1 /* Virtio ring implementation.
3 * Copyright 2007 Rusty Russell IBM Corporation
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation; either version 2 of the License, or
8 * (at your option) any later version.
10 * This program is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU General Public License for more details.
15 * You should have received a copy of the GNU General Public License
16 * along with this program; if not, write to the Free Software
17 * Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
19 #include <linux/virtio.h>
20 #include <linux/virtio_ring.h>
21 #include <linux/virtio_config.h>
22 #include <linux/device.h>
23 #include <linux/slab.h>
24 #include <linux/module.h>
25 #include <linux/hrtimer.h>
27 /* virtio guest is communicating with a virtual "device" that actually runs on
28 * a host processor. Memory barriers are used to control SMP effects. */
30 /* Where possible, use SMP barriers which are more lightweight than mandatory
31 * barriers, because mandatory barriers control MMIO effects on accesses
32 * through relaxed memory I/O windows (which virtio-pci does not use). */
33 #define virtio_mb(vq) \
34 do { if ((vq)->weak_barriers) smp_mb(); else mb(); } while(0)
35 #define virtio_rmb(vq) \
36 do { if ((vq)->weak_barriers) smp_rmb(); else rmb(); } while(0)
37 #define virtio_wmb(vq) \
38 do { if ((vq)->weak_barriers) smp_wmb(); else wmb(); } while(0)
40 /* We must force memory ordering even if guest is UP since host could be
41 * running on another CPU, but SMP barriers are defined to barrier() in that
42 * configuration. So fall back to mandatory barriers instead. */
43 #define virtio_mb(vq) mb()
44 #define virtio_rmb(vq) rmb()
45 #define virtio_wmb(vq) wmb()
49 /* For development, we want to crash whenever the ring is screwed. */
50 #define BAD_RING(_vq, fmt, args...) \
52 dev_err(&(_vq)->vq.vdev->dev, \
53 "%s:"fmt, (_vq)->vq.name, ##args); \
56 /* Caller is supposed to guarantee no reentry. */
57 #define START_USE(_vq) \
60 panic("%s:in_use = %i\n", \
61 (_vq)->vq.name, (_vq)->in_use); \
62 (_vq)->in_use = __LINE__; \
64 #define END_USE(_vq) \
65 do { BUG_ON(!(_vq)->in_use); (_vq)->in_use = 0; } while(0)
67 #define BAD_RING(_vq, fmt, args...) \
69 dev_err(&_vq->vq.vdev->dev, \
70 "%s:"fmt, (_vq)->vq.name, ##args); \
71 (_vq)->broken = true; \
77 struct vring_virtqueue
81 /* Actual memory layout for this queue */
84 /* Can we use weak barriers? */
87 /* Other side has made a mess, don't try any more. */
90 /* Host supports indirect buffers */
93 /* Host publishes avail event idx */
96 /* Number of free buffers */
97 unsigned int num_free;
98 /* Head of free buffer list. */
99 unsigned int free_head;
100 /* Number we've added since last sync. */
101 unsigned int num_added;
103 /* Last used index we've seen. */
106 /* How to notify other side. FIXME: commonalize hcalls! */
107 void (*notify)(struct virtqueue *vq);
110 /* They're supposed to lock for us. */
113 /* Figure out if their kicks are too delayed. */
114 bool last_add_time_valid;
115 ktime_t last_add_time;
118 /* Tokens for callbacks. */
122 #define to_vvq(_vq) container_of(_vq, struct vring_virtqueue, vq)
124 /* Set up an indirect table of descriptors and add it to the queue. */
125 static int vring_add_indirect(struct vring_virtqueue *vq,
126 struct scatterlist sg[],
131 struct vring_desc *desc;
135 desc = kmalloc((out + in) * sizeof(struct vring_desc), gfp);
139 /* Transfer entries from the sg list into the indirect page */
140 for (i = 0; i < out; i++) {
141 desc[i].flags = VRING_DESC_F_NEXT;
142 desc[i].addr = sg_phys(sg);
143 desc[i].len = sg->length;
147 for (; i < (out + in); i++) {
148 desc[i].flags = VRING_DESC_F_NEXT|VRING_DESC_F_WRITE;
149 desc[i].addr = sg_phys(sg);
150 desc[i].len = sg->length;
155 /* Last one doesn't continue. */
156 desc[i-1].flags &= ~VRING_DESC_F_NEXT;
159 /* We're about to use a buffer */
162 /* Use a single buffer which doesn't continue */
163 head = vq->free_head;
164 vq->vring.desc[head].flags = VRING_DESC_F_INDIRECT;
165 vq->vring.desc[head].addr = virt_to_phys(desc);
166 vq->vring.desc[head].len = i * sizeof(struct vring_desc);
168 /* Update free pointer */
169 vq->free_head = vq->vring.desc[head].next;
175 * virtqueue_add_buf - expose buffer to other end
176 * @vq: the struct virtqueue we're talking about.
177 * @sg: the description of the buffer(s).
178 * @out_num: the number of sg readable by other side
179 * @in_num: the number of sg which are writable (after readable ones)
180 * @data: the token identifying the buffer.
181 * @gfp: how to do memory allocations (if necessary).
183 * Caller must ensure we don't call this with other virtqueue operations
184 * at the same time (except where noted).
186 * Returns remaining capacity of queue or a negative error
187 * (ie. ENOSPC). Note that it only really makes sense to treat all
188 * positive return values as "available": indirect buffers mean that
189 * we can put an entire sg[] array inside a single queue entry.
191 int virtqueue_add_buf(struct virtqueue *_vq,
192 struct scatterlist sg[],
198 struct vring_virtqueue *vq = to_vvq(_vq);
199 unsigned int i, avail, uninitialized_var(prev);
204 BUG_ON(data == NULL);
208 ktime_t now = ktime_get();
210 /* No kick or get, with .1 second between? Warn. */
211 if (vq->last_add_time_valid)
212 WARN_ON(ktime_to_ms(ktime_sub(now, vq->last_add_time))
214 vq->last_add_time = now;
215 vq->last_add_time_valid = true;
219 /* If the host supports indirect descriptor tables, and we have multiple
220 * buffers, then go indirect. FIXME: tune this threshold */
221 if (vq->indirect && (out + in) > 1 && vq->num_free) {
222 head = vring_add_indirect(vq, sg, out, in, gfp);
223 if (likely(head >= 0))
227 BUG_ON(out + in > vq->vring.num);
228 BUG_ON(out + in == 0);
230 if (vq->num_free < out + in) {
231 pr_debug("Can't add buf len %i - avail = %i\n",
232 out + in, vq->num_free);
233 /* FIXME: for historical reasons, we force a notify here if
234 * there are outgoing parts to the buffer. Presumably the
235 * host should service the ring ASAP. */
242 /* We're about to use some buffers from the free list. */
243 vq->num_free -= out + in;
245 head = vq->free_head;
246 for (i = vq->free_head; out; i = vq->vring.desc[i].next, out--) {
247 vq->vring.desc[i].flags = VRING_DESC_F_NEXT;
248 vq->vring.desc[i].addr = sg_phys(sg);
249 vq->vring.desc[i].len = sg->length;
253 for (; in; i = vq->vring.desc[i].next, in--) {
254 vq->vring.desc[i].flags = VRING_DESC_F_NEXT|VRING_DESC_F_WRITE;
255 vq->vring.desc[i].addr = sg_phys(sg);
256 vq->vring.desc[i].len = sg->length;
260 /* Last one doesn't continue. */
261 vq->vring.desc[prev].flags &= ~VRING_DESC_F_NEXT;
263 /* Update free pointer */
268 vq->data[head] = data;
270 /* Put entry in available array (but don't update avail->idx until they
272 avail = (vq->vring.avail->idx & (vq->vring.num-1));
273 vq->vring.avail->ring[avail] = head;
275 /* Descriptors and available array need to be set before we expose the
276 * new available array entries. */
278 vq->vring.avail->idx++;
281 /* This is very unlikely, but theoretically possible. Kick
283 if (unlikely(vq->num_added == (1 << 16) - 1))
286 pr_debug("Added buffer head %i to %p\n", head, vq);
291 EXPORT_SYMBOL_GPL(virtqueue_add_buf);
294 * virtqueue_kick_prepare - first half of split virtqueue_kick call.
295 * @vq: the struct virtqueue
297 * Instead of virtqueue_kick(), you can do:
298 * if (virtqueue_kick_prepare(vq))
299 * virtqueue_notify(vq);
301 * This is sometimes useful because the virtqueue_kick_prepare() needs
302 * to be serialized, but the actual virtqueue_notify() call does not.
304 bool virtqueue_kick_prepare(struct virtqueue *_vq)
306 struct vring_virtqueue *vq = to_vvq(_vq);
311 /* We need to expose available array entries before checking avail
315 old = vq->vring.avail->idx - vq->num_added;
316 new = vq->vring.avail->idx;
320 if (vq->last_add_time_valid) {
321 WARN_ON(ktime_to_ms(ktime_sub(ktime_get(),
322 vq->last_add_time)) > 100);
324 vq->last_add_time_valid = false;
328 needs_kick = vring_need_event(vring_avail_event(&vq->vring),
331 needs_kick = !(vq->vring.used->flags & VRING_USED_F_NO_NOTIFY);
336 EXPORT_SYMBOL_GPL(virtqueue_kick_prepare);
339 * virtqueue_notify - second half of split virtqueue_kick call.
340 * @vq: the struct virtqueue
342 * This does not need to be serialized.
344 void virtqueue_notify(struct virtqueue *_vq)
346 struct vring_virtqueue *vq = to_vvq(_vq);
348 /* Prod other side to tell it about changes. */
351 EXPORT_SYMBOL_GPL(virtqueue_notify);
354 * virtqueue_kick - update after add_buf
355 * @vq: the struct virtqueue
357 * After one or more virtqueue_add_buf calls, invoke this to kick
360 * Caller must ensure we don't call this with other virtqueue
361 * operations at the same time (except where noted).
363 void virtqueue_kick(struct virtqueue *vq)
365 if (virtqueue_kick_prepare(vq))
366 virtqueue_notify(vq);
368 EXPORT_SYMBOL_GPL(virtqueue_kick);
370 static void detach_buf(struct vring_virtqueue *vq, unsigned int head)
374 /* Clear data ptr. */
375 vq->data[head] = NULL;
377 /* Put back on free list: find end */
380 /* Free the indirect table */
381 if (vq->vring.desc[i].flags & VRING_DESC_F_INDIRECT)
382 kfree(phys_to_virt(vq->vring.desc[i].addr));
384 while (vq->vring.desc[i].flags & VRING_DESC_F_NEXT) {
385 i = vq->vring.desc[i].next;
389 vq->vring.desc[i].next = vq->free_head;
390 vq->free_head = head;
391 /* Plus final descriptor */
395 static inline bool more_used(const struct vring_virtqueue *vq)
397 return vq->last_used_idx != vq->vring.used->idx;
401 * virtqueue_get_buf - get the next used buffer
402 * @vq: the struct virtqueue we're talking about.
403 * @len: the length written into the buffer
405 * If the driver wrote data into the buffer, @len will be set to the
406 * amount written. This means you don't need to clear the buffer
407 * beforehand to ensure there's no data leakage in the case of short
410 * Caller must ensure we don't call this with other virtqueue
411 * operations at the same time (except where noted).
413 * Returns NULL if there are no used buffers, or the "data" token
414 * handed to virtqueue_add_buf().
416 void *virtqueue_get_buf(struct virtqueue *_vq, unsigned int *len)
418 struct vring_virtqueue *vq = to_vvq(_vq);
425 if (unlikely(vq->broken)) {
430 if (!more_used(vq)) {
431 pr_debug("No more buffers in queue\n");
436 /* Only get used array entries after they have been exposed by host. */
439 last_used = (vq->last_used_idx & (vq->vring.num - 1));
440 i = vq->vring.used->ring[last_used].id;
441 *len = vq->vring.used->ring[last_used].len;
443 if (unlikely(i >= vq->vring.num)) {
444 BAD_RING(vq, "id %u out of range\n", i);
447 if (unlikely(!vq->data[i])) {
448 BAD_RING(vq, "id %u is not a head!\n", i);
452 /* detach_buf clears data, so grab it now. */
456 /* If we expect an interrupt for the next entry, tell host
457 * by writing event index and flush out the write before
458 * the read in the next get_buf call. */
459 if (!(vq->vring.avail->flags & VRING_AVAIL_F_NO_INTERRUPT)) {
460 vring_used_event(&vq->vring) = vq->last_used_idx;
465 vq->last_add_time_valid = false;
471 EXPORT_SYMBOL_GPL(virtqueue_get_buf);
474 * virtqueue_disable_cb - disable callbacks
475 * @vq: the struct virtqueue we're talking about.
477 * Note that this is not necessarily synchronous, hence unreliable and only
478 * useful as an optimization.
480 * Unlike other operations, this need not be serialized.
482 void virtqueue_disable_cb(struct virtqueue *_vq)
484 struct vring_virtqueue *vq = to_vvq(_vq);
486 vq->vring.avail->flags |= VRING_AVAIL_F_NO_INTERRUPT;
488 EXPORT_SYMBOL_GPL(virtqueue_disable_cb);
491 * virtqueue_enable_cb - restart callbacks after disable_cb.
492 * @vq: the struct virtqueue we're talking about.
494 * This re-enables callbacks; it returns "false" if there are pending
495 * buffers in the queue, to detect a possible race between the driver
496 * checking for more work, and enabling callbacks.
498 * Caller must ensure we don't call this with other virtqueue
499 * operations at the same time (except where noted).
501 bool virtqueue_enable_cb(struct virtqueue *_vq)
503 struct vring_virtqueue *vq = to_vvq(_vq);
507 /* We optimistically turn back on interrupts, then check if there was
509 /* Depending on the VIRTIO_RING_F_EVENT_IDX feature, we need to
510 * either clear the flags bit or point the event index at the next
511 * entry. Always do both to keep code simple. */
512 vq->vring.avail->flags &= ~VRING_AVAIL_F_NO_INTERRUPT;
513 vring_used_event(&vq->vring) = vq->last_used_idx;
515 if (unlikely(more_used(vq))) {
523 EXPORT_SYMBOL_GPL(virtqueue_enable_cb);
526 * virtqueue_enable_cb_delayed - restart callbacks after disable_cb.
527 * @vq: the struct virtqueue we're talking about.
529 * This re-enables callbacks but hints to the other side to delay
530 * interrupts until most of the available buffers have been processed;
531 * it returns "false" if there are many pending buffers in the queue,
532 * to detect a possible race between the driver checking for more work,
533 * and enabling callbacks.
535 * Caller must ensure we don't call this with other virtqueue
536 * operations at the same time (except where noted).
538 bool virtqueue_enable_cb_delayed(struct virtqueue *_vq)
540 struct vring_virtqueue *vq = to_vvq(_vq);
545 /* We optimistically turn back on interrupts, then check if there was
547 /* Depending on the VIRTIO_RING_F_USED_EVENT_IDX feature, we need to
548 * either clear the flags bit or point the event index at the next
549 * entry. Always do both to keep code simple. */
550 vq->vring.avail->flags &= ~VRING_AVAIL_F_NO_INTERRUPT;
551 /* TODO: tune this threshold */
552 bufs = (u16)(vq->vring.avail->idx - vq->last_used_idx) * 3 / 4;
553 vring_used_event(&vq->vring) = vq->last_used_idx + bufs;
555 if (unlikely((u16)(vq->vring.used->idx - vq->last_used_idx) > bufs)) {
563 EXPORT_SYMBOL_GPL(virtqueue_enable_cb_delayed);
566 * virtqueue_detach_unused_buf - detach first unused buffer
567 * @vq: the struct virtqueue we're talking about.
569 * Returns NULL or the "data" token handed to virtqueue_add_buf().
570 * This is not valid on an active queue; it is useful only for device
573 void *virtqueue_detach_unused_buf(struct virtqueue *_vq)
575 struct vring_virtqueue *vq = to_vvq(_vq);
581 for (i = 0; i < vq->vring.num; i++) {
584 /* detach_buf clears data, so grab it now. */
587 vq->vring.avail->idx--;
591 /* That should have freed everything. */
592 BUG_ON(vq->num_free != vq->vring.num);
597 EXPORT_SYMBOL_GPL(virtqueue_detach_unused_buf);
599 irqreturn_t vring_interrupt(int irq, void *_vq)
601 struct vring_virtqueue *vq = to_vvq(_vq);
603 if (!more_used(vq)) {
604 pr_debug("virtqueue interrupt with no work for %p\n", vq);
608 if (unlikely(vq->broken))
611 pr_debug("virtqueue callback for %p (%p)\n", vq, vq->vq.callback);
613 vq->vq.callback(&vq->vq);
617 EXPORT_SYMBOL_GPL(vring_interrupt);
619 struct virtqueue *vring_new_virtqueue(unsigned int num,
620 unsigned int vring_align,
621 struct virtio_device *vdev,
624 void (*notify)(struct virtqueue *),
625 void (*callback)(struct virtqueue *),
628 struct vring_virtqueue *vq;
631 /* We assume num is a power of 2. */
632 if (num & (num - 1)) {
633 dev_warn(&vdev->dev, "Bad virtqueue length %u\n", num);
637 vq = kmalloc(sizeof(*vq) + sizeof(void *)*num, GFP_KERNEL);
641 vring_init(&vq->vring, num, pages, vring_align);
642 vq->vq.callback = callback;
646 vq->weak_barriers = weak_barriers;
648 vq->last_used_idx = 0;
650 list_add_tail(&vq->vq.list, &vdev->vqs);
653 vq->last_add_time_valid = false;
656 vq->indirect = virtio_has_feature(vdev, VIRTIO_RING_F_INDIRECT_DESC);
657 vq->event = virtio_has_feature(vdev, VIRTIO_RING_F_EVENT_IDX);
659 /* No callback? Tell other side not to bother us. */
661 vq->vring.avail->flags |= VRING_AVAIL_F_NO_INTERRUPT;
663 /* Put everything in free lists. */
666 for (i = 0; i < num-1; i++) {
667 vq->vring.desc[i].next = i+1;
674 EXPORT_SYMBOL_GPL(vring_new_virtqueue);
676 void vring_del_virtqueue(struct virtqueue *vq)
681 EXPORT_SYMBOL_GPL(vring_del_virtqueue);
683 /* Manipulates transport-specific feature bits. */
684 void vring_transport_features(struct virtio_device *vdev)
688 for (i = VIRTIO_TRANSPORT_F_START; i < VIRTIO_TRANSPORT_F_END; i++) {
690 case VIRTIO_RING_F_INDIRECT_DESC:
692 case VIRTIO_RING_F_EVENT_IDX:
695 /* We don't understand this bit. */
696 clear_bit(i, vdev->features);
700 EXPORT_SYMBOL_GPL(vring_transport_features);
703 * virtqueue_get_vring_size - return the size of the virtqueue's vring
704 * @vq: the struct virtqueue containing the vring of interest.
706 * Returns the size of the vring. This is mainly used for boasting to
707 * userspace. Unlike other operations, this need not be serialized.
709 unsigned int virtqueue_get_vring_size(struct virtqueue *_vq)
712 struct vring_virtqueue *vq = to_vvq(_vq);
714 return vq->vring.num;
716 EXPORT_SYMBOL_GPL(virtqueue_get_vring_size);
718 MODULE_LICENSE("GPL");