 100644 hw/dataplane/virtio-blk.h
> 
> diff --git a/hw/dataplane/Makefile.objs b/hw/dataplane/Makefile.objs
> index abd408f..682aa9e 100644
> --- a/hw/dataplane/Makefile.objs
> +++ b/hw/dataplane/Makefile.objs
> @@ -1,3 +1,3 @@
>  ifeq ($(CONFIG_VIRTIO), y)
> -common-obj-$(CONFIG_VIRTIO_BLK_DATA_PLANE) += hostmem.o vring.o event-poll.o ioq.o
> +common-obj-$(CONFIG_VIRTIO_BLK_DATA_PLANE) += hostmem.o vring.o event-poll.o ioq.o virtio-blk.o
>  endif
> diff --git a/hw/dataplane/virtio-blk.c b/hw/dataplane/virtio-blk.c
> new file mode 100644
> index 0000000..5dcc872
> --- /dev/null
> +++ b/hw/dataplane/virtio-blk.c
> @@ -0,0 +1,418 @@
> +/*
> + * Dedicated thread for virtio-blk I/O processing
> + *
> + * Copyright 2012 IBM, Corp.
> + * Copyright 2012 Red Hat, Inc. and/or its affiliates
> + *
> + * Authors:
> + *   Stefan Hajnoczi <stefanha@redhat.com>
> + *
> + * This work is licensed under the terms of the GNU GPL, version 2 or later.
> + * See the COPYING file in the top-level directory.
> + *
> + */
> +
> +#include "trace.h"
> +#include "event-poll.h"
> +#include "qemu-thread.h"
> +#include "vring.h"
> +#include "ioq.h"
> +#include "hw/virtio-blk.h"
> +#include "hw/dataplane/virtio-blk.h"
> +
> +enum {
> +    SEG_MAX = 126,                  /* maximum number of I/O segments */
> +    VRING_MAX = SEG_MAX + 2,        /* maximum number of vring descriptors */
> +    REQ_MAX = VRING_MAX,            /* maximum number of requests in the vring,
> +                                     * is VRING_MAX / 2 with traditional and
> +                                     * VRING_MAX with indirect descriptors */
> +};
> +
> +typedef struct {
> +    struct iocb iocb;               /* Linux AIO control block */
> +    unsigned char *status;          /* virtio block status code */
> +    unsigned int head;              /* vring descriptor index */
> +} VirtIOBlockRequest;
> +
> +struct VirtIOBlockDataPlane {
> +    bool started;
> +    QEMUBH *start_bh;
> +    QemuThread thread;
> +
> +    int fd;                         /* image file descriptor */
> +
> +    VirtIODevice *vdev;
> +    Vring vring;                    /* virtqueue vring */
> +    EventNotifier *guest_notifier;  /* irq */
> +
> +    EventPoll event_poll;           /* event poller */
> +    EventHandler io_handler;        /* Linux AIO completion handler */
> +    EventHandler notify_handler;    /* virtqueue notify handler */
> +
> +    IOQueue ioqueue;                /* Linux AIO queue (should really be per
> +                                       dataplane thread) */
> +    VirtIOBlockRequest requests[REQ_MAX]; /* pool of requests, managed by the
> +                                             queue */
> +
> +    unsigned int num_reqs;
> +    QemuMutex num_reqs_lock;
> +    QemuCond no_reqs_cond;
> +};
> +
> +/* Raise an interrupt to signal guest, if necessary */
> +static void notify_guest(VirtIOBlockDataPlane *s)
> +{
> +    if (!vring_should_notify(s->vdev, &s->vring)) {
> +        return;
> +    }
> +
> +    event_notifier_set(s->guest_notifier);
> +}
> +
> +static void complete_request(struct iocb *iocb, ssize_t ret, void *opaque)
> +{
> +    VirtIOBlockDataPlane *s = opaque;
> +    VirtIOBlockRequest *req = container_of(iocb, VirtIOBlockRequest, iocb);
> +    int len;
> +
> +    if (likely(ret >= 0)) {
> +        *req->status = VIRTIO_BLK_S_OK;
> +        len = ret;
> +    } else {
> +        *req->status = VIRTIO_BLK_S_IOERR;
> +        len = 0;
> +    }
> +
> +    trace_virtio_blk_data_plane_complete_request(s, req->head, ret);
> +
> +    /* According to the virtio specification len should be the number of bytes
> +     * written to, but for virtio-blk it seems to be the number of bytes
> +     * transferred plus the status bytes.
> +     */
> +    vring_push(&s->vring, req->head, len + sizeof(*req->status));
> +
> +    qemu_mutex_lock(&s->num_reqs_lock);
> +    if (--s->num_reqs == 0) {
> +        qemu_cond_broadcast(&s->no_reqs_cond);
> +    }
> +    qemu_mutex_unlock(&s->num_reqs_lock);
> +}
> +
> +static void process_request(IOQueue *ioq, struct iovec iov[],
> +                            unsigned int out_num, unsigned int in_num,
> +                            unsigned int head)
> +{
> +    /* Virtio block requests look like this: */
> +    struct virtio_blk_outhdr *outhdr; /* iov[0] */
> +    /* data[]                            ... */
> +    struct virtio_blk_inhdr *inhdr;   /* iov[out_num + in_num - 1] */
> +
> +    if (unlikely(out_num == 0 || in_num == 0 ||
> +                iov[0].iov_len != sizeof *outhdr ||
> +                iov[out_num + in_num - 1].iov_len != sizeof *inhdr)) {
> +        fprintf(stderr, "virtio-blk invalid request\n");
> +        exit(1);
> +    }
> +
> +    outhdr = iov[0].iov_base;
> +    inhdr = iov[out_num + in_num - 1].iov_base;
> +

Rusty is trying to get rid of hard-coded layout assumptions,
let's not add any more if we can help it.


> +    /* TODO Linux sets the barrier bit even when not advertised! */
> +    uint32_t type = outhdr->type & ~VIRTIO_BLK_T_BARRIER;
> +    struct iocb *iocb;
> +    switch (type & (VIRTIO_BLK_T_OUT | VIRTIO_BLK_T_SCSI_CMD |
> +                    VIRTIO_BLK_T_FLUSH)) {
> +    case VIRTIO_BLK_T_IN:
> +        if (unlikely(out_num != 1)) {
> +            fprintf(stderr, "virtio-blk invalid read request\n");
> +            exit(1);
> +        }
> +        iocb = ioq_rdwr(ioq, true, &iov[1], in_num - 1, outhdr->sector * 512);
> +        break;
> +
> +    case VIRTIO_BLK_T_OUT:
> +        if (unlikely(in_num != 1)) {
> +            fprintf(stderr, "virtio-blk invalid write request\n");
> +            exit(1);
> +        }
> +        iocb = ioq_rdwr(ioq, false, &iov[1], out_num - 1, outhdr->sector * 512);
> +        break;
> +
> +    case VIRTIO_BLK_T_SCSI_CMD:
> +        if (unlikely(in_num == 0)) {
> +            fprintf(stderr, "virtio-blk invalid SCSI command request\n");
> +            exit(1);
> +        }
> +
> +        /* TODO support SCSI commands */
> +        {
> +            VirtIOBlockDataPlane *s = container_of(ioq, VirtIOBlockDataPlane,
> +                                                   ioqueue);
> +            inhdr->status = VIRTIO_BLK_S_UNSUPP;
> +            vring_push(&s->vring, head, sizeof *inhdr);
> +            notify_guest(s);
> +        }
> +        return;
> +
> +    case VIRTIO_BLK_T_FLUSH:
> +        if (unlikely(in_num != 1 || out_num != 1)) {
> +            fprintf(stderr, "virtio-blk invalid flush request\n");
> +            exit(1);
> +        }
> +
> +        /* TODO fdsync not supported by Linux AIO, do it synchronously here! */
> +        {
> +            VirtIOBlockDataPlane *s = container_of(ioq, VirtIOBlockDataPlane,
> +                                                   ioqueue);
> +            fdatasync(s->fd);
> +            inhdr->status = VIRTIO_BLK_S_OK;
> +            vring_push(&s->vring, head, sizeof *inhdr);
> +            notify_guest(s);
> +        }
> +        return;
> +
> +    default:
> +        fprintf(stderr, "virtio-blk unsupported request type %#x\n",
> +                outhdr->type);
> +        exit(1);
> +    }
> +
> +    /* Fill in virtio block metadata needed for completion */
> +    VirtIOBlockRequest *req = container_of(iocb, VirtIOBlockRequest, iocb);
> +    req->head = head;
> +    req->status = &inhdr->status;
> +}
> +
> +static bool handle_notify(EventHandler *handler)
> +{
> +    VirtIOBlockDataPlane *s = container_of(handler, VirtIOBlockDataPlane,
> +                                           notify_handler);
> +
> +    /* There is one array of iovecs into which all new requests are extracted
> +     * from the vring.  Requests are read from the vring and the translated
> +     * descriptors are written to the iovecs array.  The iovecs do not have to
> +     * persist across handle_notify() calls because the kernel copies the
> +     * iovecs on io_submit().
> +     *
> +     * Handling io_submit() EAGAIN may require storing the requests across
> +     * handle_notify() calls until the kernel has sufficient resources to
> +     * accept more I/O.  This is not implemented yet.
> +     */
> +    struct iovec iovec[VRING_MAX];
> +    struct iovec *end = &iovec[VRING_MAX];
> +    struct iovec *iov = iovec;
> +
> +    /* When a request is read from the vring, the index of the first descriptor
> +     * (aka head) is returned so that the completed request can be pushed onto
> +     * the vring later.
> +     *
> +     * The number of hypervisor read-only iovecs is out_num.  The number of
> +     * hypervisor write-only iovecs is in_num.
> +     */
> +    int head;
> +    unsigned int out_num = 0, in_num = 0;
> +    unsigned int num_queued;
> +
> +    for (;;) {
> +        /* Disable guest->host notifies to avoid unnecessary vmexits */
> +        vring_set_notification(s->vdev, &s->vring, false);
> +
> +        for (;;) {
> +            head = vring_pop(s->vdev, &s->vring, iov, end, &out_num, &in_num);
> +            if (head < 0) {
> +                break; /* no more requests */
> +            }
> +
> +            trace_virtio_blk_data_plane_process_request(s, out_num, in_num,
> +                                                        head);
> +
> +            process_request(&s->ioqueue, iov, out_num, in_num, head);
> +            iov += out_num + in_num;
> +        }
> +
> +        if (likely(head == -EAGAIN)) { /* vring emptied */
> +            /* Re-enable guest->host notifies and stop processing the vring.
> +             * But if the guest has snuck in more descriptors, keep processing.
> +             */
> +            vring_set_notification(s->vdev, &s->vring, true);
> +            smp_mb();
> +            if (!vring_more_avail(&s->vring)) {
> +                break;
> +            }
> +        } else { /* head == -ENOBUFS or fatal error, iovecs[] is depleted */
> +            /* Since there are no iovecs[] left, stop processing for now.  Do
> +             * not re-enable guest->host notifies since the I/O completion
> +             * handler knows to check for more vring descriptors anyway.
> +             */
> +            break;
> +        }
> +    }
> +
> +    num_queued = ioq_num_queued(&s->ioqueue);
> +    if (num_queued > 0) {
> +        qemu_mutex_lock(&s->num_reqs_lock);
> +        s->num_reqs += num_queued;
> +        qemu_mutex_unlock(&s->num_reqs_lock);
> +
> +        int rc = ioq_submit(&s->ioqueue);
> +        if (unlikely(rc < 0)) {
> +            fprintf(stderr, "ioq_submit failed %d\n", rc);
> +            exit(1);
> +        }
> +    }
> +    return true;
> +}
> +
> +static bool handle_io(EventHandler *handler)
> +{
> +    VirtIOBlockDataPlane *s = container_of(handler, VirtIOBlockDataPlane,
> +                                           io_handler);
> +
> +    if (ioq_run_completion(&s->ioqueue, complete_request, s) > 0) {
> +        notify_guest(s);
> +    }
> +
> +    /* If there were more requests than iovecs, the vring will not be empty yet
> +     * so check again.  There should now be enough resources to process more
> +     * requests.
> +     */
> +    if (unlikely(vring_more_avail(&s->vring))) {
> +        return handle_notify(&s->notify_handler);
> +    }
> +
> +    return true;
> +}
> +
> +static void *data_plane_thread(void *opaque)
> +{
> +    VirtIOBlockDataPlane *s = opaque;
> +    event_poll_run(&s->event_poll);
> +    return NULL;
> +}
> +
> +static void start_data_plane_bh(void *opaque)
> +{
> +    VirtIOBlockDataPlane *s = opaque;
> +
> +    qemu_bh_delete(s->start_bh);
> +    s->start_bh = NULL;
> +    qemu_thread_create(&s->thread, data_plane_thread,
> +                       s, QEMU_THREAD_JOINABLE);
> +}
> +
> +VirtIOBlockDataPlane *virtio_blk_data_plane_create(VirtIODevice *vdev, int fd)
> +{
> +    VirtIOBlockDataPlane *s;
> +
> +    s = g_new0(VirtIOBlockDataPlane, 1);
> +    s->vdev = vdev;
> +    s->fd = fd;
> +    return s;
> +}
> +
> +void virtio_blk_data_plane_destroy(VirtIOBlockDataPlane *s)
> +{
> +    if (!s) {
> +        return;
> +    }
> +    virtio_blk_data_plane_stop(s);
> +    g_free(s);
> +}
> +
> +/* Block until pending requests have completed
> + *
> + * The vring continues to be serviced so ensure no new requests will be added
> + * to avoid races.
> + */
> +void virtio_blk_data_plane_drain(VirtIOBlockDataPlane *s)
> +{
> +    qemu_mutex_lock(&s->num_reqs_lock);
> +    while (s->num_reqs > 0) {
> +        qemu_cond_wait(&s->no_reqs_cond, &s->num_reqs_lock);
> +    }
> +    qemu_mutex_unlock(&s->num_reqs_lock);
> +}
> +
> +void virtio_blk_data_plane_start(VirtIOBlockDataPlane *s)
> +{
> +    VirtQueue *vq;
> +    int i;
> +
> +    if (s->started) {
> +        return;
> +    }
> +
> +    vq = virtio_get_queue(s->vdev, 0);
> +    if (!vring_setup(&s->vring, s->vdev, 0)) {
> +        return;
> +    }
> +
> +    event_poll_init(&s->event_poll);
> +
> +    /* Set up guest notifier (irq) */
> +    if (s->vdev->binding->set_guest_notifiers(s->vdev->binding_opaque,
> +                                              true) != 0) {
> +        fprintf(stderr, "virtio-blk failed to set guest notifier, "
> +                "ensure -enable-kvm is set\n");
> +        exit(1);
> +    }
> +    s->guest_notifier = virtio_queue_get_guest_notifier(vq);
> +
> +    /* Set up virtqueue notify */
> +    if (s->vdev->binding->set_host_notifier(s->vdev->binding_opaque,
> +                                            0, true) != 0) {
> +        fprintf(stderr, "virtio-blk failed to set host notifier\n");
> +        exit(1);
> +    }
> +    event_poll_add(&s->event_poll, &s->notify_handler,
> +                   virtio_queue_get_host_notifier(vq),
> +                   handle_notify);
> +
> +    /* Set up ioqueue */
> +    ioq_init(&s->ioqueue, s->fd, REQ_MAX);
> +    for (i = 0; i < ARRAY_SIZE(s->requests); i++) {
> +        ioq_put_iocb(&s->ioqueue, &s->requests[i].iocb);
> +    }
> +    event_poll_add(&s->event_poll, &s->io_handler,
> +                   ioq_get_notifier(&s->ioqueue), handle_io);
> +
> +    s->started = true;
> +    trace_virtio_blk_data_plane_start(s);
> +
> +    /* Kick right away to begin processing requests already in vring */
> +    event_notifier_set(virtio_queue_get_host_notifier(vq));
> +
> +    /* Spawn thread in BH so it inherits iothread cpusets */
> +    s->start_bh = qemu_bh_new(start_data_plane_bh, s);
> +    qemu_bh_schedule(s->start_bh);
> +}
> +
> +void virtio_blk_data_plane_stop(VirtIOBlockDataPlane *s)
> +{
> +    if (!s->started) {
> +        return;
> +    }
> +    s->started = false;
> +    trace_virtio_blk_data_plane_stop(s);
> +
> +    /* Stop thread or cancel pending thread creation BH */
> +    if (s->start_bh) {
> +        qemu_bh_delete(s->start_bh);
> +        s->start_bh = NULL;
> +    } else {
> +        virtio_blk_data_plane_drain(s);
> +        event_poll_stop(&s->event_poll);
> +        qemu_thread_join(&s->thread);
> +    }
> +
> +    ioq_cleanup(&s->ioqueue);
> +
> +    s->vdev->binding->set_host_notifier(s->vdev->binding_opaque, 0, false);
> +
> +    event_poll_cleanup(&s->event_poll);
> +
> +    /* Clean up guest notifier (irq) */
> +    s->vdev->binding->set_guest_notifiers(s->vdev->binding_opaque, false);
> +
> +    vring_teardown(&s->vring);
> +}
> diff --git a/hw/dataplane/virtio-blk.h b/hw/dataplane/virtio-blk.h
> new file mode 100644
> index 0000000..ddf1115
> --- /dev/null
> +++ b/hw/dataplane/virtio-blk.h
> @@ -0,0 +1,41 @@
> +/*
> + * Dedicated thread for virtio-blk I/O processing
> + *
> + * Copyright 2012 IBM, Corp.
> + * Copyright 2012 Red Hat, Inc. and/or its affiliates
> + *
> + * Authors:
> + *   Stefan Hajnoczi <stefanha@redhat.com>
> + *
> + * This work is licensed under the terms of the GNU GPL, version 2 or later.
> + * See the COPYING file in the top-level directory.
> + *
> + */
> +
> +#ifndef HW_DATAPLANE_VIRTIO_BLK_H
> +#define HW_DATAPLANE_VIRTIO_BLK_H
> +
> +#include "hw/virtio.h"
> +
> +typedef struct VirtIOBlockDataPlane VirtIOBlockDataPlane;
> +
> +#ifdef CONFIG_VIRTIO_BLK_DATA_PLANE
> +VirtIOBlockDataPlane *virtio_blk_data_plane_create(VirtIODevice *vdev, int fd);
> +void virtio_blk_data_plane_destroy(VirtIOBlockDataPlane *s);
> +void virtio_blk_data_plane_start(VirtIOBlockDataPlane *s);
> +void virtio_blk_data_plane_stop(VirtIOBlockDataPlane *s);
> +void virtio_blk_data_plane_drain(VirtIOBlockDataPlane *s);
> +#else
> +static inline VirtIOBlockDataPlane *virtio_blk_data_plane_create(
> +        VirtIODevice *vdev, int fd)
> +{
> +    return NULL;
> +}
> +
> +static inline void virtio_blk_data_plane_destroy(VirtIOBlockDataPlane *s) {}
> +static inline void virtio_blk_data_plane_start(VirtIOBlockDataPlane *s) {}
> +static inline void virtio_blk_data_plane_stop(VirtIOBlockDataPlane *s) {}
> +static inline void virtio_blk_data_plane_drain(VirtIOBlockDataPlane *s) {}
> +#endif
> +
> +#endif /* HW_DATAPLANE_VIRTIO_BLK_H */
> diff --git a/trace-events b/trace-events
> index a9a791b..1edc2ae 100644
> --- a/trace-events
> +++ b/trace-events
> @@ -98,6 +98,12 @@ virtio_blk_rw_complete(void *req, int ret) "req %p ret %d"
>  virtio_blk_handle_write(void *req, uint64_t sector, size_t nsectors) "req %p sector %"PRIu64" nsectors %zu"
>  virtio_blk_handle_read(void *req, uint64_t sector, size_t nsectors) "req %p sector %"PRIu64" nsectors %zu"
>  
> +# hw/dataplane/virtio-blk.c
> +virtio_blk_data_plane_start(void *s) "dataplane %p"
> +virtio_blk_data_plane_stop(void *s) "dataplane %p"
> +virtio_blk_data_plane_process_request(void *s, unsigned int out_num, unsigned int in_num, unsigned int head) "dataplane %p out_num %u in_num %u head %u"
> +virtio_blk_data_plane_complete_request(void *s, unsigned int head, int ret) "dataplane %p head %u ret %d"
> +
>  # hw/dataplane/vring.c
>  vring_setup(uint64_t physical, void *desc, void *avail, void *used) "vring physical %#"PRIx64" desc %p avail %p used %p"
>  
> -- 
> 1.8.0

On Tue, Nov 20, 2012 at 03:04:53PM +0200, Michael S. Tsirkin wrote:
> On Tue, Nov 20, 2012 at 01:31:51PM +0100, Stefan Hajnoczi wrote:
> > +static void process_request(IOQueue *ioq, struct iovec iov[],
> > +                            unsigned int out_num, unsigned int in_num,
> > +                            unsigned int head)
> > +{
> > +    /* Virtio block requests look like this: */
> > +    struct virtio_blk_outhdr *outhdr; /* iov[0] */
> > +    /* data[]                            ... */
> > +    struct virtio_blk_inhdr *inhdr;   /* iov[out_num + in_num - 1] */
> > +
> > +    if (unlikely(out_num == 0 || in_num == 0 ||
> > +                iov[0].iov_len != sizeof *outhdr ||
> > +                iov[out_num + in_num - 1].iov_len != sizeof *inhdr)) {
> > +        fprintf(stderr, "virtio-blk invalid request\n");
> > +        exit(1);
> > +    }
> > +
> > +    outhdr = iov[0].iov_base;
> > +    inhdr = iov[out_num + in_num - 1].iov_base;
> > +
> 
> Rusty is trying to get rid of hard-coded layout assumptions,
> let's not add any more if we can help it.

Will try to abstract the iov arrays in the next version.  Still not a
fan of doing this because it complicates the code for little gain.

Stefan