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     AioContext *ctx;
  49     EventNotifier host_notifier;    /* doorbell */
  50
  51     /* Operation blocker on BDS */
  52     Error *blocker;
  53     void (*saved_complete_request)(struct VirtIOBlockReq *req,
  54                                    unsigned char status);
  55 };
  56
  57 /* Raise an interrupt to signal guest, if necessary */
  58 static void notify_guest(VirtIOBlockDataPlane *s)
  59 {
  60     if (!vring_should_notify(s->vdev, &s->vring)) {
  61         return;
  62     }
  63
  64     event_notifier_set(s->guest_notifier);
  65 }
  66
  67 static void notify_guest_bh(void *opaque)
  68 {
  69     VirtIOBlockDataPlane *s = opaque;
  70
  71     notify_guest(s);
  72 }
  73
  74 static void complete_request_vring(VirtIOBlockReq *req, unsigned char status)
  75 {
  76     VirtIOBlockDataPlane *s = req->dev->dataplane;
  77     stb_p(&req->in->status, status);
  78
  79     vring_push(s->vdev, &req->dev->dataplane->vring, &req->elem, req->in_len);
  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         int ret;
 101
 102         /* Disable guest->host notifies to avoid unnecessary vmexits */
 103         vring_disable_notification(s->vdev, &s->vring);
 104
 105         for (;;) {
 106             VirtIOBlockReq *req = virtio_blk_alloc_request(vblk);
 107
 108             ret = vring_pop(s->vdev, &s->vring, &req->elem);
 109             if (ret < 0) {
 110                 virtio_blk_free_request(req);
 111                 break; /* no more requests */
 112             }
 113
 114             trace_virtio_blk_data_plane_process_request(s, req->elem.out_num,
 115                                                         req->elem.in_num,
 116                                                         req->elem.index);
 117
 118             virtio_blk_handle_request(req, &mrb);
 119         }
 120
 121         if (mrb.num_reqs) {
 122             virtio_blk_submit_multireq(s->conf->conf.blk, &mrb);
 123         }
 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->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 dataplane "
 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     }
 182     s->ctx = iothread_get_aio_context(s->iothread);
 183     s->bh = aio_bh_new(s->ctx, notify_guest_bh, s);
 184
 185     error_setg(&s->blocker, "block device is in use by data plane");
 186     blk_op_block_all(conf->conf.blk, s->blocker);
 187     blk_op_unblock(conf->conf.blk, BLOCK_OP_TYPE_RESIZE, s->blocker);
 188     blk_op_unblock(conf->conf.blk, BLOCK_OP_TYPE_DRIVE_DEL, s->blocker);
 189     blk_op_unblock(conf->conf.blk, BLOCK_OP_TYPE_BACKUP_SOURCE, s->blocker);
 190     blk_op_unblock(conf->conf.blk, BLOCK_OP_TYPE_CHANGE, s->blocker);
 191     blk_op_unblock(conf->conf.blk, BLOCK_OP_TYPE_COMMIT_SOURCE, s->blocker);
 192     blk_op_unblock(conf->conf.blk, BLOCK_OP_TYPE_COMMIT_TARGET, s->blocker);
 193     blk_op_unblock(conf->conf.blk, BLOCK_OP_TYPE_EJECT, s->blocker);
 194     blk_op_unblock(conf->conf.blk, BLOCK_OP_TYPE_EXTERNAL_SNAPSHOT, s->blocker);
 195     blk_op_unblock(conf->conf.blk, BLOCK_OP_TYPE_INTERNAL_SNAPSHOT, s->blocker);
 196     blk_op_unblock(conf->conf.blk, BLOCK_OP_TYPE_INTERNAL_SNAPSHOT_DELETE,
 197                    s->blocker);
 198     blk_op_unblock(conf->conf.blk, BLOCK_OP_TYPE_MIRROR_SOURCE, s->blocker);
 199     blk_op_unblock(conf->conf.blk, BLOCK_OP_TYPE_STREAM, s->blocker);
 200     blk_op_unblock(conf->conf.blk, BLOCK_OP_TYPE_REPLACE, s->blocker);
 201
 202     *dataplane = s;
 203 }
 204
 205 /* Context: QEMU global mutex held */
 206 void virtio_blk_data_plane_destroy(VirtIOBlockDataPlane *s)
 207 {
 208     if (!s) {
 209         return;
 210     }
 211
 212     virtio_blk_data_plane_stop(s);
 213     blk_op_unblock_all(s->conf->conf.blk, s->blocker);
 214     error_free(s->blocker);
 215     qemu_bh_delete(s->bh);
 216     object_unref(OBJECT(s->iothread));
 217     g_free(s);
 218 }
 219
 220 /* Context: QEMU global mutex held */
 221 void virtio_blk_data_plane_start(VirtIOBlockDataPlane *s)
 222 {
 223     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(s->vdev)));
 224     VirtioBusClass *k = VIRTIO_BUS_GET_CLASS(qbus);
 225     VirtIOBlock *vblk = VIRTIO_BLK(s->vdev);
 226     VirtQueue *vq;
 227     int r;
 228
 229     if (s->started || s->disabled) {
 230         return;
 231     }
 232
 233     if (s->starting) {
 234         return;
 235     }
 236
 237     s->starting = true;
 238
 239     vq = virtio_get_queue(s->vdev, 0);
 240     if (!vring_setup(&s->vring, s->vdev, 0)) {
 241         goto fail_vring;
 242     }
 243
 244     /* Set up guest notifier (irq) */
 245     r = k->set_guest_notifiers(qbus->parent, 1, true);
 246     if (r != 0) {
 247         fprintf(stderr, "virtio-blk failed to set guest notifier (%d), "
 248                 "ensure -enable-kvm is set\n", r);
 249         goto fail_guest_notifiers;
 250     }
 251     s->guest_notifier = virtio_queue_get_guest_notifier(vq);
 252
 253     /* Set up virtqueue notify */
 254     r = k->set_host_notifier(qbus->parent, 0, true);
 255     if (r != 0) {
 256         fprintf(stderr, "virtio-blk failed to set host notifier (%d)\n", r);
 257         goto fail_host_notifier;
 258     }
 259     s->host_notifier = *virtio_queue_get_host_notifier(vq);
 260
 261     s->saved_complete_request = vblk->complete_request;
 262     vblk->complete_request = complete_request_vring;
 263
 264     s->starting = false;
 265     s->started = true;
 266     trace_virtio_blk_data_plane_start(s);
 267
 268     blk_set_aio_context(s->conf->conf.blk, s->ctx);
 269
 270     /* Kick right away to begin processing requests already in vring */
 271     event_notifier_set(virtio_queue_get_host_notifier(vq));
 272
 273     /* Get this show started by hooking up our callbacks */
 274     aio_context_acquire(s->ctx);
 275     aio_set_event_notifier(s->ctx, &s->host_notifier, true,
 276                            handle_notify);
 277     aio_context_release(s->ctx);
 278     return;
 279
 280   fail_host_notifier:
 281     k->set_guest_notifiers(qbus->parent, 1, false);
 282   fail_guest_notifiers:
 283     vring_teardown(&s->vring, s->vdev, 0);
 284     s->disabled = true;
 285   fail_vring:
 286     s->starting = false;
 287 }
 288
 289 /* Context: QEMU global mutex held */
 290 void virtio_blk_data_plane_stop(VirtIOBlockDataPlane *s)
 291 {
 292     BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(s->vdev)));
 293     VirtioBusClass *k = VIRTIO_BUS_GET_CLASS(qbus);
 294     VirtIOBlock *vblk = VIRTIO_BLK(s->vdev);
 295
 296
 297     /* Better luck next time. */
 298     if (s->disabled) {
 299         s->disabled = false;
 300         return;
 301     }
 302     if (!s->started || s->stopping) {
 303         return;
 304     }
 305     s->stopping = true;
 306     vblk->complete_request = s->saved_complete_request;
 307     trace_virtio_blk_data_plane_stop(s);
 308
 309     aio_context_acquire(s->ctx);
 310
 311     /* Stop notifications for new requests from guest */
 312     aio_set_event_notifier(s->ctx, &s->host_notifier, true, NULL);
 313
 314     /* Drain and switch bs back to the QEMU main loop */
 315     blk_set_aio_context(s->conf->conf.blk, qemu_get_aio_context());
 316
 317     aio_context_release(s->ctx);
 318
 319     /* Sync vring state back to virtqueue so that non-dataplane request
 320      * processing can continue when we disable the host notifier below.
 321      */
 322     vring_teardown(&s->vring, s->vdev, 0);
 323
 324     k->set_host_notifier(qbus->parent, 0, false);
 325
 326     /* Clean up guest notifier (irq) */
 327     k->set_guest_notifiers(qbus->parent, 1, false);
 328
 329     s->started = false;
 330     s->stopping = false;
 331 }