2
* Dedicated thread for virtio-blk I/O processing
4
* Copyright 2012 IBM, Corp.
5
* Copyright 2012 Red Hat, Inc. and/or its affiliates
8
* Stefan Hajnoczi <stefanha@redhat.com>
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 "qemu/thread.h"
18
#include "qemu/error-report.h"
19
#include "hw/virtio/dataplane/vring.h"
21
#include "block/block.h"
22
#include "hw/virtio/virtio-blk.h"
23
#include "virtio-blk.h"
24
#include "block/aio.h"
25
#include "hw/virtio/virtio-bus.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 {
51
int fd; /* image file descriptor */
54
Vring vring; /* virtqueue vring */
55
EventNotifier *guest_notifier; /* irq */
57
/* Note that these EventNotifiers are assigned by value. This is
58
* fine as long as you do not call event_notifier_cleanup on them
59
* (because you don't own the file descriptor or handle; you just
63
EventNotifier io_notifier; /* Linux AIO completion */
64
EventNotifier host_notifier; /* doorbell */
66
IOQueue ioqueue; /* Linux AIO queue (should really be per
68
VirtIOBlockRequest requests[REQ_MAX]; /* pool of requests, managed by the
71
unsigned int num_reqs;
74
/* Raise an interrupt to signal guest, if necessary */
75
static void notify_guest(VirtIOBlockDataPlane *s)
77
if (!vring_should_notify(s->vdev, &s->vring)) {
81
event_notifier_set(s->guest_notifier);
84
static void complete_request(struct iocb *iocb, ssize_t ret, void *opaque)
86
VirtIOBlockDataPlane *s = opaque;
87
VirtIOBlockRequest *req = container_of(iocb, VirtIOBlockRequest, iocb);
88
struct virtio_blk_inhdr hdr;
91
if (likely(ret >= 0)) {
92
hdr.status = VIRTIO_BLK_S_OK;
95
hdr.status = VIRTIO_BLK_S_IOERR;
99
trace_virtio_blk_data_plane_complete_request(s, req->head, ret);
101
if (req->read_qiov) {
102
assert(req->bounce_iov);
103
qemu_iovec_from_buf(req->read_qiov, 0, req->bounce_iov->iov_base, len);
104
qemu_iovec_destroy(req->read_qiov);
105
g_slice_free(QEMUIOVector, req->read_qiov);
108
if (req->bounce_iov) {
109
qemu_vfree(req->bounce_iov->iov_base);
110
g_slice_free(struct iovec, req->bounce_iov);
113
qemu_iovec_from_buf(req->inhdr, 0, &hdr, sizeof(hdr));
114
qemu_iovec_destroy(req->inhdr);
115
g_slice_free(QEMUIOVector, req->inhdr);
117
/* According to the virtio specification len should be the number of bytes
118
* written to, but for virtio-blk it seems to be the number of bytes
119
* transferred plus the status bytes.
121
vring_push(&s->vring, req->head, len + sizeof(hdr));
126
static void complete_request_early(VirtIOBlockDataPlane *s, unsigned int head,
127
QEMUIOVector *inhdr, unsigned char status)
129
struct virtio_blk_inhdr hdr = {
133
qemu_iovec_from_buf(inhdr, 0, &hdr, sizeof(hdr));
134
qemu_iovec_destroy(inhdr);
135
g_slice_free(QEMUIOVector, inhdr);
137
vring_push(&s->vring, head, sizeof(hdr));
141
/* Get disk serial number */
142
static void do_get_id_cmd(VirtIOBlockDataPlane *s,
143
struct iovec *iov, unsigned int iov_cnt,
144
unsigned int head, QEMUIOVector *inhdr)
146
char id[VIRTIO_BLK_ID_BYTES];
148
/* Serial number not NUL-terminated when shorter than buffer */
149
strncpy(id, s->blk->serial ? s->blk->serial : "", sizeof(id));
150
iov_from_buf(iov, iov_cnt, 0, id, sizeof(id));
151
complete_request_early(s, head, inhdr, VIRTIO_BLK_S_OK);
154
static int do_rdwr_cmd(VirtIOBlockDataPlane *s, bool read,
155
struct iovec *iov, unsigned int iov_cnt,
156
long long offset, unsigned int head,
161
struct iovec *bounce_iov = NULL;
162
QEMUIOVector *read_qiov = NULL;
164
qemu_iovec_init_external(&qiov, iov, iov_cnt);
165
if (!bdrv_qiov_is_aligned(s->blk->conf.bs, &qiov)) {
166
void *bounce_buffer = qemu_blockalign(s->blk->conf.bs, qiov.size);
169
/* Need to copy back from bounce buffer on completion */
170
read_qiov = g_slice_new(QEMUIOVector);
171
qemu_iovec_init(read_qiov, iov_cnt);
172
qemu_iovec_concat_iov(read_qiov, iov, iov_cnt, 0, qiov.size);
174
qemu_iovec_to_buf(&qiov, 0, bounce_buffer, qiov.size);
177
/* Redirect I/O to aligned bounce buffer */
178
bounce_iov = g_slice_new(struct iovec);
179
bounce_iov->iov_base = bounce_buffer;
180
bounce_iov->iov_len = qiov.size;
185
iocb = ioq_rdwr(&s->ioqueue, read, iov, iov_cnt, offset);
187
/* Fill in virtio block metadata needed for completion */
188
VirtIOBlockRequest *req = container_of(iocb, VirtIOBlockRequest, iocb);
191
req->bounce_iov = bounce_iov;
192
req->read_qiov = read_qiov;
196
static int process_request(IOQueue *ioq, struct iovec iov[],
197
unsigned int out_num, unsigned int in_num,
200
VirtIOBlockDataPlane *s = container_of(ioq, VirtIOBlockDataPlane, ioqueue);
201
struct iovec *in_iov = &iov[out_num];
202
struct virtio_blk_outhdr outhdr;
207
if (unlikely(iov_to_buf(iov, out_num, 0, &outhdr,
208
sizeof(outhdr)) != sizeof(outhdr))) {
209
error_report("virtio-blk request outhdr too short");
212
iov_discard_front(&iov, &out_num, sizeof(outhdr));
214
/* Grab inhdr for later */
215
in_size = iov_size(in_iov, in_num);
216
if (in_size < sizeof(struct virtio_blk_inhdr)) {
217
error_report("virtio_blk request inhdr too short");
220
inhdr = g_slice_new(QEMUIOVector);
221
qemu_iovec_init(inhdr, 1);
222
qemu_iovec_concat_iov(inhdr, in_iov, in_num,
223
in_size - sizeof(struct virtio_blk_inhdr),
224
sizeof(struct virtio_blk_inhdr));
225
iov_discard_back(in_iov, &in_num, sizeof(struct virtio_blk_inhdr));
227
/* TODO Linux sets the barrier bit even when not advertised! */
228
outhdr.type &= ~VIRTIO_BLK_T_BARRIER;
230
switch (outhdr.type) {
231
case VIRTIO_BLK_T_IN:
232
do_rdwr_cmd(s, true, in_iov, in_num, outhdr.sector * 512, head, inhdr);
235
case VIRTIO_BLK_T_OUT:
236
do_rdwr_cmd(s, false, iov, out_num, outhdr.sector * 512, head, inhdr);
239
case VIRTIO_BLK_T_SCSI_CMD:
240
/* TODO support SCSI commands */
241
complete_request_early(s, head, inhdr, VIRTIO_BLK_S_UNSUPP);
244
case VIRTIO_BLK_T_FLUSH:
245
/* TODO fdsync not supported by Linux AIO, do it synchronously here! */
246
if (qemu_fdatasync(s->fd) < 0) {
247
complete_request_early(s, head, inhdr, VIRTIO_BLK_S_IOERR);
249
complete_request_early(s, head, inhdr, VIRTIO_BLK_S_OK);
253
case VIRTIO_BLK_T_GET_ID:
254
do_get_id_cmd(s, in_iov, in_num, head, inhdr);
258
error_report("virtio-blk unsupported request type %#x", outhdr.type);
259
qemu_iovec_destroy(inhdr);
260
g_slice_free(QEMUIOVector, inhdr);
265
static void handle_notify(EventNotifier *e)
267
VirtIOBlockDataPlane *s = container_of(e, VirtIOBlockDataPlane,
270
/* There is one array of iovecs into which all new requests are extracted
271
* from the vring. Requests are read from the vring and the translated
272
* descriptors are written to the iovecs array. The iovecs do not have to
273
* persist across handle_notify() calls because the kernel copies the
274
* iovecs on io_submit().
276
* Handling io_submit() EAGAIN may require storing the requests across
277
* handle_notify() calls until the kernel has sufficient resources to
278
* accept more I/O. This is not implemented yet.
280
struct iovec iovec[VRING_MAX];
281
struct iovec *end = &iovec[VRING_MAX];
282
struct iovec *iov = iovec;
284
/* When a request is read from the vring, the index of the first descriptor
285
* (aka head) is returned so that the completed request can be pushed onto
288
* The number of hypervisor read-only iovecs is out_num. The number of
289
* hypervisor write-only iovecs is in_num.
292
unsigned int out_num = 0, in_num = 0;
293
unsigned int num_queued;
295
event_notifier_test_and_clear(&s->host_notifier);
297
/* Disable guest->host notifies to avoid unnecessary vmexits */
298
vring_disable_notification(s->vdev, &s->vring);
301
head = vring_pop(s->vdev, &s->vring, iov, end, &out_num, &in_num);
303
break; /* no more requests */
306
trace_virtio_blk_data_plane_process_request(s, out_num, in_num,
309
if (process_request(&s->ioqueue, iov, out_num, in_num, head) < 0) {
310
vring_set_broken(&s->vring);
313
iov += out_num + in_num;
316
if (likely(head == -EAGAIN)) { /* vring emptied */
317
/* Re-enable guest->host notifies and stop processing the vring.
318
* But if the guest has snuck in more descriptors, keep processing.
320
if (vring_enable_notification(s->vdev, &s->vring)) {
323
} else { /* head == -ENOBUFS or fatal error, iovecs[] is depleted */
324
/* Since there are no iovecs[] left, stop processing for now. Do
325
* not re-enable guest->host notifies since the I/O completion
326
* handler knows to check for more vring descriptors anyway.
332
num_queued = ioq_num_queued(&s->ioqueue);
333
if (num_queued > 0) {
334
s->num_reqs += num_queued;
336
int rc = ioq_submit(&s->ioqueue);
337
if (unlikely(rc < 0)) {
338
fprintf(stderr, "ioq_submit failed %d\n", rc);
344
static void handle_io(EventNotifier *e)
346
VirtIOBlockDataPlane *s = container_of(e, VirtIOBlockDataPlane,
349
event_notifier_test_and_clear(&s->io_notifier);
350
if (ioq_run_completion(&s->ioqueue, complete_request, s) > 0) {
354
/* If there were more requests than iovecs, the vring will not be empty yet
355
* so check again. There should now be enough resources to process more
358
if (unlikely(vring_more_avail(&s->vring))) {
359
handle_notify(&s->host_notifier);
363
static void *data_plane_thread(void *opaque)
365
VirtIOBlockDataPlane *s = opaque;
367
while (!s->stopping || s->num_reqs > 0) {
368
aio_poll(s->ctx, true);
373
static void start_data_plane_bh(void *opaque)
375
VirtIOBlockDataPlane *s = opaque;
377
qemu_bh_delete(s->start_bh);
379
qemu_thread_create(&s->thread, data_plane_thread,
380
s, QEMU_THREAD_JOINABLE);
383
bool virtio_blk_data_plane_create(VirtIODevice *vdev, VirtIOBlkConf *blk,
384
VirtIOBlockDataPlane **dataplane)
386
VirtIOBlockDataPlane *s;
391
if (!blk->data_plane) {
396
error_report("device is incompatible with x-data-plane, use scsi=off");
400
if (blk->config_wce) {
401
error_report("device is incompatible with x-data-plane, "
402
"use config-wce=off");
406
/* If dataplane is (re-)enabled while the guest is running there could be
407
* block jobs that can conflict.
409
if (bdrv_in_use(blk->conf.bs)) {
410
error_report("cannot start dataplane thread while device is in use");
414
fd = raw_get_aio_fd(blk->conf.bs);
416
error_report("drive is incompatible with x-data-plane, "
417
"use format=raw,cache=none,aio=native");
421
s = g_new0(VirtIOBlockDataPlane, 1);
426
/* Prevent block operations that conflict with data plane thread */
427
bdrv_set_in_use(blk->conf.bs, 1);
433
void virtio_blk_data_plane_destroy(VirtIOBlockDataPlane *s)
439
virtio_blk_data_plane_stop(s);
440
bdrv_set_in_use(s->blk->conf.bs, 0);
444
void virtio_blk_data_plane_start(VirtIOBlockDataPlane *s)
446
BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(s->vdev)));
447
VirtioBusClass *k = VIRTIO_BUS_GET_CLASS(qbus);
461
vq = virtio_get_queue(s->vdev, 0);
462
if (!vring_setup(&s->vring, s->vdev, 0)) {
467
s->ctx = aio_context_new();
469
/* Set up guest notifier (irq) */
470
if (k->set_guest_notifiers(qbus->parent, 1, true) != 0) {
471
fprintf(stderr, "virtio-blk failed to set guest notifier, "
472
"ensure -enable-kvm is set\n");
475
s->guest_notifier = virtio_queue_get_guest_notifier(vq);
477
/* Set up virtqueue notify */
478
if (k->set_host_notifier(qbus->parent, 0, true) != 0) {
479
fprintf(stderr, "virtio-blk failed to set host notifier\n");
482
s->host_notifier = *virtio_queue_get_host_notifier(vq);
483
aio_set_event_notifier(s->ctx, &s->host_notifier, handle_notify);
486
ioq_init(&s->ioqueue, s->fd, REQ_MAX);
487
for (i = 0; i < ARRAY_SIZE(s->requests); i++) {
488
ioq_put_iocb(&s->ioqueue, &s->requests[i].iocb);
490
s->io_notifier = *ioq_get_notifier(&s->ioqueue);
491
aio_set_event_notifier(s->ctx, &s->io_notifier, handle_io);
495
trace_virtio_blk_data_plane_start(s);
497
/* Kick right away to begin processing requests already in vring */
498
event_notifier_set(virtio_queue_get_host_notifier(vq));
500
/* Spawn thread in BH so it inherits iothread cpusets */
501
s->start_bh = qemu_bh_new(start_data_plane_bh, s);
502
qemu_bh_schedule(s->start_bh);
505
void virtio_blk_data_plane_stop(VirtIOBlockDataPlane *s)
507
BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(s->vdev)));
508
VirtioBusClass *k = VIRTIO_BUS_GET_CLASS(qbus);
509
if (!s->started || s->stopping) {
513
trace_virtio_blk_data_plane_stop(s);
515
/* Stop thread or cancel pending thread creation BH */
517
qemu_bh_delete(s->start_bh);
521
qemu_thread_join(&s->thread);
524
aio_set_event_notifier(s->ctx, &s->io_notifier, NULL);
525
ioq_cleanup(&s->ioqueue);
527
aio_set_event_notifier(s->ctx, &s->host_notifier, NULL);
528
k->set_host_notifier(qbus->parent, 0, false);
530
aio_context_unref(s->ctx);
532
/* Clean up guest notifier (irq) */
533
k->set_guest_notifiers(qbus->parent, 1, false);
535
vring_teardown(&s->vring, s->vdev, 0);