Loading...
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#include <linux/slab.h>
24
25/* virtio guest is communicating with a virtual "device" that actually runs on
26 * a host processor. Memory barriers are used to control SMP effects. */
27#ifdef CONFIG_SMP
28/* Where possible, use SMP barriers which are more lightweight than mandatory
29 * barriers, because mandatory barriers control MMIO effects on accesses
30 * through relaxed memory I/O windows (which virtio does not use). */
31#define virtio_mb() smp_mb()
32#define virtio_rmb() smp_rmb()
33#define virtio_wmb() smp_wmb()
34#else
35/* We must force memory ordering even if guest is UP since host could be
36 * running on another CPU, but SMP barriers are defined to barrier() in that
37 * configuration. So fall back to mandatory barriers instead. */
38#define virtio_mb() mb()
39#define virtio_rmb() rmb()
40#define virtio_wmb() wmb()
41#endif
42
43#ifdef DEBUG
44/* For development, we want to crash whenever the ring is screwed. */
45#define BAD_RING(_vq, fmt, args...) \
46 do { \
47 dev_err(&(_vq)->vq.vdev->dev, \
48 "%s:"fmt, (_vq)->vq.name, ##args); \
49 BUG(); \
50 } while (0)
51/* Caller is supposed to guarantee no reentry. */
52#define START_USE(_vq) \
53 do { \
54 if ((_vq)->in_use) \
55 panic("%s:in_use = %i\n", \
56 (_vq)->vq.name, (_vq)->in_use); \
57 (_vq)->in_use = __LINE__; \
58 } while (0)
59#define END_USE(_vq) \
60 do { BUG_ON(!(_vq)->in_use); (_vq)->in_use = 0; } while(0)
61#else
62#define BAD_RING(_vq, fmt, args...) \
63 do { \
64 dev_err(&_vq->vq.vdev->dev, \
65 "%s:"fmt, (_vq)->vq.name, ##args); \
66 (_vq)->broken = true; \
67 } while (0)
68#define START_USE(vq)
69#define END_USE(vq)
70#endif
71
72struct vring_virtqueue
73{
74 struct virtqueue vq;
75
76 /* Actual memory layout for this queue */
77 struct vring vring;
78
79 /* Other side has made a mess, don't try any more. */
80 bool broken;
81
82 /* Host supports indirect buffers */
83 bool indirect;
84
85 /* Host publishes avail event idx */
86 bool event;
87
88 /* Number of free buffers */
89 unsigned int num_free;
90 /* Head of free buffer list. */
91 unsigned int free_head;
92 /* Number we've added since last sync. */
93 unsigned int num_added;
94
95 /* Last used index we've seen. */
96 u16 last_used_idx;
97
98 /* How to notify other side. FIXME: commonalize hcalls! */
99 void (*notify)(struct virtqueue *vq);
100
101#ifdef DEBUG
102 /* They're supposed to lock for us. */
103 unsigned int in_use;
104#endif
105
106 /* Tokens for callbacks. */
107 void *data[];
108};
109
110#define to_vvq(_vq) container_of(_vq, struct vring_virtqueue, vq)
111
112/* Set up an indirect table of descriptors and add it to the queue. */
113static int vring_add_indirect(struct vring_virtqueue *vq,
114 struct scatterlist sg[],
115 unsigned int out,
116 unsigned int in,
117 gfp_t gfp)
118{
119 struct vring_desc *desc;
120 unsigned head;
121 int i;
122
123 desc = kmalloc((out + in) * sizeof(struct vring_desc), gfp);
124 if (!desc)
125 return -ENOMEM;
126
127 /* Transfer entries from the sg list into the indirect page */
128 for (i = 0; i < out; i++) {
129 desc[i].flags = VRING_DESC_F_NEXT;
130 desc[i].addr = sg_phys(sg);
131 desc[i].len = sg->length;
132 desc[i].next = i+1;
133 sg++;
134 }
135 for (; i < (out + in); i++) {
136 desc[i].flags = VRING_DESC_F_NEXT|VRING_DESC_F_WRITE;
137 desc[i].addr = sg_phys(sg);
138 desc[i].len = sg->length;
139 desc[i].next = i+1;
140 sg++;
141 }
142
143 /* Last one doesn't continue. */
144 desc[i-1].flags &= ~VRING_DESC_F_NEXT;
145 desc[i-1].next = 0;
146
147 /* We're about to use a buffer */
148 vq->num_free--;
149
150 /* Use a single buffer which doesn't continue */
151 head = vq->free_head;
152 vq->vring.desc[head].flags = VRING_DESC_F_INDIRECT;
153 vq->vring.desc[head].addr = virt_to_phys(desc);
154 vq->vring.desc[head].len = i * sizeof(struct vring_desc);
155
156 /* Update free pointer */
157 vq->free_head = vq->vring.desc[head].next;
158
159 return head;
160}
161
162int virtqueue_add_buf_gfp(struct virtqueue *_vq,
163 struct scatterlist sg[],
164 unsigned int out,
165 unsigned int in,
166 void *data,
167 gfp_t gfp)
168{
169 struct vring_virtqueue *vq = to_vvq(_vq);
170 unsigned int i, avail, uninitialized_var(prev);
171 int head;
172
173 START_USE(vq);
174
175 BUG_ON(data == NULL);
176
177 /* If the host supports indirect descriptor tables, and we have multiple
178 * buffers, then go indirect. FIXME: tune this threshold */
179 if (vq->indirect && (out + in) > 1 && vq->num_free) {
180 head = vring_add_indirect(vq, sg, out, in, gfp);
181 if (likely(head >= 0))
182 goto add_head;
183 }
184
185 BUG_ON(out + in > vq->vring.num);
186 BUG_ON(out + in == 0);
187
188 if (vq->num_free < out + in) {
189 pr_debug("Can't add buf len %i - avail = %i\n",
190 out + in, vq->num_free);
191 /* FIXME: for historical reasons, we force a notify here if
192 * there are outgoing parts to the buffer. Presumably the
193 * host should service the ring ASAP. */
194 if (out)
195 vq->notify(&vq->vq);
196 END_USE(vq);
197 return -ENOSPC;
198 }
199
200 /* We're about to use some buffers from the free list. */
201 vq->num_free -= out + in;
202
203 head = vq->free_head;
204 for (i = vq->free_head; out; i = vq->vring.desc[i].next, out--) {
205 vq->vring.desc[i].flags = VRING_DESC_F_NEXT;
206 vq->vring.desc[i].addr = sg_phys(sg);
207 vq->vring.desc[i].len = sg->length;
208 prev = i;
209 sg++;
210 }
211 for (; in; i = vq->vring.desc[i].next, in--) {
212 vq->vring.desc[i].flags = VRING_DESC_F_NEXT|VRING_DESC_F_WRITE;
213 vq->vring.desc[i].addr = sg_phys(sg);
214 vq->vring.desc[i].len = sg->length;
215 prev = i;
216 sg++;
217 }
218 /* Last one doesn't continue. */
219 vq->vring.desc[prev].flags &= ~VRING_DESC_F_NEXT;
220
221 /* Update free pointer */
222 vq->free_head = i;
223
224add_head:
225 /* Set token. */
226 vq->data[head] = data;
227
228 /* Put entry in available array (but don't update avail->idx until they
229 * do sync). FIXME: avoid modulus here? */
230 avail = (vq->vring.avail->idx + vq->num_added++) % vq->vring.num;
231 vq->vring.avail->ring[avail] = head;
232
233 pr_debug("Added buffer head %i to %p\n", head, vq);
234 END_USE(vq);
235
236 return vq->num_free;
237}
238EXPORT_SYMBOL_GPL(virtqueue_add_buf_gfp);
239
240void virtqueue_kick(struct virtqueue *_vq)
241{
242 struct vring_virtqueue *vq = to_vvq(_vq);
243 u16 new, old;
244 START_USE(vq);
245 /* Descriptors and available array need to be set before we expose the
246 * new available array entries. */
247 virtio_wmb();
248
249 old = vq->vring.avail->idx;
250 new = vq->vring.avail->idx = old + vq->num_added;
251 vq->num_added = 0;
252
253 /* Need to update avail index before checking if we should notify */
254 virtio_mb();
255
256 if (vq->event ?
257 vring_need_event(vring_avail_event(&vq->vring), new, old) :
258 !(vq->vring.used->flags & VRING_USED_F_NO_NOTIFY))
259 /* Prod other side to tell it about changes. */
260 vq->notify(&vq->vq);
261
262 END_USE(vq);
263}
264EXPORT_SYMBOL_GPL(virtqueue_kick);
265
266static void detach_buf(struct vring_virtqueue *vq, unsigned int head)
267{
268 unsigned int i;
269
270 /* Clear data ptr. */
271 vq->data[head] = NULL;
272
273 /* Put back on free list: find end */
274 i = head;
275
276 /* Free the indirect table */
277 if (vq->vring.desc[i].flags & VRING_DESC_F_INDIRECT)
278 kfree(phys_to_virt(vq->vring.desc[i].addr));
279
280 while (vq->vring.desc[i].flags & VRING_DESC_F_NEXT) {
281 i = vq->vring.desc[i].next;
282 vq->num_free++;
283 }
284
285 vq->vring.desc[i].next = vq->free_head;
286 vq->free_head = head;
287 /* Plus final descriptor */
288 vq->num_free++;
289}
290
291static inline bool more_used(const struct vring_virtqueue *vq)
292{
293 return vq->last_used_idx != vq->vring.used->idx;
294}
295
296void *virtqueue_get_buf(struct virtqueue *_vq, unsigned int *len)
297{
298 struct vring_virtqueue *vq = to_vvq(_vq);
299 void *ret;
300 unsigned int i;
301
302 START_USE(vq);
303
304 if (unlikely(vq->broken)) {
305 END_USE(vq);
306 return NULL;
307 }
308
309 if (!more_used(vq)) {
310 pr_debug("No more buffers in queue\n");
311 END_USE(vq);
312 return NULL;
313 }
314
315 /* Only get used array entries after they have been exposed by host. */
316 virtio_rmb();
317
318 i = vq->vring.used->ring[vq->last_used_idx%vq->vring.num].id;
319 *len = vq->vring.used->ring[vq->last_used_idx%vq->vring.num].len;
320
321 if (unlikely(i >= vq->vring.num)) {
322 BAD_RING(vq, "id %u out of range\n", i);
323 return NULL;
324 }
325 if (unlikely(!vq->data[i])) {
326 BAD_RING(vq, "id %u is not a head!\n", i);
327 return NULL;
328 }
329
330 /* detach_buf clears data, so grab it now. */
331 ret = vq->data[i];
332 detach_buf(vq, i);
333 vq->last_used_idx++;
334 /* If we expect an interrupt for the next entry, tell host
335 * by writing event index and flush out the write before
336 * the read in the next get_buf call. */
337 if (!(vq->vring.avail->flags & VRING_AVAIL_F_NO_INTERRUPT)) {
338 vring_used_event(&vq->vring) = vq->last_used_idx;
339 virtio_mb();
340 }
341
342 END_USE(vq);
343 return ret;
344}
345EXPORT_SYMBOL_GPL(virtqueue_get_buf);
346
347void virtqueue_disable_cb(struct virtqueue *_vq)
348{
349 struct vring_virtqueue *vq = to_vvq(_vq);
350
351 vq->vring.avail->flags |= VRING_AVAIL_F_NO_INTERRUPT;
352}
353EXPORT_SYMBOL_GPL(virtqueue_disable_cb);
354
355bool virtqueue_enable_cb(struct virtqueue *_vq)
356{
357 struct vring_virtqueue *vq = to_vvq(_vq);
358
359 START_USE(vq);
360
361 /* We optimistically turn back on interrupts, then check if there was
362 * more to do. */
363 /* Depending on the VIRTIO_RING_F_EVENT_IDX feature, we need to
364 * either clear the flags bit or point the event index at the next
365 * entry. Always do both to keep code simple. */
366 vq->vring.avail->flags &= ~VRING_AVAIL_F_NO_INTERRUPT;
367 vring_used_event(&vq->vring) = vq->last_used_idx;
368 virtio_mb();
369 if (unlikely(more_used(vq))) {
370 END_USE(vq);
371 return false;
372 }
373
374 END_USE(vq);
375 return true;
376}
377EXPORT_SYMBOL_GPL(virtqueue_enable_cb);
378
379bool virtqueue_enable_cb_delayed(struct virtqueue *_vq)
380{
381 struct vring_virtqueue *vq = to_vvq(_vq);
382 u16 bufs;
383
384 START_USE(vq);
385
386 /* We optimistically turn back on interrupts, then check if there was
387 * more to do. */
388 /* Depending on the VIRTIO_RING_F_USED_EVENT_IDX feature, we need to
389 * either clear the flags bit or point the event index at the next
390 * entry. Always do both to keep code simple. */
391 vq->vring.avail->flags &= ~VRING_AVAIL_F_NO_INTERRUPT;
392 /* TODO: tune this threshold */
393 bufs = (u16)(vq->vring.avail->idx - vq->last_used_idx) * 3 / 4;
394 vring_used_event(&vq->vring) = vq->last_used_idx + bufs;
395 virtio_mb();
396 if (unlikely((u16)(vq->vring.used->idx - vq->last_used_idx) > bufs)) {
397 END_USE(vq);
398 return false;
399 }
400
401 END_USE(vq);
402 return true;
403}
404EXPORT_SYMBOL_GPL(virtqueue_enable_cb_delayed);
405
406void *virtqueue_detach_unused_buf(struct virtqueue *_vq)
407{
408 struct vring_virtqueue *vq = to_vvq(_vq);
409 unsigned int i;
410 void *buf;
411
412 START_USE(vq);
413
414 for (i = 0; i < vq->vring.num; i++) {
415 if (!vq->data[i])
416 continue;
417 /* detach_buf clears data, so grab it now. */
418 buf = vq->data[i];
419 detach_buf(vq, i);
420 vq->vring.avail->idx--;
421 END_USE(vq);
422 return buf;
423 }
424 /* That should have freed everything. */
425 BUG_ON(vq->num_free != vq->vring.num);
426
427 END_USE(vq);
428 return NULL;
429}
430EXPORT_SYMBOL_GPL(virtqueue_detach_unused_buf);
431
432irqreturn_t vring_interrupt(int irq, void *_vq)
433{
434 struct vring_virtqueue *vq = to_vvq(_vq);
435
436 if (!more_used(vq)) {
437 pr_debug("virtqueue interrupt with no work for %p\n", vq);
438 return IRQ_NONE;
439 }
440
441 if (unlikely(vq->broken))
442 return IRQ_HANDLED;
443
444 pr_debug("virtqueue callback for %p (%p)\n", vq, vq->vq.callback);
445 if (vq->vq.callback)
446 vq->vq.callback(&vq->vq);
447
448 return IRQ_HANDLED;
449}
450EXPORT_SYMBOL_GPL(vring_interrupt);
451
452struct virtqueue *vring_new_virtqueue(unsigned int num,
453 unsigned int vring_align,
454 struct virtio_device *vdev,
455 void *pages,
456 void (*notify)(struct virtqueue *),
457 void (*callback)(struct virtqueue *),
458 const char *name)
459{
460 struct vring_virtqueue *vq;
461 unsigned int i;
462
463 /* We assume num is a power of 2. */
464 if (num & (num - 1)) {
465 dev_warn(&vdev->dev, "Bad virtqueue length %u\n", num);
466 return NULL;
467 }
468
469 vq = kmalloc(sizeof(*vq) + sizeof(void *)*num, GFP_KERNEL);
470 if (!vq)
471 return NULL;
472
473 vring_init(&vq->vring, num, pages, vring_align);
474 vq->vq.callback = callback;
475 vq->vq.vdev = vdev;
476 vq->vq.name = name;
477 vq->notify = notify;
478 vq->broken = false;
479 vq->last_used_idx = 0;
480 vq->num_added = 0;
481 list_add_tail(&vq->vq.list, &vdev->vqs);
482#ifdef DEBUG
483 vq->in_use = false;
484#endif
485
486 vq->indirect = virtio_has_feature(vdev, VIRTIO_RING_F_INDIRECT_DESC);
487 vq->event = virtio_has_feature(vdev, VIRTIO_RING_F_EVENT_IDX);
488
489 /* No callback? Tell other side not to bother us. */
490 if (!callback)
491 vq->vring.avail->flags |= VRING_AVAIL_F_NO_INTERRUPT;
492
493 /* Put everything in free lists. */
494 vq->num_free = num;
495 vq->free_head = 0;
496 for (i = 0; i < num-1; i++) {
497 vq->vring.desc[i].next = i+1;
498 vq->data[i] = NULL;
499 }
500 vq->data[i] = NULL;
501
502 return &vq->vq;
503}
504EXPORT_SYMBOL_GPL(vring_new_virtqueue);
505
506void vring_del_virtqueue(struct virtqueue *vq)
507{
508 list_del(&vq->list);
509 kfree(to_vvq(vq));
510}
511EXPORT_SYMBOL_GPL(vring_del_virtqueue);
512
513/* Manipulates transport-specific feature bits. */
514void vring_transport_features(struct virtio_device *vdev)
515{
516 unsigned int i;
517
518 for (i = VIRTIO_TRANSPORT_F_START; i < VIRTIO_TRANSPORT_F_END; i++) {
519 switch (i) {
520 case VIRTIO_RING_F_INDIRECT_DESC:
521 break;
522 case VIRTIO_RING_F_EVENT_IDX:
523 break;
524 default:
525 /* We don't understand this bit. */
526 clear_bit(i, vdev->features);
527 }
528 }
529}
530EXPORT_SYMBOL_GPL(vring_transport_features);
531
532MODULE_LICENSE("GPL");
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#include <linux/slab.h>
24#include <linux/module.h>
25#include <linux/hrtimer.h>
26
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. */
29#ifdef CONFIG_SMP
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)
39#else
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()
46#endif
47
48#ifdef DEBUG
49/* For development, we want to crash whenever the ring is screwed. */
50#define BAD_RING(_vq, fmt, args...) \
51 do { \
52 dev_err(&(_vq)->vq.vdev->dev, \
53 "%s:"fmt, (_vq)->vq.name, ##args); \
54 BUG(); \
55 } while (0)
56/* Caller is supposed to guarantee no reentry. */
57#define START_USE(_vq) \
58 do { \
59 if ((_vq)->in_use) \
60 panic("%s:in_use = %i\n", \
61 (_vq)->vq.name, (_vq)->in_use); \
62 (_vq)->in_use = __LINE__; \
63 } while (0)
64#define END_USE(_vq) \
65 do { BUG_ON(!(_vq)->in_use); (_vq)->in_use = 0; } while(0)
66#else
67#define BAD_RING(_vq, fmt, args...) \
68 do { \
69 dev_err(&_vq->vq.vdev->dev, \
70 "%s:"fmt, (_vq)->vq.name, ##args); \
71 (_vq)->broken = true; \
72 } while (0)
73#define START_USE(vq)
74#define END_USE(vq)
75#endif
76
77struct vring_virtqueue
78{
79 struct virtqueue vq;
80
81 /* Actual memory layout for this queue */
82 struct vring vring;
83
84 /* Can we use weak barriers? */
85 bool weak_barriers;
86
87 /* Other side has made a mess, don't try any more. */
88 bool broken;
89
90 /* Host supports indirect buffers */
91 bool indirect;
92
93 /* Host publishes avail event idx */
94 bool event;
95
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;
102
103 /* Last used index we've seen. */
104 u16 last_used_idx;
105
106 /* How to notify other side. FIXME: commonalize hcalls! */
107 void (*notify)(struct virtqueue *vq);
108
109#ifdef DEBUG
110 /* They're supposed to lock for us. */
111 unsigned int in_use;
112
113 /* Figure out if their kicks are too delayed. */
114 bool last_add_time_valid;
115 ktime_t last_add_time;
116#endif
117
118 /* Tokens for callbacks. */
119 void *data[];
120};
121
122#define to_vvq(_vq) container_of(_vq, struct vring_virtqueue, vq)
123
124/* Set up an indirect table of descriptors and add it to the queue. */
125static int vring_add_indirect(struct vring_virtqueue *vq,
126 struct scatterlist sg[],
127 unsigned int out,
128 unsigned int in,
129 gfp_t gfp)
130{
131 struct vring_desc *desc;
132 unsigned head;
133 int i;
134
135 desc = kmalloc((out + in) * sizeof(struct vring_desc), gfp);
136 if (!desc)
137 return -ENOMEM;
138
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;
144 desc[i].next = i+1;
145 sg++;
146 }
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;
151 desc[i].next = i+1;
152 sg++;
153 }
154
155 /* Last one doesn't continue. */
156 desc[i-1].flags &= ~VRING_DESC_F_NEXT;
157 desc[i-1].next = 0;
158
159 /* We're about to use a buffer */
160 vq->num_free--;
161
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);
167
168 /* Update free pointer */
169 vq->free_head = vq->vring.desc[head].next;
170
171 return head;
172}
173
174/**
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).
182 *
183 * Caller must ensure we don't call this with other virtqueue operations
184 * at the same time (except where noted).
185 *
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.
190 */
191int virtqueue_add_buf(struct virtqueue *_vq,
192 struct scatterlist sg[],
193 unsigned int out,
194 unsigned int in,
195 void *data,
196 gfp_t gfp)
197{
198 struct vring_virtqueue *vq = to_vvq(_vq);
199 unsigned int i, avail, uninitialized_var(prev);
200 int head;
201
202 START_USE(vq);
203
204 BUG_ON(data == NULL);
205
206#ifdef DEBUG
207 {
208 ktime_t now = ktime_get();
209
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))
213 > 100);
214 vq->last_add_time = now;
215 vq->last_add_time_valid = true;
216 }
217#endif
218
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))
224 goto add_head;
225 }
226
227 BUG_ON(out + in > vq->vring.num);
228 BUG_ON(out + in == 0);
229
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. */
236 if (out)
237 vq->notify(&vq->vq);
238 END_USE(vq);
239 return -ENOSPC;
240 }
241
242 /* We're about to use some buffers from the free list. */
243 vq->num_free -= out + in;
244
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;
250 prev = i;
251 sg++;
252 }
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;
257 prev = i;
258 sg++;
259 }
260 /* Last one doesn't continue. */
261 vq->vring.desc[prev].flags &= ~VRING_DESC_F_NEXT;
262
263 /* Update free pointer */
264 vq->free_head = i;
265
266add_head:
267 /* Set token. */
268 vq->data[head] = data;
269
270 /* Put entry in available array (but don't update avail->idx until they
271 * do sync). */
272 avail = (vq->vring.avail->idx & (vq->vring.num-1));
273 vq->vring.avail->ring[avail] = head;
274
275 /* Descriptors and available array need to be set before we expose the
276 * new available array entries. */
277 virtio_wmb(vq);
278 vq->vring.avail->idx++;
279 vq->num_added++;
280
281 /* This is very unlikely, but theoretically possible. Kick
282 * just in case. */
283 if (unlikely(vq->num_added == (1 << 16) - 1))
284 virtqueue_kick(_vq);
285
286 pr_debug("Added buffer head %i to %p\n", head, vq);
287 END_USE(vq);
288
289 return vq->num_free;
290}
291EXPORT_SYMBOL_GPL(virtqueue_add_buf);
292
293/**
294 * virtqueue_kick_prepare - first half of split virtqueue_kick call.
295 * @vq: the struct virtqueue
296 *
297 * Instead of virtqueue_kick(), you can do:
298 * if (virtqueue_kick_prepare(vq))
299 * virtqueue_notify(vq);
300 *
301 * This is sometimes useful because the virtqueue_kick_prepare() needs
302 * to be serialized, but the actual virtqueue_notify() call does not.
303 */
304bool virtqueue_kick_prepare(struct virtqueue *_vq)
305{
306 struct vring_virtqueue *vq = to_vvq(_vq);
307 u16 new, old;
308 bool needs_kick;
309
310 START_USE(vq);
311 /* We need to expose available array entries before checking avail
312 * event. */
313 virtio_mb(vq);
314
315 old = vq->vring.avail->idx - vq->num_added;
316 new = vq->vring.avail->idx;
317 vq->num_added = 0;
318
319#ifdef DEBUG
320 if (vq->last_add_time_valid) {
321 WARN_ON(ktime_to_ms(ktime_sub(ktime_get(),
322 vq->last_add_time)) > 100);
323 }
324 vq->last_add_time_valid = false;
325#endif
326
327 if (vq->event) {
328 needs_kick = vring_need_event(vring_avail_event(&vq->vring),
329 new, old);
330 } else {
331 needs_kick = !(vq->vring.used->flags & VRING_USED_F_NO_NOTIFY);
332 }
333 END_USE(vq);
334 return needs_kick;
335}
336EXPORT_SYMBOL_GPL(virtqueue_kick_prepare);
337
338/**
339 * virtqueue_notify - second half of split virtqueue_kick call.
340 * @vq: the struct virtqueue
341 *
342 * This does not need to be serialized.
343 */
344void virtqueue_notify(struct virtqueue *_vq)
345{
346 struct vring_virtqueue *vq = to_vvq(_vq);
347
348 /* Prod other side to tell it about changes. */
349 vq->notify(_vq);
350}
351EXPORT_SYMBOL_GPL(virtqueue_notify);
352
353/**
354 * virtqueue_kick - update after add_buf
355 * @vq: the struct virtqueue
356 *
357 * After one or more virtqueue_add_buf calls, invoke this to kick
358 * the other side.
359 *
360 * Caller must ensure we don't call this with other virtqueue
361 * operations at the same time (except where noted).
362 */
363void virtqueue_kick(struct virtqueue *vq)
364{
365 if (virtqueue_kick_prepare(vq))
366 virtqueue_notify(vq);
367}
368EXPORT_SYMBOL_GPL(virtqueue_kick);
369
370static void detach_buf(struct vring_virtqueue *vq, unsigned int head)
371{
372 unsigned int i;
373
374 /* Clear data ptr. */
375 vq->data[head] = NULL;
376
377 /* Put back on free list: find end */
378 i = head;
379
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));
383
384 while (vq->vring.desc[i].flags & VRING_DESC_F_NEXT) {
385 i = vq->vring.desc[i].next;
386 vq->num_free++;
387 }
388
389 vq->vring.desc[i].next = vq->free_head;
390 vq->free_head = head;
391 /* Plus final descriptor */
392 vq->num_free++;
393}
394
395static inline bool more_used(const struct vring_virtqueue *vq)
396{
397 return vq->last_used_idx != vq->vring.used->idx;
398}
399
400/**
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
404 *
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
408 * writes.
409 *
410 * Caller must ensure we don't call this with other virtqueue
411 * operations at the same time (except where noted).
412 *
413 * Returns NULL if there are no used buffers, or the "data" token
414 * handed to virtqueue_add_buf().
415 */
416void *virtqueue_get_buf(struct virtqueue *_vq, unsigned int *len)
417{
418 struct vring_virtqueue *vq = to_vvq(_vq);
419 void *ret;
420 unsigned int i;
421 u16 last_used;
422
423 START_USE(vq);
424
425 if (unlikely(vq->broken)) {
426 END_USE(vq);
427 return NULL;
428 }
429
430 if (!more_used(vq)) {
431 pr_debug("No more buffers in queue\n");
432 END_USE(vq);
433 return NULL;
434 }
435
436 /* Only get used array entries after they have been exposed by host. */
437 virtio_rmb(vq);
438
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;
442
443 if (unlikely(i >= vq->vring.num)) {
444 BAD_RING(vq, "id %u out of range\n", i);
445 return NULL;
446 }
447 if (unlikely(!vq->data[i])) {
448 BAD_RING(vq, "id %u is not a head!\n", i);
449 return NULL;
450 }
451
452 /* detach_buf clears data, so grab it now. */
453 ret = vq->data[i];
454 detach_buf(vq, i);
455 vq->last_used_idx++;
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;
461 virtio_mb(vq);
462 }
463
464#ifdef DEBUG
465 vq->last_add_time_valid = false;
466#endif
467
468 END_USE(vq);
469 return ret;
470}
471EXPORT_SYMBOL_GPL(virtqueue_get_buf);
472
473/**
474 * virtqueue_disable_cb - disable callbacks
475 * @vq: the struct virtqueue we're talking about.
476 *
477 * Note that this is not necessarily synchronous, hence unreliable and only
478 * useful as an optimization.
479 *
480 * Unlike other operations, this need not be serialized.
481 */
482void virtqueue_disable_cb(struct virtqueue *_vq)
483{
484 struct vring_virtqueue *vq = to_vvq(_vq);
485
486 vq->vring.avail->flags |= VRING_AVAIL_F_NO_INTERRUPT;
487}
488EXPORT_SYMBOL_GPL(virtqueue_disable_cb);
489
490/**
491 * virtqueue_enable_cb - restart callbacks after disable_cb.
492 * @vq: the struct virtqueue we're talking about.
493 *
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.
497 *
498 * Caller must ensure we don't call this with other virtqueue
499 * operations at the same time (except where noted).
500 */
501bool virtqueue_enable_cb(struct virtqueue *_vq)
502{
503 struct vring_virtqueue *vq = to_vvq(_vq);
504
505 START_USE(vq);
506
507 /* We optimistically turn back on interrupts, then check if there was
508 * more to do. */
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;
514 virtio_mb(vq);
515 if (unlikely(more_used(vq))) {
516 END_USE(vq);
517 return false;
518 }
519
520 END_USE(vq);
521 return true;
522}
523EXPORT_SYMBOL_GPL(virtqueue_enable_cb);
524
525/**
526 * virtqueue_enable_cb_delayed - restart callbacks after disable_cb.
527 * @vq: the struct virtqueue we're talking about.
528 *
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.
534 *
535 * Caller must ensure we don't call this with other virtqueue
536 * operations at the same time (except where noted).
537 */
538bool virtqueue_enable_cb_delayed(struct virtqueue *_vq)
539{
540 struct vring_virtqueue *vq = to_vvq(_vq);
541 u16 bufs;
542
543 START_USE(vq);
544
545 /* We optimistically turn back on interrupts, then check if there was
546 * more to do. */
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;
554 virtio_mb(vq);
555 if (unlikely((u16)(vq->vring.used->idx - vq->last_used_idx) > bufs)) {
556 END_USE(vq);
557 return false;
558 }
559
560 END_USE(vq);
561 return true;
562}
563EXPORT_SYMBOL_GPL(virtqueue_enable_cb_delayed);
564
565/**
566 * virtqueue_detach_unused_buf - detach first unused buffer
567 * @vq: the struct virtqueue we're talking about.
568 *
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
571 * shutdown.
572 */
573void *virtqueue_detach_unused_buf(struct virtqueue *_vq)
574{
575 struct vring_virtqueue *vq = to_vvq(_vq);
576 unsigned int i;
577 void *buf;
578
579 START_USE(vq);
580
581 for (i = 0; i < vq->vring.num; i++) {
582 if (!vq->data[i])
583 continue;
584 /* detach_buf clears data, so grab it now. */
585 buf = vq->data[i];
586 detach_buf(vq, i);
587 vq->vring.avail->idx--;
588 END_USE(vq);
589 return buf;
590 }
591 /* That should have freed everything. */
592 BUG_ON(vq->num_free != vq->vring.num);
593
594 END_USE(vq);
595 return NULL;
596}
597EXPORT_SYMBOL_GPL(virtqueue_detach_unused_buf);
598
599irqreturn_t vring_interrupt(int irq, void *_vq)
600{
601 struct vring_virtqueue *vq = to_vvq(_vq);
602
603 if (!more_used(vq)) {
604 pr_debug("virtqueue interrupt with no work for %p\n", vq);
605 return IRQ_NONE;
606 }
607
608 if (unlikely(vq->broken))
609 return IRQ_HANDLED;
610
611 pr_debug("virtqueue callback for %p (%p)\n", vq, vq->vq.callback);
612 if (vq->vq.callback)
613 vq->vq.callback(&vq->vq);
614
615 return IRQ_HANDLED;
616}
617EXPORT_SYMBOL_GPL(vring_interrupt);
618
619struct virtqueue *vring_new_virtqueue(unsigned int num,
620 unsigned int vring_align,
621 struct virtio_device *vdev,
622 bool weak_barriers,
623 void *pages,
624 void (*notify)(struct virtqueue *),
625 void (*callback)(struct virtqueue *),
626 const char *name)
627{
628 struct vring_virtqueue *vq;
629 unsigned int i;
630
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);
634 return NULL;
635 }
636
637 vq = kmalloc(sizeof(*vq) + sizeof(void *)*num, GFP_KERNEL);
638 if (!vq)
639 return NULL;
640
641 vring_init(&vq->vring, num, pages, vring_align);
642 vq->vq.callback = callback;
643 vq->vq.vdev = vdev;
644 vq->vq.name = name;
645 vq->notify = notify;
646 vq->weak_barriers = weak_barriers;
647 vq->broken = false;
648 vq->last_used_idx = 0;
649 vq->num_added = 0;
650 list_add_tail(&vq->vq.list, &vdev->vqs);
651#ifdef DEBUG
652 vq->in_use = false;
653 vq->last_add_time_valid = false;
654#endif
655
656 vq->indirect = virtio_has_feature(vdev, VIRTIO_RING_F_INDIRECT_DESC);
657 vq->event = virtio_has_feature(vdev, VIRTIO_RING_F_EVENT_IDX);
658
659 /* No callback? Tell other side not to bother us. */
660 if (!callback)
661 vq->vring.avail->flags |= VRING_AVAIL_F_NO_INTERRUPT;
662
663 /* Put everything in free lists. */
664 vq->num_free = num;
665 vq->free_head = 0;
666 for (i = 0; i < num-1; i++) {
667 vq->vring.desc[i].next = i+1;
668 vq->data[i] = NULL;
669 }
670 vq->data[i] = NULL;
671
672 return &vq->vq;
673}
674EXPORT_SYMBOL_GPL(vring_new_virtqueue);
675
676void vring_del_virtqueue(struct virtqueue *vq)
677{
678 list_del(&vq->list);
679 kfree(to_vvq(vq));
680}
681EXPORT_SYMBOL_GPL(vring_del_virtqueue);
682
683/* Manipulates transport-specific feature bits. */
684void vring_transport_features(struct virtio_device *vdev)
685{
686 unsigned int i;
687
688 for (i = VIRTIO_TRANSPORT_F_START; i < VIRTIO_TRANSPORT_F_END; i++) {
689 switch (i) {
690 case VIRTIO_RING_F_INDIRECT_DESC:
691 break;
692 case VIRTIO_RING_F_EVENT_IDX:
693 break;
694 default:
695 /* We don't understand this bit. */
696 clear_bit(i, vdev->features);
697 }
698 }
699}
700EXPORT_SYMBOL_GPL(vring_transport_features);
701
702/**
703 * virtqueue_get_vring_size - return the size of the virtqueue's vring
704 * @vq: the struct virtqueue containing the vring of interest.
705 *
706 * Returns the size of the vring. This is mainly used for boasting to
707 * userspace. Unlike other operations, this need not be serialized.
708 */
709unsigned int virtqueue_get_vring_size(struct virtqueue *_vq)
710{
711
712 struct vring_virtqueue *vq = to_vvq(_vq);
713
714 return vq->vring.num;
715}
716EXPORT_SYMBOL_GPL(virtqueue_get_vring_size);
717
718MODULE_LICENSE("GPL");