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