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 /* Head of free buffer list. */
97 unsigned int free_head;
98 /* Number we've added since last sync. */
99 unsigned int num_added;
101 /* Last used index we've seen. */
104 /* How to notify other side. FIXME: commonalize hcalls! */
105 void (*notify)(struct virtqueue *vq);
108 /* They're supposed to lock for us. */
111 /* Figure out if their kicks are too delayed. */
112 bool last_add_time_valid;
113 ktime_t last_add_time;
116 /* Tokens for callbacks. */
120 #define to_vvq(_vq) container_of(_vq, struct vring_virtqueue, vq)
122 /* Set up an indirect table of descriptors and add it to the queue. */
123 static int vring_add_indirect(struct vring_virtqueue *vq,
124 struct scatterlist sg[],
129 struct vring_desc *desc;
134 * We require lowmem mappings for the descriptors because
135 * otherwise virt_to_phys will give us bogus addresses in the
138 gfp &= ~(__GFP_HIGHMEM | __GFP_HIGH);
140 desc = kmalloc((out + in) * sizeof(struct vring_desc), gfp);
144 /* Transfer entries from the sg list into the indirect page */
145 for (i = 0; i < out; i++) {
146 desc[i].flags = VRING_DESC_F_NEXT;
147 desc[i].addr = sg_phys(sg);
148 desc[i].len = sg->length;
152 for (; i < (out + in); i++) {
153 desc[i].flags = VRING_DESC_F_NEXT|VRING_DESC_F_WRITE;
154 desc[i].addr = sg_phys(sg);
155 desc[i].len = sg->length;
160 /* Last one doesn't continue. */
161 desc[i-1].flags &= ~VRING_DESC_F_NEXT;
164 /* We're about to use a buffer */
167 /* Use a single buffer which doesn't continue */
168 head = vq->free_head;
169 vq->vring.desc[head].flags = VRING_DESC_F_INDIRECT;
170 vq->vring.desc[head].addr = virt_to_phys(desc);
171 vq->vring.desc[head].len = i * sizeof(struct vring_desc);
173 /* Update free pointer */
174 vq->free_head = vq->vring.desc[head].next;
180 * virtqueue_add_buf - expose buffer to other end
181 * @vq: the struct virtqueue we're talking about.
182 * @sg: the description of the buffer(s).
183 * @out_num: the number of sg readable by other side
184 * @in_num: the number of sg which are writable (after readable ones)
185 * @data: the token identifying the buffer.
186 * @gfp: how to do memory allocations (if necessary).
188 * Caller must ensure we don't call this with other virtqueue operations
189 * at the same time (except where noted).
191 * Returns remaining capacity of queue or a negative error
192 * (ie. ENOSPC). Note that it only really makes sense to treat all
193 * positive return values as "available": indirect buffers mean that
194 * we can put an entire sg[] array inside a single queue entry.
196 int virtqueue_add_buf(struct virtqueue *_vq,
197 struct scatterlist sg[],
203 struct vring_virtqueue *vq = to_vvq(_vq);
204 unsigned int i, avail, uninitialized_var(prev);
209 BUG_ON(data == NULL);
213 ktime_t now = ktime_get();
215 /* No kick or get, with .1 second between? Warn. */
216 if (vq->last_add_time_valid)
217 WARN_ON(ktime_to_ms(ktime_sub(now, vq->last_add_time))
219 vq->last_add_time = now;
220 vq->last_add_time_valid = true;
224 /* If the host supports indirect descriptor tables, and we have multiple
225 * buffers, then go indirect. FIXME: tune this threshold */
226 if (vq->indirect && (out + in) > 1 && vq->vq.num_free) {
227 head = vring_add_indirect(vq, sg, out, in, gfp);
228 if (likely(head >= 0))
232 BUG_ON(out + in > vq->vring.num);
233 BUG_ON(out + in == 0);
235 if (vq->vq.num_free < out + in) {
236 pr_debug("Can't add buf len %i - avail = %i\n",
237 out + in, vq->vq.num_free);
238 /* FIXME: for historical reasons, we force a notify here if
239 * there are outgoing parts to the buffer. Presumably the
240 * host should service the ring ASAP. */
247 /* We're about to use some buffers from the free list. */
248 vq->vq.num_free -= out + in;
250 head = vq->free_head;
251 for (i = vq->free_head; out; i = vq->vring.desc[i].next, out--) {
252 vq->vring.desc[i].flags = VRING_DESC_F_NEXT;
253 vq->vring.desc[i].addr = sg_phys(sg);
254 vq->vring.desc[i].len = sg->length;
258 for (; in; i = vq->vring.desc[i].next, in--) {
259 vq->vring.desc[i].flags = VRING_DESC_F_NEXT|VRING_DESC_F_WRITE;
260 vq->vring.desc[i].addr = sg_phys(sg);
261 vq->vring.desc[i].len = sg->length;
265 /* Last one doesn't continue. */
266 vq->vring.desc[prev].flags &= ~VRING_DESC_F_NEXT;
268 /* Update free pointer */
273 vq->data[head] = data;
275 /* Put entry in available array (but don't update avail->idx until they
277 avail = (vq->vring.avail->idx & (vq->vring.num-1));
278 vq->vring.avail->ring[avail] = head;
280 /* Descriptors and available array need to be set before we expose the
281 * new available array entries. */
283 vq->vring.avail->idx++;
286 /* This is very unlikely, but theoretically possible. Kick
288 if (unlikely(vq->num_added == (1 << 16) - 1))
291 pr_debug("Added buffer head %i to %p\n", head, vq);
294 return vq->vq.num_free;
296 EXPORT_SYMBOL_GPL(virtqueue_add_buf);
299 * virtqueue_kick_prepare - first half of split virtqueue_kick call.
300 * @vq: the struct virtqueue
302 * Instead of virtqueue_kick(), you can do:
303 * if (virtqueue_kick_prepare(vq))
304 * virtqueue_notify(vq);
306 * This is sometimes useful because the virtqueue_kick_prepare() needs
307 * to be serialized, but the actual virtqueue_notify() call does not.
309 bool virtqueue_kick_prepare(struct virtqueue *_vq)
311 struct vring_virtqueue *vq = to_vvq(_vq);
316 /* We need to expose available array entries before checking avail
320 old = vq->vring.avail->idx - vq->num_added;
321 new = vq->vring.avail->idx;
325 if (vq->last_add_time_valid) {
326 WARN_ON(ktime_to_ms(ktime_sub(ktime_get(),
327 vq->last_add_time)) > 100);
329 vq->last_add_time_valid = false;
333 needs_kick = vring_need_event(vring_avail_event(&vq->vring),
336 needs_kick = !(vq->vring.used->flags & VRING_USED_F_NO_NOTIFY);
341 EXPORT_SYMBOL_GPL(virtqueue_kick_prepare);
344 * virtqueue_notify - second half of split virtqueue_kick call.
345 * @vq: the struct virtqueue
347 * This does not need to be serialized.
349 void virtqueue_notify(struct virtqueue *_vq)
351 struct vring_virtqueue *vq = to_vvq(_vq);
353 /* Prod other side to tell it about changes. */
356 EXPORT_SYMBOL_GPL(virtqueue_notify);
359 * virtqueue_kick - update after add_buf
360 * @vq: the struct virtqueue
362 * After one or more virtqueue_add_buf calls, invoke this to kick
365 * Caller must ensure we don't call this with other virtqueue
366 * operations at the same time (except where noted).
368 void virtqueue_kick(struct virtqueue *vq)
370 if (virtqueue_kick_prepare(vq))
371 virtqueue_notify(vq);
373 EXPORT_SYMBOL_GPL(virtqueue_kick);
375 static void detach_buf(struct vring_virtqueue *vq, unsigned int head)
379 /* Clear data ptr. */
380 vq->data[head] = NULL;
382 /* Put back on free list: find end */
385 /* Free the indirect table */
386 if (vq->vring.desc[i].flags & VRING_DESC_F_INDIRECT)
387 kfree(phys_to_virt(vq->vring.desc[i].addr));
389 while (vq->vring.desc[i].flags & VRING_DESC_F_NEXT) {
390 i = vq->vring.desc[i].next;
394 vq->vring.desc[i].next = vq->free_head;
395 vq->free_head = head;
396 /* Plus final descriptor */
400 static inline bool more_used(const struct vring_virtqueue *vq)
402 return vq->last_used_idx != vq->vring.used->idx;
406 * virtqueue_get_buf - get the next used buffer
407 * @vq: the struct virtqueue we're talking about.
408 * @len: the length written into the buffer
410 * If the driver wrote data into the buffer, @len will be set to the
411 * amount written. This means you don't need to clear the buffer
412 * beforehand to ensure there's no data leakage in the case of short
415 * Caller must ensure we don't call this with other virtqueue
416 * operations at the same time (except where noted).
418 * Returns NULL if there are no used buffers, or the "data" token
419 * handed to virtqueue_add_buf().
421 void *virtqueue_get_buf(struct virtqueue *_vq, unsigned int *len)
423 struct vring_virtqueue *vq = to_vvq(_vq);
430 if (unlikely(vq->broken)) {
435 if (!more_used(vq)) {
436 pr_debug("No more buffers in queue\n");
441 /* Only get used array entries after they have been exposed by host. */
444 last_used = (vq->last_used_idx & (vq->vring.num - 1));
445 i = vq->vring.used->ring[last_used].id;
446 *len = vq->vring.used->ring[last_used].len;
448 if (unlikely(i >= vq->vring.num)) {
449 BAD_RING(vq, "id %u out of range\n", i);
452 if (unlikely(!vq->data[i])) {
453 BAD_RING(vq, "id %u is not a head!\n", i);
457 /* detach_buf clears data, so grab it now. */
461 /* If we expect an interrupt for the next entry, tell host
462 * by writing event index and flush out the write before
463 * the read in the next get_buf call. */
464 if (!(vq->vring.avail->flags & VRING_AVAIL_F_NO_INTERRUPT)) {
465 vring_used_event(&vq->vring) = vq->last_used_idx;
470 vq->last_add_time_valid = false;
476 EXPORT_SYMBOL_GPL(virtqueue_get_buf);
479 * virtqueue_disable_cb - disable callbacks
480 * @vq: the struct virtqueue we're talking about.
482 * Note that this is not necessarily synchronous, hence unreliable and only
483 * useful as an optimization.
485 * Unlike other operations, this need not be serialized.
487 void virtqueue_disable_cb(struct virtqueue *_vq)
489 struct vring_virtqueue *vq = to_vvq(_vq);
491 vq->vring.avail->flags |= VRING_AVAIL_F_NO_INTERRUPT;
493 EXPORT_SYMBOL_GPL(virtqueue_disable_cb);
496 * virtqueue_enable_cb - restart callbacks after disable_cb.
497 * @vq: the struct virtqueue we're talking about.
499 * This re-enables callbacks; it returns "false" if there are pending
500 * buffers in the queue, to detect a possible race between the driver
501 * checking for more work, and enabling callbacks.
503 * Caller must ensure we don't call this with other virtqueue
504 * operations at the same time (except where noted).
506 bool virtqueue_enable_cb(struct virtqueue *_vq)
508 struct vring_virtqueue *vq = to_vvq(_vq);
512 /* We optimistically turn back on interrupts, then check if there was
514 /* Depending on the VIRTIO_RING_F_EVENT_IDX feature, we need to
515 * either clear the flags bit or point the event index at the next
516 * entry. Always do both to keep code simple. */
517 vq->vring.avail->flags &= ~VRING_AVAIL_F_NO_INTERRUPT;
518 vring_used_event(&vq->vring) = vq->last_used_idx;
520 if (unlikely(more_used(vq))) {
528 EXPORT_SYMBOL_GPL(virtqueue_enable_cb);
531 * virtqueue_enable_cb_delayed - restart callbacks after disable_cb.
532 * @vq: the struct virtqueue we're talking about.
534 * This re-enables callbacks but hints to the other side to delay
535 * interrupts until most of the available buffers have been processed;
536 * it returns "false" if there are many pending buffers in the queue,
537 * to detect a possible race between the driver checking for more work,
538 * and enabling callbacks.
540 * Caller must ensure we don't call this with other virtqueue
541 * operations at the same time (except where noted).
543 bool virtqueue_enable_cb_delayed(struct virtqueue *_vq)
545 struct vring_virtqueue *vq = to_vvq(_vq);
550 /* We optimistically turn back on interrupts, then check if there was
552 /* Depending on the VIRTIO_RING_F_USED_EVENT_IDX feature, we need to
553 * either clear the flags bit or point the event index at the next
554 * entry. Always do both to keep code simple. */
555 vq->vring.avail->flags &= ~VRING_AVAIL_F_NO_INTERRUPT;
556 /* TODO: tune this threshold */
557 bufs = (u16)(vq->vring.avail->idx - vq->last_used_idx) * 3 / 4;
558 vring_used_event(&vq->vring) = vq->last_used_idx + bufs;
560 if (unlikely((u16)(vq->vring.used->idx - vq->last_used_idx) > bufs)) {
568 EXPORT_SYMBOL_GPL(virtqueue_enable_cb_delayed);
571 * virtqueue_detach_unused_buf - detach first unused buffer
572 * @vq: the struct virtqueue we're talking about.
574 * Returns NULL or the "data" token handed to virtqueue_add_buf().
575 * This is not valid on an active queue; it is useful only for device
578 void *virtqueue_detach_unused_buf(struct virtqueue *_vq)
580 struct vring_virtqueue *vq = to_vvq(_vq);
586 for (i = 0; i < vq->vring.num; i++) {
589 /* detach_buf clears data, so grab it now. */
592 vq->vring.avail->idx--;
596 /* That should have freed everything. */
597 BUG_ON(vq->vq.num_free != vq->vring.num);
602 EXPORT_SYMBOL_GPL(virtqueue_detach_unused_buf);
604 irqreturn_t vring_interrupt(int irq, void *_vq)
606 struct vring_virtqueue *vq = to_vvq(_vq);
608 if (!more_used(vq)) {
609 pr_debug("virtqueue interrupt with no work for %p\n", vq);
613 if (unlikely(vq->broken))
616 pr_debug("virtqueue callback for %p (%p)\n", vq, vq->vq.callback);
618 vq->vq.callback(&vq->vq);
622 EXPORT_SYMBOL_GPL(vring_interrupt);
624 struct virtqueue *vring_new_virtqueue(unsigned int index,
626 unsigned int vring_align,
627 struct virtio_device *vdev,
630 void (*notify)(struct virtqueue *),
631 void (*callback)(struct virtqueue *),
634 struct vring_virtqueue *vq;
637 /* We assume num is a power of 2. */
638 if (num & (num - 1)) {
639 dev_warn(&vdev->dev, "Bad virtqueue length %u\n", num);
643 vq = kmalloc(sizeof(*vq) + sizeof(void *)*num, GFP_KERNEL);
647 vring_init(&vq->vring, num, pages, vring_align);
648 vq->vq.callback = callback;
651 vq->vq.num_free = num;
652 vq->vq.index = index;
654 vq->weak_barriers = weak_barriers;
656 vq->last_used_idx = 0;
658 list_add_tail(&vq->vq.list, &vdev->vqs);
661 vq->last_add_time_valid = false;
664 vq->indirect = virtio_has_feature(vdev, VIRTIO_RING_F_INDIRECT_DESC);
665 vq->event = virtio_has_feature(vdev, VIRTIO_RING_F_EVENT_IDX);
667 /* No callback? Tell other side not to bother us. */
669 vq->vring.avail->flags |= VRING_AVAIL_F_NO_INTERRUPT;
671 /* Put everything in free lists. */
673 for (i = 0; i < num-1; i++) {
674 vq->vring.desc[i].next = i+1;
681 EXPORT_SYMBOL_GPL(vring_new_virtqueue);
683 void vring_del_virtqueue(struct virtqueue *vq)
688 EXPORT_SYMBOL_GPL(vring_del_virtqueue);
690 /* Manipulates transport-specific feature bits. */
691 void vring_transport_features(struct virtio_device *vdev)
695 for (i = VIRTIO_TRANSPORT_F_START; i < VIRTIO_TRANSPORT_F_END; i++) {
697 case VIRTIO_RING_F_INDIRECT_DESC:
699 case VIRTIO_RING_F_EVENT_IDX:
702 /* We don't understand this bit. */
703 clear_bit(i, vdev->features);
707 EXPORT_SYMBOL_GPL(vring_transport_features);
710 * virtqueue_get_vring_size - return the size of the virtqueue's vring
711 * @vq: the struct virtqueue containing the vring of interest.
713 * Returns the size of the vring. This is mainly used for boasting to
714 * userspace. Unlike other operations, this need not be serialized.
716 unsigned int virtqueue_get_vring_size(struct virtqueue *_vq)
719 struct vring_virtqueue *vq = to_vvq(_vq);
721 return vq->vring.num;
723 EXPORT_SYMBOL_GPL(virtqueue_get_vring_size);
725 MODULE_LICENSE("GPL");