virtio_ring.c revision 3b8706240ee6084ccb46e53cd3a554356b7eeec8 
1/* Virtio ring implementation.
2 *
3 *  Copyright 2007 Rusty Russell IBM Corporation
4 *
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.
9 *
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.
14 *
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
18 */
19#include <linux/virtio.h>
20#include <linux/virtio_ring.h>
21#include <linux/virtio_config.h>
22#include <linux/device.h>
23
24/* virtio guest is communicating with a virtual "device" that actually runs on
25 * a host processor.  Memory barriers are used to control SMP effects. */
26#ifdef CONFIG_SMP
27/* Where possible, use SMP barriers which are more lightweight than mandatory
28 * barriers, because mandatory barriers control MMIO effects on accesses
29 * through relaxed memory I/O windows (which virtio does not use). */
30#define virtio_mb() smp_mb()
31#define virtio_rmb() smp_rmb()
32#define virtio_wmb() smp_wmb()
33#else
34/* We must force memory ordering even if guest is UP since host could be
35 * running on another CPU, but SMP barriers are defined to barrier() in that
36 * configuration. So fall back to mandatory barriers instead. */
37#define virtio_mb() mb()
38#define virtio_rmb() rmb()
39#define virtio_wmb() wmb()
40#endif
41
42#ifdef DEBUG
43/* For development, we want to crash whenever the ring is screwed. */
44#define BAD_RING(_vq, fmt, args...)				\
45	do {							\
46		dev_err(&(_vq)->vq.vdev->dev,			\
47			"%s:"fmt, (_vq)->vq.name, ##args);	\
48		BUG();						\
49	} while (0)
50/* Caller is supposed to guarantee no reentry. */
51#define START_USE(_vq)						\
52	do {							\
53		if ((_vq)->in_use)				\
54			panic("%s:in_use = %i\n",		\
55			      (_vq)->vq.name, (_vq)->in_use);	\
56		(_vq)->in_use = __LINE__;			\
57	} while (0)
58#define END_USE(_vq) \
59	do { BUG_ON(!(_vq)->in_use); (_vq)->in_use = 0; } while(0)
60#else
61#define BAD_RING(_vq, fmt, args...)				\
62	do {							\
63		dev_err(&_vq->vq.vdev->dev,			\
64			"%s:"fmt, (_vq)->vq.name, ##args);	\
65		(_vq)->broken = true;				\
66	} while (0)
67#define START_USE(vq)
68#define END_USE(vq)
69#endif
70
71struct vring_virtqueue
72{
73	struct virtqueue vq;
74
75	/* Actual memory layout for this queue */
76	struct vring vring;
77
78	/* Other side has made a mess, don't try any more. */
79	bool broken;
80
81	/* Host supports indirect buffers */
82	bool indirect;
83
84	/* Number of free buffers */
85	unsigned int num_free;
86	/* Head of free buffer list. */
87	unsigned int free_head;
88	/* Number we've added since last sync. */
89	unsigned int num_added;
90
91	/* Last used index we've seen. */
92	u16 last_used_idx;
93
94	/* How to notify other side. FIXME: commonalize hcalls! */
95	void (*notify)(struct virtqueue *vq);
96
97#ifdef DEBUG
98	/* They're supposed to lock for us. */
99	unsigned int in_use;
100#endif
101
102	/* Tokens for callbacks. */
103	void *data[];
104};
105
106#define to_vvq(_vq) container_of(_vq, struct vring_virtqueue, vq)
107
108/* Set up an indirect table of descriptors and add it to the queue. */
109static int vring_add_indirect(struct vring_virtqueue *vq,
110			      struct scatterlist sg[],
111			      unsigned int out,
112			      unsigned int in)
113{
114	struct vring_desc *desc;
115	unsigned head;
116	int i;
117
118	desc = kmalloc((out + in) * sizeof(struct vring_desc), GFP_ATOMIC);
119	if (!desc)
120		return vq->vring.num;
121
122	/* Transfer entries from the sg list into the indirect page */
123	for (i = 0; i < out; i++) {
124		desc[i].flags = VRING_DESC_F_NEXT;
125		desc[i].addr = sg_phys(sg);
126		desc[i].len = sg->length;
127		desc[i].next = i+1;
128		sg++;
129	}
130	for (; i < (out + in); i++) {
131		desc[i].flags = VRING_DESC_F_NEXT|VRING_DESC_F_WRITE;
132		desc[i].addr = sg_phys(sg);
133		desc[i].len = sg->length;
134		desc[i].next = i+1;
135		sg++;
136	}
137
138	/* Last one doesn't continue. */
139	desc[i-1].flags &= ~VRING_DESC_F_NEXT;
140	desc[i-1].next = 0;
141
142	/* We're about to use a buffer */
143	vq->num_free--;
144
145	/* Use a single buffer which doesn't continue */
146	head = vq->free_head;
147	vq->vring.desc[head].flags = VRING_DESC_F_INDIRECT;
148	vq->vring.desc[head].addr = virt_to_phys(desc);
149	vq->vring.desc[head].len = i * sizeof(struct vring_desc);
150
151	/* Update free pointer */
152	vq->free_head = vq->vring.desc[head].next;
153
154	return head;
155}
156
157static int vring_add_buf(struct virtqueue *_vq,
158			 struct scatterlist sg[],
159			 unsigned int out,
160			 unsigned int in,
161			 void *data)
162{
163	struct vring_virtqueue *vq = to_vvq(_vq);
164	unsigned int i, avail, head, uninitialized_var(prev);
165
166	START_USE(vq);
167
168	BUG_ON(data == NULL);
169
170	/* If the host supports indirect descriptor tables, and we have multiple
171	 * buffers, then go indirect. FIXME: tune this threshold */
172	if (vq->indirect && (out + in) > 1 && vq->num_free) {
173		head = vring_add_indirect(vq, sg, out, in);
174		if (head != vq->vring.num)
175			goto add_head;
176	}
177
178	BUG_ON(out + in > vq->vring.num);
179	BUG_ON(out + in == 0);
180
181	if (vq->num_free < out + in) {
182		pr_debug("Can't add buf len %i - avail = %i\n",
183			 out + in, vq->num_free);
184		/* FIXME: for historical reasons, we force a notify here if
185		 * there are outgoing parts to the buffer.  Presumably the
186		 * host should service the ring ASAP. */
187		if (out)
188			vq->notify(&vq->vq);
189		END_USE(vq);
190		return -ENOSPC;
191	}
192
193	/* We're about to use some buffers from the free list. */
194	vq->num_free -= out + in;
195
196	head = vq->free_head;
197	for (i = vq->free_head; out; i = vq->vring.desc[i].next, out--) {
198		vq->vring.desc[i].flags = VRING_DESC_F_NEXT;
199		vq->vring.desc[i].addr = sg_phys(sg);
200		vq->vring.desc[i].len = sg->length;
201		prev = i;
202		sg++;
203	}
204	for (; in; i = vq->vring.desc[i].next, in--) {
205		vq->vring.desc[i].flags = VRING_DESC_F_NEXT|VRING_DESC_F_WRITE;
206		vq->vring.desc[i].addr = sg_phys(sg);
207		vq->vring.desc[i].len = sg->length;
208		prev = i;
209		sg++;
210	}
211	/* Last one doesn't continue. */
212	vq->vring.desc[prev].flags &= ~VRING_DESC_F_NEXT;
213
214	/* Update free pointer */
215	vq->free_head = i;
216
217add_head:
218	/* Set token. */
219	vq->data[head] = data;
220
221	/* Put entry in available array (but don't update avail->idx until they
222	 * do sync).  FIXME: avoid modulus here? */
223	avail = (vq->vring.avail->idx + vq->num_added++) % vq->vring.num;
224	vq->vring.avail->ring[avail] = head;
225
226	pr_debug("Added buffer head %i to %p\n", head, vq);
227	END_USE(vq);
228
229	/* If we're indirect, we can fit many (assuming not OOM). */
230	if (vq->indirect)
231		return vq->num_free ? vq->vring.num : 0;
232	return vq->num_free;
233}
234
235static void vring_kick(struct virtqueue *_vq)
236{
237	struct vring_virtqueue *vq = to_vvq(_vq);
238	START_USE(vq);
239	/* Descriptors and available array need to be set before we expose the
240	 * new available array entries. */
241	virtio_wmb();
242
243	vq->vring.avail->idx += vq->num_added;
244	vq->num_added = 0;
245
246	/* Need to update avail index before checking if we should notify */
247	virtio_mb();
248
249	if (!(vq->vring.used->flags & VRING_USED_F_NO_NOTIFY))
250		/* Prod other side to tell it about changes. */
251		vq->notify(&vq->vq);
252
253	END_USE(vq);
254}
255
256static void detach_buf(struct vring_virtqueue *vq, unsigned int head)
257{
258	unsigned int i;
259
260	/* Clear data ptr. */
261	vq->data[head] = NULL;
262
263	/* Put back on free list: find end */
264	i = head;
265
266	/* Free the indirect table */
267	if (vq->vring.desc[i].flags & VRING_DESC_F_INDIRECT)
268		kfree(phys_to_virt(vq->vring.desc[i].addr));
269
270	while (vq->vring.desc[i].flags & VRING_DESC_F_NEXT) {
271		i = vq->vring.desc[i].next;
272		vq->num_free++;
273	}
274
275	vq->vring.desc[i].next = vq->free_head;
276	vq->free_head = head;
277	/* Plus final descriptor */
278	vq->num_free++;
279}
280
281static inline bool more_used(const struct vring_virtqueue *vq)
282{
283	return vq->last_used_idx != vq->vring.used->idx;
284}
285
286static void *vring_get_buf(struct virtqueue *_vq, unsigned int *len)
287{
288	struct vring_virtqueue *vq = to_vvq(_vq);
289	void *ret;
290	unsigned int i;
291
292	START_USE(vq);
293
294	if (unlikely(vq->broken)) {
295		END_USE(vq);
296		return NULL;
297	}
298
299	if (!more_used(vq)) {
300		pr_debug("No more buffers in queue\n");
301		END_USE(vq);
302		return NULL;
303	}
304
305	/* Only get used array entries after they have been exposed by host. */
306	virtio_rmb();
307
308	i = vq->vring.used->ring[vq->last_used_idx%vq->vring.num].id;
309	*len = vq->vring.used->ring[vq->last_used_idx%vq->vring.num].len;
310
311	if (unlikely(i >= vq->vring.num)) {
312		BAD_RING(vq, "id %u out of range\n", i);
313		return NULL;
314	}
315	if (unlikely(!vq->data[i])) {
316		BAD_RING(vq, "id %u is not a head!\n", i);
317		return NULL;
318	}
319
320	/* detach_buf clears data, so grab it now. */
321	ret = vq->data[i];
322	detach_buf(vq, i);
323	vq->last_used_idx++;
324	END_USE(vq);
325	return ret;
326}
327
328static void vring_disable_cb(struct virtqueue *_vq)
329{
330	struct vring_virtqueue *vq = to_vvq(_vq);
331
332	vq->vring.avail->flags |= VRING_AVAIL_F_NO_INTERRUPT;
333}
334
335static bool vring_enable_cb(struct virtqueue *_vq)
336{
337	struct vring_virtqueue *vq = to_vvq(_vq);
338
339	START_USE(vq);
340
341	/* We optimistically turn back on interrupts, then check if there was
342	 * more to do. */
343	vq->vring.avail->flags &= ~VRING_AVAIL_F_NO_INTERRUPT;
344	virtio_mb();
345	if (unlikely(more_used(vq))) {
346		END_USE(vq);
347		return false;
348	}
349
350	END_USE(vq);
351	return true;
352}
353
354static void *vring_detach_unused_buf(struct virtqueue *_vq)
355{
356	struct vring_virtqueue *vq = to_vvq(_vq);
357	unsigned int i;
358	void *buf;
359
360	START_USE(vq);
361
362	for (i = 0; i < vq->vring.num; i++) {
363		if (!vq->data[i])
364			continue;
365		/* detach_buf clears data, so grab it now. */
366		buf = vq->data[i];
367		detach_buf(vq, i);
368		END_USE(vq);
369		return buf;
370	}
371	/* That should have freed everything. */
372	BUG_ON(vq->num_free != vq->vring.num);
373
374	END_USE(vq);
375	return NULL;
376}
377
378irqreturn_t vring_interrupt(int irq, void *_vq)
379{
380	struct vring_virtqueue *vq = to_vvq(_vq);
381
382	if (!more_used(vq)) {
383		pr_debug("virtqueue interrupt with no work for %p\n", vq);
384		return IRQ_NONE;
385	}
386
387	if (unlikely(vq->broken))
388		return IRQ_HANDLED;
389
390	pr_debug("virtqueue callback for %p (%p)\n", vq, vq->vq.callback);
391	if (vq->vq.callback)
392		vq->vq.callback(&vq->vq);
393
394	return IRQ_HANDLED;
395}
396EXPORT_SYMBOL_GPL(vring_interrupt);
397
398static struct virtqueue_ops vring_vq_ops = {
399	.add_buf = vring_add_buf,
400	.get_buf = vring_get_buf,
401	.kick = vring_kick,
402	.disable_cb = vring_disable_cb,
403	.enable_cb = vring_enable_cb,
404	.detach_unused_buf = vring_detach_unused_buf,
405};
406
407struct virtqueue *vring_new_virtqueue(unsigned int num,
408				      unsigned int vring_align,
409				      struct virtio_device *vdev,
410				      void *pages,
411				      void (*notify)(struct virtqueue *),
412				      void (*callback)(struct virtqueue *),
413				      const char *name)
414{
415	struct vring_virtqueue *vq;
416	unsigned int i;
417
418	/* We assume num is a power of 2. */
419	if (num & (num - 1)) {
420		dev_warn(&vdev->dev, "Bad virtqueue length %u\n", num);
421		return NULL;
422	}
423
424	vq = kmalloc(sizeof(*vq) + sizeof(void *)*num, GFP_KERNEL);
425	if (!vq)
426		return NULL;
427
428	vring_init(&vq->vring, num, pages, vring_align);
429	vq->vq.callback = callback;
430	vq->vq.vdev = vdev;
431	vq->vq.vq_ops = &vring_vq_ops;
432	vq->vq.name = name;
433	vq->notify = notify;
434	vq->broken = false;
435	vq->last_used_idx = 0;
436	vq->num_added = 0;
437	list_add_tail(&vq->vq.list, &vdev->vqs);
438#ifdef DEBUG
439	vq->in_use = false;
440#endif
441
442	vq->indirect = virtio_has_feature(vdev, VIRTIO_RING_F_INDIRECT_DESC);
443
444	/* No callback?  Tell other side not to bother us. */
445	if (!callback)
446		vq->vring.avail->flags |= VRING_AVAIL_F_NO_INTERRUPT;
447
448	/* Put everything in free lists. */
449	vq->num_free = num;
450	vq->free_head = 0;
451	for (i = 0; i < num-1; i++) {
452		vq->vring.desc[i].next = i+1;
453		vq->data[i] = NULL;
454	}
455	vq->data[i] = NULL;
456
457	return &vq->vq;
458}
459EXPORT_SYMBOL_GPL(vring_new_virtqueue);
460
461void vring_del_virtqueue(struct virtqueue *vq)
462{
463	list_del(&vq->list);
464	kfree(to_vvq(vq));
465}
466EXPORT_SYMBOL_GPL(vring_del_virtqueue);
467
468/* Manipulates transport-specific feature bits. */
469void vring_transport_features(struct virtio_device *vdev)
470{
471	unsigned int i;
472
473	for (i = VIRTIO_TRANSPORT_F_START; i < VIRTIO_TRANSPORT_F_END; i++) {
474		switch (i) {
475		case VIRTIO_RING_F_INDIRECT_DESC:
476			break;
477		default:
478			/* We don't understand this bit. */
479			clear_bit(i, vdev->features);
480		}
481	}
482}
483EXPORT_SYMBOL_GPL(vring_transport_features);
484
485MODULE_LICENSE("GPL");
486