hw/block/dataplane/virtio-blk.c

   1 /*
   2  * Dedicated thread for virtio-blk I/O processing
   3  *
   4  * Copyright 2012 IBM, Corp.
   5  * Copyright 2012 Red Hat, Inc. and/or its affiliates
   6  *
   7  * Authors:
   8  *   Stefan Hajnoczi <stefanha@redhat.com>
   9  *
  10  * This work is licensed under the terms of the GNU GPL, version 2 or later.
  11  * See the COPYING file in the top-level directory.
  12  *
  13  */
  14
  15 #include "trace.h"
  16 #include "qemu/iov.h"
  17 #include "qemu/thread.h"
  18 #include "qemu/error-report.h"
  19 #include "hw/virtio/virtio-access.h"
  20 #include "hw/virtio/dataplane/vring.h"
  21 #include "hw/virtio/dataplane/vring-accessors.h"
  22 #include "sysemu/block-backend.h"
  23 #include "hw/virtio/virtio-blk.h"
  24 #include "virtio-blk.h"
  25 #include "block/aio.h"
  26 #include "hw/virtio/virtio-bus.h"
  27 #include "qom/object_interfaces.h"
  28
  29 struct VirtIOBlockDataPlane {
  30     bool started;
  31     bool starting;
  32     bool stopping;
  33     bool disabled;
  34
  35     VirtIOBlkConf *conf;
  36
  37     VirtIODevice *vdev;
  38     Vring vring;                    /* virtqueue vring */
  39     EventNotifier *guest_notifier;  /* irq */
  40     QEMUBH *bh;                     /* bh for guest notification */
  41
  42     /* Note that these EventNotifiers are assigned by value.  This is
  43      * fine as long as you do not call event_notifier_cleanup on them
  44      * (because you don't own the file descriptor or handle; you just
  45      * use it).
  46      */
  47     IOThread *iothread;
  48     IOThread internal_iothread_obj;
  49     AioContext *ctx;
  50     EventNotifier host_notifier;    /* doorbell */
  51
  52     /* Operation blocker on BDS */
  53     Error *blocker;
  54     void (*saved_complete_request)(struct VirtIOBlockReq *req,
  55                                    unsigned char status);
  56 };
  57
  58 /* Raise an interrupt to signal guest, if necessary */
  59 static void notify_guest(VirtIOBlockDataPlane *s)
  60 {
  61     if (!vring_should_notify(s->vdev, &s->vring)) {
  62         return;
  63     }
  64
  65     event_notifier_set(s->guest_notifier);
  66 }
  67
  68 static void notify_guest_bh(void *opaque)
  69 {
  70     VirtIOBlockDataPlane *s = opaque;
  71
  72     notify_guest(s);
  73 }
  74
  75 static void complete_request_vring(VirtIOBlockReq *req, unsigned char status)
  76 {
  77     VirtIOBlockDataPlane *s = req->dev->dataplane;
  78     stb_p(&req->in->status, status);
  79
  80     vring_push(s->vdev, &req->dev->dataplane->vring, &req->elem,
  81                req->qiov.size + sizeof(*req->in));
  82
  83     /* Suppress notification to guest by BH and its scheduled
  84      * flag because requests are completed as a batch after io
  85      * plug & unplug is introduced, and the BH can still be
  86      * executed in dataplane aio context even after it is
  87      * stopped, so needn't worry about notification loss with BH.
  88      */
  89     qemu_bh_schedule(s->bh);
  90 }
  91
  92 static void handle_notify(EventNotifier *e)
  93 {
  94     VirtIOBlockDataPlane *s = container_of(e, VirtIOBlockDataPlane,
  95                                            host_notifier);
  96     VirtIOBlock *vblk = VIRTIO_BLK(s->vdev);
  97
  98     event_notifier_test_and_clear(&s->host_notifier);
  99     blk_io_plug(s->conf->conf.blk);
 100     for (;;) {
 101         MultiReqBuffer mrb = {};
 102         int ret;
 103
 104         /* Disable guest->host notifies to avoid unnecessary vmexits */
 105         vring_disable_notification(s->vdev, &s->vring);
 106
 107         for (;;) {
 108             VirtIOBlockReq *req = virtio_blk_alloc_request(vblk);
 109
 110             ret = vring_pop(s->vdev, &s->vring, &req->elem);
 111             if (ret < 0) {
 112                 virtio_blk_free_request(req);
 113                 break; /* no more requests */
 114             }
 115
 116             trace_virtio_blk_data_plane_process_request(s, req->elem.out_num,
 117                                                         req->elem.in_num,
 118                                                         req->elem.index);
 119
 120             virtio_blk_handle_request(req, &mrb);
 121         }
 122
 123         if (mrb.num_reqs) {
 124             virtio_blk_submit_multireq(s->conf->conf.blk, &mrb);
 125         }
 126
 127         if (likely(ret == -EAGAIN)) { /* vring emptied */
 128             /* Re-enable guest->host notifies and stop processing the vring.
 129              * But if the guest has snuck in more descriptors, keep processing.
 130              */
 131             if (vring_enable_notification(s->vdev, &s->vring)) {
 132                 break;
 133             }
 134         } else { /* fatal error */
 135             break;
 136         }
 137     }
 138     blk_io_unplug(s->conf->conf.blk);
 139 }
 140
 141 /* Context: QEMU global mutex held */
 142 void virtio_blk_data_plane_create(VirtIODevice *vdev, VirtIOBlkConf *conf,
 143                                   VirtIOBlockDataPlane **dataplane,
 144                                   Error **errp)
 145 {
 146     VirtIOBlockDataPlane *s;
 147     Error *local_err = NULL;
 148     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev)));
 149     VirtioBusClass *k = VIRTIO_BUS_GET_CLASS(qbus);
 150
 151     *dataplane = NULL;
 152
 153     if (!conf->data_plane && !conf->iothread) {
 154         return;
 155     }
 156
 157     /* Don't try if transport does not support notifiers. */
 158     if (!k->set_guest_notifiers || !k->set_host_notifier) {
 159         error_setg(errp,
 160                    "device is incompatible with x-data-plane "
 161                    "(transport does not support notifiers)");
 162         return;
 163     }
 164
 165     /* If dataplane is (re-)enabled while the guest is running there could be
 166      * block jobs that can conflict.
 167      */
 168     if (blk_op_is_blocked(conf->conf.blk, BLOCK_OP_TYPE_DATAPLANE,
 169                           &local_err)) {
 170         error_setg(errp, "cannot start dataplane thread: %s",
 171                    error_get_pretty(local_err));
 172         error_free(local_err);
 173         return;
 174     }
 175
 176     s = g_new0(VirtIOBlockDataPlane, 1);
 177     s->vdev = vdev;
 178     s->conf = conf;
 179
 180     if (conf->iothread) {
 181         s->iothread = conf->iothread;
 182         object_ref(OBJECT(s->iothread));
 183     } else {
 184         /* Create per-device IOThread if none specified.  This is for
 185          * x-data-plane option compatibility.  If x-data-plane is removed we
 186          * can drop this.
 187          */
 188         object_initialize(&s->internal_iothread_obj,
 189                           sizeof(s->internal_iothread_obj),
 190                           TYPE_IOTHREAD);
 191         user_creatable_complete(OBJECT(&s->internal_iothread_obj), &error_abort);
 192         s->iothread = &s->internal_iothread_obj;
 193     }
 194     s->ctx = iothread_get_aio_context(s->iothread);
 195     s->bh = aio_bh_new(s->ctx, notify_guest_bh, s);
 196
 197     error_setg(&s->blocker, "block device is in use by data plane");
 198     blk_op_block_all(conf->conf.blk, s->blocker);
 199     blk_op_unblock(conf->conf.blk, BLOCK_OP_TYPE_RESIZE, s->blocker);
 200     blk_op_unblock(conf->conf.blk, BLOCK_OP_TYPE_DRIVE_DEL, s->blocker);
 201     blk_op_unblock(conf->conf.blk, BLOCK_OP_TYPE_BACKUP_SOURCE, s->blocker);
 202     blk_op_unblock(conf->conf.blk, BLOCK_OP_TYPE_CHANGE, s->blocker);
 203     blk_op_unblock(conf->conf.blk, BLOCK_OP_TYPE_COMMIT_SOURCE, s->blocker);
 204     blk_op_unblock(conf->conf.blk, BLOCK_OP_TYPE_COMMIT_TARGET, s->blocker);
 205     blk_op_unblock(conf->conf.blk, BLOCK_OP_TYPE_EJECT, s->blocker);
 206     blk_op_unblock(conf->conf.blk, BLOCK_OP_TYPE_EXTERNAL_SNAPSHOT, s->blocker);
 207     blk_op_unblock(conf->conf.blk, BLOCK_OP_TYPE_INTERNAL_SNAPSHOT, s->blocker);
 208     blk_op_unblock(conf->conf.blk, BLOCK_OP_TYPE_INTERNAL_SNAPSHOT_DELETE,
 209                    s->blocker);
 210     blk_op_unblock(conf->conf.blk, BLOCK_OP_TYPE_MIRROR, s->blocker);
 211     blk_op_unblock(conf->conf.blk, BLOCK_OP_TYPE_STREAM, s->blocker);
 212     blk_op_unblock(conf->conf.blk, BLOCK_OP_TYPE_REPLACE, s->blocker);
 213
 214     *dataplane = s;
 215 }
 216
 217 /* Context: QEMU global mutex held */
 218 void virtio_blk_data_plane_destroy(VirtIOBlockDataPlane *s)
 219 {
 220     if (!s) {
 221         return;
 222     }
 223
 224     virtio_blk_data_plane_stop(s);
 225     blk_op_unblock_all(s->conf->conf.blk, s->blocker);
 226     error_free(s->blocker);
 227     object_unref(OBJECT(s->iothread));
 228     qemu_bh_delete(s->bh);
 229     g_free(s);
 230 }
 231
 232 /* Context: QEMU global mutex held */
 233 void virtio_blk_data_plane_start(VirtIOBlockDataPlane *s)
 234 {
 235     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(s->vdev)));
 236     VirtioBusClass *k = VIRTIO_BUS_GET_CLASS(qbus);
 237     VirtIOBlock *vblk = VIRTIO_BLK(s->vdev);
 238     VirtQueue *vq;
 239     int r;
 240
 241     if (s->started || s->disabled) {
 242         return;
 243     }
 244
 245     if (s->starting) {
 246         return;
 247     }
 248
 249     s->starting = true;
 250
 251     vq = virtio_get_queue(s->vdev, 0);
 252     if (!vring_setup(&s->vring, s->vdev, 0)) {
 253         goto fail_vring;
 254     }
 255
 256     /* Set up guest notifier (irq) */
 257     r = k->set_guest_notifiers(qbus->parent, 1, true);
 258     if (r != 0) {
 259         fprintf(stderr, "virtio-blk failed to set guest notifier (%d), "
 260                 "ensure -enable-kvm is set\n", r);
 261         goto fail_guest_notifiers;
 262     }
 263     s->guest_notifier = virtio_queue_get_guest_notifier(vq);
 264
 265     /* Set up virtqueue notify */
 266     r = k->set_host_notifier(qbus->parent, 0, true);
 267     if (r != 0) {
 268         fprintf(stderr, "virtio-blk failed to set host notifier (%d)\n", r);
 269         goto fail_host_notifier;
 270     }
 271     s->host_notifier = *virtio_queue_get_host_notifier(vq);
 272
 273     s->saved_complete_request = vblk->complete_request;
 274     vblk->complete_request = complete_request_vring;
 275
 276     s->starting = false;
 277     s->started = true;
 278     trace_virtio_blk_data_plane_start(s);
 279
 280     blk_set_aio_context(s->conf->conf.blk, s->ctx);
 281
 282     /* Kick right away to begin processing requests already in vring */
 283     event_notifier_set(virtio_queue_get_host_notifier(vq));
 284
 285     /* Get this show started by hooking up our callbacks */
 286     aio_context_acquire(s->ctx);
 287     aio_set_event_notifier(s->ctx, &s->host_notifier, handle_notify);
 288     aio_context_release(s->ctx);
 289     return;
 290
 291   fail_host_notifier:
 292     k->set_guest_notifiers(qbus->parent, 1, false);
 293   fail_guest_notifiers:
 294     vring_teardown(&s->vring, s->vdev, 0);
 295     s->disabled = true;
 296   fail_vring:
 297     s->starting = false;
 298 }
 299
 300 /* Context: QEMU global mutex held */
 301 void virtio_blk_data_plane_stop(VirtIOBlockDataPlane *s)
 302 {
 303     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(s->vdev)));
 304     VirtioBusClass *k = VIRTIO_BUS_GET_CLASS(qbus);
 305     VirtIOBlock *vblk = VIRTIO_BLK(s->vdev);
 306
 307
 308     /* Better luck next time. */
 309     if (s->disabled) {
 310         s->disabled = false;
 311         return;
 312     }
 313     if (!s->started || s->stopping) {
 314         return;
 315     }
 316     s->stopping = true;
 317     vblk->complete_request = s->saved_complete_request;
 318     trace_virtio_blk_data_plane_stop(s);
 319
 320     aio_context_acquire(s->ctx);
 321
 322     /* Stop notifications for new requests from guest */
 323     aio_set_event_notifier(s->ctx, &s->host_notifier, NULL);
 324
 325     /* Drain and switch bs back to the QEMU main loop */
 326     blk_set_aio_context(s->conf->conf.blk, qemu_get_aio_context());
 327
 328     aio_context_release(s->ctx);
 329
 330     /* Sync vring state back to virtqueue so that non-dataplane request
 331      * processing can continue when we disable the host notifier below.
 332      */
 333     vring_teardown(&s->vring, s->vdev, 0);
 334
 335     k->set_host_notifier(qbus->parent, 0, false);
 336
 337     /* Clean up guest notifier (irq) */
 338     k->set_guest_notifiers(qbus->parent, 1, false);
 339
 340     s->started = false;
 341     s->stopping = false;
 342 }