Loading...
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
28enum {
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
39struct vhost_vdpa_as {
40 struct hlist_node hash_link;
41 struct vhost_iotlb iotlb;
42 u32 id;
43};
44
45struct 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 bool suspended;
63};
64
65static DEFINE_IDA(vhost_vdpa_ida);
66
67static dev_t vhost_vdpa_major;
68
69static void vhost_vdpa_iotlb_unmap(struct vhost_vdpa *v,
70 struct vhost_iotlb *iotlb, u64 start,
71 u64 last, u32 asid);
72
73static inline u32 iotlb_to_asid(struct vhost_iotlb *iotlb)
74{
75 struct vhost_vdpa_as *as = container_of(iotlb, struct
76 vhost_vdpa_as, iotlb);
77 return as->id;
78}
79
80static struct vhost_vdpa_as *asid_to_as(struct vhost_vdpa *v, u32 asid)
81{
82 struct hlist_head *head = &v->as[asid % VHOST_VDPA_IOTLB_BUCKETS];
83 struct vhost_vdpa_as *as;
84
85 hlist_for_each_entry(as, head, hash_link)
86 if (as->id == asid)
87 return as;
88
89 return NULL;
90}
91
92static struct vhost_iotlb *asid_to_iotlb(struct vhost_vdpa *v, u32 asid)
93{
94 struct vhost_vdpa_as *as = asid_to_as(v, asid);
95
96 if (!as)
97 return NULL;
98
99 return &as->iotlb;
100}
101
102static struct vhost_vdpa_as *vhost_vdpa_alloc_as(struct vhost_vdpa *v, u32 asid)
103{
104 struct hlist_head *head = &v->as[asid % VHOST_VDPA_IOTLB_BUCKETS];
105 struct vhost_vdpa_as *as;
106
107 if (asid_to_as(v, asid))
108 return NULL;
109
110 if (asid >= v->vdpa->nas)
111 return NULL;
112
113 as = kmalloc(sizeof(*as), GFP_KERNEL);
114 if (!as)
115 return NULL;
116
117 vhost_iotlb_init(&as->iotlb, 0, 0);
118 as->id = asid;
119 hlist_add_head(&as->hash_link, head);
120
121 return as;
122}
123
124static struct vhost_vdpa_as *vhost_vdpa_find_alloc_as(struct vhost_vdpa *v,
125 u32 asid)
126{
127 struct vhost_vdpa_as *as = asid_to_as(v, asid);
128
129 if (as)
130 return as;
131
132 return vhost_vdpa_alloc_as(v, asid);
133}
134
135static void vhost_vdpa_reset_map(struct vhost_vdpa *v, u32 asid)
136{
137 struct vdpa_device *vdpa = v->vdpa;
138 const struct vdpa_config_ops *ops = vdpa->config;
139
140 if (ops->reset_map)
141 ops->reset_map(vdpa, asid);
142}
143
144static int vhost_vdpa_remove_as(struct vhost_vdpa *v, u32 asid)
145{
146 struct vhost_vdpa_as *as = asid_to_as(v, asid);
147
148 if (!as)
149 return -EINVAL;
150
151 hlist_del(&as->hash_link);
152 vhost_vdpa_iotlb_unmap(v, &as->iotlb, 0ULL, 0ULL - 1, asid);
153 /*
154 * Devices with vendor specific IOMMU may need to restore
155 * iotlb to the initial or default state, which cannot be
156 * cleaned up in the all range unmap call above. Give them
157 * a chance to clean up or reset the map to the desired
158 * state.
159 */
160 vhost_vdpa_reset_map(v, asid);
161 kfree(as);
162
163 return 0;
164}
165
166static void handle_vq_kick(struct vhost_work *work)
167{
168 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
169 poll.work);
170 struct vhost_vdpa *v = container_of(vq->dev, struct vhost_vdpa, vdev);
171 const struct vdpa_config_ops *ops = v->vdpa->config;
172
173 ops->kick_vq(v->vdpa, vq - v->vqs);
174}
175
176static irqreturn_t vhost_vdpa_virtqueue_cb(void *private)
177{
178 struct vhost_virtqueue *vq = private;
179 struct eventfd_ctx *call_ctx = vq->call_ctx.ctx;
180
181 if (call_ctx)
182 eventfd_signal(call_ctx);
183
184 return IRQ_HANDLED;
185}
186
187static irqreturn_t vhost_vdpa_config_cb(void *private)
188{
189 struct vhost_vdpa *v = private;
190 struct eventfd_ctx *config_ctx = v->config_ctx;
191
192 if (config_ctx)
193 eventfd_signal(config_ctx);
194
195 return IRQ_HANDLED;
196}
197
198static void vhost_vdpa_setup_vq_irq(struct vhost_vdpa *v, u16 qid)
199{
200 struct vhost_virtqueue *vq = &v->vqs[qid];
201 const struct vdpa_config_ops *ops = v->vdpa->config;
202 struct vdpa_device *vdpa = v->vdpa;
203 int ret, irq;
204
205 if (!ops->get_vq_irq)
206 return;
207
208 irq = ops->get_vq_irq(vdpa, qid);
209 if (irq < 0)
210 return;
211
212 if (!vq->call_ctx.ctx)
213 return;
214
215 vq->call_ctx.producer.irq = irq;
216 ret = irq_bypass_register_producer(&vq->call_ctx.producer);
217 if (unlikely(ret))
218 dev_info(&v->dev, "vq %u, irq bypass producer (token %p) registration fails, ret = %d\n",
219 qid, vq->call_ctx.producer.token, ret);
220}
221
222static void vhost_vdpa_unsetup_vq_irq(struct vhost_vdpa *v, u16 qid)
223{
224 struct vhost_virtqueue *vq = &v->vqs[qid];
225
226 irq_bypass_unregister_producer(&vq->call_ctx.producer);
227}
228
229static int _compat_vdpa_reset(struct vhost_vdpa *v)
230{
231 struct vdpa_device *vdpa = v->vdpa;
232 u32 flags = 0;
233
234 v->suspended = false;
235
236 if (v->vdev.vqs) {
237 flags |= !vhost_backend_has_feature(v->vdev.vqs[0],
238 VHOST_BACKEND_F_IOTLB_PERSIST) ?
239 VDPA_RESET_F_CLEAN_MAP : 0;
240 }
241
242 return vdpa_reset(vdpa, flags);
243}
244
245static int vhost_vdpa_reset(struct vhost_vdpa *v)
246{
247 v->in_batch = 0;
248 return _compat_vdpa_reset(v);
249}
250
251static long vhost_vdpa_bind_mm(struct vhost_vdpa *v)
252{
253 struct vdpa_device *vdpa = v->vdpa;
254 const struct vdpa_config_ops *ops = vdpa->config;
255
256 if (!vdpa->use_va || !ops->bind_mm)
257 return 0;
258
259 return ops->bind_mm(vdpa, v->vdev.mm);
260}
261
262static void vhost_vdpa_unbind_mm(struct vhost_vdpa *v)
263{
264 struct vdpa_device *vdpa = v->vdpa;
265 const struct vdpa_config_ops *ops = vdpa->config;
266
267 if (!vdpa->use_va || !ops->unbind_mm)
268 return;
269
270 ops->unbind_mm(vdpa);
271}
272
273static long vhost_vdpa_get_device_id(struct vhost_vdpa *v, u8 __user *argp)
274{
275 struct vdpa_device *vdpa = v->vdpa;
276 const struct vdpa_config_ops *ops = vdpa->config;
277 u32 device_id;
278
279 device_id = ops->get_device_id(vdpa);
280
281 if (copy_to_user(argp, &device_id, sizeof(device_id)))
282 return -EFAULT;
283
284 return 0;
285}
286
287static long vhost_vdpa_get_status(struct vhost_vdpa *v, u8 __user *statusp)
288{
289 struct vdpa_device *vdpa = v->vdpa;
290 const struct vdpa_config_ops *ops = vdpa->config;
291 u8 status;
292
293 status = ops->get_status(vdpa);
294
295 if (copy_to_user(statusp, &status, sizeof(status)))
296 return -EFAULT;
297
298 return 0;
299}
300
301static long vhost_vdpa_set_status(struct vhost_vdpa *v, u8 __user *statusp)
302{
303 struct vdpa_device *vdpa = v->vdpa;
304 const struct vdpa_config_ops *ops = vdpa->config;
305 u8 status, status_old;
306 u32 nvqs = v->nvqs;
307 int ret;
308 u16 i;
309
310 if (copy_from_user(&status, statusp, sizeof(status)))
311 return -EFAULT;
312
313 status_old = ops->get_status(vdpa);
314
315 /*
316 * Userspace shouldn't remove status bits unless reset the
317 * status to 0.
318 */
319 if (status != 0 && (status_old & ~status) != 0)
320 return -EINVAL;
321
322 if ((status_old & VIRTIO_CONFIG_S_DRIVER_OK) && !(status & VIRTIO_CONFIG_S_DRIVER_OK))
323 for (i = 0; i < nvqs; i++)
324 vhost_vdpa_unsetup_vq_irq(v, i);
325
326 if (status == 0) {
327 ret = _compat_vdpa_reset(v);
328 if (ret)
329 return ret;
330 } else
331 vdpa_set_status(vdpa, status);
332
333 if ((status & VIRTIO_CONFIG_S_DRIVER_OK) && !(status_old & VIRTIO_CONFIG_S_DRIVER_OK))
334 for (i = 0; i < nvqs; i++)
335 vhost_vdpa_setup_vq_irq(v, i);
336
337 return 0;
338}
339
340static int vhost_vdpa_config_validate(struct vhost_vdpa *v,
341 struct vhost_vdpa_config *c)
342{
343 struct vdpa_device *vdpa = v->vdpa;
344 size_t size = vdpa->config->get_config_size(vdpa);
345
346 if (c->len == 0 || c->off > size)
347 return -EINVAL;
348
349 if (c->len > size - c->off)
350 return -E2BIG;
351
352 return 0;
353}
354
355static long vhost_vdpa_get_config(struct vhost_vdpa *v,
356 struct vhost_vdpa_config __user *c)
357{
358 struct vdpa_device *vdpa = v->vdpa;
359 struct vhost_vdpa_config config;
360 unsigned long size = offsetof(struct vhost_vdpa_config, buf);
361 u8 *buf;
362
363 if (copy_from_user(&config, c, size))
364 return -EFAULT;
365 if (vhost_vdpa_config_validate(v, &config))
366 return -EINVAL;
367 buf = kvzalloc(config.len, GFP_KERNEL);
368 if (!buf)
369 return -ENOMEM;
370
371 vdpa_get_config(vdpa, config.off, buf, config.len);
372
373 if (copy_to_user(c->buf, buf, config.len)) {
374 kvfree(buf);
375 return -EFAULT;
376 }
377
378 kvfree(buf);
379 return 0;
380}
381
382static long vhost_vdpa_set_config(struct vhost_vdpa *v,
383 struct vhost_vdpa_config __user *c)
384{
385 struct vdpa_device *vdpa = v->vdpa;
386 struct vhost_vdpa_config config;
387 unsigned long size = offsetof(struct vhost_vdpa_config, buf);
388 u8 *buf;
389
390 if (copy_from_user(&config, c, size))
391 return -EFAULT;
392 if (vhost_vdpa_config_validate(v, &config))
393 return -EINVAL;
394
395 buf = vmemdup_user(c->buf, config.len);
396 if (IS_ERR(buf))
397 return PTR_ERR(buf);
398
399 vdpa_set_config(vdpa, config.off, buf, config.len);
400
401 kvfree(buf);
402 return 0;
403}
404
405static bool vhost_vdpa_can_suspend(const struct vhost_vdpa *v)
406{
407 struct vdpa_device *vdpa = v->vdpa;
408 const struct vdpa_config_ops *ops = vdpa->config;
409
410 return ops->suspend;
411}
412
413static bool vhost_vdpa_can_resume(const struct vhost_vdpa *v)
414{
415 struct vdpa_device *vdpa = v->vdpa;
416 const struct vdpa_config_ops *ops = vdpa->config;
417
418 return ops->resume;
419}
420
421static bool vhost_vdpa_has_desc_group(const struct vhost_vdpa *v)
422{
423 struct vdpa_device *vdpa = v->vdpa;
424 const struct vdpa_config_ops *ops = vdpa->config;
425
426 return ops->get_vq_desc_group;
427}
428
429static long vhost_vdpa_get_features(struct vhost_vdpa *v, u64 __user *featurep)
430{
431 struct vdpa_device *vdpa = v->vdpa;
432 const struct vdpa_config_ops *ops = vdpa->config;
433 u64 features;
434
435 features = ops->get_device_features(vdpa);
436
437 if (copy_to_user(featurep, &features, sizeof(features)))
438 return -EFAULT;
439
440 return 0;
441}
442
443static u64 vhost_vdpa_get_backend_features(const struct vhost_vdpa *v)
444{
445 struct vdpa_device *vdpa = v->vdpa;
446 const struct vdpa_config_ops *ops = vdpa->config;
447
448 if (!ops->get_backend_features)
449 return 0;
450 else
451 return ops->get_backend_features(vdpa);
452}
453
454static bool vhost_vdpa_has_persistent_map(const struct vhost_vdpa *v)
455{
456 struct vdpa_device *vdpa = v->vdpa;
457 const struct vdpa_config_ops *ops = vdpa->config;
458
459 return (!ops->set_map && !ops->dma_map) || ops->reset_map ||
460 vhost_vdpa_get_backend_features(v) & BIT_ULL(VHOST_BACKEND_F_IOTLB_PERSIST);
461}
462
463static long vhost_vdpa_set_features(struct vhost_vdpa *v, u64 __user *featurep)
464{
465 struct vdpa_device *vdpa = v->vdpa;
466 const struct vdpa_config_ops *ops = vdpa->config;
467 struct vhost_dev *d = &v->vdev;
468 u64 actual_features;
469 u64 features;
470 int i;
471
472 /*
473 * It's not allowed to change the features after they have
474 * been negotiated.
475 */
476 if (ops->get_status(vdpa) & VIRTIO_CONFIG_S_FEATURES_OK)
477 return -EBUSY;
478
479 if (copy_from_user(&features, featurep, sizeof(features)))
480 return -EFAULT;
481
482 if (vdpa_set_features(vdpa, features))
483 return -EINVAL;
484
485 /* let the vqs know what has been configured */
486 actual_features = ops->get_driver_features(vdpa);
487 for (i = 0; i < d->nvqs; ++i) {
488 struct vhost_virtqueue *vq = d->vqs[i];
489
490 mutex_lock(&vq->mutex);
491 vq->acked_features = actual_features;
492 mutex_unlock(&vq->mutex);
493 }
494
495 return 0;
496}
497
498static long vhost_vdpa_get_vring_num(struct vhost_vdpa *v, u16 __user *argp)
499{
500 struct vdpa_device *vdpa = v->vdpa;
501 const struct vdpa_config_ops *ops = vdpa->config;
502 u16 num;
503
504 num = ops->get_vq_num_max(vdpa);
505
506 if (copy_to_user(argp, &num, sizeof(num)))
507 return -EFAULT;
508
509 return 0;
510}
511
512static void vhost_vdpa_config_put(struct vhost_vdpa *v)
513{
514 if (v->config_ctx) {
515 eventfd_ctx_put(v->config_ctx);
516 v->config_ctx = NULL;
517 }
518}
519
520static long vhost_vdpa_set_config_call(struct vhost_vdpa *v, u32 __user *argp)
521{
522 struct vdpa_callback cb;
523 int fd;
524 struct eventfd_ctx *ctx;
525
526 cb.callback = vhost_vdpa_config_cb;
527 cb.private = v;
528 if (copy_from_user(&fd, argp, sizeof(fd)))
529 return -EFAULT;
530
531 ctx = fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(fd);
532 swap(ctx, v->config_ctx);
533
534 if (!IS_ERR_OR_NULL(ctx))
535 eventfd_ctx_put(ctx);
536
537 if (IS_ERR(v->config_ctx)) {
538 long ret = PTR_ERR(v->config_ctx);
539
540 v->config_ctx = NULL;
541 return ret;
542 }
543
544 v->vdpa->config->set_config_cb(v->vdpa, &cb);
545
546 return 0;
547}
548
549static long vhost_vdpa_get_iova_range(struct vhost_vdpa *v, u32 __user *argp)
550{
551 struct vhost_vdpa_iova_range range = {
552 .first = v->range.first,
553 .last = v->range.last,
554 };
555
556 if (copy_to_user(argp, &range, sizeof(range)))
557 return -EFAULT;
558 return 0;
559}
560
561static long vhost_vdpa_get_config_size(struct vhost_vdpa *v, u32 __user *argp)
562{
563 struct vdpa_device *vdpa = v->vdpa;
564 const struct vdpa_config_ops *ops = vdpa->config;
565 u32 size;
566
567 size = ops->get_config_size(vdpa);
568
569 if (copy_to_user(argp, &size, sizeof(size)))
570 return -EFAULT;
571
572 return 0;
573}
574
575static long vhost_vdpa_get_vqs_count(struct vhost_vdpa *v, u32 __user *argp)
576{
577 struct vdpa_device *vdpa = v->vdpa;
578
579 if (copy_to_user(argp, &vdpa->nvqs, sizeof(vdpa->nvqs)))
580 return -EFAULT;
581
582 return 0;
583}
584
585/* After a successful return of ioctl the device must not process more
586 * virtqueue descriptors. The device can answer to read or writes of config
587 * fields as if it were not suspended. In particular, writing to "queue_enable"
588 * with a value of 1 will not make the device start processing buffers.
589 */
590static long vhost_vdpa_suspend(struct vhost_vdpa *v)
591{
592 struct vdpa_device *vdpa = v->vdpa;
593 const struct vdpa_config_ops *ops = vdpa->config;
594 int ret;
595
596 if (!(ops->get_status(vdpa) & VIRTIO_CONFIG_S_DRIVER_OK))
597 return 0;
598
599 if (!ops->suspend)
600 return -EOPNOTSUPP;
601
602 ret = ops->suspend(vdpa);
603 if (!ret)
604 v->suspended = true;
605
606 return ret;
607}
608
609/* After a successful return of this ioctl the device resumes processing
610 * virtqueue descriptors. The device becomes fully operational the same way it
611 * was before it was suspended.
612 */
613static long vhost_vdpa_resume(struct vhost_vdpa *v)
614{
615 struct vdpa_device *vdpa = v->vdpa;
616 const struct vdpa_config_ops *ops = vdpa->config;
617 int ret;
618
619 if (!(ops->get_status(vdpa) & VIRTIO_CONFIG_S_DRIVER_OK))
620 return 0;
621
622 if (!ops->resume)
623 return -EOPNOTSUPP;
624
625 ret = ops->resume(vdpa);
626 if (!ret)
627 v->suspended = false;
628
629 return ret;
630}
631
632static long vhost_vdpa_vring_ioctl(struct vhost_vdpa *v, unsigned int cmd,
633 void __user *argp)
634{
635 struct vdpa_device *vdpa = v->vdpa;
636 const struct vdpa_config_ops *ops = vdpa->config;
637 struct vdpa_vq_state vq_state;
638 struct vdpa_callback cb;
639 struct vhost_virtqueue *vq;
640 struct vhost_vring_state s;
641 u32 idx;
642 long r;
643
644 r = get_user(idx, (u32 __user *)argp);
645 if (r < 0)
646 return r;
647
648 if (idx >= v->nvqs)
649 return -ENOBUFS;
650
651 idx = array_index_nospec(idx, v->nvqs);
652 vq = &v->vqs[idx];
653
654 switch (cmd) {
655 case VHOST_VDPA_SET_VRING_ENABLE:
656 if (copy_from_user(&s, argp, sizeof(s)))
657 return -EFAULT;
658 ops->set_vq_ready(vdpa, idx, s.num);
659 return 0;
660 case VHOST_VDPA_GET_VRING_GROUP:
661 if (!ops->get_vq_group)
662 return -EOPNOTSUPP;
663 s.index = idx;
664 s.num = ops->get_vq_group(vdpa, idx);
665 if (s.num >= vdpa->ngroups)
666 return -EIO;
667 else if (copy_to_user(argp, &s, sizeof(s)))
668 return -EFAULT;
669 return 0;
670 case VHOST_VDPA_GET_VRING_DESC_GROUP:
671 if (!vhost_vdpa_has_desc_group(v))
672 return -EOPNOTSUPP;
673 s.index = idx;
674 s.num = ops->get_vq_desc_group(vdpa, idx);
675 if (s.num >= vdpa->ngroups)
676 return -EIO;
677 else if (copy_to_user(argp, &s, sizeof(s)))
678 return -EFAULT;
679 return 0;
680 case VHOST_VDPA_SET_GROUP_ASID:
681 if (copy_from_user(&s, argp, sizeof(s)))
682 return -EFAULT;
683 if (s.num >= vdpa->nas)
684 return -EINVAL;
685 if (!ops->set_group_asid)
686 return -EOPNOTSUPP;
687 return ops->set_group_asid(vdpa, idx, s.num);
688 case VHOST_VDPA_GET_VRING_SIZE:
689 if (!ops->get_vq_size)
690 return -EOPNOTSUPP;
691 s.index = idx;
692 s.num = ops->get_vq_size(vdpa, idx);
693 if (copy_to_user(argp, &s, sizeof(s)))
694 return -EFAULT;
695 return 0;
696 case VHOST_GET_VRING_BASE:
697 r = ops->get_vq_state(v->vdpa, idx, &vq_state);
698 if (r)
699 return r;
700
701 if (vhost_has_feature(vq, VIRTIO_F_RING_PACKED)) {
702 vq->last_avail_idx = vq_state.packed.last_avail_idx |
703 (vq_state.packed.last_avail_counter << 15);
704 vq->last_used_idx = vq_state.packed.last_used_idx |
705 (vq_state.packed.last_used_counter << 15);
706 } else {
707 vq->last_avail_idx = vq_state.split.avail_index;
708 }
709 break;
710 case VHOST_SET_VRING_CALL:
711 if (vq->call_ctx.ctx) {
712 if (ops->get_status(vdpa) &
713 VIRTIO_CONFIG_S_DRIVER_OK)
714 vhost_vdpa_unsetup_vq_irq(v, idx);
715 vq->call_ctx.producer.token = NULL;
716 }
717 break;
718 }
719
720 r = vhost_vring_ioctl(&v->vdev, cmd, argp);
721 if (r)
722 return r;
723
724 switch (cmd) {
725 case VHOST_SET_VRING_ADDR:
726 if ((ops->get_status(vdpa) & VIRTIO_CONFIG_S_DRIVER_OK) && !v->suspended)
727 return -EINVAL;
728
729 if (ops->set_vq_address(vdpa, idx,
730 (u64)(uintptr_t)vq->desc,
731 (u64)(uintptr_t)vq->avail,
732 (u64)(uintptr_t)vq->used))
733 r = -EINVAL;
734 break;
735
736 case VHOST_SET_VRING_BASE:
737 if ((ops->get_status(vdpa) & VIRTIO_CONFIG_S_DRIVER_OK) && !v->suspended)
738 return -EINVAL;
739
740 if (vhost_has_feature(vq, VIRTIO_F_RING_PACKED)) {
741 vq_state.packed.last_avail_idx = vq->last_avail_idx & 0x7fff;
742 vq_state.packed.last_avail_counter = !!(vq->last_avail_idx & 0x8000);
743 vq_state.packed.last_used_idx = vq->last_used_idx & 0x7fff;
744 vq_state.packed.last_used_counter = !!(vq->last_used_idx & 0x8000);
745 } else {
746 vq_state.split.avail_index = vq->last_avail_idx;
747 }
748 r = ops->set_vq_state(vdpa, idx, &vq_state);
749 break;
750
751 case VHOST_SET_VRING_CALL:
752 if (vq->call_ctx.ctx) {
753 cb.callback = vhost_vdpa_virtqueue_cb;
754 cb.private = vq;
755 cb.trigger = vq->call_ctx.ctx;
756 vq->call_ctx.producer.token = vq->call_ctx.ctx;
757 if (ops->get_status(vdpa) &
758 VIRTIO_CONFIG_S_DRIVER_OK)
759 vhost_vdpa_setup_vq_irq(v, idx);
760 } else {
761 cb.callback = NULL;
762 cb.private = NULL;
763 cb.trigger = NULL;
764 }
765 ops->set_vq_cb(vdpa, idx, &cb);
766 break;
767
768 case VHOST_SET_VRING_NUM:
769 ops->set_vq_num(vdpa, idx, vq->num);
770 break;
771 }
772
773 return r;
774}
775
776static long vhost_vdpa_unlocked_ioctl(struct file *filep,
777 unsigned int cmd, unsigned long arg)
778{
779 struct vhost_vdpa *v = filep->private_data;
780 struct vhost_dev *d = &v->vdev;
781 void __user *argp = (void __user *)arg;
782 u64 __user *featurep = argp;
783 u64 features;
784 long r = 0;
785
786 if (cmd == VHOST_SET_BACKEND_FEATURES) {
787 if (copy_from_user(&features, featurep, sizeof(features)))
788 return -EFAULT;
789 if (features & ~(VHOST_VDPA_BACKEND_FEATURES |
790 BIT_ULL(VHOST_BACKEND_F_DESC_ASID) |
791 BIT_ULL(VHOST_BACKEND_F_IOTLB_PERSIST) |
792 BIT_ULL(VHOST_BACKEND_F_SUSPEND) |
793 BIT_ULL(VHOST_BACKEND_F_RESUME) |
794 BIT_ULL(VHOST_BACKEND_F_ENABLE_AFTER_DRIVER_OK)))
795 return -EOPNOTSUPP;
796 if ((features & BIT_ULL(VHOST_BACKEND_F_SUSPEND)) &&
797 !vhost_vdpa_can_suspend(v))
798 return -EOPNOTSUPP;
799 if ((features & BIT_ULL(VHOST_BACKEND_F_RESUME)) &&
800 !vhost_vdpa_can_resume(v))
801 return -EOPNOTSUPP;
802 if ((features & BIT_ULL(VHOST_BACKEND_F_DESC_ASID)) &&
803 !(features & BIT_ULL(VHOST_BACKEND_F_IOTLB_ASID)))
804 return -EINVAL;
805 if ((features & BIT_ULL(VHOST_BACKEND_F_DESC_ASID)) &&
806 !vhost_vdpa_has_desc_group(v))
807 return -EOPNOTSUPP;
808 if ((features & BIT_ULL(VHOST_BACKEND_F_IOTLB_PERSIST)) &&
809 !vhost_vdpa_has_persistent_map(v))
810 return -EOPNOTSUPP;
811 vhost_set_backend_features(&v->vdev, features);
812 return 0;
813 }
814
815 mutex_lock(&d->mutex);
816
817 switch (cmd) {
818 case VHOST_VDPA_GET_DEVICE_ID:
819 r = vhost_vdpa_get_device_id(v, argp);
820 break;
821 case VHOST_VDPA_GET_STATUS:
822 r = vhost_vdpa_get_status(v, argp);
823 break;
824 case VHOST_VDPA_SET_STATUS:
825 r = vhost_vdpa_set_status(v, argp);
826 break;
827 case VHOST_VDPA_GET_CONFIG:
828 r = vhost_vdpa_get_config(v, argp);
829 break;
830 case VHOST_VDPA_SET_CONFIG:
831 r = vhost_vdpa_set_config(v, argp);
832 break;
833 case VHOST_GET_FEATURES:
834 r = vhost_vdpa_get_features(v, argp);
835 break;
836 case VHOST_SET_FEATURES:
837 r = vhost_vdpa_set_features(v, argp);
838 break;
839 case VHOST_VDPA_GET_VRING_NUM:
840 r = vhost_vdpa_get_vring_num(v, argp);
841 break;
842 case VHOST_VDPA_GET_GROUP_NUM:
843 if (copy_to_user(argp, &v->vdpa->ngroups,
844 sizeof(v->vdpa->ngroups)))
845 r = -EFAULT;
846 break;
847 case VHOST_VDPA_GET_AS_NUM:
848 if (copy_to_user(argp, &v->vdpa->nas, sizeof(v->vdpa->nas)))
849 r = -EFAULT;
850 break;
851 case VHOST_SET_LOG_BASE:
852 case VHOST_SET_LOG_FD:
853 r = -ENOIOCTLCMD;
854 break;
855 case VHOST_VDPA_SET_CONFIG_CALL:
856 r = vhost_vdpa_set_config_call(v, argp);
857 break;
858 case VHOST_GET_BACKEND_FEATURES:
859 features = VHOST_VDPA_BACKEND_FEATURES;
860 if (vhost_vdpa_can_suspend(v))
861 features |= BIT_ULL(VHOST_BACKEND_F_SUSPEND);
862 if (vhost_vdpa_can_resume(v))
863 features |= BIT_ULL(VHOST_BACKEND_F_RESUME);
864 if (vhost_vdpa_has_desc_group(v))
865 features |= BIT_ULL(VHOST_BACKEND_F_DESC_ASID);
866 if (vhost_vdpa_has_persistent_map(v))
867 features |= BIT_ULL(VHOST_BACKEND_F_IOTLB_PERSIST);
868 features |= vhost_vdpa_get_backend_features(v);
869 if (copy_to_user(featurep, &features, sizeof(features)))
870 r = -EFAULT;
871 break;
872 case VHOST_VDPA_GET_IOVA_RANGE:
873 r = vhost_vdpa_get_iova_range(v, argp);
874 break;
875 case VHOST_VDPA_GET_CONFIG_SIZE:
876 r = vhost_vdpa_get_config_size(v, argp);
877 break;
878 case VHOST_VDPA_GET_VQS_COUNT:
879 r = vhost_vdpa_get_vqs_count(v, argp);
880 break;
881 case VHOST_VDPA_SUSPEND:
882 r = vhost_vdpa_suspend(v);
883 break;
884 case VHOST_VDPA_RESUME:
885 r = vhost_vdpa_resume(v);
886 break;
887 default:
888 r = vhost_dev_ioctl(&v->vdev, cmd, argp);
889 if (r == -ENOIOCTLCMD)
890 r = vhost_vdpa_vring_ioctl(v, cmd, argp);
891 break;
892 }
893
894 if (r)
895 goto out;
896
897 switch (cmd) {
898 case VHOST_SET_OWNER:
899 r = vhost_vdpa_bind_mm(v);
900 if (r)
901 vhost_dev_reset_owner(d, NULL);
902 break;
903 }
904out:
905 mutex_unlock(&d->mutex);
906 return r;
907}
908static void vhost_vdpa_general_unmap(struct vhost_vdpa *v,
909 struct vhost_iotlb_map *map, u32 asid)
910{
911 struct vdpa_device *vdpa = v->vdpa;
912 const struct vdpa_config_ops *ops = vdpa->config;
913 if (ops->dma_map) {
914 ops->dma_unmap(vdpa, asid, map->start, map->size);
915 } else if (ops->set_map == NULL) {
916 iommu_unmap(v->domain, map->start, map->size);
917 }
918}
919
920static void vhost_vdpa_pa_unmap(struct vhost_vdpa *v, struct vhost_iotlb *iotlb,
921 u64 start, u64 last, u32 asid)
922{
923 struct vhost_dev *dev = &v->vdev;
924 struct vhost_iotlb_map *map;
925 struct page *page;
926 unsigned long pfn, pinned;
927
928 while ((map = vhost_iotlb_itree_first(iotlb, start, last)) != NULL) {
929 pinned = PFN_DOWN(map->size);
930 for (pfn = PFN_DOWN(map->addr);
931 pinned > 0; pfn++, pinned--) {
932 page = pfn_to_page(pfn);
933 if (map->perm & VHOST_ACCESS_WO)
934 set_page_dirty_lock(page);
935 unpin_user_page(page);
936 }
937 atomic64_sub(PFN_DOWN(map->size), &dev->mm->pinned_vm);
938 vhost_vdpa_general_unmap(v, map, asid);
939 vhost_iotlb_map_free(iotlb, map);
940 }
941}
942
943static void vhost_vdpa_va_unmap(struct vhost_vdpa *v, struct vhost_iotlb *iotlb,
944 u64 start, u64 last, u32 asid)
945{
946 struct vhost_iotlb_map *map;
947 struct vdpa_map_file *map_file;
948
949 while ((map = vhost_iotlb_itree_first(iotlb, start, last)) != NULL) {
950 map_file = (struct vdpa_map_file *)map->opaque;
951 fput(map_file->file);
952 kfree(map_file);
953 vhost_vdpa_general_unmap(v, map, asid);
954 vhost_iotlb_map_free(iotlb, map);
955 }
956}
957
958static void vhost_vdpa_iotlb_unmap(struct vhost_vdpa *v,
959 struct vhost_iotlb *iotlb, u64 start,
960 u64 last, u32 asid)
961{
962 struct vdpa_device *vdpa = v->vdpa;
963
964 if (vdpa->use_va)
965 return vhost_vdpa_va_unmap(v, iotlb, start, last, asid);
966
967 return vhost_vdpa_pa_unmap(v, iotlb, start, last, asid);
968}
969
970static int perm_to_iommu_flags(u32 perm)
971{
972 int flags = 0;
973
974 switch (perm) {
975 case VHOST_ACCESS_WO:
976 flags |= IOMMU_WRITE;
977 break;
978 case VHOST_ACCESS_RO:
979 flags |= IOMMU_READ;
980 break;
981 case VHOST_ACCESS_RW:
982 flags |= (IOMMU_WRITE | IOMMU_READ);
983 break;
984 default:
985 WARN(1, "invalidate vhost IOTLB permission\n");
986 break;
987 }
988
989 return flags | IOMMU_CACHE;
990}
991
992static int vhost_vdpa_map(struct vhost_vdpa *v, struct vhost_iotlb *iotlb,
993 u64 iova, u64 size, u64 pa, u32 perm, void *opaque)
994{
995 struct vhost_dev *dev = &v->vdev;
996 struct vdpa_device *vdpa = v->vdpa;
997 const struct vdpa_config_ops *ops = vdpa->config;
998 u32 asid = iotlb_to_asid(iotlb);
999 int r = 0;
1000
1001 r = vhost_iotlb_add_range_ctx(iotlb, iova, iova + size - 1,
1002 pa, perm, opaque);
1003 if (r)
1004 return r;
1005
1006 if (ops->dma_map) {
1007 r = ops->dma_map(vdpa, asid, iova, size, pa, perm, opaque);
1008 } else if (ops->set_map) {
1009 if (!v->in_batch)
1010 r = ops->set_map(vdpa, asid, iotlb);
1011 } else {
1012 r = iommu_map(v->domain, iova, pa, size,
1013 perm_to_iommu_flags(perm),
1014 GFP_KERNEL_ACCOUNT);
1015 }
1016 if (r) {
1017 vhost_iotlb_del_range(iotlb, iova, iova + size - 1);
1018 return r;
1019 }
1020
1021 if (!vdpa->use_va)
1022 atomic64_add(PFN_DOWN(size), &dev->mm->pinned_vm);
1023
1024 return 0;
1025}
1026
1027static void vhost_vdpa_unmap(struct vhost_vdpa *v,
1028 struct vhost_iotlb *iotlb,
1029 u64 iova, u64 size)
1030{
1031 struct vdpa_device *vdpa = v->vdpa;
1032 const struct vdpa_config_ops *ops = vdpa->config;
1033 u32 asid = iotlb_to_asid(iotlb);
1034
1035 vhost_vdpa_iotlb_unmap(v, iotlb, iova, iova + size - 1, asid);
1036
1037 if (ops->set_map) {
1038 if (!v->in_batch)
1039 ops->set_map(vdpa, asid, iotlb);
1040 }
1041
1042}
1043
1044static int vhost_vdpa_va_map(struct vhost_vdpa *v,
1045 struct vhost_iotlb *iotlb,
1046 u64 iova, u64 size, u64 uaddr, u32 perm)
1047{
1048 struct vhost_dev *dev = &v->vdev;
1049 u64 offset, map_size, map_iova = iova;
1050 struct vdpa_map_file *map_file;
1051 struct vm_area_struct *vma;
1052 int ret = 0;
1053
1054 mmap_read_lock(dev->mm);
1055
1056 while (size) {
1057 vma = find_vma(dev->mm, uaddr);
1058 if (!vma) {
1059 ret = -EINVAL;
1060 break;
1061 }
1062 map_size = min(size, vma->vm_end - uaddr);
1063 if (!(vma->vm_file && (vma->vm_flags & VM_SHARED) &&
1064 !(vma->vm_flags & (VM_IO | VM_PFNMAP))))
1065 goto next;
1066
1067 map_file = kzalloc(sizeof(*map_file), GFP_KERNEL);
1068 if (!map_file) {
1069 ret = -ENOMEM;
1070 break;
1071 }
1072 offset = (vma->vm_pgoff << PAGE_SHIFT) + uaddr - vma->vm_start;
1073 map_file->offset = offset;
1074 map_file->file = get_file(vma->vm_file);
1075 ret = vhost_vdpa_map(v, iotlb, map_iova, map_size, uaddr,
1076 perm, map_file);
1077 if (ret) {
1078 fput(map_file->file);
1079 kfree(map_file);
1080 break;
1081 }
1082next:
1083 size -= map_size;
1084 uaddr += map_size;
1085 map_iova += map_size;
1086 }
1087 if (ret)
1088 vhost_vdpa_unmap(v, iotlb, iova, map_iova - iova);
1089
1090 mmap_read_unlock(dev->mm);
1091
1092 return ret;
1093}
1094
1095static int vhost_vdpa_pa_map(struct vhost_vdpa *v,
1096 struct vhost_iotlb *iotlb,
1097 u64 iova, u64 size, u64 uaddr, u32 perm)
1098{
1099 struct vhost_dev *dev = &v->vdev;
1100 struct page **page_list;
1101 unsigned long list_size = PAGE_SIZE / sizeof(struct page *);
1102 unsigned int gup_flags = FOLL_LONGTERM;
1103 unsigned long npages, cur_base, map_pfn, last_pfn = 0;
1104 unsigned long lock_limit, sz2pin, nchunks, i;
1105 u64 start = iova;
1106 long pinned;
1107 int ret = 0;
1108
1109 /* Limit the use of memory for bookkeeping */
1110 page_list = (struct page **) __get_free_page(GFP_KERNEL);
1111 if (!page_list)
1112 return -ENOMEM;
1113
1114 if (perm & VHOST_ACCESS_WO)
1115 gup_flags |= FOLL_WRITE;
1116
1117 npages = PFN_UP(size + (iova & ~PAGE_MASK));
1118 if (!npages) {
1119 ret = -EINVAL;
1120 goto free;
1121 }
1122
1123 mmap_read_lock(dev->mm);
1124
1125 lock_limit = PFN_DOWN(rlimit(RLIMIT_MEMLOCK));
1126 if (npages + atomic64_read(&dev->mm->pinned_vm) > lock_limit) {
1127 ret = -ENOMEM;
1128 goto unlock;
1129 }
1130
1131 cur_base = uaddr & PAGE_MASK;
1132 iova &= PAGE_MASK;
1133 nchunks = 0;
1134
1135 while (npages) {
1136 sz2pin = min_t(unsigned long, npages, list_size);
1137 pinned = pin_user_pages(cur_base, sz2pin,
1138 gup_flags, page_list);
1139 if (sz2pin != pinned) {
1140 if (pinned < 0) {
1141 ret = pinned;
1142 } else {
1143 unpin_user_pages(page_list, pinned);
1144 ret = -ENOMEM;
1145 }
1146 goto out;
1147 }
1148 nchunks++;
1149
1150 if (!last_pfn)
1151 map_pfn = page_to_pfn(page_list[0]);
1152
1153 for (i = 0; i < pinned; i++) {
1154 unsigned long this_pfn = page_to_pfn(page_list[i]);
1155 u64 csize;
1156
1157 if (last_pfn && (this_pfn != last_pfn + 1)) {
1158 /* Pin a contiguous chunk of memory */
1159 csize = PFN_PHYS(last_pfn - map_pfn + 1);
1160 ret = vhost_vdpa_map(v, iotlb, iova, csize,
1161 PFN_PHYS(map_pfn),
1162 perm, NULL);
1163 if (ret) {
1164 /*
1165 * Unpin the pages that are left unmapped
1166 * from this point on in the current
1167 * page_list. The remaining outstanding
1168 * ones which may stride across several
1169 * chunks will be covered in the common
1170 * error path subsequently.
1171 */
1172 unpin_user_pages(&page_list[i],
1173 pinned - i);
1174 goto out;
1175 }
1176
1177 map_pfn = this_pfn;
1178 iova += csize;
1179 nchunks = 0;
1180 }
1181
1182 last_pfn = this_pfn;
1183 }
1184
1185 cur_base += PFN_PHYS(pinned);
1186 npages -= pinned;
1187 }
1188
1189 /* Pin the rest chunk */
1190 ret = vhost_vdpa_map(v, iotlb, iova, PFN_PHYS(last_pfn - map_pfn + 1),
1191 PFN_PHYS(map_pfn), perm, NULL);
1192out:
1193 if (ret) {
1194 if (nchunks) {
1195 unsigned long pfn;
1196
1197 /*
1198 * Unpin the outstanding pages which are yet to be
1199 * mapped but haven't due to vdpa_map() or
1200 * pin_user_pages() failure.
1201 *
1202 * Mapped pages are accounted in vdpa_map(), hence
1203 * the corresponding unpinning will be handled by
1204 * vdpa_unmap().
1205 */
1206 WARN_ON(!last_pfn);
1207 for (pfn = map_pfn; pfn <= last_pfn; pfn++)
1208 unpin_user_page(pfn_to_page(pfn));
1209 }
1210 vhost_vdpa_unmap(v, iotlb, start, size);
1211 }
1212unlock:
1213 mmap_read_unlock(dev->mm);
1214free:
1215 free_page((unsigned long)page_list);
1216 return ret;
1217
1218}
1219
1220static int vhost_vdpa_process_iotlb_update(struct vhost_vdpa *v,
1221 struct vhost_iotlb *iotlb,
1222 struct vhost_iotlb_msg *msg)
1223{
1224 struct vdpa_device *vdpa = v->vdpa;
1225
1226 if (msg->iova < v->range.first || !msg->size ||
1227 msg->iova > U64_MAX - msg->size + 1 ||
1228 msg->iova + msg->size - 1 > v->range.last)
1229 return -EINVAL;
1230
1231 if (vhost_iotlb_itree_first(iotlb, msg->iova,
1232 msg->iova + msg->size - 1))
1233 return -EEXIST;
1234
1235 if (vdpa->use_va)
1236 return vhost_vdpa_va_map(v, iotlb, msg->iova, msg->size,
1237 msg->uaddr, msg->perm);
1238
1239 return vhost_vdpa_pa_map(v, iotlb, msg->iova, msg->size, msg->uaddr,
1240 msg->perm);
1241}
1242
1243static int vhost_vdpa_process_iotlb_msg(struct vhost_dev *dev, u32 asid,
1244 struct vhost_iotlb_msg *msg)
1245{
1246 struct vhost_vdpa *v = container_of(dev, struct vhost_vdpa, vdev);
1247 struct vdpa_device *vdpa = v->vdpa;
1248 const struct vdpa_config_ops *ops = vdpa->config;
1249 struct vhost_iotlb *iotlb = NULL;
1250 struct vhost_vdpa_as *as = NULL;
1251 int r = 0;
1252
1253 mutex_lock(&dev->mutex);
1254
1255 r = vhost_dev_check_owner(dev);
1256 if (r)
1257 goto unlock;
1258
1259 if (msg->type == VHOST_IOTLB_UPDATE ||
1260 msg->type == VHOST_IOTLB_BATCH_BEGIN) {
1261 as = vhost_vdpa_find_alloc_as(v, asid);
1262 if (!as) {
1263 dev_err(&v->dev, "can't find and alloc asid %d\n",
1264 asid);
1265 r = -EINVAL;
1266 goto unlock;
1267 }
1268 iotlb = &as->iotlb;
1269 } else
1270 iotlb = asid_to_iotlb(v, asid);
1271
1272 if ((v->in_batch && v->batch_asid != asid) || !iotlb) {
1273 if (v->in_batch && v->batch_asid != asid) {
1274 dev_info(&v->dev, "batch id %d asid %d\n",
1275 v->batch_asid, asid);
1276 }
1277 if (!iotlb)
1278 dev_err(&v->dev, "no iotlb for asid %d\n", asid);
1279 r = -EINVAL;
1280 goto unlock;
1281 }
1282
1283 switch (msg->type) {
1284 case VHOST_IOTLB_UPDATE:
1285 r = vhost_vdpa_process_iotlb_update(v, iotlb, msg);
1286 break;
1287 case VHOST_IOTLB_INVALIDATE:
1288 vhost_vdpa_unmap(v, iotlb, msg->iova, msg->size);
1289 break;
1290 case VHOST_IOTLB_BATCH_BEGIN:
1291 v->batch_asid = asid;
1292 v->in_batch = true;
1293 break;
1294 case VHOST_IOTLB_BATCH_END:
1295 if (v->in_batch && ops->set_map)
1296 ops->set_map(vdpa, asid, iotlb);
1297 v->in_batch = false;
1298 break;
1299 default:
1300 r = -EINVAL;
1301 break;
1302 }
1303unlock:
1304 mutex_unlock(&dev->mutex);
1305
1306 return r;
1307}
1308
1309static ssize_t vhost_vdpa_chr_write_iter(struct kiocb *iocb,
1310 struct iov_iter *from)
1311{
1312 struct file *file = iocb->ki_filp;
1313 struct vhost_vdpa *v = file->private_data;
1314 struct vhost_dev *dev = &v->vdev;
1315
1316 return vhost_chr_write_iter(dev, from);
1317}
1318
1319static int vhost_vdpa_alloc_domain(struct vhost_vdpa *v)
1320{
1321 struct vdpa_device *vdpa = v->vdpa;
1322 const struct vdpa_config_ops *ops = vdpa->config;
1323 struct device *dma_dev = vdpa_get_dma_dev(vdpa);
1324 int ret;
1325
1326 /* Device want to do DMA by itself */
1327 if (ops->set_map || ops->dma_map)
1328 return 0;
1329
1330 if (!device_iommu_capable(dma_dev, IOMMU_CAP_CACHE_COHERENCY)) {
1331 dev_warn_once(&v->dev,
1332 "Failed to allocate domain, device is not IOMMU cache coherent capable\n");
1333 return -ENOTSUPP;
1334 }
1335
1336 v->domain = iommu_paging_domain_alloc(dma_dev);
1337 if (IS_ERR(v->domain)) {
1338 ret = PTR_ERR(v->domain);
1339 v->domain = NULL;
1340 return ret;
1341 }
1342
1343 ret = iommu_attach_device(v->domain, dma_dev);
1344 if (ret)
1345 goto err_attach;
1346
1347 return 0;
1348
1349err_attach:
1350 iommu_domain_free(v->domain);
1351 v->domain = NULL;
1352 return ret;
1353}
1354
1355static void vhost_vdpa_free_domain(struct vhost_vdpa *v)
1356{
1357 struct vdpa_device *vdpa = v->vdpa;
1358 struct device *dma_dev = vdpa_get_dma_dev(vdpa);
1359
1360 if (v->domain) {
1361 iommu_detach_device(v->domain, dma_dev);
1362 iommu_domain_free(v->domain);
1363 }
1364
1365 v->domain = NULL;
1366}
1367
1368static void vhost_vdpa_set_iova_range(struct vhost_vdpa *v)
1369{
1370 struct vdpa_iova_range *range = &v->range;
1371 struct vdpa_device *vdpa = v->vdpa;
1372 const struct vdpa_config_ops *ops = vdpa->config;
1373
1374 if (ops->get_iova_range) {
1375 *range = ops->get_iova_range(vdpa);
1376 } else if (v->domain && v->domain->geometry.force_aperture) {
1377 range->first = v->domain->geometry.aperture_start;
1378 range->last = v->domain->geometry.aperture_end;
1379 } else {
1380 range->first = 0;
1381 range->last = ULLONG_MAX;
1382 }
1383}
1384
1385static void vhost_vdpa_cleanup(struct vhost_vdpa *v)
1386{
1387 struct vhost_vdpa_as *as;
1388 u32 asid;
1389
1390 for (asid = 0; asid < v->vdpa->nas; asid++) {
1391 as = asid_to_as(v, asid);
1392 if (as)
1393 vhost_vdpa_remove_as(v, asid);
1394 }
1395
1396 vhost_vdpa_free_domain(v);
1397 vhost_dev_cleanup(&v->vdev);
1398 kfree(v->vdev.vqs);
1399 v->vdev.vqs = NULL;
1400}
1401
1402static int vhost_vdpa_open(struct inode *inode, struct file *filep)
1403{
1404 struct vhost_vdpa *v;
1405 struct vhost_dev *dev;
1406 struct vhost_virtqueue **vqs;
1407 int r, opened;
1408 u32 i, nvqs;
1409
1410 v = container_of(inode->i_cdev, struct vhost_vdpa, cdev);
1411
1412 opened = atomic_cmpxchg(&v->opened, 0, 1);
1413 if (opened)
1414 return -EBUSY;
1415
1416 nvqs = v->nvqs;
1417 r = vhost_vdpa_reset(v);
1418 if (r)
1419 goto err;
1420
1421 vqs = kmalloc_array(nvqs, sizeof(*vqs), GFP_KERNEL);
1422 if (!vqs) {
1423 r = -ENOMEM;
1424 goto err;
1425 }
1426
1427 dev = &v->vdev;
1428 for (i = 0; i < nvqs; i++) {
1429 vqs[i] = &v->vqs[i];
1430 vqs[i]->handle_kick = handle_vq_kick;
1431 vqs[i]->call_ctx.ctx = NULL;
1432 }
1433 vhost_dev_init(dev, vqs, nvqs, 0, 0, 0, false,
1434 vhost_vdpa_process_iotlb_msg);
1435
1436 r = vhost_vdpa_alloc_domain(v);
1437 if (r)
1438 goto err_alloc_domain;
1439
1440 vhost_vdpa_set_iova_range(v);
1441
1442 filep->private_data = v;
1443
1444 return 0;
1445
1446err_alloc_domain:
1447 vhost_vdpa_cleanup(v);
1448err:
1449 atomic_dec(&v->opened);
1450 return r;
1451}
1452
1453static void vhost_vdpa_clean_irq(struct vhost_vdpa *v)
1454{
1455 u32 i;
1456
1457 for (i = 0; i < v->nvqs; i++)
1458 vhost_vdpa_unsetup_vq_irq(v, i);
1459}
1460
1461static int vhost_vdpa_release(struct inode *inode, struct file *filep)
1462{
1463 struct vhost_vdpa *v = filep->private_data;
1464 struct vhost_dev *d = &v->vdev;
1465
1466 mutex_lock(&d->mutex);
1467 filep->private_data = NULL;
1468 vhost_vdpa_clean_irq(v);
1469 vhost_vdpa_reset(v);
1470 vhost_dev_stop(&v->vdev);
1471 vhost_vdpa_unbind_mm(v);
1472 vhost_vdpa_config_put(v);
1473 vhost_vdpa_cleanup(v);
1474 mutex_unlock(&d->mutex);
1475
1476 atomic_dec(&v->opened);
1477 complete(&v->completion);
1478
1479 return 0;
1480}
1481
1482#ifdef CONFIG_MMU
1483static vm_fault_t vhost_vdpa_fault(struct vm_fault *vmf)
1484{
1485 struct vhost_vdpa *v = vmf->vma->vm_file->private_data;
1486 struct vdpa_device *vdpa = v->vdpa;
1487 const struct vdpa_config_ops *ops = vdpa->config;
1488 struct vdpa_notification_area notify;
1489 struct vm_area_struct *vma = vmf->vma;
1490 u16 index = vma->vm_pgoff;
1491
1492 notify = ops->get_vq_notification(vdpa, index);
1493
1494 return vmf_insert_pfn(vma, vmf->address & PAGE_MASK, PFN_DOWN(notify.addr));
1495}
1496
1497static const struct vm_operations_struct vhost_vdpa_vm_ops = {
1498 .fault = vhost_vdpa_fault,
1499};
1500
1501static int vhost_vdpa_mmap(struct file *file, struct vm_area_struct *vma)
1502{
1503 struct vhost_vdpa *v = vma->vm_file->private_data;
1504 struct vdpa_device *vdpa = v->vdpa;
1505 const struct vdpa_config_ops *ops = vdpa->config;
1506 struct vdpa_notification_area notify;
1507 unsigned long index = vma->vm_pgoff;
1508
1509 if (vma->vm_end - vma->vm_start != PAGE_SIZE)
1510 return -EINVAL;
1511 if ((vma->vm_flags & VM_SHARED) == 0)
1512 return -EINVAL;
1513 if (vma->vm_flags & VM_READ)
1514 return -EINVAL;
1515 if (index > 65535)
1516 return -EINVAL;
1517 if (!ops->get_vq_notification)
1518 return -ENOTSUPP;
1519
1520 /* To be safe and easily modelled by userspace, We only
1521 * support the doorbell which sits on the page boundary and
1522 * does not share the page with other registers.
1523 */
1524 notify = ops->get_vq_notification(vdpa, index);
1525 if (notify.addr & (PAGE_SIZE - 1))
1526 return -EINVAL;
1527 if (vma->vm_end - vma->vm_start != notify.size)
1528 return -ENOTSUPP;
1529
1530 vm_flags_set(vma, VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP);
1531 vma->vm_ops = &vhost_vdpa_vm_ops;
1532 return 0;
1533}
1534#endif /* CONFIG_MMU */
1535
1536static const struct file_operations vhost_vdpa_fops = {
1537 .owner = THIS_MODULE,
1538 .open = vhost_vdpa_open,
1539 .release = vhost_vdpa_release,
1540 .write_iter = vhost_vdpa_chr_write_iter,
1541 .unlocked_ioctl = vhost_vdpa_unlocked_ioctl,
1542#ifdef CONFIG_MMU
1543 .mmap = vhost_vdpa_mmap,
1544#endif /* CONFIG_MMU */
1545 .compat_ioctl = compat_ptr_ioctl,
1546};
1547
1548static void vhost_vdpa_release_dev(struct device *device)
1549{
1550 struct vhost_vdpa *v =
1551 container_of(device, struct vhost_vdpa, dev);
1552
1553 ida_free(&vhost_vdpa_ida, v->minor);
1554 kfree(v->vqs);
1555 kfree(v);
1556}
1557
1558static int vhost_vdpa_probe(struct vdpa_device *vdpa)
1559{
1560 const struct vdpa_config_ops *ops = vdpa->config;
1561 struct vhost_vdpa *v;
1562 int minor;
1563 int i, r;
1564
1565 /* We can't support platform IOMMU device with more than 1
1566 * group or as
1567 */
1568 if (!ops->set_map && !ops->dma_map &&
1569 (vdpa->ngroups > 1 || vdpa->nas > 1))
1570 return -EOPNOTSUPP;
1571
1572 v = kzalloc(sizeof(*v), GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1573 if (!v)
1574 return -ENOMEM;
1575
1576 minor = ida_alloc_max(&vhost_vdpa_ida, VHOST_VDPA_DEV_MAX - 1,
1577 GFP_KERNEL);
1578 if (minor < 0) {
1579 kfree(v);
1580 return minor;
1581 }
1582
1583 atomic_set(&v->opened, 0);
1584 v->minor = minor;
1585 v->vdpa = vdpa;
1586 v->nvqs = vdpa->nvqs;
1587 v->virtio_id = ops->get_device_id(vdpa);
1588
1589 device_initialize(&v->dev);
1590 v->dev.release = vhost_vdpa_release_dev;
1591 v->dev.parent = &vdpa->dev;
1592 v->dev.devt = MKDEV(MAJOR(vhost_vdpa_major), minor);
1593 v->vqs = kmalloc_array(v->nvqs, sizeof(struct vhost_virtqueue),
1594 GFP_KERNEL);
1595 if (!v->vqs) {
1596 r = -ENOMEM;
1597 goto err;
1598 }
1599
1600 r = dev_set_name(&v->dev, "vhost-vdpa-%u", minor);
1601 if (r)
1602 goto err;
1603
1604 cdev_init(&v->cdev, &vhost_vdpa_fops);
1605 v->cdev.owner = THIS_MODULE;
1606
1607 r = cdev_device_add(&v->cdev, &v->dev);
1608 if (r)
1609 goto err;
1610
1611 init_completion(&v->completion);
1612 vdpa_set_drvdata(vdpa, v);
1613
1614 for (i = 0; i < VHOST_VDPA_IOTLB_BUCKETS; i++)
1615 INIT_HLIST_HEAD(&v->as[i]);
1616
1617 return 0;
1618
1619err:
1620 put_device(&v->dev);
1621 return r;
1622}
1623
1624static void vhost_vdpa_remove(struct vdpa_device *vdpa)
1625{
1626 struct vhost_vdpa *v = vdpa_get_drvdata(vdpa);
1627 int opened;
1628
1629 cdev_device_del(&v->cdev, &v->dev);
1630
1631 do {
1632 opened = atomic_cmpxchg(&v->opened, 0, 1);
1633 if (!opened)
1634 break;
1635 wait_for_completion(&v->completion);
1636 } while (1);
1637
1638 put_device(&v->dev);
1639}
1640
1641static struct vdpa_driver vhost_vdpa_driver = {
1642 .driver = {
1643 .name = "vhost_vdpa",
1644 },
1645 .probe = vhost_vdpa_probe,
1646 .remove = vhost_vdpa_remove,
1647};
1648
1649static int __init vhost_vdpa_init(void)
1650{
1651 int r;
1652
1653 r = alloc_chrdev_region(&vhost_vdpa_major, 0, VHOST_VDPA_DEV_MAX,
1654 "vhost-vdpa");
1655 if (r)
1656 goto err_alloc_chrdev;
1657
1658 r = vdpa_register_driver(&vhost_vdpa_driver);
1659 if (r)
1660 goto err_vdpa_register_driver;
1661
1662 return 0;
1663
1664err_vdpa_register_driver:
1665 unregister_chrdev_region(vhost_vdpa_major, VHOST_VDPA_DEV_MAX);
1666err_alloc_chrdev:
1667 return r;
1668}
1669module_init(vhost_vdpa_init);
1670
1671static void __exit vhost_vdpa_exit(void)
1672{
1673 vdpa_unregister_driver(&vhost_vdpa_driver);
1674 unregister_chrdev_region(vhost_vdpa_major, VHOST_VDPA_DEV_MAX);
1675}
1676module_exit(vhost_vdpa_exit);
1677
1678MODULE_VERSION("0.0.1");
1679MODULE_LICENSE("GPL v2");
1680MODULE_AUTHOR("Intel Corporation");
1681MODULE_DESCRIPTION("vDPA-based vhost backend for virtio");
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/iommu.h>
20#include <linux/uuid.h>
21#include <linux/vdpa.h>
22#include <linux/nospec.h>
23#include <linux/vhost.h>
24#include <linux/virtio_net.h>
25#include <linux/kernel.h>
26
27#include "vhost.h"
28
29enum {
30 VHOST_VDPA_BACKEND_FEATURES =
31 (1ULL << VHOST_BACKEND_F_IOTLB_MSG_V2) |
32 (1ULL << VHOST_BACKEND_F_IOTLB_BATCH),
33};
34
35#define VHOST_VDPA_DEV_MAX (1U << MINORBITS)
36
37struct vhost_vdpa {
38 struct vhost_dev vdev;
39 struct iommu_domain *domain;
40 struct vhost_virtqueue *vqs;
41 struct completion completion;
42 struct vdpa_device *vdpa;
43 struct device dev;
44 struct cdev cdev;
45 atomic_t opened;
46 int nvqs;
47 int virtio_id;
48 int minor;
49 struct eventfd_ctx *config_ctx;
50 int in_batch;
51};
52
53static DEFINE_IDA(vhost_vdpa_ida);
54
55static dev_t vhost_vdpa_major;
56
57static void handle_vq_kick(struct vhost_work *work)
58{
59 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
60 poll.work);
61 struct vhost_vdpa *v = container_of(vq->dev, struct vhost_vdpa, vdev);
62 const struct vdpa_config_ops *ops = v->vdpa->config;
63
64 ops->kick_vq(v->vdpa, vq - v->vqs);
65}
66
67static irqreturn_t vhost_vdpa_virtqueue_cb(void *private)
68{
69 struct vhost_virtqueue *vq = private;
70 struct eventfd_ctx *call_ctx = vq->call_ctx.ctx;
71
72 if (call_ctx)
73 eventfd_signal(call_ctx, 1);
74
75 return IRQ_HANDLED;
76}
77
78static irqreturn_t vhost_vdpa_config_cb(void *private)
79{
80 struct vhost_vdpa *v = private;
81 struct eventfd_ctx *config_ctx = v->config_ctx;
82
83 if (config_ctx)
84 eventfd_signal(config_ctx, 1);
85
86 return IRQ_HANDLED;
87}
88
89static void vhost_vdpa_setup_vq_irq(struct vhost_vdpa *v, u16 qid)
90{
91 struct vhost_virtqueue *vq = &v->vqs[qid];
92 const struct vdpa_config_ops *ops = v->vdpa->config;
93 struct vdpa_device *vdpa = v->vdpa;
94 int ret, irq;
95
96 if (!ops->get_vq_irq)
97 return;
98
99 irq = ops->get_vq_irq(vdpa, qid);
100 spin_lock(&vq->call_ctx.ctx_lock);
101 irq_bypass_unregister_producer(&vq->call_ctx.producer);
102 if (!vq->call_ctx.ctx || irq < 0) {
103 spin_unlock(&vq->call_ctx.ctx_lock);
104 return;
105 }
106
107 vq->call_ctx.producer.token = vq->call_ctx.ctx;
108 vq->call_ctx.producer.irq = irq;
109 ret = irq_bypass_register_producer(&vq->call_ctx.producer);
110 spin_unlock(&vq->call_ctx.ctx_lock);
111}
112
113static void vhost_vdpa_unsetup_vq_irq(struct vhost_vdpa *v, u16 qid)
114{
115 struct vhost_virtqueue *vq = &v->vqs[qid];
116
117 spin_lock(&vq->call_ctx.ctx_lock);
118 irq_bypass_unregister_producer(&vq->call_ctx.producer);
119 spin_unlock(&vq->call_ctx.ctx_lock);
120}
121
122static void vhost_vdpa_reset(struct vhost_vdpa *v)
123{
124 struct vdpa_device *vdpa = v->vdpa;
125
126 vdpa_reset(vdpa);
127 v->in_batch = 0;
128}
129
130static long vhost_vdpa_get_device_id(struct vhost_vdpa *v, u8 __user *argp)
131{
132 struct vdpa_device *vdpa = v->vdpa;
133 const struct vdpa_config_ops *ops = vdpa->config;
134 u32 device_id;
135
136 device_id = ops->get_device_id(vdpa);
137
138 if (copy_to_user(argp, &device_id, sizeof(device_id)))
139 return -EFAULT;
140
141 return 0;
142}
143
144static long vhost_vdpa_get_status(struct vhost_vdpa *v, u8 __user *statusp)
145{
146 struct vdpa_device *vdpa = v->vdpa;
147 const struct vdpa_config_ops *ops = vdpa->config;
148 u8 status;
149
150 status = ops->get_status(vdpa);
151
152 if (copy_to_user(statusp, &status, sizeof(status)))
153 return -EFAULT;
154
155 return 0;
156}
157
158static long vhost_vdpa_set_status(struct vhost_vdpa *v, u8 __user *statusp)
159{
160 struct vdpa_device *vdpa = v->vdpa;
161 const struct vdpa_config_ops *ops = vdpa->config;
162 u8 status, status_old;
163 int nvqs = v->nvqs;
164 u16 i;
165
166 if (copy_from_user(&status, statusp, sizeof(status)))
167 return -EFAULT;
168
169 status_old = ops->get_status(vdpa);
170
171 /*
172 * Userspace shouldn't remove status bits unless reset the
173 * status to 0.
174 */
175 if (status != 0 && (ops->get_status(vdpa) & ~status) != 0)
176 return -EINVAL;
177
178 ops->set_status(vdpa, status);
179
180 if ((status & VIRTIO_CONFIG_S_DRIVER_OK) && !(status_old & VIRTIO_CONFIG_S_DRIVER_OK))
181 for (i = 0; i < nvqs; i++)
182 vhost_vdpa_setup_vq_irq(v, i);
183
184 if ((status_old & VIRTIO_CONFIG_S_DRIVER_OK) && !(status & VIRTIO_CONFIG_S_DRIVER_OK))
185 for (i = 0; i < nvqs; i++)
186 vhost_vdpa_unsetup_vq_irq(v, i);
187
188 return 0;
189}
190
191static int vhost_vdpa_config_validate(struct vhost_vdpa *v,
192 struct vhost_vdpa_config *c)
193{
194 long size = 0;
195
196 switch (v->virtio_id) {
197 case VIRTIO_ID_NET:
198 size = sizeof(struct virtio_net_config);
199 break;
200 }
201
202 if (c->len == 0)
203 return -EINVAL;
204
205 if (c->len > size - c->off)
206 return -E2BIG;
207
208 return 0;
209}
210
211static long vhost_vdpa_get_config(struct vhost_vdpa *v,
212 struct vhost_vdpa_config __user *c)
213{
214 struct vdpa_device *vdpa = v->vdpa;
215 struct vhost_vdpa_config config;
216 unsigned long size = offsetof(struct vhost_vdpa_config, buf);
217 u8 *buf;
218
219 if (copy_from_user(&config, c, size))
220 return -EFAULT;
221 if (vhost_vdpa_config_validate(v, &config))
222 return -EINVAL;
223 buf = kvzalloc(config.len, GFP_KERNEL);
224 if (!buf)
225 return -ENOMEM;
226
227 vdpa_get_config(vdpa, config.off, buf, config.len);
228
229 if (copy_to_user(c->buf, buf, config.len)) {
230 kvfree(buf);
231 return -EFAULT;
232 }
233
234 kvfree(buf);
235 return 0;
236}
237
238static long vhost_vdpa_set_config(struct vhost_vdpa *v,
239 struct vhost_vdpa_config __user *c)
240{
241 struct vdpa_device *vdpa = v->vdpa;
242 const struct vdpa_config_ops *ops = vdpa->config;
243 struct vhost_vdpa_config config;
244 unsigned long size = offsetof(struct vhost_vdpa_config, buf);
245 u8 *buf;
246
247 if (copy_from_user(&config, c, size))
248 return -EFAULT;
249 if (vhost_vdpa_config_validate(v, &config))
250 return -EINVAL;
251 buf = kvzalloc(config.len, GFP_KERNEL);
252 if (!buf)
253 return -ENOMEM;
254
255 if (copy_from_user(buf, c->buf, config.len)) {
256 kvfree(buf);
257 return -EFAULT;
258 }
259
260 ops->set_config(vdpa, config.off, buf, config.len);
261
262 kvfree(buf);
263 return 0;
264}
265
266static long vhost_vdpa_get_features(struct vhost_vdpa *v, u64 __user *featurep)
267{
268 struct vdpa_device *vdpa = v->vdpa;
269 const struct vdpa_config_ops *ops = vdpa->config;
270 u64 features;
271
272 features = ops->get_features(vdpa);
273
274 if (copy_to_user(featurep, &features, sizeof(features)))
275 return -EFAULT;
276
277 return 0;
278}
279
280static long vhost_vdpa_set_features(struct vhost_vdpa *v, u64 __user *featurep)
281{
282 struct vdpa_device *vdpa = v->vdpa;
283 const struct vdpa_config_ops *ops = vdpa->config;
284 u64 features;
285
286 /*
287 * It's not allowed to change the features after they have
288 * been negotiated.
289 */
290 if (ops->get_status(vdpa) & VIRTIO_CONFIG_S_FEATURES_OK)
291 return -EBUSY;
292
293 if (copy_from_user(&features, featurep, sizeof(features)))
294 return -EFAULT;
295
296 if (vdpa_set_features(vdpa, features))
297 return -EINVAL;
298
299 return 0;
300}
301
302static long vhost_vdpa_get_vring_num(struct vhost_vdpa *v, u16 __user *argp)
303{
304 struct vdpa_device *vdpa = v->vdpa;
305 const struct vdpa_config_ops *ops = vdpa->config;
306 u16 num;
307
308 num = ops->get_vq_num_max(vdpa);
309
310 if (copy_to_user(argp, &num, sizeof(num)))
311 return -EFAULT;
312
313 return 0;
314}
315
316static void vhost_vdpa_config_put(struct vhost_vdpa *v)
317{
318 if (v->config_ctx)
319 eventfd_ctx_put(v->config_ctx);
320}
321
322static long vhost_vdpa_set_config_call(struct vhost_vdpa *v, u32 __user *argp)
323{
324 struct vdpa_callback cb;
325 int fd;
326 struct eventfd_ctx *ctx;
327
328 cb.callback = vhost_vdpa_config_cb;
329 cb.private = v->vdpa;
330 if (copy_from_user(&fd, argp, sizeof(fd)))
331 return -EFAULT;
332
333 ctx = fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(fd);
334 swap(ctx, v->config_ctx);
335
336 if (!IS_ERR_OR_NULL(ctx))
337 eventfd_ctx_put(ctx);
338
339 if (IS_ERR(v->config_ctx))
340 return PTR_ERR(v->config_ctx);
341
342 v->vdpa->config->set_config_cb(v->vdpa, &cb);
343
344 return 0;
345}
346
347static long vhost_vdpa_vring_ioctl(struct vhost_vdpa *v, unsigned int cmd,
348 void __user *argp)
349{
350 struct vdpa_device *vdpa = v->vdpa;
351 const struct vdpa_config_ops *ops = vdpa->config;
352 struct vdpa_vq_state vq_state;
353 struct vdpa_callback cb;
354 struct vhost_virtqueue *vq;
355 struct vhost_vring_state s;
356 u32 idx;
357 long r;
358
359 r = get_user(idx, (u32 __user *)argp);
360 if (r < 0)
361 return r;
362
363 if (idx >= v->nvqs)
364 return -ENOBUFS;
365
366 idx = array_index_nospec(idx, v->nvqs);
367 vq = &v->vqs[idx];
368
369 switch (cmd) {
370 case VHOST_VDPA_SET_VRING_ENABLE:
371 if (copy_from_user(&s, argp, sizeof(s)))
372 return -EFAULT;
373 ops->set_vq_ready(vdpa, idx, s.num);
374 return 0;
375 case VHOST_GET_VRING_BASE:
376 r = ops->get_vq_state(v->vdpa, idx, &vq_state);
377 if (r)
378 return r;
379
380 vq->last_avail_idx = vq_state.avail_index;
381 break;
382 }
383
384 r = vhost_vring_ioctl(&v->vdev, cmd, argp);
385 if (r)
386 return r;
387
388 switch (cmd) {
389 case VHOST_SET_VRING_ADDR:
390 if (ops->set_vq_address(vdpa, idx,
391 (u64)(uintptr_t)vq->desc,
392 (u64)(uintptr_t)vq->avail,
393 (u64)(uintptr_t)vq->used))
394 r = -EINVAL;
395 break;
396
397 case VHOST_SET_VRING_BASE:
398 vq_state.avail_index = vq->last_avail_idx;
399 if (ops->set_vq_state(vdpa, idx, &vq_state))
400 r = -EINVAL;
401 break;
402
403 case VHOST_SET_VRING_CALL:
404 if (vq->call_ctx.ctx) {
405 cb.callback = vhost_vdpa_virtqueue_cb;
406 cb.private = vq;
407 } else {
408 cb.callback = NULL;
409 cb.private = NULL;
410 }
411 ops->set_vq_cb(vdpa, idx, &cb);
412 vhost_vdpa_setup_vq_irq(v, idx);
413 break;
414
415 case VHOST_SET_VRING_NUM:
416 ops->set_vq_num(vdpa, idx, vq->num);
417 break;
418 }
419
420 return r;
421}
422
423static long vhost_vdpa_unlocked_ioctl(struct file *filep,
424 unsigned int cmd, unsigned long arg)
425{
426 struct vhost_vdpa *v = filep->private_data;
427 struct vhost_dev *d = &v->vdev;
428 void __user *argp = (void __user *)arg;
429 u64 __user *featurep = argp;
430 u64 features;
431 long r;
432
433 if (cmd == VHOST_SET_BACKEND_FEATURES) {
434 r = copy_from_user(&features, featurep, sizeof(features));
435 if (r)
436 return r;
437 if (features & ~VHOST_VDPA_BACKEND_FEATURES)
438 return -EOPNOTSUPP;
439 vhost_set_backend_features(&v->vdev, features);
440 return 0;
441 }
442
443 mutex_lock(&d->mutex);
444
445 switch (cmd) {
446 case VHOST_VDPA_GET_DEVICE_ID:
447 r = vhost_vdpa_get_device_id(v, argp);
448 break;
449 case VHOST_VDPA_GET_STATUS:
450 r = vhost_vdpa_get_status(v, argp);
451 break;
452 case VHOST_VDPA_SET_STATUS:
453 r = vhost_vdpa_set_status(v, argp);
454 break;
455 case VHOST_VDPA_GET_CONFIG:
456 r = vhost_vdpa_get_config(v, argp);
457 break;
458 case VHOST_VDPA_SET_CONFIG:
459 r = vhost_vdpa_set_config(v, argp);
460 break;
461 case VHOST_GET_FEATURES:
462 r = vhost_vdpa_get_features(v, argp);
463 break;
464 case VHOST_SET_FEATURES:
465 r = vhost_vdpa_set_features(v, argp);
466 break;
467 case VHOST_VDPA_GET_VRING_NUM:
468 r = vhost_vdpa_get_vring_num(v, argp);
469 break;
470 case VHOST_SET_LOG_BASE:
471 case VHOST_SET_LOG_FD:
472 r = -ENOIOCTLCMD;
473 break;
474 case VHOST_VDPA_SET_CONFIG_CALL:
475 r = vhost_vdpa_set_config_call(v, argp);
476 break;
477 case VHOST_GET_BACKEND_FEATURES:
478 features = VHOST_VDPA_BACKEND_FEATURES;
479 r = copy_to_user(featurep, &features, sizeof(features));
480 break;
481 default:
482 r = vhost_dev_ioctl(&v->vdev, cmd, argp);
483 if (r == -ENOIOCTLCMD)
484 r = vhost_vdpa_vring_ioctl(v, cmd, argp);
485 break;
486 }
487
488 mutex_unlock(&d->mutex);
489 return r;
490}
491
492static void vhost_vdpa_iotlb_unmap(struct vhost_vdpa *v, u64 start, u64 last)
493{
494 struct vhost_dev *dev = &v->vdev;
495 struct vhost_iotlb *iotlb = dev->iotlb;
496 struct vhost_iotlb_map *map;
497 struct page *page;
498 unsigned long pfn, pinned;
499
500 while ((map = vhost_iotlb_itree_first(iotlb, start, last)) != NULL) {
501 pinned = map->size >> PAGE_SHIFT;
502 for (pfn = map->addr >> PAGE_SHIFT;
503 pinned > 0; pfn++, pinned--) {
504 page = pfn_to_page(pfn);
505 if (map->perm & VHOST_ACCESS_WO)
506 set_page_dirty_lock(page);
507 unpin_user_page(page);
508 }
509 atomic64_sub(map->size >> PAGE_SHIFT, &dev->mm->pinned_vm);
510 vhost_iotlb_map_free(iotlb, map);
511 }
512}
513
514static void vhost_vdpa_iotlb_free(struct vhost_vdpa *v)
515{
516 struct vhost_dev *dev = &v->vdev;
517
518 vhost_vdpa_iotlb_unmap(v, 0ULL, 0ULL - 1);
519 kfree(dev->iotlb);
520 dev->iotlb = NULL;
521}
522
523static int perm_to_iommu_flags(u32 perm)
524{
525 int flags = 0;
526
527 switch (perm) {
528 case VHOST_ACCESS_WO:
529 flags |= IOMMU_WRITE;
530 break;
531 case VHOST_ACCESS_RO:
532 flags |= IOMMU_READ;
533 break;
534 case VHOST_ACCESS_RW:
535 flags |= (IOMMU_WRITE | IOMMU_READ);
536 break;
537 default:
538 WARN(1, "invalidate vhost IOTLB permission\n");
539 break;
540 }
541
542 return flags | IOMMU_CACHE;
543}
544
545static int vhost_vdpa_map(struct vhost_vdpa *v,
546 u64 iova, u64 size, u64 pa, u32 perm)
547{
548 struct vhost_dev *dev = &v->vdev;
549 struct vdpa_device *vdpa = v->vdpa;
550 const struct vdpa_config_ops *ops = vdpa->config;
551 int r = 0;
552
553 r = vhost_iotlb_add_range(dev->iotlb, iova, iova + size - 1,
554 pa, perm);
555 if (r)
556 return r;
557
558 if (ops->dma_map) {
559 r = ops->dma_map(vdpa, iova, size, pa, perm);
560 } else if (ops->set_map) {
561 if (!v->in_batch)
562 r = ops->set_map(vdpa, dev->iotlb);
563 } else {
564 r = iommu_map(v->domain, iova, pa, size,
565 perm_to_iommu_flags(perm));
566 }
567
568 if (r)
569 vhost_iotlb_del_range(dev->iotlb, iova, iova + size - 1);
570
571 return r;
572}
573
574static void vhost_vdpa_unmap(struct vhost_vdpa *v, u64 iova, u64 size)
575{
576 struct vhost_dev *dev = &v->vdev;
577 struct vdpa_device *vdpa = v->vdpa;
578 const struct vdpa_config_ops *ops = vdpa->config;
579
580 vhost_vdpa_iotlb_unmap(v, iova, iova + size - 1);
581
582 if (ops->dma_map) {
583 ops->dma_unmap(vdpa, iova, size);
584 } else if (ops->set_map) {
585 if (!v->in_batch)
586 ops->set_map(vdpa, dev->iotlb);
587 } else {
588 iommu_unmap(v->domain, iova, size);
589 }
590}
591
592static int vhost_vdpa_process_iotlb_update(struct vhost_vdpa *v,
593 struct vhost_iotlb_msg *msg)
594{
595 struct vhost_dev *dev = &v->vdev;
596 struct vhost_iotlb *iotlb = dev->iotlb;
597 struct page **page_list;
598 struct vm_area_struct **vmas;
599 unsigned int gup_flags = FOLL_LONGTERM;
600 unsigned long map_pfn, last_pfn = 0;
601 unsigned long npages, lock_limit;
602 unsigned long i, nmap = 0;
603 u64 iova = msg->iova;
604 long pinned;
605 int ret = 0;
606
607 if (vhost_iotlb_itree_first(iotlb, msg->iova,
608 msg->iova + msg->size - 1))
609 return -EEXIST;
610
611 if (msg->perm & VHOST_ACCESS_WO)
612 gup_flags |= FOLL_WRITE;
613
614 npages = PAGE_ALIGN(msg->size + (iova & ~PAGE_MASK)) >> PAGE_SHIFT;
615 if (!npages)
616 return -EINVAL;
617
618 page_list = kvmalloc_array(npages, sizeof(struct page *), GFP_KERNEL);
619 vmas = kvmalloc_array(npages, sizeof(struct vm_area_struct *),
620 GFP_KERNEL);
621 if (!page_list || !vmas) {
622 ret = -ENOMEM;
623 goto free;
624 }
625
626 mmap_read_lock(dev->mm);
627
628 lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
629 if (npages + atomic64_read(&dev->mm->pinned_vm) > lock_limit) {
630 ret = -ENOMEM;
631 goto unlock;
632 }
633
634 pinned = pin_user_pages(msg->uaddr & PAGE_MASK, npages, gup_flags,
635 page_list, vmas);
636 if (npages != pinned) {
637 if (pinned < 0) {
638 ret = pinned;
639 } else {
640 unpin_user_pages(page_list, pinned);
641 ret = -ENOMEM;
642 }
643 goto unlock;
644 }
645
646 iova &= PAGE_MASK;
647 map_pfn = page_to_pfn(page_list[0]);
648
649 /* One more iteration to avoid extra vdpa_map() call out of loop. */
650 for (i = 0; i <= npages; i++) {
651 unsigned long this_pfn;
652 u64 csize;
653
654 /* The last chunk may have no valid PFN next to it */
655 this_pfn = i < npages ? page_to_pfn(page_list[i]) : -1UL;
656
657 if (last_pfn && (this_pfn == -1UL ||
658 this_pfn != last_pfn + 1)) {
659 /* Pin a contiguous chunk of memory */
660 csize = last_pfn - map_pfn + 1;
661 ret = vhost_vdpa_map(v, iova, csize << PAGE_SHIFT,
662 map_pfn << PAGE_SHIFT,
663 msg->perm);
664 if (ret) {
665 /*
666 * Unpin the rest chunks of memory on the
667 * flight with no corresponding vdpa_map()
668 * calls having been made yet. On the other
669 * hand, vdpa_unmap() in the failure path
670 * is in charge of accounting the number of
671 * pinned pages for its own.
672 * This asymmetrical pattern of accounting
673 * is for efficiency to pin all pages at
674 * once, while there is no other callsite
675 * of vdpa_map() than here above.
676 */
677 unpin_user_pages(&page_list[nmap],
678 npages - nmap);
679 goto out;
680 }
681 atomic64_add(csize, &dev->mm->pinned_vm);
682 nmap += csize;
683 iova += csize << PAGE_SHIFT;
684 map_pfn = this_pfn;
685 }
686 last_pfn = this_pfn;
687 }
688
689 WARN_ON(nmap != npages);
690out:
691 if (ret)
692 vhost_vdpa_unmap(v, msg->iova, msg->size);
693unlock:
694 mmap_read_unlock(dev->mm);
695free:
696 kvfree(vmas);
697 kvfree(page_list);
698 return ret;
699}
700
701static int vhost_vdpa_process_iotlb_msg(struct vhost_dev *dev,
702 struct vhost_iotlb_msg *msg)
703{
704 struct vhost_vdpa *v = container_of(dev, struct vhost_vdpa, vdev);
705 struct vdpa_device *vdpa = v->vdpa;
706 const struct vdpa_config_ops *ops = vdpa->config;
707 int r = 0;
708
709 r = vhost_dev_check_owner(dev);
710 if (r)
711 return r;
712
713 switch (msg->type) {
714 case VHOST_IOTLB_UPDATE:
715 r = vhost_vdpa_process_iotlb_update(v, msg);
716 break;
717 case VHOST_IOTLB_INVALIDATE:
718 vhost_vdpa_unmap(v, msg->iova, msg->size);
719 break;
720 case VHOST_IOTLB_BATCH_BEGIN:
721 v->in_batch = true;
722 break;
723 case VHOST_IOTLB_BATCH_END:
724 if (v->in_batch && ops->set_map)
725 ops->set_map(vdpa, dev->iotlb);
726 v->in_batch = false;
727 break;
728 default:
729 r = -EINVAL;
730 break;
731 }
732
733 return r;
734}
735
736static ssize_t vhost_vdpa_chr_write_iter(struct kiocb *iocb,
737 struct iov_iter *from)
738{
739 struct file *file = iocb->ki_filp;
740 struct vhost_vdpa *v = file->private_data;
741 struct vhost_dev *dev = &v->vdev;
742
743 return vhost_chr_write_iter(dev, from);
744}
745
746static int vhost_vdpa_alloc_domain(struct vhost_vdpa *v)
747{
748 struct vdpa_device *vdpa = v->vdpa;
749 const struct vdpa_config_ops *ops = vdpa->config;
750 struct device *dma_dev = vdpa_get_dma_dev(vdpa);
751 struct bus_type *bus;
752 int ret;
753
754 /* Device want to do DMA by itself */
755 if (ops->set_map || ops->dma_map)
756 return 0;
757
758 bus = dma_dev->bus;
759 if (!bus)
760 return -EFAULT;
761
762 if (!iommu_capable(bus, IOMMU_CAP_CACHE_COHERENCY))
763 return -ENOTSUPP;
764
765 v->domain = iommu_domain_alloc(bus);
766 if (!v->domain)
767 return -EIO;
768
769 ret = iommu_attach_device(v->domain, dma_dev);
770 if (ret)
771 goto err_attach;
772
773 return 0;
774
775err_attach:
776 iommu_domain_free(v->domain);
777 return ret;
778}
779
780static void vhost_vdpa_free_domain(struct vhost_vdpa *v)
781{
782 struct vdpa_device *vdpa = v->vdpa;
783 struct device *dma_dev = vdpa_get_dma_dev(vdpa);
784
785 if (v->domain) {
786 iommu_detach_device(v->domain, dma_dev);
787 iommu_domain_free(v->domain);
788 }
789
790 v->domain = NULL;
791}
792
793static int vhost_vdpa_open(struct inode *inode, struct file *filep)
794{
795 struct vhost_vdpa *v;
796 struct vhost_dev *dev;
797 struct vhost_virtqueue **vqs;
798 int nvqs, i, r, opened;
799
800 v = container_of(inode->i_cdev, struct vhost_vdpa, cdev);
801
802 opened = atomic_cmpxchg(&v->opened, 0, 1);
803 if (opened)
804 return -EBUSY;
805
806 nvqs = v->nvqs;
807 vhost_vdpa_reset(v);
808
809 vqs = kmalloc_array(nvqs, sizeof(*vqs), GFP_KERNEL);
810 if (!vqs) {
811 r = -ENOMEM;
812 goto err;
813 }
814
815 dev = &v->vdev;
816 for (i = 0; i < nvqs; i++) {
817 vqs[i] = &v->vqs[i];
818 vqs[i]->handle_kick = handle_vq_kick;
819 }
820 vhost_dev_init(dev, vqs, nvqs, 0, 0, 0, false,
821 vhost_vdpa_process_iotlb_msg);
822
823 dev->iotlb = vhost_iotlb_alloc(0, 0);
824 if (!dev->iotlb) {
825 r = -ENOMEM;
826 goto err_init_iotlb;
827 }
828
829 r = vhost_vdpa_alloc_domain(v);
830 if (r)
831 goto err_init_iotlb;
832
833 filep->private_data = v;
834
835 return 0;
836
837err_init_iotlb:
838 vhost_dev_cleanup(&v->vdev);
839 kfree(vqs);
840err:
841 atomic_dec(&v->opened);
842 return r;
843}
844
845static void vhost_vdpa_clean_irq(struct vhost_vdpa *v)
846{
847 struct vhost_virtqueue *vq;
848 int i;
849
850 for (i = 0; i < v->nvqs; i++) {
851 vq = &v->vqs[i];
852 if (vq->call_ctx.producer.irq)
853 irq_bypass_unregister_producer(&vq->call_ctx.producer);
854 }
855}
856
857static int vhost_vdpa_release(struct inode *inode, struct file *filep)
858{
859 struct vhost_vdpa *v = filep->private_data;
860 struct vhost_dev *d = &v->vdev;
861
862 mutex_lock(&d->mutex);
863 filep->private_data = NULL;
864 vhost_vdpa_reset(v);
865 vhost_dev_stop(&v->vdev);
866 vhost_vdpa_iotlb_free(v);
867 vhost_vdpa_free_domain(v);
868 vhost_vdpa_config_put(v);
869 vhost_vdpa_clean_irq(v);
870 vhost_dev_cleanup(&v->vdev);
871 kfree(v->vdev.vqs);
872 mutex_unlock(&d->mutex);
873
874 atomic_dec(&v->opened);
875 complete(&v->completion);
876
877 return 0;
878}
879
880#ifdef CONFIG_MMU
881static vm_fault_t vhost_vdpa_fault(struct vm_fault *vmf)
882{
883 struct vhost_vdpa *v = vmf->vma->vm_file->private_data;
884 struct vdpa_device *vdpa = v->vdpa;
885 const struct vdpa_config_ops *ops = vdpa->config;
886 struct vdpa_notification_area notify;
887 struct vm_area_struct *vma = vmf->vma;
888 u16 index = vma->vm_pgoff;
889
890 notify = ops->get_vq_notification(vdpa, index);
891
892 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
893 if (remap_pfn_range(vma, vmf->address & PAGE_MASK,
894 notify.addr >> PAGE_SHIFT, PAGE_SIZE,
895 vma->vm_page_prot))
896 return VM_FAULT_SIGBUS;
897
898 return VM_FAULT_NOPAGE;
899}
900
901static const struct vm_operations_struct vhost_vdpa_vm_ops = {
902 .fault = vhost_vdpa_fault,
903};
904
905static int vhost_vdpa_mmap(struct file *file, struct vm_area_struct *vma)
906{
907 struct vhost_vdpa *v = vma->vm_file->private_data;
908 struct vdpa_device *vdpa = v->vdpa;
909 const struct vdpa_config_ops *ops = vdpa->config;
910 struct vdpa_notification_area notify;
911 unsigned long index = vma->vm_pgoff;
912
913 if (vma->vm_end - vma->vm_start != PAGE_SIZE)
914 return -EINVAL;
915 if ((vma->vm_flags & VM_SHARED) == 0)
916 return -EINVAL;
917 if (vma->vm_flags & VM_READ)
918 return -EINVAL;
919 if (index > 65535)
920 return -EINVAL;
921 if (!ops->get_vq_notification)
922 return -ENOTSUPP;
923
924 /* To be safe and easily modelled by userspace, We only
925 * support the doorbell which sits on the page boundary and
926 * does not share the page with other registers.
927 */
928 notify = ops->get_vq_notification(vdpa, index);
929 if (notify.addr & (PAGE_SIZE - 1))
930 return -EINVAL;
931 if (vma->vm_end - vma->vm_start != notify.size)
932 return -ENOTSUPP;
933
934 vma->vm_ops = &vhost_vdpa_vm_ops;
935 return 0;
936}
937#endif /* CONFIG_MMU */
938
939static const struct file_operations vhost_vdpa_fops = {
940 .owner = THIS_MODULE,
941 .open = vhost_vdpa_open,
942 .release = vhost_vdpa_release,
943 .write_iter = vhost_vdpa_chr_write_iter,
944 .unlocked_ioctl = vhost_vdpa_unlocked_ioctl,
945#ifdef CONFIG_MMU
946 .mmap = vhost_vdpa_mmap,
947#endif /* CONFIG_MMU */
948 .compat_ioctl = compat_ptr_ioctl,
949};
950
951static void vhost_vdpa_release_dev(struct device *device)
952{
953 struct vhost_vdpa *v =
954 container_of(device, struct vhost_vdpa, dev);
955
956 ida_simple_remove(&vhost_vdpa_ida, v->minor);
957 kfree(v->vqs);
958 kfree(v);
959}
960
961static int vhost_vdpa_probe(struct vdpa_device *vdpa)
962{
963 const struct vdpa_config_ops *ops = vdpa->config;
964 struct vhost_vdpa *v;
965 int minor;
966 int r;
967
968 /* Currently, we only accept the network devices. */
969 if (ops->get_device_id(vdpa) != VIRTIO_ID_NET)
970 return -ENOTSUPP;
971
972 v = kzalloc(sizeof(*v), GFP_KERNEL | __GFP_RETRY_MAYFAIL);
973 if (!v)
974 return -ENOMEM;
975
976 minor = ida_simple_get(&vhost_vdpa_ida, 0,
977 VHOST_VDPA_DEV_MAX, GFP_KERNEL);
978 if (minor < 0) {
979 kfree(v);
980 return minor;
981 }
982
983 atomic_set(&v->opened, 0);
984 v->minor = minor;
985 v->vdpa = vdpa;
986 v->nvqs = vdpa->nvqs;
987 v->virtio_id = ops->get_device_id(vdpa);
988
989 device_initialize(&v->dev);
990 v->dev.release = vhost_vdpa_release_dev;
991 v->dev.parent = &vdpa->dev;
992 v->dev.devt = MKDEV(MAJOR(vhost_vdpa_major), minor);
993 v->vqs = kmalloc_array(v->nvqs, sizeof(struct vhost_virtqueue),
994 GFP_KERNEL);
995 if (!v->vqs) {
996 r = -ENOMEM;
997 goto err;
998 }
999
1000 r = dev_set_name(&v->dev, "vhost-vdpa-%u", minor);
1001 if (r)
1002 goto err;
1003
1004 cdev_init(&v->cdev, &vhost_vdpa_fops);
1005 v->cdev.owner = THIS_MODULE;
1006
1007 r = cdev_device_add(&v->cdev, &v->dev);
1008 if (r)
1009 goto err;
1010
1011 init_completion(&v->completion);
1012 vdpa_set_drvdata(vdpa, v);
1013
1014 return 0;
1015
1016err:
1017 put_device(&v->dev);
1018 return r;
1019}
1020
1021static void vhost_vdpa_remove(struct vdpa_device *vdpa)
1022{
1023 struct vhost_vdpa *v = vdpa_get_drvdata(vdpa);
1024 int opened;
1025
1026 cdev_device_del(&v->cdev, &v->dev);
1027
1028 do {
1029 opened = atomic_cmpxchg(&v->opened, 0, 1);
1030 if (!opened)
1031 break;
1032 wait_for_completion(&v->completion);
1033 } while (1);
1034
1035 put_device(&v->dev);
1036}
1037
1038static struct vdpa_driver vhost_vdpa_driver = {
1039 .driver = {
1040 .name = "vhost_vdpa",
1041 },
1042 .probe = vhost_vdpa_probe,
1043 .remove = vhost_vdpa_remove,
1044};
1045
1046static int __init vhost_vdpa_init(void)
1047{
1048 int r;
1049
1050 r = alloc_chrdev_region(&vhost_vdpa_major, 0, VHOST_VDPA_DEV_MAX,
1051 "vhost-vdpa");
1052 if (r)
1053 goto err_alloc_chrdev;
1054
1055 r = vdpa_register_driver(&vhost_vdpa_driver);
1056 if (r)
1057 goto err_vdpa_register_driver;
1058
1059 return 0;
1060
1061err_vdpa_register_driver:
1062 unregister_chrdev_region(vhost_vdpa_major, VHOST_VDPA_DEV_MAX);
1063err_alloc_chrdev:
1064 return r;
1065}
1066module_init(vhost_vdpa_init);
1067
1068static void __exit vhost_vdpa_exit(void)
1069{
1070 vdpa_unregister_driver(&vhost_vdpa_driver);
1071 unregister_chrdev_region(vhost_vdpa_major, VHOST_VDPA_DEV_MAX);
1072}
1073module_exit(vhost_vdpa_exit);
1074
1075MODULE_VERSION("0.0.1");
1076MODULE_LICENSE("GPL v2");
1077MODULE_AUTHOR("Intel Corporation");
1078MODULE_DESCRIPTION("vDPA-based vhost backend for virtio");