1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * Copyright (C) 2018-2020 Intel Corporation.
4 * Copyright (C) 2020 Red Hat, Inc.
5 *
6 * Author: Tiwei Bie <tiwei.bie@intel.com>
7 * Jason Wang <jasowang@redhat.com>
8 *
9 * Thanks Michael S. Tsirkin for the valuable comments and
10 * suggestions. And thanks to Cunming Liang and Zhihong Wang for all
11 * their supports.
12 */
13
14 #include <linux/kernel.h>
15 #include <linux/module.h>
16 #include <linux/cdev.h>
17 #include <linux/device.h>
18 #include <linux/mm.h>
19 #include <linux/slab.h>
20 #include <linux/iommu.h>
21 #include <linux/uuid.h>
22 #include <linux/vdpa.h>
23 #include <linux/nospec.h>
24 #include <linux/vhost.h>
25
26 #include "vhost.h"
27
28 enum {
29 VHOST_VDPA_BACKEND_FEATURES =
30 (1ULL << VHOST_BACKEND_F_IOTLB_MSG_V2) |
31 (1ULL << VHOST_BACKEND_F_IOTLB_BATCH) |
32 (1ULL << VHOST_BACKEND_F_IOTLB_ASID),
33 };
34
35 #define VHOST_VDPA_DEV_MAX (1U << MINORBITS)
36
37 #define VHOST_VDPA_IOTLB_BUCKETS 16
38
39 struct vhost_vdpa_as {
40 struct hlist_node hash_link;
41 struct vhost_iotlb iotlb;
42 u32 id;
43 };
44
45 struct vhost_vdpa {
46 struct vhost_dev vdev;
47 struct iommu_domain *domain;
48 struct vhost_virtqueue *vqs;
49 struct completion completion;
50 struct vdpa_device *vdpa;
51 struct hlist_head as[VHOST_VDPA_IOTLB_BUCKETS];
52 struct device dev;
53 struct cdev cdev;
54 atomic_t opened;
55 u32 nvqs;
56 int virtio_id;
57 int minor;
58 struct eventfd_ctx *config_ctx;
59 int in_batch;
60 struct vdpa_iova_range range;
61 u32 batch_asid;
62 };
63
64 static DEFINE_IDA(vhost_vdpa_ida);
65
66 static dev_t vhost_vdpa_major;
67
68 static void vhost_vdpa_iotlb_unmap(struct vhost_vdpa *v,
69 struct vhost_iotlb *iotlb, u64 start,
70 u64 last, u32 asid);
71
iotlb_to_asid(struct vhost_iotlb * iotlb)72 static inline u32 iotlb_to_asid(struct vhost_iotlb *iotlb)
73 {
74 struct vhost_vdpa_as *as = container_of(iotlb, struct
75 vhost_vdpa_as, iotlb);
76 return as->id;
77 }
78
asid_to_as(struct vhost_vdpa * v,u32 asid)79 static struct vhost_vdpa_as *asid_to_as(struct vhost_vdpa *v, u32 asid)
80 {
81 struct hlist_head *head = &v->as[asid % VHOST_VDPA_IOTLB_BUCKETS];
82 struct vhost_vdpa_as *as;
83
84 hlist_for_each_entry(as, head, hash_link)
85 if (as->id == asid)
86 return as;
87
88 return NULL;
89 }
90
asid_to_iotlb(struct vhost_vdpa * v,u32 asid)91 static struct vhost_iotlb *asid_to_iotlb(struct vhost_vdpa *v, u32 asid)
92 {
93 struct vhost_vdpa_as *as = asid_to_as(v, asid);
94
95 if (!as)
96 return NULL;
97
98 return &as->iotlb;
99 }
100
vhost_vdpa_alloc_as(struct vhost_vdpa * v,u32 asid)101 static struct vhost_vdpa_as *vhost_vdpa_alloc_as(struct vhost_vdpa *v, u32 asid)
102 {
103 struct hlist_head *head = &v->as[asid % VHOST_VDPA_IOTLB_BUCKETS];
104 struct vhost_vdpa_as *as;
105
106 if (asid_to_as(v, asid))
107 return NULL;
108
109 if (asid >= v->vdpa->nas)
110 return NULL;
111
112 as = kmalloc(sizeof(*as), GFP_KERNEL);
113 if (!as)
114 return NULL;
115
116 vhost_iotlb_init(&as->iotlb, 0, 0);
117 as->id = asid;
118 hlist_add_head(&as->hash_link, head);
119
120 return as;
121 }
122
vhost_vdpa_find_alloc_as(struct vhost_vdpa * v,u32 asid)123 static struct vhost_vdpa_as *vhost_vdpa_find_alloc_as(struct vhost_vdpa *v,
124 u32 asid)
125 {
126 struct vhost_vdpa_as *as = asid_to_as(v, asid);
127
128 if (as)
129 return as;
130
131 return vhost_vdpa_alloc_as(v, asid);
132 }
133
vhost_vdpa_remove_as(struct vhost_vdpa * v,u32 asid)134 static int vhost_vdpa_remove_as(struct vhost_vdpa *v, u32 asid)
135 {
136 struct vhost_vdpa_as *as = asid_to_as(v, asid);
137
138 if (!as)
139 return -EINVAL;
140
141 hlist_del(&as->hash_link);
142 vhost_vdpa_iotlb_unmap(v, &as->iotlb, 0ULL, 0ULL - 1, asid);
143 kfree(as);
144
145 return 0;
146 }
147
handle_vq_kick(struct vhost_work * work)148 static void handle_vq_kick(struct vhost_work *work)
149 {
150 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
151 poll.work);
152 struct vhost_vdpa *v = container_of(vq->dev, struct vhost_vdpa, vdev);
153 const struct vdpa_config_ops *ops = v->vdpa->config;
154
155 ops->kick_vq(v->vdpa, vq - v->vqs);
156 }
157
vhost_vdpa_virtqueue_cb(void * private)158 static irqreturn_t vhost_vdpa_virtqueue_cb(void *private)
159 {
160 struct vhost_virtqueue *vq = private;
161 struct eventfd_ctx *call_ctx = vq->call_ctx.ctx;
162
163 if (call_ctx)
164 eventfd_signal(call_ctx, 1);
165
166 return IRQ_HANDLED;
167 }
168
vhost_vdpa_config_cb(void * private)169 static irqreturn_t vhost_vdpa_config_cb(void *private)
170 {
171 struct vhost_vdpa *v = private;
172 struct eventfd_ctx *config_ctx = v->config_ctx;
173
174 if (config_ctx)
175 eventfd_signal(config_ctx, 1);
176
177 return IRQ_HANDLED;
178 }
179
vhost_vdpa_setup_vq_irq(struct vhost_vdpa * v,u16 qid)180 static void vhost_vdpa_setup_vq_irq(struct vhost_vdpa *v, u16 qid)
181 {
182 struct vhost_virtqueue *vq = &v->vqs[qid];
183 const struct vdpa_config_ops *ops = v->vdpa->config;
184 struct vdpa_device *vdpa = v->vdpa;
185 int ret, irq;
186
187 if (!ops->get_vq_irq)
188 return;
189
190 irq = ops->get_vq_irq(vdpa, qid);
191 if (irq < 0)
192 return;
193
194 irq_bypass_unregister_producer(&vq->call_ctx.producer);
195 if (!vq->call_ctx.ctx)
196 return;
197
198 vq->call_ctx.producer.token = vq->call_ctx.ctx;
199 vq->call_ctx.producer.irq = irq;
200 ret = irq_bypass_register_producer(&vq->call_ctx.producer);
201 if (unlikely(ret))
202 dev_info(&v->dev, "vq %u, irq bypass producer (token %p) registration fails, ret = %d\n",
203 qid, vq->call_ctx.producer.token, ret);
204 }
205
vhost_vdpa_unsetup_vq_irq(struct vhost_vdpa * v,u16 qid)206 static void vhost_vdpa_unsetup_vq_irq(struct vhost_vdpa *v, u16 qid)
207 {
208 struct vhost_virtqueue *vq = &v->vqs[qid];
209
210 irq_bypass_unregister_producer(&vq->call_ctx.producer);
211 }
212
vhost_vdpa_reset(struct vhost_vdpa * v)213 static int vhost_vdpa_reset(struct vhost_vdpa *v)
214 {
215 struct vdpa_device *vdpa = v->vdpa;
216
217 v->in_batch = 0;
218
219 return vdpa_reset(vdpa);
220 }
221
vhost_vdpa_get_device_id(struct vhost_vdpa * v,u8 __user * argp)222 static long vhost_vdpa_get_device_id(struct vhost_vdpa *v, u8 __user *argp)
223 {
224 struct vdpa_device *vdpa = v->vdpa;
225 const struct vdpa_config_ops *ops = vdpa->config;
226 u32 device_id;
227
228 device_id = ops->get_device_id(vdpa);
229
230 if (copy_to_user(argp, &device_id, sizeof(device_id)))
231 return -EFAULT;
232
233 return 0;
234 }
235
vhost_vdpa_get_status(struct vhost_vdpa * v,u8 __user * statusp)236 static long vhost_vdpa_get_status(struct vhost_vdpa *v, u8 __user *statusp)
237 {
238 struct vdpa_device *vdpa = v->vdpa;
239 const struct vdpa_config_ops *ops = vdpa->config;
240 u8 status;
241
242 status = ops->get_status(vdpa);
243
244 if (copy_to_user(statusp, &status, sizeof(status)))
245 return -EFAULT;
246
247 return 0;
248 }
249
vhost_vdpa_set_status(struct vhost_vdpa * v,u8 __user * statusp)250 static long vhost_vdpa_set_status(struct vhost_vdpa *v, u8 __user *statusp)
251 {
252 struct vdpa_device *vdpa = v->vdpa;
253 const struct vdpa_config_ops *ops = vdpa->config;
254 u8 status, status_old;
255 u32 nvqs = v->nvqs;
256 int ret;
257 u16 i;
258
259 if (copy_from_user(&status, statusp, sizeof(status)))
260 return -EFAULT;
261
262 status_old = ops->get_status(vdpa);
263
264 /*
265 * Userspace shouldn't remove status bits unless reset the
266 * status to 0.
267 */
268 if (status != 0 && (status_old & ~status) != 0)
269 return -EINVAL;
270
271 if ((status_old & VIRTIO_CONFIG_S_DRIVER_OK) && !(status & VIRTIO_CONFIG_S_DRIVER_OK))
272 for (i = 0; i < nvqs; i++)
273 vhost_vdpa_unsetup_vq_irq(v, i);
274
275 if (status == 0) {
276 ret = vdpa_reset(vdpa);
277 if (ret)
278 return ret;
279 } else
280 vdpa_set_status(vdpa, status);
281
282 if ((status & VIRTIO_CONFIG_S_DRIVER_OK) && !(status_old & VIRTIO_CONFIG_S_DRIVER_OK))
283 for (i = 0; i < nvqs; i++)
284 vhost_vdpa_setup_vq_irq(v, i);
285
286 return 0;
287 }
288
vhost_vdpa_config_validate(struct vhost_vdpa * v,struct vhost_vdpa_config * c)289 static int vhost_vdpa_config_validate(struct vhost_vdpa *v,
290 struct vhost_vdpa_config *c)
291 {
292 struct vdpa_device *vdpa = v->vdpa;
293 size_t size = vdpa->config->get_config_size(vdpa);
294
295 if (c->len == 0 || c->off > size)
296 return -EINVAL;
297
298 if (c->len > size - c->off)
299 return -E2BIG;
300
301 return 0;
302 }
303
vhost_vdpa_get_config(struct vhost_vdpa * v,struct vhost_vdpa_config __user * c)304 static long vhost_vdpa_get_config(struct vhost_vdpa *v,
305 struct vhost_vdpa_config __user *c)
306 {
307 struct vdpa_device *vdpa = v->vdpa;
308 struct vhost_vdpa_config config;
309 unsigned long size = offsetof(struct vhost_vdpa_config, buf);
310 u8 *buf;
311
312 if (copy_from_user(&config, c, size))
313 return -EFAULT;
314 if (vhost_vdpa_config_validate(v, &config))
315 return -EINVAL;
316 buf = kvzalloc(config.len, GFP_KERNEL);
317 if (!buf)
318 return -ENOMEM;
319
320 vdpa_get_config(vdpa, config.off, buf, config.len);
321
322 if (copy_to_user(c->buf, buf, config.len)) {
323 kvfree(buf);
324 return -EFAULT;
325 }
326
327 kvfree(buf);
328 return 0;
329 }
330
vhost_vdpa_set_config(struct vhost_vdpa * v,struct vhost_vdpa_config __user * c)331 static long vhost_vdpa_set_config(struct vhost_vdpa *v,
332 struct vhost_vdpa_config __user *c)
333 {
334 struct vdpa_device *vdpa = v->vdpa;
335 struct vhost_vdpa_config config;
336 unsigned long size = offsetof(struct vhost_vdpa_config, buf);
337 u8 *buf;
338
339 if (copy_from_user(&config, c, size))
340 return -EFAULT;
341 if (vhost_vdpa_config_validate(v, &config))
342 return -EINVAL;
343
344 buf = vmemdup_user(c->buf, config.len);
345 if (IS_ERR(buf))
346 return PTR_ERR(buf);
347
348 vdpa_set_config(vdpa, config.off, buf, config.len);
349
350 kvfree(buf);
351 return 0;
352 }
353
vhost_vdpa_can_suspend(const struct vhost_vdpa * v)354 static bool vhost_vdpa_can_suspend(const struct vhost_vdpa *v)
355 {
356 struct vdpa_device *vdpa = v->vdpa;
357 const struct vdpa_config_ops *ops = vdpa->config;
358
359 return ops->suspend;
360 }
361
vhost_vdpa_get_features(struct vhost_vdpa * v,u64 __user * featurep)362 static long vhost_vdpa_get_features(struct vhost_vdpa *v, u64 __user *featurep)
363 {
364 struct vdpa_device *vdpa = v->vdpa;
365 const struct vdpa_config_ops *ops = vdpa->config;
366 u64 features;
367
368 features = ops->get_device_features(vdpa);
369
370 if (copy_to_user(featurep, &features, sizeof(features)))
371 return -EFAULT;
372
373 return 0;
374 }
375
vhost_vdpa_set_features(struct vhost_vdpa * v,u64 __user * featurep)376 static long vhost_vdpa_set_features(struct vhost_vdpa *v, u64 __user *featurep)
377 {
378 struct vdpa_device *vdpa = v->vdpa;
379 const struct vdpa_config_ops *ops = vdpa->config;
380 struct vhost_dev *d = &v->vdev;
381 u64 actual_features;
382 u64 features;
383 int i;
384
385 /*
386 * It's not allowed to change the features after they have
387 * been negotiated.
388 */
389 if (ops->get_status(vdpa) & VIRTIO_CONFIG_S_FEATURES_OK)
390 return -EBUSY;
391
392 if (copy_from_user(&features, featurep, sizeof(features)))
393 return -EFAULT;
394
395 if (vdpa_set_features(vdpa, features))
396 return -EINVAL;
397
398 /* let the vqs know what has been configured */
399 actual_features = ops->get_driver_features(vdpa);
400 for (i = 0; i < d->nvqs; ++i) {
401 struct vhost_virtqueue *vq = d->vqs[i];
402
403 mutex_lock(&vq->mutex);
404 vq->acked_features = actual_features;
405 mutex_unlock(&vq->mutex);
406 }
407
408 return 0;
409 }
410
vhost_vdpa_get_vring_num(struct vhost_vdpa * v,u16 __user * argp)411 static long vhost_vdpa_get_vring_num(struct vhost_vdpa *v, u16 __user *argp)
412 {
413 struct vdpa_device *vdpa = v->vdpa;
414 const struct vdpa_config_ops *ops = vdpa->config;
415 u16 num;
416
417 num = ops->get_vq_num_max(vdpa);
418
419 if (copy_to_user(argp, &num, sizeof(num)))
420 return -EFAULT;
421
422 return 0;
423 }
424
vhost_vdpa_config_put(struct vhost_vdpa * v)425 static void vhost_vdpa_config_put(struct vhost_vdpa *v)
426 {
427 if (v->config_ctx) {
428 eventfd_ctx_put(v->config_ctx);
429 v->config_ctx = NULL;
430 }
431 }
432
vhost_vdpa_set_config_call(struct vhost_vdpa * v,u32 __user * argp)433 static long vhost_vdpa_set_config_call(struct vhost_vdpa *v, u32 __user *argp)
434 {
435 struct vdpa_callback cb;
436 int fd;
437 struct eventfd_ctx *ctx;
438
439 cb.callback = vhost_vdpa_config_cb;
440 cb.private = v;
441 if (copy_from_user(&fd, argp, sizeof(fd)))
442 return -EFAULT;
443
444 ctx = fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(fd);
445 swap(ctx, v->config_ctx);
446
447 if (!IS_ERR_OR_NULL(ctx))
448 eventfd_ctx_put(ctx);
449
450 if (IS_ERR(v->config_ctx)) {
451 long ret = PTR_ERR(v->config_ctx);
452
453 v->config_ctx = NULL;
454 return ret;
455 }
456
457 v->vdpa->config->set_config_cb(v->vdpa, &cb);
458
459 return 0;
460 }
461
vhost_vdpa_get_iova_range(struct vhost_vdpa * v,u32 __user * argp)462 static long vhost_vdpa_get_iova_range(struct vhost_vdpa *v, u32 __user *argp)
463 {
464 struct vhost_vdpa_iova_range range = {
465 .first = v->range.first,
466 .last = v->range.last,
467 };
468
469 if (copy_to_user(argp, &range, sizeof(range)))
470 return -EFAULT;
471 return 0;
472 }
473
vhost_vdpa_get_config_size(struct vhost_vdpa * v,u32 __user * argp)474 static long vhost_vdpa_get_config_size(struct vhost_vdpa *v, u32 __user *argp)
475 {
476 struct vdpa_device *vdpa = v->vdpa;
477 const struct vdpa_config_ops *ops = vdpa->config;
478 u32 size;
479
480 size = ops->get_config_size(vdpa);
481
482 if (copy_to_user(argp, &size, sizeof(size)))
483 return -EFAULT;
484
485 return 0;
486 }
487
vhost_vdpa_get_vqs_count(struct vhost_vdpa * v,u32 __user * argp)488 static long vhost_vdpa_get_vqs_count(struct vhost_vdpa *v, u32 __user *argp)
489 {
490 struct vdpa_device *vdpa = v->vdpa;
491
492 if (copy_to_user(argp, &vdpa->nvqs, sizeof(vdpa->nvqs)))
493 return -EFAULT;
494
495 return 0;
496 }
497
498 /* After a successful return of ioctl the device must not process more
499 * virtqueue descriptors. The device can answer to read or writes of config
500 * fields as if it were not suspended. In particular, writing to "queue_enable"
501 * with a value of 1 will not make the device start processing buffers.
502 */
vhost_vdpa_suspend(struct vhost_vdpa * v)503 static long vhost_vdpa_suspend(struct vhost_vdpa *v)
504 {
505 struct vdpa_device *vdpa = v->vdpa;
506 const struct vdpa_config_ops *ops = vdpa->config;
507
508 if (!ops->suspend)
509 return -EOPNOTSUPP;
510
511 return ops->suspend(vdpa);
512 }
513
vhost_vdpa_vring_ioctl(struct vhost_vdpa * v,unsigned int cmd,void __user * argp)514 static long vhost_vdpa_vring_ioctl(struct vhost_vdpa *v, unsigned int cmd,
515 void __user *argp)
516 {
517 struct vdpa_device *vdpa = v->vdpa;
518 const struct vdpa_config_ops *ops = vdpa->config;
519 struct vdpa_vq_state vq_state;
520 struct vdpa_callback cb;
521 struct vhost_virtqueue *vq;
522 struct vhost_vring_state s;
523 u32 idx;
524 long r;
525
526 r = get_user(idx, (u32 __user *)argp);
527 if (r < 0)
528 return r;
529
530 if (idx >= v->nvqs)
531 return -ENOBUFS;
532
533 idx = array_index_nospec(idx, v->nvqs);
534 vq = &v->vqs[idx];
535
536 switch (cmd) {
537 case VHOST_VDPA_SET_VRING_ENABLE:
538 if (copy_from_user(&s, argp, sizeof(s)))
539 return -EFAULT;
540 ops->set_vq_ready(vdpa, idx, s.num);
541 return 0;
542 case VHOST_VDPA_GET_VRING_GROUP:
543 if (!ops->get_vq_group)
544 return -EOPNOTSUPP;
545 s.index = idx;
546 s.num = ops->get_vq_group(vdpa, idx);
547 if (s.num >= vdpa->ngroups)
548 return -EIO;
549 else if (copy_to_user(argp, &s, sizeof(s)))
550 return -EFAULT;
551 return 0;
552 case VHOST_VDPA_SET_GROUP_ASID:
553 if (copy_from_user(&s, argp, sizeof(s)))
554 return -EFAULT;
555 if (s.num >= vdpa->nas)
556 return -EINVAL;
557 if (!ops->set_group_asid)
558 return -EOPNOTSUPP;
559 return ops->set_group_asid(vdpa, idx, s.num);
560 case VHOST_GET_VRING_BASE:
561 r = ops->get_vq_state(v->vdpa, idx, &vq_state);
562 if (r)
563 return r;
564
565 if (vhost_has_feature(vq, VIRTIO_F_RING_PACKED)) {
566 vq->last_avail_idx = vq_state.packed.last_avail_idx |
567 (vq_state.packed.last_avail_counter << 15);
568 vq->last_used_idx = vq_state.packed.last_used_idx |
569 (vq_state.packed.last_used_counter << 15);
570 } else {
571 vq->last_avail_idx = vq_state.split.avail_index;
572 }
573 break;
574 }
575
576 r = vhost_vring_ioctl(&v->vdev, cmd, argp);
577 if (r)
578 return r;
579
580 switch (cmd) {
581 case VHOST_SET_VRING_ADDR:
582 if (ops->set_vq_address(vdpa, idx,
583 (u64)(uintptr_t)vq->desc,
584 (u64)(uintptr_t)vq->avail,
585 (u64)(uintptr_t)vq->used))
586 r = -EINVAL;
587 break;
588
589 case VHOST_SET_VRING_BASE:
590 if (vhost_has_feature(vq, VIRTIO_F_RING_PACKED)) {
591 vq_state.packed.last_avail_idx = vq->last_avail_idx & 0x7fff;
592 vq_state.packed.last_avail_counter = !!(vq->last_avail_idx & 0x8000);
593 vq_state.packed.last_used_idx = vq->last_used_idx & 0x7fff;
594 vq_state.packed.last_used_counter = !!(vq->last_used_idx & 0x8000);
595 } else {
596 vq_state.split.avail_index = vq->last_avail_idx;
597 }
598 r = ops->set_vq_state(vdpa, idx, &vq_state);
599 break;
600
601 case VHOST_SET_VRING_CALL:
602 if (vq->call_ctx.ctx) {
603 cb.callback = vhost_vdpa_virtqueue_cb;
604 cb.private = vq;
605 } else {
606 cb.callback = NULL;
607 cb.private = NULL;
608 }
609 ops->set_vq_cb(vdpa, idx, &cb);
610 vhost_vdpa_setup_vq_irq(v, idx);
611 break;
612
613 case VHOST_SET_VRING_NUM:
614 ops->set_vq_num(vdpa, idx, vq->num);
615 break;
616 }
617
618 return r;
619 }
620
vhost_vdpa_unlocked_ioctl(struct file * filep,unsigned int cmd,unsigned long arg)621 static long vhost_vdpa_unlocked_ioctl(struct file *filep,
622 unsigned int cmd, unsigned long arg)
623 {
624 struct vhost_vdpa *v = filep->private_data;
625 struct vhost_dev *d = &v->vdev;
626 void __user *argp = (void __user *)arg;
627 u64 __user *featurep = argp;
628 u64 features;
629 long r = 0;
630
631 if (cmd == VHOST_SET_BACKEND_FEATURES) {
632 if (copy_from_user(&features, featurep, sizeof(features)))
633 return -EFAULT;
634 if (features & ~(VHOST_VDPA_BACKEND_FEATURES |
635 BIT_ULL(VHOST_BACKEND_F_SUSPEND)))
636 return -EOPNOTSUPP;
637 if ((features & BIT_ULL(VHOST_BACKEND_F_SUSPEND)) &&
638 !vhost_vdpa_can_suspend(v))
639 return -EOPNOTSUPP;
640 vhost_set_backend_features(&v->vdev, features);
641 return 0;
642 }
643
644 mutex_lock(&d->mutex);
645
646 switch (cmd) {
647 case VHOST_VDPA_GET_DEVICE_ID:
648 r = vhost_vdpa_get_device_id(v, argp);
649 break;
650 case VHOST_VDPA_GET_STATUS:
651 r = vhost_vdpa_get_status(v, argp);
652 break;
653 case VHOST_VDPA_SET_STATUS:
654 r = vhost_vdpa_set_status(v, argp);
655 break;
656 case VHOST_VDPA_GET_CONFIG:
657 r = vhost_vdpa_get_config(v, argp);
658 break;
659 case VHOST_VDPA_SET_CONFIG:
660 r = vhost_vdpa_set_config(v, argp);
661 break;
662 case VHOST_GET_FEATURES:
663 r = vhost_vdpa_get_features(v, argp);
664 break;
665 case VHOST_SET_FEATURES:
666 r = vhost_vdpa_set_features(v, argp);
667 break;
668 case VHOST_VDPA_GET_VRING_NUM:
669 r = vhost_vdpa_get_vring_num(v, argp);
670 break;
671 case VHOST_VDPA_GET_GROUP_NUM:
672 if (copy_to_user(argp, &v->vdpa->ngroups,
673 sizeof(v->vdpa->ngroups)))
674 r = -EFAULT;
675 break;
676 case VHOST_VDPA_GET_AS_NUM:
677 if (copy_to_user(argp, &v->vdpa->nas, sizeof(v->vdpa->nas)))
678 r = -EFAULT;
679 break;
680 case VHOST_SET_LOG_BASE:
681 case VHOST_SET_LOG_FD:
682 r = -ENOIOCTLCMD;
683 break;
684 case VHOST_VDPA_SET_CONFIG_CALL:
685 r = vhost_vdpa_set_config_call(v, argp);
686 break;
687 case VHOST_GET_BACKEND_FEATURES:
688 features = VHOST_VDPA_BACKEND_FEATURES;
689 if (vhost_vdpa_can_suspend(v))
690 features |= BIT_ULL(VHOST_BACKEND_F_SUSPEND);
691 if (copy_to_user(featurep, &features, sizeof(features)))
692 r = -EFAULT;
693 break;
694 case VHOST_VDPA_GET_IOVA_RANGE:
695 r = vhost_vdpa_get_iova_range(v, argp);
696 break;
697 case VHOST_VDPA_GET_CONFIG_SIZE:
698 r = vhost_vdpa_get_config_size(v, argp);
699 break;
700 case VHOST_VDPA_GET_VQS_COUNT:
701 r = vhost_vdpa_get_vqs_count(v, argp);
702 break;
703 case VHOST_VDPA_SUSPEND:
704 r = vhost_vdpa_suspend(v);
705 break;
706 default:
707 r = vhost_dev_ioctl(&v->vdev, cmd, argp);
708 if (r == -ENOIOCTLCMD)
709 r = vhost_vdpa_vring_ioctl(v, cmd, argp);
710 break;
711 }
712
713 mutex_unlock(&d->mutex);
714 return r;
715 }
vhost_vdpa_general_unmap(struct vhost_vdpa * v,struct vhost_iotlb_map * map,u32 asid)716 static void vhost_vdpa_general_unmap(struct vhost_vdpa *v,
717 struct vhost_iotlb_map *map, u32 asid)
718 {
719 struct vdpa_device *vdpa = v->vdpa;
720 const struct vdpa_config_ops *ops = vdpa->config;
721 if (ops->dma_map) {
722 ops->dma_unmap(vdpa, asid, map->start, map->size);
723 } else if (ops->set_map == NULL) {
724 iommu_unmap(v->domain, map->start, map->size);
725 }
726 }
727
vhost_vdpa_pa_unmap(struct vhost_vdpa * v,struct vhost_iotlb * iotlb,u64 start,u64 last,u32 asid)728 static void vhost_vdpa_pa_unmap(struct vhost_vdpa *v, struct vhost_iotlb *iotlb,
729 u64 start, u64 last, u32 asid)
730 {
731 struct vhost_dev *dev = &v->vdev;
732 struct vhost_iotlb_map *map;
733 struct page *page;
734 unsigned long pfn, pinned;
735
736 while ((map = vhost_iotlb_itree_first(iotlb, start, last)) != NULL) {
737 pinned = PFN_DOWN(map->size);
738 for (pfn = PFN_DOWN(map->addr);
739 pinned > 0; pfn++, pinned--) {
740 page = pfn_to_page(pfn);
741 if (map->perm & VHOST_ACCESS_WO)
742 set_page_dirty_lock(page);
743 unpin_user_page(page);
744 }
745 atomic64_sub(PFN_DOWN(map->size), &dev->mm->pinned_vm);
746 vhost_vdpa_general_unmap(v, map, asid);
747 vhost_iotlb_map_free(iotlb, map);
748 }
749 }
750
vhost_vdpa_va_unmap(struct vhost_vdpa * v,struct vhost_iotlb * iotlb,u64 start,u64 last,u32 asid)751 static void vhost_vdpa_va_unmap(struct vhost_vdpa *v, struct vhost_iotlb *iotlb,
752 u64 start, u64 last, u32 asid)
753 {
754 struct vhost_iotlb_map *map;
755 struct vdpa_map_file *map_file;
756
757 while ((map = vhost_iotlb_itree_first(iotlb, start, last)) != NULL) {
758 map_file = (struct vdpa_map_file *)map->opaque;
759 fput(map_file->file);
760 kfree(map_file);
761 vhost_vdpa_general_unmap(v, map, asid);
762 vhost_iotlb_map_free(iotlb, map);
763 }
764 }
765
vhost_vdpa_iotlb_unmap(struct vhost_vdpa * v,struct vhost_iotlb * iotlb,u64 start,u64 last,u32 asid)766 static void vhost_vdpa_iotlb_unmap(struct vhost_vdpa *v,
767 struct vhost_iotlb *iotlb, u64 start,
768 u64 last, u32 asid)
769 {
770 struct vdpa_device *vdpa = v->vdpa;
771
772 if (vdpa->use_va)
773 return vhost_vdpa_va_unmap(v, iotlb, start, last, asid);
774
775 return vhost_vdpa_pa_unmap(v, iotlb, start, last, asid);
776 }
777
perm_to_iommu_flags(u32 perm)778 static int perm_to_iommu_flags(u32 perm)
779 {
780 int flags = 0;
781
782 switch (perm) {
783 case VHOST_ACCESS_WO:
784 flags |= IOMMU_WRITE;
785 break;
786 case VHOST_ACCESS_RO:
787 flags |= IOMMU_READ;
788 break;
789 case VHOST_ACCESS_RW:
790 flags |= (IOMMU_WRITE | IOMMU_READ);
791 break;
792 default:
793 WARN(1, "invalidate vhost IOTLB permission\n");
794 break;
795 }
796
797 return flags | IOMMU_CACHE;
798 }
799
vhost_vdpa_map(struct vhost_vdpa * v,struct vhost_iotlb * iotlb,u64 iova,u64 size,u64 pa,u32 perm,void * opaque)800 static int vhost_vdpa_map(struct vhost_vdpa *v, struct vhost_iotlb *iotlb,
801 u64 iova, u64 size, u64 pa, u32 perm, void *opaque)
802 {
803 struct vhost_dev *dev = &v->vdev;
804 struct vdpa_device *vdpa = v->vdpa;
805 const struct vdpa_config_ops *ops = vdpa->config;
806 u32 asid = iotlb_to_asid(iotlb);
807 int r = 0;
808
809 r = vhost_iotlb_add_range_ctx(iotlb, iova, iova + size - 1,
810 pa, perm, opaque);
811 if (r)
812 return r;
813
814 if (ops->dma_map) {
815 r = ops->dma_map(vdpa, asid, iova, size, pa, perm, opaque);
816 } else if (ops->set_map) {
817 if (!v->in_batch)
818 r = ops->set_map(vdpa, asid, iotlb);
819 } else {
820 r = iommu_map(v->domain, iova, pa, size,
821 perm_to_iommu_flags(perm));
822 }
823 if (r) {
824 vhost_iotlb_del_range(iotlb, iova, iova + size - 1);
825 return r;
826 }
827
828 if (!vdpa->use_va)
829 atomic64_add(PFN_DOWN(size), &dev->mm->pinned_vm);
830
831 return 0;
832 }
833
vhost_vdpa_unmap(struct vhost_vdpa * v,struct vhost_iotlb * iotlb,u64 iova,u64 size)834 static void vhost_vdpa_unmap(struct vhost_vdpa *v,
835 struct vhost_iotlb *iotlb,
836 u64 iova, u64 size)
837 {
838 struct vdpa_device *vdpa = v->vdpa;
839 const struct vdpa_config_ops *ops = vdpa->config;
840 u32 asid = iotlb_to_asid(iotlb);
841
842 vhost_vdpa_iotlb_unmap(v, iotlb, iova, iova + size - 1, asid);
843
844 if (ops->set_map) {
845 if (!v->in_batch)
846 ops->set_map(vdpa, asid, iotlb);
847 }
848
849 }
850
vhost_vdpa_va_map(struct vhost_vdpa * v,struct vhost_iotlb * iotlb,u64 iova,u64 size,u64 uaddr,u32 perm)851 static int vhost_vdpa_va_map(struct vhost_vdpa *v,
852 struct vhost_iotlb *iotlb,
853 u64 iova, u64 size, u64 uaddr, u32 perm)
854 {
855 struct vhost_dev *dev = &v->vdev;
856 u64 offset, map_size, map_iova = iova;
857 struct vdpa_map_file *map_file;
858 struct vm_area_struct *vma;
859 int ret = 0;
860
861 mmap_read_lock(dev->mm);
862
863 while (size) {
864 vma = find_vma(dev->mm, uaddr);
865 if (!vma) {
866 ret = -EINVAL;
867 break;
868 }
869 map_size = min(size, vma->vm_end - uaddr);
870 if (!(vma->vm_file && (vma->vm_flags & VM_SHARED) &&
871 !(vma->vm_flags & (VM_IO | VM_PFNMAP))))
872 goto next;
873
874 map_file = kzalloc(sizeof(*map_file), GFP_KERNEL);
875 if (!map_file) {
876 ret = -ENOMEM;
877 break;
878 }
879 offset = (vma->vm_pgoff << PAGE_SHIFT) + uaddr - vma->vm_start;
880 map_file->offset = offset;
881 map_file->file = get_file(vma->vm_file);
882 ret = vhost_vdpa_map(v, iotlb, map_iova, map_size, uaddr,
883 perm, map_file);
884 if (ret) {
885 fput(map_file->file);
886 kfree(map_file);
887 break;
888 }
889 next:
890 size -= map_size;
891 uaddr += map_size;
892 map_iova += map_size;
893 }
894 if (ret)
895 vhost_vdpa_unmap(v, iotlb, iova, map_iova - iova);
896
897 mmap_read_unlock(dev->mm);
898
899 return ret;
900 }
901
vhost_vdpa_pa_map(struct vhost_vdpa * v,struct vhost_iotlb * iotlb,u64 iova,u64 size,u64 uaddr,u32 perm)902 static int vhost_vdpa_pa_map(struct vhost_vdpa *v,
903 struct vhost_iotlb *iotlb,
904 u64 iova, u64 size, u64 uaddr, u32 perm)
905 {
906 struct vhost_dev *dev = &v->vdev;
907 struct page **page_list;
908 unsigned long list_size = PAGE_SIZE / sizeof(struct page *);
909 unsigned int gup_flags = FOLL_LONGTERM;
910 unsigned long npages, cur_base, map_pfn, last_pfn = 0;
911 unsigned long lock_limit, sz2pin, nchunks, i;
912 u64 start = iova;
913 long pinned;
914 int ret = 0;
915
916 /* Limit the use of memory for bookkeeping */
917 page_list = (struct page **) __get_free_page(GFP_KERNEL);
918 if (!page_list)
919 return -ENOMEM;
920
921 if (perm & VHOST_ACCESS_WO)
922 gup_flags |= FOLL_WRITE;
923
924 npages = PFN_UP(size + (iova & ~PAGE_MASK));
925 if (!npages) {
926 ret = -EINVAL;
927 goto free;
928 }
929
930 mmap_read_lock(dev->mm);
931
932 lock_limit = PFN_DOWN(rlimit(RLIMIT_MEMLOCK));
933 if (npages + atomic64_read(&dev->mm->pinned_vm) > lock_limit) {
934 ret = -ENOMEM;
935 goto unlock;
936 }
937
938 cur_base = uaddr & PAGE_MASK;
939 iova &= PAGE_MASK;
940 nchunks = 0;
941
942 while (npages) {
943 sz2pin = min_t(unsigned long, npages, list_size);
944 pinned = pin_user_pages(cur_base, sz2pin,
945 gup_flags, page_list, NULL);
946 if (sz2pin != pinned) {
947 if (pinned < 0) {
948 ret = pinned;
949 } else {
950 unpin_user_pages(page_list, pinned);
951 ret = -ENOMEM;
952 }
953 goto out;
954 }
955 nchunks++;
956
957 if (!last_pfn)
958 map_pfn = page_to_pfn(page_list[0]);
959
960 for (i = 0; i < pinned; i++) {
961 unsigned long this_pfn = page_to_pfn(page_list[i]);
962 u64 csize;
963
964 if (last_pfn && (this_pfn != last_pfn + 1)) {
965 /* Pin a contiguous chunk of memory */
966 csize = PFN_PHYS(last_pfn - map_pfn + 1);
967 ret = vhost_vdpa_map(v, iotlb, iova, csize,
968 PFN_PHYS(map_pfn),
969 perm, NULL);
970 if (ret) {
971 /*
972 * Unpin the pages that are left unmapped
973 * from this point on in the current
974 * page_list. The remaining outstanding
975 * ones which may stride across several
976 * chunks will be covered in the common
977 * error path subsequently.
978 */
979 unpin_user_pages(&page_list[i],
980 pinned - i);
981 goto out;
982 }
983
984 map_pfn = this_pfn;
985 iova += csize;
986 nchunks = 0;
987 }
988
989 last_pfn = this_pfn;
990 }
991
992 cur_base += PFN_PHYS(pinned);
993 npages -= pinned;
994 }
995
996 /* Pin the rest chunk */
997 ret = vhost_vdpa_map(v, iotlb, iova, PFN_PHYS(last_pfn - map_pfn + 1),
998 PFN_PHYS(map_pfn), perm, NULL);
999 out:
1000 if (ret) {
1001 if (nchunks) {
1002 unsigned long pfn;
1003
1004 /*
1005 * Unpin the outstanding pages which are yet to be
1006 * mapped but haven't due to vdpa_map() or
1007 * pin_user_pages() failure.
1008 *
1009 * Mapped pages are accounted in vdpa_map(), hence
1010 * the corresponding unpinning will be handled by
1011 * vdpa_unmap().
1012 */
1013 WARN_ON(!last_pfn);
1014 for (pfn = map_pfn; pfn <= last_pfn; pfn++)
1015 unpin_user_page(pfn_to_page(pfn));
1016 }
1017 vhost_vdpa_unmap(v, iotlb, start, size);
1018 }
1019 unlock:
1020 mmap_read_unlock(dev->mm);
1021 free:
1022 free_page((unsigned long)page_list);
1023 return ret;
1024
1025 }
1026
vhost_vdpa_process_iotlb_update(struct vhost_vdpa * v,struct vhost_iotlb * iotlb,struct vhost_iotlb_msg * msg)1027 static int vhost_vdpa_process_iotlb_update(struct vhost_vdpa *v,
1028 struct vhost_iotlb *iotlb,
1029 struct vhost_iotlb_msg *msg)
1030 {
1031 struct vdpa_device *vdpa = v->vdpa;
1032
1033 if (msg->iova < v->range.first || !msg->size ||
1034 msg->iova > U64_MAX - msg->size + 1 ||
1035 msg->iova + msg->size - 1 > v->range.last)
1036 return -EINVAL;
1037
1038 if (vhost_iotlb_itree_first(iotlb, msg->iova,
1039 msg->iova + msg->size - 1))
1040 return -EEXIST;
1041
1042 if (vdpa->use_va)
1043 return vhost_vdpa_va_map(v, iotlb, msg->iova, msg->size,
1044 msg->uaddr, msg->perm);
1045
1046 return vhost_vdpa_pa_map(v, iotlb, msg->iova, msg->size, msg->uaddr,
1047 msg->perm);
1048 }
1049
vhost_vdpa_process_iotlb_msg(struct vhost_dev * dev,u32 asid,struct vhost_iotlb_msg * msg)1050 static int vhost_vdpa_process_iotlb_msg(struct vhost_dev *dev, u32 asid,
1051 struct vhost_iotlb_msg *msg)
1052 {
1053 struct vhost_vdpa *v = container_of(dev, struct vhost_vdpa, vdev);
1054 struct vdpa_device *vdpa = v->vdpa;
1055 const struct vdpa_config_ops *ops = vdpa->config;
1056 struct vhost_iotlb *iotlb = NULL;
1057 struct vhost_vdpa_as *as = NULL;
1058 int r = 0;
1059
1060 mutex_lock(&dev->mutex);
1061
1062 r = vhost_dev_check_owner(dev);
1063 if (r)
1064 goto unlock;
1065
1066 if (msg->type == VHOST_IOTLB_UPDATE ||
1067 msg->type == VHOST_IOTLB_BATCH_BEGIN) {
1068 as = vhost_vdpa_find_alloc_as(v, asid);
1069 if (!as) {
1070 dev_err(&v->dev, "can't find and alloc asid %d\n",
1071 asid);
1072 r = -EINVAL;
1073 goto unlock;
1074 }
1075 iotlb = &as->iotlb;
1076 } else
1077 iotlb = asid_to_iotlb(v, asid);
1078
1079 if ((v->in_batch && v->batch_asid != asid) || !iotlb) {
1080 if (v->in_batch && v->batch_asid != asid) {
1081 dev_info(&v->dev, "batch id %d asid %d\n",
1082 v->batch_asid, asid);
1083 }
1084 if (!iotlb)
1085 dev_err(&v->dev, "no iotlb for asid %d\n", asid);
1086 r = -EINVAL;
1087 goto unlock;
1088 }
1089
1090 switch (msg->type) {
1091 case VHOST_IOTLB_UPDATE:
1092 r = vhost_vdpa_process_iotlb_update(v, iotlb, msg);
1093 break;
1094 case VHOST_IOTLB_INVALIDATE:
1095 vhost_vdpa_unmap(v, iotlb, msg->iova, msg->size);
1096 break;
1097 case VHOST_IOTLB_BATCH_BEGIN:
1098 v->batch_asid = asid;
1099 v->in_batch = true;
1100 break;
1101 case VHOST_IOTLB_BATCH_END:
1102 if (v->in_batch && ops->set_map)
1103 ops->set_map(vdpa, asid, iotlb);
1104 v->in_batch = false;
1105 break;
1106 default:
1107 r = -EINVAL;
1108 break;
1109 }
1110 unlock:
1111 mutex_unlock(&dev->mutex);
1112
1113 return r;
1114 }
1115
vhost_vdpa_chr_write_iter(struct kiocb * iocb,struct iov_iter * from)1116 static ssize_t vhost_vdpa_chr_write_iter(struct kiocb *iocb,
1117 struct iov_iter *from)
1118 {
1119 struct file *file = iocb->ki_filp;
1120 struct vhost_vdpa *v = file->private_data;
1121 struct vhost_dev *dev = &v->vdev;
1122
1123 return vhost_chr_write_iter(dev, from);
1124 }
1125
vhost_vdpa_alloc_domain(struct vhost_vdpa * v)1126 static int vhost_vdpa_alloc_domain(struct vhost_vdpa *v)
1127 {
1128 struct vdpa_device *vdpa = v->vdpa;
1129 const struct vdpa_config_ops *ops = vdpa->config;
1130 struct device *dma_dev = vdpa_get_dma_dev(vdpa);
1131 struct bus_type *bus;
1132 int ret;
1133
1134 /* Device want to do DMA by itself */
1135 if (ops->set_map || ops->dma_map)
1136 return 0;
1137
1138 bus = dma_dev->bus;
1139 if (!bus)
1140 return -EFAULT;
1141
1142 if (!device_iommu_capable(dma_dev, IOMMU_CAP_CACHE_COHERENCY))
1143 return -ENOTSUPP;
1144
1145 v->domain = iommu_domain_alloc(bus);
1146 if (!v->domain)
1147 return -EIO;
1148
1149 ret = iommu_attach_device(v->domain, dma_dev);
1150 if (ret)
1151 goto err_attach;
1152
1153 return 0;
1154
1155 err_attach:
1156 iommu_domain_free(v->domain);
1157 v->domain = NULL;
1158 return ret;
1159 }
1160
vhost_vdpa_free_domain(struct vhost_vdpa * v)1161 static void vhost_vdpa_free_domain(struct vhost_vdpa *v)
1162 {
1163 struct vdpa_device *vdpa = v->vdpa;
1164 struct device *dma_dev = vdpa_get_dma_dev(vdpa);
1165
1166 if (v->domain) {
1167 iommu_detach_device(v->domain, dma_dev);
1168 iommu_domain_free(v->domain);
1169 }
1170
1171 v->domain = NULL;
1172 }
1173
vhost_vdpa_set_iova_range(struct vhost_vdpa * v)1174 static void vhost_vdpa_set_iova_range(struct vhost_vdpa *v)
1175 {
1176 struct vdpa_iova_range *range = &v->range;
1177 struct vdpa_device *vdpa = v->vdpa;
1178 const struct vdpa_config_ops *ops = vdpa->config;
1179
1180 if (ops->get_iova_range) {
1181 *range = ops->get_iova_range(vdpa);
1182 } else if (v->domain && v->domain->geometry.force_aperture) {
1183 range->first = v->domain->geometry.aperture_start;
1184 range->last = v->domain->geometry.aperture_end;
1185 } else {
1186 range->first = 0;
1187 range->last = ULLONG_MAX;
1188 }
1189 }
1190
vhost_vdpa_cleanup(struct vhost_vdpa * v)1191 static void vhost_vdpa_cleanup(struct vhost_vdpa *v)
1192 {
1193 struct vhost_vdpa_as *as;
1194 u32 asid;
1195
1196 for (asid = 0; asid < v->vdpa->nas; asid++) {
1197 as = asid_to_as(v, asid);
1198 if (as)
1199 vhost_vdpa_remove_as(v, asid);
1200 }
1201
1202 vhost_vdpa_free_domain(v);
1203 vhost_dev_cleanup(&v->vdev);
1204 kfree(v->vdev.vqs);
1205 }
1206
vhost_vdpa_open(struct inode * inode,struct file * filep)1207 static int vhost_vdpa_open(struct inode *inode, struct file *filep)
1208 {
1209 struct vhost_vdpa *v;
1210 struct vhost_dev *dev;
1211 struct vhost_virtqueue **vqs;
1212 int r, opened;
1213 u32 i, nvqs;
1214
1215 v = container_of(inode->i_cdev, struct vhost_vdpa, cdev);
1216
1217 opened = atomic_cmpxchg(&v->opened, 0, 1);
1218 if (opened)
1219 return -EBUSY;
1220
1221 nvqs = v->nvqs;
1222 r = vhost_vdpa_reset(v);
1223 if (r)
1224 goto err;
1225
1226 vqs = kmalloc_array(nvqs, sizeof(*vqs), GFP_KERNEL);
1227 if (!vqs) {
1228 r = -ENOMEM;
1229 goto err;
1230 }
1231
1232 dev = &v->vdev;
1233 for (i = 0; i < nvqs; i++) {
1234 vqs[i] = &v->vqs[i];
1235 vqs[i]->handle_kick = handle_vq_kick;
1236 }
1237 vhost_dev_init(dev, vqs, nvqs, 0, 0, 0, false,
1238 vhost_vdpa_process_iotlb_msg);
1239
1240 r = vhost_vdpa_alloc_domain(v);
1241 if (r)
1242 goto err_alloc_domain;
1243
1244 vhost_vdpa_set_iova_range(v);
1245
1246 filep->private_data = v;
1247
1248 return 0;
1249
1250 err_alloc_domain:
1251 vhost_vdpa_cleanup(v);
1252 err:
1253 atomic_dec(&v->opened);
1254 return r;
1255 }
1256
vhost_vdpa_clean_irq(struct vhost_vdpa * v)1257 static void vhost_vdpa_clean_irq(struct vhost_vdpa *v)
1258 {
1259 u32 i;
1260
1261 for (i = 0; i < v->nvqs; i++)
1262 vhost_vdpa_unsetup_vq_irq(v, i);
1263 }
1264
vhost_vdpa_release(struct inode * inode,struct file * filep)1265 static int vhost_vdpa_release(struct inode *inode, struct file *filep)
1266 {
1267 struct vhost_vdpa *v = filep->private_data;
1268 struct vhost_dev *d = &v->vdev;
1269
1270 mutex_lock(&d->mutex);
1271 filep->private_data = NULL;
1272 vhost_vdpa_clean_irq(v);
1273 vhost_vdpa_reset(v);
1274 vhost_dev_stop(&v->vdev);
1275 vhost_vdpa_config_put(v);
1276 vhost_vdpa_cleanup(v);
1277 mutex_unlock(&d->mutex);
1278
1279 atomic_dec(&v->opened);
1280 complete(&v->completion);
1281
1282 return 0;
1283 }
1284
1285 #ifdef CONFIG_MMU
vhost_vdpa_fault(struct vm_fault * vmf)1286 static vm_fault_t vhost_vdpa_fault(struct vm_fault *vmf)
1287 {
1288 struct vhost_vdpa *v = vmf->vma->vm_file->private_data;
1289 struct vdpa_device *vdpa = v->vdpa;
1290 const struct vdpa_config_ops *ops = vdpa->config;
1291 struct vdpa_notification_area notify;
1292 struct vm_area_struct *vma = vmf->vma;
1293 u16 index = vma->vm_pgoff;
1294
1295 notify = ops->get_vq_notification(vdpa, index);
1296
1297 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1298 if (remap_pfn_range(vma, vmf->address & PAGE_MASK,
1299 PFN_DOWN(notify.addr), PAGE_SIZE,
1300 vma->vm_page_prot))
1301 return VM_FAULT_SIGBUS;
1302
1303 return VM_FAULT_NOPAGE;
1304 }
1305
1306 static const struct vm_operations_struct vhost_vdpa_vm_ops = {
1307 .fault = vhost_vdpa_fault,
1308 };
1309
vhost_vdpa_mmap(struct file * file,struct vm_area_struct * vma)1310 static int vhost_vdpa_mmap(struct file *file, struct vm_area_struct *vma)
1311 {
1312 struct vhost_vdpa *v = vma->vm_file->private_data;
1313 struct vdpa_device *vdpa = v->vdpa;
1314 const struct vdpa_config_ops *ops = vdpa->config;
1315 struct vdpa_notification_area notify;
1316 unsigned long index = vma->vm_pgoff;
1317
1318 if (vma->vm_end - vma->vm_start != PAGE_SIZE)
1319 return -EINVAL;
1320 if ((vma->vm_flags & VM_SHARED) == 0)
1321 return -EINVAL;
1322 if (vma->vm_flags & VM_READ)
1323 return -EINVAL;
1324 if (index > 65535)
1325 return -EINVAL;
1326 if (!ops->get_vq_notification)
1327 return -ENOTSUPP;
1328
1329 /* To be safe and easily modelled by userspace, We only
1330 * support the doorbell which sits on the page boundary and
1331 * does not share the page with other registers.
1332 */
1333 notify = ops->get_vq_notification(vdpa, index);
1334 if (notify.addr & (PAGE_SIZE - 1))
1335 return -EINVAL;
1336 if (vma->vm_end - vma->vm_start != notify.size)
1337 return -ENOTSUPP;
1338
1339 vm_flags_set(vma, VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP);
1340 vma->vm_ops = &vhost_vdpa_vm_ops;
1341 return 0;
1342 }
1343 #endif /* CONFIG_MMU */
1344
1345 static const struct file_operations vhost_vdpa_fops = {
1346 .owner = THIS_MODULE,
1347 .open = vhost_vdpa_open,
1348 .release = vhost_vdpa_release,
1349 .write_iter = vhost_vdpa_chr_write_iter,
1350 .unlocked_ioctl = vhost_vdpa_unlocked_ioctl,
1351 #ifdef CONFIG_MMU
1352 .mmap = vhost_vdpa_mmap,
1353 #endif /* CONFIG_MMU */
1354 .compat_ioctl = compat_ptr_ioctl,
1355 };
1356
vhost_vdpa_release_dev(struct device * device)1357 static void vhost_vdpa_release_dev(struct device *device)
1358 {
1359 struct vhost_vdpa *v =
1360 container_of(device, struct vhost_vdpa, dev);
1361
1362 ida_simple_remove(&vhost_vdpa_ida, v->minor);
1363 kfree(v->vqs);
1364 kfree(v);
1365 }
1366
vhost_vdpa_probe(struct vdpa_device * vdpa)1367 static int vhost_vdpa_probe(struct vdpa_device *vdpa)
1368 {
1369 const struct vdpa_config_ops *ops = vdpa->config;
1370 struct vhost_vdpa *v;
1371 int minor;
1372 int i, r;
1373
1374 /* We can't support platform IOMMU device with more than 1
1375 * group or as
1376 */
1377 if (!ops->set_map && !ops->dma_map &&
1378 (vdpa->ngroups > 1 || vdpa->nas > 1))
1379 return -EOPNOTSUPP;
1380
1381 v = kzalloc(sizeof(*v), GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1382 if (!v)
1383 return -ENOMEM;
1384
1385 minor = ida_simple_get(&vhost_vdpa_ida, 0,
1386 VHOST_VDPA_DEV_MAX, GFP_KERNEL);
1387 if (minor < 0) {
1388 kfree(v);
1389 return minor;
1390 }
1391
1392 atomic_set(&v->opened, 0);
1393 v->minor = minor;
1394 v->vdpa = vdpa;
1395 v->nvqs = vdpa->nvqs;
1396 v->virtio_id = ops->get_device_id(vdpa);
1397
1398 device_initialize(&v->dev);
1399 v->dev.release = vhost_vdpa_release_dev;
1400 v->dev.parent = &vdpa->dev;
1401 v->dev.devt = MKDEV(MAJOR(vhost_vdpa_major), minor);
1402 v->vqs = kmalloc_array(v->nvqs, sizeof(struct vhost_virtqueue),
1403 GFP_KERNEL);
1404 if (!v->vqs) {
1405 r = -ENOMEM;
1406 goto err;
1407 }
1408
1409 r = dev_set_name(&v->dev, "vhost-vdpa-%u", minor);
1410 if (r)
1411 goto err;
1412
1413 cdev_init(&v->cdev, &vhost_vdpa_fops);
1414 v->cdev.owner = THIS_MODULE;
1415
1416 r = cdev_device_add(&v->cdev, &v->dev);
1417 if (r)
1418 goto err;
1419
1420 init_completion(&v->completion);
1421 vdpa_set_drvdata(vdpa, v);
1422
1423 for (i = 0; i < VHOST_VDPA_IOTLB_BUCKETS; i++)
1424 INIT_HLIST_HEAD(&v->as[i]);
1425
1426 return 0;
1427
1428 err:
1429 put_device(&v->dev);
1430 return r;
1431 }
1432
vhost_vdpa_remove(struct vdpa_device * vdpa)1433 static void vhost_vdpa_remove(struct vdpa_device *vdpa)
1434 {
1435 struct vhost_vdpa *v = vdpa_get_drvdata(vdpa);
1436 int opened;
1437
1438 cdev_device_del(&v->cdev, &v->dev);
1439
1440 do {
1441 opened = atomic_cmpxchg(&v->opened, 0, 1);
1442 if (!opened)
1443 break;
1444 wait_for_completion(&v->completion);
1445 } while (1);
1446
1447 put_device(&v->dev);
1448 }
1449
1450 static struct vdpa_driver vhost_vdpa_driver = {
1451 .driver = {
1452 .name = "vhost_vdpa",
1453 },
1454 .probe = vhost_vdpa_probe,
1455 .remove = vhost_vdpa_remove,
1456 };
1457
vhost_vdpa_init(void)1458 static int __init vhost_vdpa_init(void)
1459 {
1460 int r;
1461
1462 r = alloc_chrdev_region(&vhost_vdpa_major, 0, VHOST_VDPA_DEV_MAX,
1463 "vhost-vdpa");
1464 if (r)
1465 goto err_alloc_chrdev;
1466
1467 r = vdpa_register_driver(&vhost_vdpa_driver);
1468 if (r)
1469 goto err_vdpa_register_driver;
1470
1471 return 0;
1472
1473 err_vdpa_register_driver:
1474 unregister_chrdev_region(vhost_vdpa_major, VHOST_VDPA_DEV_MAX);
1475 err_alloc_chrdev:
1476 return r;
1477 }
1478 module_init(vhost_vdpa_init);
1479
vhost_vdpa_exit(void)1480 static void __exit vhost_vdpa_exit(void)
1481 {
1482 vdpa_unregister_driver(&vhost_vdpa_driver);
1483 unregister_chrdev_region(vhost_vdpa_major, VHOST_VDPA_DEV_MAX);
1484 }
1485 module_exit(vhost_vdpa_exit);
1486
1487 MODULE_VERSION("0.0.1");
1488 MODULE_LICENSE("GPL v2");
1489 MODULE_AUTHOR("Intel Corporation");
1490 MODULE_DESCRIPTION("vDPA-based vhost backend for virtio");
1491