2 * Dedicated thread for virtio-blk I/O processing
4 * Copyright 2012 IBM, Corp.
5 * Copyright 2012 Red Hat, Inc. and/or its affiliates
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.
17 #include "event-poll.h"
18 #include "qemu/thread.h"
19 #include "qemu/error-report.h"
22 #include "migration/migration.h"
23 #include "block/block.h"
24 #include "hw/virtio-blk.h"
25 #include "hw/dataplane/virtio-blk.h"
28 SEG_MAX = 126, /* maximum number of I/O segments */
29 VRING_MAX = SEG_MAX + 2, /* maximum number of vring descriptors */
30 REQ_MAX = VRING_MAX, /* maximum number of requests in the vring,
31 * is VRING_MAX / 2 with traditional and
32 * VRING_MAX with indirect descriptors */
36 struct iocb iocb; /* Linux AIO control block */
37 QEMUIOVector *inhdr; /* iovecs for virtio_blk_inhdr */
38 unsigned int head; /* vring descriptor index */
39 struct iovec *bounce_iov; /* used if guest buffers are unaligned */
40 QEMUIOVector *read_qiov; /* for read completion /w bounce buffer */
43 struct VirtIOBlockDataPlane {
50 int fd; /* image file descriptor */
53 Vring vring; /* virtqueue vring */
54 EventNotifier *guest_notifier; /* irq */
56 EventPoll event_poll; /* event poller */
57 EventHandler io_handler; /* Linux AIO completion handler */
58 EventHandler notify_handler; /* virtqueue notify handler */
60 IOQueue ioqueue; /* Linux AIO queue (should really be per
62 VirtIOBlockRequest requests[REQ_MAX]; /* pool of requests, managed by the
65 unsigned int num_reqs;
67 Error *migration_blocker;
70 /* Raise an interrupt to signal guest, if necessary */
71 static void notify_guest(VirtIOBlockDataPlane *s)
73 if (!vring_should_notify(s->vdev, &s->vring)) {
77 event_notifier_set(s->guest_notifier);
80 static void complete_request(struct iocb *iocb, ssize_t ret, void *opaque)
82 VirtIOBlockDataPlane *s = opaque;
83 VirtIOBlockRequest *req = container_of(iocb, VirtIOBlockRequest, iocb);
84 struct virtio_blk_inhdr hdr;
87 if (likely(ret >= 0)) {
88 hdr.status = VIRTIO_BLK_S_OK;
91 hdr.status = VIRTIO_BLK_S_IOERR;
95 trace_virtio_blk_data_plane_complete_request(s, req->head, ret);
98 assert(req->bounce_iov);
99 qemu_iovec_from_buf(req->read_qiov, 0, req->bounce_iov->iov_base, len);
100 qemu_iovec_destroy(req->read_qiov);
101 g_slice_free(QEMUIOVector, req->read_qiov);
104 if (req->bounce_iov) {
105 qemu_vfree(req->bounce_iov->iov_base);
106 g_slice_free(struct iovec, req->bounce_iov);
109 qemu_iovec_from_buf(req->inhdr, 0, &hdr, sizeof(hdr));
110 qemu_iovec_destroy(req->inhdr);
111 g_slice_free(QEMUIOVector, req->inhdr);
113 /* According to the virtio specification len should be the number of bytes
114 * written to, but for virtio-blk it seems to be the number of bytes
115 * transferred plus the status bytes.
117 vring_push(&s->vring, req->head, len + sizeof(hdr));
122 static void complete_request_early(VirtIOBlockDataPlane *s, unsigned int head,
123 QEMUIOVector *inhdr, unsigned char status)
125 struct virtio_blk_inhdr hdr = {
129 qemu_iovec_from_buf(inhdr, 0, &hdr, sizeof(hdr));
130 qemu_iovec_destroy(inhdr);
131 g_slice_free(QEMUIOVector, inhdr);
133 vring_push(&s->vring, head, sizeof(hdr));
137 /* Get disk serial number */
138 static void do_get_id_cmd(VirtIOBlockDataPlane *s,
139 struct iovec *iov, unsigned int iov_cnt,
140 unsigned int head, QEMUIOVector *inhdr)
142 char id[VIRTIO_BLK_ID_BYTES];
144 /* Serial number not NUL-terminated when shorter than buffer */
145 strncpy(id, s->blk->serial ? s->blk->serial : "", sizeof(id));
146 iov_from_buf(iov, iov_cnt, 0, id, sizeof(id));
147 complete_request_early(s, head, inhdr, VIRTIO_BLK_S_OK);
150 static int do_rdwr_cmd(VirtIOBlockDataPlane *s, bool read,
151 struct iovec *iov, unsigned int iov_cnt,
152 long long offset, unsigned int head,
157 struct iovec *bounce_iov = NULL;
158 QEMUIOVector *read_qiov = NULL;
160 qemu_iovec_init_external(&qiov, iov, iov_cnt);
161 if (!bdrv_qiov_is_aligned(s->blk->conf.bs, &qiov)) {
162 void *bounce_buffer = qemu_blockalign(s->blk->conf.bs, qiov.size);
165 /* Need to copy back from bounce buffer on completion */
166 read_qiov = g_slice_new(QEMUIOVector);
167 qemu_iovec_init(read_qiov, iov_cnt);
168 qemu_iovec_concat_iov(read_qiov, iov, iov_cnt, 0, qiov.size);
170 qemu_iovec_to_buf(&qiov, 0, bounce_buffer, qiov.size);
173 /* Redirect I/O to aligned bounce buffer */
174 bounce_iov = g_slice_new(struct iovec);
175 bounce_iov->iov_base = bounce_buffer;
176 bounce_iov->iov_len = qiov.size;
181 iocb = ioq_rdwr(&s->ioqueue, read, iov, iov_cnt, offset);
183 /* Fill in virtio block metadata needed for completion */
184 VirtIOBlockRequest *req = container_of(iocb, VirtIOBlockRequest, iocb);
187 req->bounce_iov = bounce_iov;
188 req->read_qiov = read_qiov;
192 static int process_request(IOQueue *ioq, struct iovec iov[],
193 unsigned int out_num, unsigned int in_num,
196 VirtIOBlockDataPlane *s = container_of(ioq, VirtIOBlockDataPlane, ioqueue);
197 struct iovec *in_iov = &iov[out_num];
198 struct virtio_blk_outhdr outhdr;
203 if (unlikely(iov_to_buf(iov, out_num, 0, &outhdr,
204 sizeof(outhdr)) != sizeof(outhdr))) {
205 error_report("virtio-blk request outhdr too short");
208 iov_discard_front(&iov, &out_num, sizeof(outhdr));
210 /* Grab inhdr for later */
211 in_size = iov_size(in_iov, in_num);
212 if (in_size < sizeof(struct virtio_blk_inhdr)) {
213 error_report("virtio_blk request inhdr too short");
216 inhdr = g_slice_new(QEMUIOVector);
217 qemu_iovec_init(inhdr, 1);
218 qemu_iovec_concat_iov(inhdr, in_iov, in_num,
219 in_size - sizeof(struct virtio_blk_inhdr),
220 sizeof(struct virtio_blk_inhdr));
221 iov_discard_back(in_iov, &in_num, sizeof(struct virtio_blk_inhdr));
223 /* TODO Linux sets the barrier bit even when not advertised! */
224 outhdr.type &= ~VIRTIO_BLK_T_BARRIER;
226 switch (outhdr.type) {
227 case VIRTIO_BLK_T_IN:
228 do_rdwr_cmd(s, true, in_iov, in_num, outhdr.sector * 512, head, inhdr);
231 case VIRTIO_BLK_T_OUT:
232 do_rdwr_cmd(s, false, iov, out_num, outhdr.sector * 512, head, inhdr);
235 case VIRTIO_BLK_T_SCSI_CMD:
236 /* TODO support SCSI commands */
237 complete_request_early(s, head, inhdr, VIRTIO_BLK_S_UNSUPP);
240 case VIRTIO_BLK_T_FLUSH:
241 /* TODO fdsync not supported by Linux AIO, do it synchronously here! */
242 if (qemu_fdatasync(s->fd) < 0) {
243 complete_request_early(s, head, inhdr, VIRTIO_BLK_S_IOERR);
245 complete_request_early(s, head, inhdr, VIRTIO_BLK_S_OK);
249 case VIRTIO_BLK_T_GET_ID:
250 do_get_id_cmd(s, in_iov, in_num, head, inhdr);
254 error_report("virtio-blk unsupported request type %#x", outhdr.type);
255 qemu_iovec_destroy(inhdr);
256 g_slice_free(QEMUIOVector, inhdr);
261 static void handle_notify(EventHandler *handler)
263 VirtIOBlockDataPlane *s = container_of(handler, VirtIOBlockDataPlane,
266 /* There is one array of iovecs into which all new requests are extracted
267 * from the vring. Requests are read from the vring and the translated
268 * descriptors are written to the iovecs array. The iovecs do not have to
269 * persist across handle_notify() calls because the kernel copies the
270 * iovecs on io_submit().
272 * Handling io_submit() EAGAIN may require storing the requests across
273 * handle_notify() calls until the kernel has sufficient resources to
274 * accept more I/O. This is not implemented yet.
276 struct iovec iovec[VRING_MAX];
277 struct iovec *end = &iovec[VRING_MAX];
278 struct iovec *iov = iovec;
280 /* When a request is read from the vring, the index of the first descriptor
281 * (aka head) is returned so that the completed request can be pushed onto
284 * The number of hypervisor read-only iovecs is out_num. The number of
285 * hypervisor write-only iovecs is in_num.
288 unsigned int out_num = 0, in_num = 0;
289 unsigned int num_queued;
292 /* Disable guest->host notifies to avoid unnecessary vmexits */
293 vring_disable_notification(s->vdev, &s->vring);
296 head = vring_pop(s->vdev, &s->vring, iov, end, &out_num, &in_num);
298 break; /* no more requests */
301 trace_virtio_blk_data_plane_process_request(s, out_num, in_num,
304 if (process_request(&s->ioqueue, iov, out_num, in_num, head) < 0) {
305 vring_set_broken(&s->vring);
308 iov += out_num + in_num;
311 if (likely(head == -EAGAIN)) { /* vring emptied */
312 /* Re-enable guest->host notifies and stop processing the vring.
313 * But if the guest has snuck in more descriptors, keep processing.
315 if (vring_enable_notification(s->vdev, &s->vring)) {
318 } else { /* head == -ENOBUFS or fatal error, iovecs[] is depleted */
319 /* Since there are no iovecs[] left, stop processing for now. Do
320 * not re-enable guest->host notifies since the I/O completion
321 * handler knows to check for more vring descriptors anyway.
327 num_queued = ioq_num_queued(&s->ioqueue);
328 if (num_queued > 0) {
329 s->num_reqs += num_queued;
331 int rc = ioq_submit(&s->ioqueue);
332 if (unlikely(rc < 0)) {
333 fprintf(stderr, "ioq_submit failed %d\n", rc);
339 static void handle_io(EventHandler *handler)
341 VirtIOBlockDataPlane *s = container_of(handler, VirtIOBlockDataPlane,
344 if (ioq_run_completion(&s->ioqueue, complete_request, s) > 0) {
348 /* If there were more requests than iovecs, the vring will not be empty yet
349 * so check again. There should now be enough resources to process more
352 if (unlikely(vring_more_avail(&s->vring))) {
353 handle_notify(&s->notify_handler);
357 static void *data_plane_thread(void *opaque)
359 VirtIOBlockDataPlane *s = opaque;
362 event_poll(&s->event_poll);
363 } while (!s->stopping || s->num_reqs > 0);
367 static void start_data_plane_bh(void *opaque)
369 VirtIOBlockDataPlane *s = opaque;
371 qemu_bh_delete(s->start_bh);
373 qemu_thread_create(&s->thread, data_plane_thread,
374 s, QEMU_THREAD_JOINABLE);
377 bool virtio_blk_data_plane_create(VirtIODevice *vdev, VirtIOBlkConf *blk,
378 VirtIOBlockDataPlane **dataplane)
380 VirtIOBlockDataPlane *s;
385 if (!blk->data_plane) {
390 error_report("device is incompatible with x-data-plane, use scsi=off");
394 if (blk->config_wce) {
395 error_report("device is incompatible with x-data-plane, "
396 "use config-wce=off");
400 fd = raw_get_aio_fd(blk->conf.bs);
402 error_report("drive is incompatible with x-data-plane, "
403 "use format=raw,cache=none,aio=native");
407 s = g_new0(VirtIOBlockDataPlane, 1);
412 /* Prevent block operations that conflict with data plane thread */
413 bdrv_set_in_use(blk->conf.bs, 1);
415 error_setg(&s->migration_blocker,
416 "x-data-plane does not support migration");
417 migrate_add_blocker(s->migration_blocker);
423 void virtio_blk_data_plane_destroy(VirtIOBlockDataPlane *s)
429 virtio_blk_data_plane_stop(s);
430 migrate_del_blocker(s->migration_blocker);
431 error_free(s->migration_blocker);
432 bdrv_set_in_use(s->blk->conf.bs, 0);
436 void virtio_blk_data_plane_start(VirtIOBlockDataPlane *s)
445 vq = virtio_get_queue(s->vdev, 0);
446 if (!vring_setup(&s->vring, s->vdev, 0)) {
450 event_poll_init(&s->event_poll);
452 /* Set up guest notifier (irq) */
453 if (s->vdev->binding->set_guest_notifiers(s->vdev->binding_opaque, 1,
455 fprintf(stderr, "virtio-blk failed to set guest notifier, "
456 "ensure -enable-kvm is set\n");
459 s->guest_notifier = virtio_queue_get_guest_notifier(vq);
461 /* Set up virtqueue notify */
462 if (s->vdev->binding->set_host_notifier(s->vdev->binding_opaque,
464 fprintf(stderr, "virtio-blk failed to set host notifier\n");
467 event_poll_add(&s->event_poll, &s->notify_handler,
468 virtio_queue_get_host_notifier(vq),
472 ioq_init(&s->ioqueue, s->fd, REQ_MAX);
473 for (i = 0; i < ARRAY_SIZE(s->requests); i++) {
474 ioq_put_iocb(&s->ioqueue, &s->requests[i].iocb);
476 event_poll_add(&s->event_poll, &s->io_handler,
477 ioq_get_notifier(&s->ioqueue), handle_io);
480 trace_virtio_blk_data_plane_start(s);
482 /* Kick right away to begin processing requests already in vring */
483 event_notifier_set(virtio_queue_get_host_notifier(vq));
485 /* Spawn thread in BH so it inherits iothread cpusets */
486 s->start_bh = qemu_bh_new(start_data_plane_bh, s);
487 qemu_bh_schedule(s->start_bh);
490 void virtio_blk_data_plane_stop(VirtIOBlockDataPlane *s)
492 if (!s->started || s->stopping) {
496 trace_virtio_blk_data_plane_stop(s);
498 /* Stop thread or cancel pending thread creation BH */
500 qemu_bh_delete(s->start_bh);
503 event_poll_notify(&s->event_poll);
504 qemu_thread_join(&s->thread);
507 ioq_cleanup(&s->ioqueue);
509 s->vdev->binding->set_host_notifier(s->vdev->binding_opaque, 0, false);
511 event_poll_cleanup(&s->event_poll);
513 /* Clean up guest notifier (irq) */
514 s->vdev->binding->set_guest_notifiers(s->vdev->binding_opaque, 1, false);
516 vring_teardown(&s->vring);