1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * VIRTIO based driver for vDPA device
4 *
5 * Copyright (c) 2020, Red Hat. All rights reserved.
6 * Author: Jason Wang <jasowang@redhat.com>
7 *
8 */
9
10#include <linux/init.h>
11#include <linux/module.h>
12#include <linux/device.h>
13#include <linux/kernel.h>
14#include <linux/slab.h>
15#include <linux/uuid.h>
16#include <linux/group_cpus.h>
17#include <linux/virtio.h>
18#include <linux/vdpa.h>
19#include <linux/virtio_config.h>
20#include <linux/virtio_ring.h>
21
22#define MOD_VERSION "0.1"
23#define MOD_AUTHOR "Jason Wang <jasowang@redhat.com>"
24#define MOD_DESC "vDPA bus driver for virtio devices"
25#define MOD_LICENSE "GPL v2"
26
27struct virtio_vdpa_device {
28 struct virtio_device vdev;
29 struct vdpa_device *vdpa;
30 u64 features;
31
32 /* The lock to protect virtqueue list */
33 spinlock_t lock;
34 /* List of virtio_vdpa_vq_info */
35 struct list_head virtqueues;
36};
37
38struct virtio_vdpa_vq_info {
39 /* the actual virtqueue */
40 struct virtqueue *vq;
41
42 /* the list node for the virtqueues list */
43 struct list_head node;
44};
45
46static inline struct virtio_vdpa_device *
47to_virtio_vdpa_device(struct virtio_device *dev)
48{
49 return container_of(dev, struct virtio_vdpa_device, vdev);
50}
51
52static struct vdpa_device *vd_get_vdpa(struct virtio_device *vdev)
53{
54 return to_virtio_vdpa_device(dev: vdev)->vdpa;
55}
56
57static void virtio_vdpa_get(struct virtio_device *vdev, unsigned int offset,
58 void *buf, unsigned int len)
59{
60 struct vdpa_device *vdpa = vd_get_vdpa(vdev);
61
62 vdpa_get_config(vdev: vdpa, offset, buf, len);
63}
64
65static void virtio_vdpa_set(struct virtio_device *vdev, unsigned int offset,
66 const void *buf, unsigned int len)
67{
68 struct vdpa_device *vdpa = vd_get_vdpa(vdev);
69
70 vdpa_set_config(dev: vdpa, offset, buf, length: len);
71}
72
73static u32 virtio_vdpa_generation(struct virtio_device *vdev)
74{
75 struct vdpa_device *vdpa = vd_get_vdpa(vdev);
76 const struct vdpa_config_ops *ops = vdpa->config;
77
78 if (ops->get_generation)
79 return ops->get_generation(vdpa);
80
81 return 0;
82}
83
84static u8 virtio_vdpa_get_status(struct virtio_device *vdev)
85{
86 struct vdpa_device *vdpa = vd_get_vdpa(vdev);
87 const struct vdpa_config_ops *ops = vdpa->config;
88
89 return ops->get_status(vdpa);
90}
91
92static void virtio_vdpa_set_status(struct virtio_device *vdev, u8 status)
93{
94 struct vdpa_device *vdpa = vd_get_vdpa(vdev);
95
96 return vdpa_set_status(vdev: vdpa, status);
97}
98
99static void virtio_vdpa_reset(struct virtio_device *vdev)
100{
101 struct vdpa_device *vdpa = vd_get_vdpa(vdev);
102
103 vdpa_reset(vdev: vdpa, flags: 0);
104}
105
106static bool virtio_vdpa_notify(struct virtqueue *vq)
107{
108 struct vdpa_device *vdpa = vd_get_vdpa(vdev: vq->vdev);
109 const struct vdpa_config_ops *ops = vdpa->config;
110
111 ops->kick_vq(vdpa, vq->index);
112
113 return true;
114}
115
116static bool virtio_vdpa_notify_with_data(struct virtqueue *vq)
117{
118 struct vdpa_device *vdpa = vd_get_vdpa(vdev: vq->vdev);
119 const struct vdpa_config_ops *ops = vdpa->config;
120 u32 data = vring_notification_data(vq: vq);
121
122 ops->kick_vq_with_data(vdpa, data);
123
124 return true;
125}
126
127static irqreturn_t virtio_vdpa_config_cb(void *private)
128{
129 struct virtio_vdpa_device *vd_dev = private;
130
131 virtio_config_changed(dev: &vd_dev->vdev);
132
133 return IRQ_HANDLED;
134}
135
136static irqreturn_t virtio_vdpa_virtqueue_cb(void *private)
137{
138 struct virtio_vdpa_vq_info *info = private;
139
140 return vring_interrupt(irq: 0, vq: info->vq);
141}
142
143static struct virtqueue *
144virtio_vdpa_setup_vq(struct virtio_device *vdev, unsigned int index,
145 void (*callback)(struct virtqueue *vq),
146 const char *name, bool ctx)
147{
148 struct virtio_vdpa_device *vd_dev = to_virtio_vdpa_device(dev: vdev);
149 struct vdpa_device *vdpa = vd_get_vdpa(vdev);
150 struct device *dma_dev;
151 const struct vdpa_config_ops *ops = vdpa->config;
152 struct virtio_vdpa_vq_info *info;
153 bool (*notify)(struct virtqueue *vq) = virtio_vdpa_notify;
154 struct vdpa_callback cb;
155 struct virtqueue *vq;
156 u64 desc_addr, driver_addr, device_addr;
157 /* Assume split virtqueue, switch to packed if necessary */
158 struct vdpa_vq_state state = {0};
159 unsigned long flags;
160 u32 align, max_num, min_num = 1;
161 bool may_reduce_num = true;
162 int err;
163
164 if (!name)
165 return NULL;
166
167 if (index >= vdpa->nvqs)
168 return ERR_PTR(error: -ENOENT);
169
170 /* We cannot accept VIRTIO_F_NOTIFICATION_DATA without kick_vq_with_data */
171 if (__virtio_test_bit(vdev, VIRTIO_F_NOTIFICATION_DATA)) {
172 if (ops->kick_vq_with_data)
173 notify = virtio_vdpa_notify_with_data;
174 else
175 __virtio_clear_bit(vdev, VIRTIO_F_NOTIFICATION_DATA);
176 }
177
178 /* Queue shouldn't already be set up. */
179 if (ops->get_vq_ready(vdpa, index))
180 return ERR_PTR(error: -ENOENT);
181
182 /* Allocate and fill out our active queue description */
183 info = kmalloc(size: sizeof(*info), GFP_KERNEL);
184 if (!info)
185 return ERR_PTR(error: -ENOMEM);
186
187 max_num = ops->get_vq_num_max(vdpa);
188 if (max_num == 0) {
189 err = -ENOENT;
190 goto error_new_virtqueue;
191 }
192
193 if (ops->get_vq_num_min)
194 min_num = ops->get_vq_num_min(vdpa);
195
196 may_reduce_num = (max_num == min_num) ? false : true;
197
198 /* Create the vring */
199 align = ops->get_vq_align(vdpa);
200
201 if (ops->get_vq_dma_dev)
202 dma_dev = ops->get_vq_dma_dev(vdpa, index);
203 else
204 dma_dev = vdpa_get_dma_dev(vdev: vdpa);
205 vq = vring_create_virtqueue_dma(index, num: max_num, vring_align: align, vdev,
206 weak_barriers: true, may_reduce_num, ctx,
207 notify, callback, name, dma_dev);
208 if (!vq) {
209 err = -ENOMEM;
210 goto error_new_virtqueue;
211 }
212
213 vq->num_max = max_num;
214
215 /* Setup virtqueue callback */
216 cb.callback = callback ? virtio_vdpa_virtqueue_cb : NULL;
217 cb.private = info;
218 cb.trigger = NULL;
219 ops->set_vq_cb(vdpa, index, &cb);
220 ops->set_vq_num(vdpa, index, virtqueue_get_vring_size(vq));
221
222 desc_addr = virtqueue_get_desc_addr(vq);
223 driver_addr = virtqueue_get_avail_addr(vq);
224 device_addr = virtqueue_get_used_addr(vq);
225
226 if (ops->set_vq_address(vdpa, index,
227 desc_addr, driver_addr,
228 device_addr)) {
229 err = -EINVAL;
230 goto err_vq;
231 }
232
233 /* reset virtqueue state index */
234 if (virtio_has_feature(vdev, VIRTIO_F_RING_PACKED)) {
235 struct vdpa_vq_state_packed *s = &state.packed;
236
237 s->last_avail_counter = 1;
238 s->last_avail_idx = 0;
239 s->last_used_counter = 1;
240 s->last_used_idx = 0;
241 }
242 err = ops->set_vq_state(vdpa, index, &state);
243 if (err)
244 goto err_vq;
245
246 ops->set_vq_ready(vdpa, index, 1);
247
248 vq->priv = info;
249 info->vq = vq;
250
251 spin_lock_irqsave(&vd_dev->lock, flags);
252 list_add(new: &info->node, head: &vd_dev->virtqueues);
253 spin_unlock_irqrestore(lock: &vd_dev->lock, flags);
254
255 return vq;
256
257err_vq:
258 vring_del_virtqueue(vq);
259error_new_virtqueue:
260 ops->set_vq_ready(vdpa, index, 0);
261 /* VDPA driver should make sure vq is stopeed here */
262 WARN_ON(ops->get_vq_ready(vdpa, index));
263 kfree(objp: info);
264 return ERR_PTR(error: err);
265}
266
267static void virtio_vdpa_del_vq(struct virtqueue *vq)
268{
269 struct virtio_vdpa_device *vd_dev = to_virtio_vdpa_device(dev: vq->vdev);
270 struct vdpa_device *vdpa = vd_dev->vdpa;
271 const struct vdpa_config_ops *ops = vdpa->config;
272 struct virtio_vdpa_vq_info *info = vq->priv;
273 unsigned int index = vq->index;
274 unsigned long flags;
275
276 spin_lock_irqsave(&vd_dev->lock, flags);
277 list_del(entry: &info->node);
278 spin_unlock_irqrestore(lock: &vd_dev->lock, flags);
279
280 /* Select and deactivate the queue (best effort) */
281 ops->set_vq_ready(vdpa, index, 0);
282
283 vring_del_virtqueue(vq);
284
285 kfree(objp: info);
286}
287
288static void virtio_vdpa_del_vqs(struct virtio_device *vdev)
289{
290 struct virtqueue *vq, *n;
291
292 list_for_each_entry_safe(vq, n, &vdev->vqs, list)
293 virtio_vdpa_del_vq(vq);
294}
295
296static void default_calc_sets(struct irq_affinity *affd, unsigned int affvecs)
297{
298 affd->nr_sets = 1;
299 affd->set_size[0] = affvecs;
300}
301
302static struct cpumask *
303create_affinity_masks(unsigned int nvecs, struct irq_affinity *affd)
304{
305 unsigned int affvecs = 0, curvec, usedvecs, i;
306 struct cpumask *masks = NULL;
307
308 if (nvecs > affd->pre_vectors + affd->post_vectors)
309 affvecs = nvecs - affd->pre_vectors - affd->post_vectors;
310
311 if (!affd->calc_sets)
312 affd->calc_sets = default_calc_sets;
313
314 affd->calc_sets(affd, affvecs);
315
316 if (!affvecs)
317 return NULL;
318
319 masks = kcalloc(n: nvecs, size: sizeof(*masks), GFP_KERNEL);
320 if (!masks)
321 return NULL;
322
323 /* Fill out vectors at the beginning that don't need affinity */
324 for (curvec = 0; curvec < affd->pre_vectors; curvec++)
325 cpumask_setall(dstp: &masks[curvec]);
326
327 for (i = 0, usedvecs = 0; i < affd->nr_sets; i++) {
328 unsigned int this_vecs = affd->set_size[i];
329 int j;
330 struct cpumask *result = group_cpus_evenly(numgrps: this_vecs);
331
332 if (!result) {
333 kfree(objp: masks);
334 return NULL;
335 }
336
337 for (j = 0; j < this_vecs; j++)
338 cpumask_copy(dstp: &masks[curvec + j], srcp: &result[j]);
339 kfree(objp: result);
340
341 curvec += this_vecs;
342 usedvecs += this_vecs;
343 }
344
345 /* Fill out vectors at the end that don't need affinity */
346 if (usedvecs >= affvecs)
347 curvec = affd->pre_vectors + affvecs;
348 else
349 curvec = affd->pre_vectors + usedvecs;
350 for (; curvec < nvecs; curvec++)
351 cpumask_setall(dstp: &masks[curvec]);
352
353 return masks;
354}
355
356static int virtio_vdpa_find_vqs(struct virtio_device *vdev, unsigned int nvqs,
357 struct virtqueue *vqs[],
358 vq_callback_t *callbacks[],
359 const char * const names[],
360 const bool *ctx,
361 struct irq_affinity *desc)
362{
363 struct virtio_vdpa_device *vd_dev = to_virtio_vdpa_device(dev: vdev);
364 struct vdpa_device *vdpa = vd_get_vdpa(vdev);
365 const struct vdpa_config_ops *ops = vdpa->config;
366 struct irq_affinity default_affd = { 0 };
367 struct cpumask *masks;
368 struct vdpa_callback cb;
369 bool has_affinity = desc && ops->set_vq_affinity;
370 int i, err, queue_idx = 0;
371
372 if (has_affinity) {
373 masks = create_affinity_masks(nvecs: nvqs, affd: desc ? desc : &default_affd);
374 if (!masks)
375 return -ENOMEM;
376 }
377
378 for (i = 0; i < nvqs; ++i) {
379 if (!names[i]) {
380 vqs[i] = NULL;
381 continue;
382 }
383
384 vqs[i] = virtio_vdpa_setup_vq(vdev, index: queue_idx++,
385 callback: callbacks[i], name: names[i], ctx: ctx ?
386 ctx[i] : false);
387 if (IS_ERR(ptr: vqs[i])) {
388 err = PTR_ERR(ptr: vqs[i]);
389 goto err_setup_vq;
390 }
391
392 if (has_affinity)
393 ops->set_vq_affinity(vdpa, i, &masks[i]);
394 }
395
396 cb.callback = virtio_vdpa_config_cb;
397 cb.private = vd_dev;
398 ops->set_config_cb(vdpa, &cb);
399 if (has_affinity)
400 kfree(objp: masks);
401
402 return 0;
403
404err_setup_vq:
405 virtio_vdpa_del_vqs(vdev);
406 if (has_affinity)
407 kfree(objp: masks);
408 return err;
409}
410
411static u64 virtio_vdpa_get_features(struct virtio_device *vdev)
412{
413 struct vdpa_device *vdpa = vd_get_vdpa(vdev);
414 const struct vdpa_config_ops *ops = vdpa->config;
415
416 return ops->get_device_features(vdpa);
417}
418
419static int virtio_vdpa_finalize_features(struct virtio_device *vdev)
420{
421 struct vdpa_device *vdpa = vd_get_vdpa(vdev);
422
423 /* Give virtio_ring a chance to accept features. */
424 vring_transport_features(vdev);
425
426 return vdpa_set_features(vdev: vdpa, features: vdev->features);
427}
428
429static const char *virtio_vdpa_bus_name(struct virtio_device *vdev)
430{
431 struct virtio_vdpa_device *vd_dev = to_virtio_vdpa_device(dev: vdev);
432 struct vdpa_device *vdpa = vd_dev->vdpa;
433
434 return dev_name(dev: &vdpa->dev);
435}
436
437static int virtio_vdpa_set_vq_affinity(struct virtqueue *vq,
438 const struct cpumask *cpu_mask)
439{
440 struct virtio_vdpa_device *vd_dev = to_virtio_vdpa_device(dev: vq->vdev);
441 struct vdpa_device *vdpa = vd_dev->vdpa;
442 const struct vdpa_config_ops *ops = vdpa->config;
443 unsigned int index = vq->index;
444
445 if (ops->set_vq_affinity)
446 return ops->set_vq_affinity(vdpa, index, cpu_mask);
447
448 return 0;
449}
450
451static const struct cpumask *
452virtio_vdpa_get_vq_affinity(struct virtio_device *vdev, int index)
453{
454 struct vdpa_device *vdpa = vd_get_vdpa(vdev);
455 const struct vdpa_config_ops *ops = vdpa->config;
456
457 if (ops->get_vq_affinity)
458 return ops->get_vq_affinity(vdpa, index);
459
460 return NULL;
461}
462
463static const struct virtio_config_ops virtio_vdpa_config_ops = {
464 .get = virtio_vdpa_get,
465 .set = virtio_vdpa_set,
466 .generation = virtio_vdpa_generation,
467 .get_status = virtio_vdpa_get_status,
468 .set_status = virtio_vdpa_set_status,
469 .reset = virtio_vdpa_reset,
470 .find_vqs = virtio_vdpa_find_vqs,
471 .del_vqs = virtio_vdpa_del_vqs,
472 .get_features = virtio_vdpa_get_features,
473 .finalize_features = virtio_vdpa_finalize_features,
474 .bus_name = virtio_vdpa_bus_name,
475 .set_vq_affinity = virtio_vdpa_set_vq_affinity,
476 .get_vq_affinity = virtio_vdpa_get_vq_affinity,
477};
478
479static void virtio_vdpa_release_dev(struct device *_d)
480{
481 struct virtio_device *vdev =
482 container_of(_d, struct virtio_device, dev);
483 struct virtio_vdpa_device *vd_dev =
484 container_of(vdev, struct virtio_vdpa_device, vdev);
485
486 kfree(objp: vd_dev);
487}
488
489static int virtio_vdpa_probe(struct vdpa_device *vdpa)
490{
491 const struct vdpa_config_ops *ops = vdpa->config;
492 struct virtio_vdpa_device *vd_dev, *reg_dev = NULL;
493 int ret = -EINVAL;
494
495 vd_dev = kzalloc(size: sizeof(*vd_dev), GFP_KERNEL);
496 if (!vd_dev)
497 return -ENOMEM;
498
499 vd_dev->vdev.dev.parent = vdpa_get_dma_dev(vdev: vdpa);
500 vd_dev->vdev.dev.release = virtio_vdpa_release_dev;
501 vd_dev->vdev.config = &virtio_vdpa_config_ops;
502 vd_dev->vdpa = vdpa;
503 INIT_LIST_HEAD(list: &vd_dev->virtqueues);
504 spin_lock_init(&vd_dev->lock);
505
506 vd_dev->vdev.id.device = ops->get_device_id(vdpa);
507 if (vd_dev->vdev.id.device == 0)
508 goto err;
509
510 vd_dev->vdev.id.vendor = ops->get_vendor_id(vdpa);
511 ret = register_virtio_device(dev: &vd_dev->vdev);
512 reg_dev = vd_dev;
513 if (ret)
514 goto err;
515
516 vdpa_set_drvdata(vdev: vdpa, data: vd_dev);
517
518 return 0;
519
520err:
521 if (reg_dev)
522 put_device(dev: &vd_dev->vdev.dev);
523 else
524 kfree(objp: vd_dev);
525 return ret;
526}
527
528static void virtio_vdpa_remove(struct vdpa_device *vdpa)
529{
530 struct virtio_vdpa_device *vd_dev = vdpa_get_drvdata(vdev: vdpa);
531
532 unregister_virtio_device(dev: &vd_dev->vdev);
533}
534
535static struct vdpa_driver virtio_vdpa_driver = {
536 .driver = {
537 .name = "virtio_vdpa",
538 },
539 .probe = virtio_vdpa_probe,
540 .remove = virtio_vdpa_remove,
541};
542
543module_vdpa_driver(virtio_vdpa_driver);
544
545MODULE_VERSION(MOD_VERSION);
546MODULE_LICENSE(MOD_LICENSE);
547MODULE_AUTHOR(MOD_AUTHOR);
548MODULE_DESCRIPTION(MOD_DESC);
549

source code of linux/drivers/virtio/virtio_vdpa.c