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.
15 #include "qemu/osdep.h"
16 #include "qapi/error.h"
19 #include "qemu/main-loop.h"
20 #include "qemu/thread.h"
21 #include "qemu/error-report.h"
22 #include "hw/virtio/virtio-access.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"
29 struct VirtIOBlockDataPlane
{
35 QEMUBH
*bh
; /* bh for guest notification */
36 unsigned long *batch_notify_vqs
;
37 bool batch_notifications
;
39 /* Note that these EventNotifiers are assigned by value. This is
40 * fine as long as you do not call event_notifier_cleanup on them
41 * (because you don't own the file descriptor or handle; you just
48 /* Raise an interrupt to signal guest, if necessary */
49 void virtio_blk_data_plane_notify(VirtIOBlockDataPlane
*s
, VirtQueue
*vq
)
51 if (s
->batch_notifications
) {
52 set_bit(virtio_get_queue_index(vq
), s
->batch_notify_vqs
);
53 qemu_bh_schedule(s
->bh
);
55 virtio_notify_irqfd(s
->vdev
, vq
);
59 static void notify_guest_bh(void *opaque
)
61 VirtIOBlockDataPlane
*s
= opaque
;
62 unsigned nvqs
= s
->conf
->num_queues
;
63 unsigned long bitmap
[BITS_TO_LONGS(nvqs
)];
66 memcpy(bitmap
, s
->batch_notify_vqs
, sizeof(bitmap
));
67 memset(s
->batch_notify_vqs
, 0, sizeof(bitmap
));
69 for (j
= 0; j
< nvqs
; j
+= BITS_PER_LONG
) {
70 unsigned long bits
= bitmap
[j
/ BITS_PER_LONG
];
73 unsigned i
= j
+ ctzl(bits
);
74 VirtQueue
*vq
= virtio_get_queue(s
->vdev
, i
);
76 virtio_notify_irqfd(s
->vdev
, vq
);
78 bits
&= bits
- 1; /* clear right-most bit */
83 /* Context: QEMU global mutex held */
84 bool virtio_blk_data_plane_create(VirtIODevice
*vdev
, VirtIOBlkConf
*conf
,
85 VirtIOBlockDataPlane
**dataplane
,
88 VirtIOBlockDataPlane
*s
;
89 BusState
*qbus
= BUS(qdev_get_parent_bus(DEVICE(vdev
)));
90 VirtioBusClass
*k
= VIRTIO_BUS_GET_CLASS(qbus
);
95 if (!k
->set_guest_notifiers
|| !k
->ioeventfd_assign
) {
97 "device is incompatible with iothread "
98 "(transport does not support notifiers)");
101 if (!virtio_device_ioeventfd_enabled(vdev
)) {
102 error_setg(errp
, "ioeventfd is required for iothread");
106 /* If dataplane is (re-)enabled while the guest is running there could
107 * be block jobs that can conflict.
109 if (blk_op_is_blocked(conf
->conf
.blk
, BLOCK_OP_TYPE_DATAPLANE
, errp
)) {
110 error_prepend(errp
, "cannot start virtio-blk dataplane: ");
114 /* Don't try if transport does not support notifiers. */
115 if (!virtio_device_ioeventfd_enabled(vdev
)) {
119 s
= g_new0(VirtIOBlockDataPlane
, 1);
123 if (conf
->iothread
) {
124 s
->iothread
= conf
->iothread
;
125 object_ref(OBJECT(s
->iothread
));
126 s
->ctx
= iothread_get_aio_context(s
->iothread
);
128 s
->ctx
= qemu_get_aio_context();
130 s
->bh
= aio_bh_new_guarded(s
->ctx
, notify_guest_bh
, s
,
131 &DEVICE(vdev
)->mem_reentrancy_guard
);
132 s
->batch_notify_vqs
= bitmap_new(conf
->num_queues
);
139 /* Context: QEMU global mutex held */
140 void virtio_blk_data_plane_destroy(VirtIOBlockDataPlane
*s
)
148 vblk
= VIRTIO_BLK(s
->vdev
);
149 assert(!vblk
->dataplane_started
);
150 g_free(s
->batch_notify_vqs
);
151 qemu_bh_delete(s
->bh
);
153 object_unref(OBJECT(s
->iothread
));
158 /* Context: QEMU global mutex held */
159 int virtio_blk_data_plane_start(VirtIODevice
*vdev
)
161 VirtIOBlock
*vblk
= VIRTIO_BLK(vdev
);
162 VirtIOBlockDataPlane
*s
= vblk
->dataplane
;
163 BusState
*qbus
= BUS(qdev_get_parent_bus(DEVICE(vblk
)));
164 VirtioBusClass
*k
= VIRTIO_BUS_GET_CLASS(qbus
);
165 AioContext
*old_context
;
167 unsigned nvqs
= s
->conf
->num_queues
;
168 Error
*local_err
= NULL
;
171 if (vblk
->dataplane_started
|| s
->starting
) {
177 if (!virtio_vdev_has_feature(vdev
, VIRTIO_RING_F_EVENT_IDX
)) {
178 s
->batch_notifications
= true;
180 s
->batch_notifications
= false;
183 /* Set up guest notifier (irq) */
184 r
= k
->set_guest_notifiers(qbus
->parent
, nvqs
, true);
186 error_report("virtio-blk failed to set guest notifier (%d), "
187 "ensure -accel kvm is set.", r
);
188 goto fail_guest_notifiers
;
192 * Batch all the host notifiers in a single transaction to avoid
193 * quadratic time complexity in address_space_update_ioeventfds().
195 memory_region_transaction_begin();
197 /* Set up virtqueue notify */
198 for (i
= 0; i
< nvqs
; i
++) {
199 r
= virtio_bus_set_host_notifier(VIRTIO_BUS(qbus
), i
, true);
203 fprintf(stderr
, "virtio-blk failed to set host notifier (%d)\n", r
);
205 virtio_bus_set_host_notifier(VIRTIO_BUS(qbus
), i
, false);
209 * The transaction expects the ioeventfds to be open when it
210 * commits. Do it now, before the cleanup loop.
212 memory_region_transaction_commit();
215 virtio_bus_cleanup_host_notifier(VIRTIO_BUS(qbus
), j
);
217 goto fail_host_notifiers
;
221 memory_region_transaction_commit();
224 * These fields are visible to the IOThread so we rely on implicit barriers
225 * in aio_context_acquire() on the write side and aio_notify_accept() on
229 vblk
->dataplane_started
= true;
230 trace_virtio_blk_data_plane_start(s
);
232 old_context
= blk_get_aio_context(s
->conf
->conf
.blk
);
233 aio_context_acquire(old_context
);
234 r
= blk_set_aio_context(s
->conf
->conf
.blk
, s
->ctx
, &local_err
);
235 aio_context_release(old_context
);
237 error_report_err(local_err
);
238 goto fail_aio_context
;
241 /* Kick right away to begin processing requests already in vring */
242 for (i
= 0; i
< nvqs
; i
++) {
243 VirtQueue
*vq
= virtio_get_queue(s
->vdev
, i
);
245 event_notifier_set(virtio_queue_get_host_notifier(vq
));
248 /* Get this show started by hooking up our callbacks */
249 if (!blk_in_drain(s
->conf
->conf
.blk
)) {
250 aio_context_acquire(s
->ctx
);
251 for (i
= 0; i
< nvqs
; i
++) {
252 VirtQueue
*vq
= virtio_get_queue(s
->vdev
, i
);
254 virtio_queue_aio_attach_host_notifier(vq
, s
->ctx
);
256 aio_context_release(s
->ctx
);
261 memory_region_transaction_begin();
263 for (i
= 0; i
< nvqs
; i
++) {
264 virtio_bus_set_host_notifier(VIRTIO_BUS(qbus
), i
, false);
267 memory_region_transaction_commit();
269 for (i
= 0; i
< nvqs
; i
++) {
270 virtio_bus_cleanup_host_notifier(VIRTIO_BUS(qbus
), i
);
273 k
->set_guest_notifiers(qbus
->parent
, nvqs
, false);
274 fail_guest_notifiers
:
275 vblk
->dataplane_disabled
= true;
277 vblk
->dataplane_started
= true;
281 /* Stop notifications for new requests from guest.
283 * Context: BH in IOThread
285 static void virtio_blk_data_plane_stop_bh(void *opaque
)
287 VirtIOBlockDataPlane
*s
= opaque
;
290 for (i
= 0; i
< s
->conf
->num_queues
; i
++) {
291 VirtQueue
*vq
= virtio_get_queue(s
->vdev
, i
);
292 EventNotifier
*host_notifier
= virtio_queue_get_host_notifier(vq
);
294 virtio_queue_aio_detach_host_notifier(vq
, s
->ctx
);
297 * Test and clear notifier after disabling event, in case poll callback
298 * didn't have time to run.
300 virtio_queue_host_notifier_read(host_notifier
);
304 /* Context: QEMU global mutex held */
305 void virtio_blk_data_plane_stop(VirtIODevice
*vdev
)
307 VirtIOBlock
*vblk
= VIRTIO_BLK(vdev
);
308 VirtIOBlockDataPlane
*s
= vblk
->dataplane
;
309 BusState
*qbus
= qdev_get_parent_bus(DEVICE(vblk
));
310 VirtioBusClass
*k
= VIRTIO_BUS_GET_CLASS(qbus
);
312 unsigned nvqs
= s
->conf
->num_queues
;
314 if (!vblk
->dataplane_started
|| s
->stopping
) {
318 /* Better luck next time. */
319 if (vblk
->dataplane_disabled
) {
320 vblk
->dataplane_disabled
= false;
321 vblk
->dataplane_started
= false;
325 trace_virtio_blk_data_plane_stop(s
);
327 if (!blk_in_drain(s
->conf
->conf
.blk
)) {
328 aio_wait_bh_oneshot(s
->ctx
, virtio_blk_data_plane_stop_bh
, s
);
331 aio_context_acquire(s
->ctx
);
333 /* Wait for virtio_blk_dma_restart_bh() and in flight I/O to complete */
334 blk_drain(s
->conf
->conf
.blk
);
337 * Try to switch bs back to the QEMU main loop. If other users keep the
338 * BlockBackend in the iothread, that's ok
340 blk_set_aio_context(s
->conf
->conf
.blk
, qemu_get_aio_context(), NULL
);
342 aio_context_release(s
->ctx
);
345 * Batch all the host notifiers in a single transaction to avoid
346 * quadratic time complexity in address_space_update_ioeventfds().
348 memory_region_transaction_begin();
350 for (i
= 0; i
< nvqs
; i
++) {
351 virtio_bus_set_host_notifier(VIRTIO_BUS(qbus
), i
, false);
355 * The transaction expects the ioeventfds to be open when it
356 * commits. Do it now, before the cleanup loop.
358 memory_region_transaction_commit();
360 for (i
= 0; i
< nvqs
; i
++) {
361 virtio_bus_cleanup_host_notifier(VIRTIO_BUS(qbus
), i
);
364 qemu_bh_cancel(s
->bh
);
365 notify_guest_bh(s
); /* final chance to notify guest */
367 /* Clean up guest notifier (irq) */
368 k
->set_guest_notifiers(qbus
->parent
, nvqs
, false);
370 vblk
->dataplane_started
= false;