1 // SPDX-License-Identifier: GPL-2.0
3 * Copyright (C) 2018-2020 Intel Corporation.
4 * Copyright (C) 2020 Red Hat, Inc.
6 * Author: Tiwei Bie <tiwei.bie@intel.com>
7 * Jason Wang <jasowang@redhat.com>
9 * Thanks Michael S. Tsirkin for the valuable comments and
10 * suggestions. And thanks to Cunming Liang and Zhihong Wang for all
14 #include <linux/kernel.h>
15 #include <linux/module.h>
16 #include <linux/cdev.h>
17 #include <linux/device.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>
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),
35 #define VHOST_VDPA_DEV_MAX (1U << MINORBITS)
37 #define VHOST_VDPA_IOTLB_BUCKETS 16
39 struct vhost_vdpa_as {
40 struct hlist_node hash_link;
41 struct vhost_iotlb iotlb;
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];
58 struct eventfd_ctx *config_ctx;
60 struct vdpa_iova_range range;
64 static DEFINE_IDA(vhost_vdpa_ida);
66 static dev_t vhost_vdpa_major;
68 static void vhost_vdpa_iotlb_unmap(struct vhost_vdpa *v,
69 struct vhost_iotlb *iotlb, u64 start,
72 static inline u32 iotlb_to_asid(struct vhost_iotlb *iotlb)
74 struct vhost_vdpa_as *as = container_of(iotlb, struct
75 vhost_vdpa_as, iotlb);
79 static struct vhost_vdpa_as *asid_to_as(struct vhost_vdpa *v, u32 asid)
81 struct hlist_head *head = &v->as[asid % VHOST_VDPA_IOTLB_BUCKETS];
82 struct vhost_vdpa_as *as;
84 hlist_for_each_entry(as, head, hash_link)
91 static struct vhost_iotlb *asid_to_iotlb(struct vhost_vdpa *v, u32 asid)
93 struct vhost_vdpa_as *as = asid_to_as(v, asid);
101 static struct vhost_vdpa_as *vhost_vdpa_alloc_as(struct vhost_vdpa *v, u32 asid)
103 struct hlist_head *head = &v->as[asid % VHOST_VDPA_IOTLB_BUCKETS];
104 struct vhost_vdpa_as *as;
106 if (asid_to_as(v, asid))
109 if (asid >= v->vdpa->nas)
112 as = kmalloc(sizeof(*as), GFP_KERNEL);
116 vhost_iotlb_init(&as->iotlb, 0, 0);
118 hlist_add_head(&as->hash_link, head);
123 static struct vhost_vdpa_as *vhost_vdpa_find_alloc_as(struct vhost_vdpa *v,
126 struct vhost_vdpa_as *as = asid_to_as(v, asid);
131 return vhost_vdpa_alloc_as(v, asid);
134 static int vhost_vdpa_remove_as(struct vhost_vdpa *v, u32 asid)
136 struct vhost_vdpa_as *as = asid_to_as(v, asid);
141 hlist_del(&as->hash_link);
142 vhost_vdpa_iotlb_unmap(v, &as->iotlb, 0ULL, 0ULL - 1, asid);
148 static void handle_vq_kick(struct vhost_work *work)
150 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
152 struct vhost_vdpa *v = container_of(vq->dev, struct vhost_vdpa, vdev);
153 const struct vdpa_config_ops *ops = v->vdpa->config;
155 ops->kick_vq(v->vdpa, vq - v->vqs);
158 static irqreturn_t vhost_vdpa_virtqueue_cb(void *private)
160 struct vhost_virtqueue *vq = private;
161 struct eventfd_ctx *call_ctx = vq->call_ctx.ctx;
164 eventfd_signal(call_ctx, 1);
169 static irqreturn_t vhost_vdpa_config_cb(void *private)
171 struct vhost_vdpa *v = private;
172 struct eventfd_ctx *config_ctx = v->config_ctx;
175 eventfd_signal(config_ctx, 1);
180 static void vhost_vdpa_setup_vq_irq(struct vhost_vdpa *v, u16 qid)
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;
187 if (!ops->get_vq_irq)
190 irq = ops->get_vq_irq(vdpa, qid);
194 irq_bypass_unregister_producer(&vq->call_ctx.producer);
195 if (!vq->call_ctx.ctx)
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);
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);
206 static void vhost_vdpa_unsetup_vq_irq(struct vhost_vdpa *v, u16 qid)
208 struct vhost_virtqueue *vq = &v->vqs[qid];
210 irq_bypass_unregister_producer(&vq->call_ctx.producer);
213 static int vhost_vdpa_reset(struct vhost_vdpa *v)
215 struct vdpa_device *vdpa = v->vdpa;
219 return vdpa_reset(vdpa);
222 static long vhost_vdpa_get_device_id(struct vhost_vdpa *v, u8 __user *argp)
224 struct vdpa_device *vdpa = v->vdpa;
225 const struct vdpa_config_ops *ops = vdpa->config;
228 device_id = ops->get_device_id(vdpa);
230 if (copy_to_user(argp, &device_id, sizeof(device_id)))
236 static long vhost_vdpa_get_status(struct vhost_vdpa *v, u8 __user *statusp)
238 struct vdpa_device *vdpa = v->vdpa;
239 const struct vdpa_config_ops *ops = vdpa->config;
242 status = ops->get_status(vdpa);
244 if (copy_to_user(statusp, &status, sizeof(status)))
250 static long vhost_vdpa_set_status(struct vhost_vdpa *v, u8 __user *statusp)
252 struct vdpa_device *vdpa = v->vdpa;
253 const struct vdpa_config_ops *ops = vdpa->config;
254 u8 status, status_old;
259 if (copy_from_user(&status, statusp, sizeof(status)))
262 status_old = ops->get_status(vdpa);
265 * Userspace shouldn't remove status bits unless reset the
268 if (status != 0 && (status_old & ~status) != 0)
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);
276 ret = vdpa_reset(vdpa);
280 vdpa_set_status(vdpa, status);
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);
289 static int vhost_vdpa_config_validate(struct vhost_vdpa *v,
290 struct vhost_vdpa_config *c)
292 struct vdpa_device *vdpa = v->vdpa;
293 size_t size = vdpa->config->get_config_size(vdpa);
295 if (c->len == 0 || c->off > size)
298 if (c->len > size - c->off)
304 static long vhost_vdpa_get_config(struct vhost_vdpa *v,
305 struct vhost_vdpa_config __user *c)
307 struct vdpa_device *vdpa = v->vdpa;
308 struct vhost_vdpa_config config;
309 unsigned long size = offsetof(struct vhost_vdpa_config, buf);
312 if (copy_from_user(&config, c, size))
314 if (vhost_vdpa_config_validate(v, &config))
316 buf = kvzalloc(config.len, GFP_KERNEL);
320 vdpa_get_config(vdpa, config.off, buf, config.len);
322 if (copy_to_user(c->buf, buf, config.len)) {
331 static long vhost_vdpa_set_config(struct vhost_vdpa *v,
332 struct vhost_vdpa_config __user *c)
334 struct vdpa_device *vdpa = v->vdpa;
335 struct vhost_vdpa_config config;
336 unsigned long size = offsetof(struct vhost_vdpa_config, buf);
339 if (copy_from_user(&config, c, size))
341 if (vhost_vdpa_config_validate(v, &config))
344 buf = vmemdup_user(c->buf, config.len);
348 vdpa_set_config(vdpa, config.off, buf, config.len);
354 static bool vhost_vdpa_can_suspend(const struct vhost_vdpa *v)
356 struct vdpa_device *vdpa = v->vdpa;
357 const struct vdpa_config_ops *ops = vdpa->config;
362 static long vhost_vdpa_get_features(struct vhost_vdpa *v, u64 __user *featurep)
364 struct vdpa_device *vdpa = v->vdpa;
365 const struct vdpa_config_ops *ops = vdpa->config;
368 features = ops->get_device_features(vdpa);
370 if (copy_to_user(featurep, &features, sizeof(features)))
376 static long vhost_vdpa_set_features(struct vhost_vdpa *v, u64 __user *featurep)
378 struct vdpa_device *vdpa = v->vdpa;
379 const struct vdpa_config_ops *ops = vdpa->config;
383 * It's not allowed to change the features after they have
386 if (ops->get_status(vdpa) & VIRTIO_CONFIG_S_FEATURES_OK)
389 if (copy_from_user(&features, featurep, sizeof(features)))
392 if (vdpa_set_features(vdpa, features))
398 static long vhost_vdpa_get_vring_num(struct vhost_vdpa *v, u16 __user *argp)
400 struct vdpa_device *vdpa = v->vdpa;
401 const struct vdpa_config_ops *ops = vdpa->config;
404 num = ops->get_vq_num_max(vdpa);
406 if (copy_to_user(argp, &num, sizeof(num)))
412 static void vhost_vdpa_config_put(struct vhost_vdpa *v)
415 eventfd_ctx_put(v->config_ctx);
416 v->config_ctx = NULL;
420 static long vhost_vdpa_set_config_call(struct vhost_vdpa *v, u32 __user *argp)
422 struct vdpa_callback cb;
424 struct eventfd_ctx *ctx;
426 cb.callback = vhost_vdpa_config_cb;
428 if (copy_from_user(&fd, argp, sizeof(fd)))
431 ctx = fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(fd);
432 swap(ctx, v->config_ctx);
434 if (!IS_ERR_OR_NULL(ctx))
435 eventfd_ctx_put(ctx);
437 if (IS_ERR(v->config_ctx)) {
438 long ret = PTR_ERR(v->config_ctx);
440 v->config_ctx = NULL;
444 v->vdpa->config->set_config_cb(v->vdpa, &cb);
449 static long vhost_vdpa_get_iova_range(struct vhost_vdpa *v, u32 __user *argp)
451 struct vhost_vdpa_iova_range range = {
452 .first = v->range.first,
453 .last = v->range.last,
456 if (copy_to_user(argp, &range, sizeof(range)))
461 static long vhost_vdpa_get_config_size(struct vhost_vdpa *v, u32 __user *argp)
463 struct vdpa_device *vdpa = v->vdpa;
464 const struct vdpa_config_ops *ops = vdpa->config;
467 size = ops->get_config_size(vdpa);
469 if (copy_to_user(argp, &size, sizeof(size)))
475 static long vhost_vdpa_get_vqs_count(struct vhost_vdpa *v, u32 __user *argp)
477 struct vdpa_device *vdpa = v->vdpa;
479 if (copy_to_user(argp, &vdpa->nvqs, sizeof(vdpa->nvqs)))
485 /* After a successful return of ioctl the device must not process more
486 * virtqueue descriptors. The device can answer to read or writes of config
487 * fields as if it were not suspended. In particular, writing to "queue_enable"
488 * with a value of 1 will not make the device start processing buffers.
490 static long vhost_vdpa_suspend(struct vhost_vdpa *v)
492 struct vdpa_device *vdpa = v->vdpa;
493 const struct vdpa_config_ops *ops = vdpa->config;
498 return ops->suspend(vdpa);
501 static long vhost_vdpa_vring_ioctl(struct vhost_vdpa *v, unsigned int cmd,
504 struct vdpa_device *vdpa = v->vdpa;
505 const struct vdpa_config_ops *ops = vdpa->config;
506 struct vdpa_vq_state vq_state;
507 struct vdpa_callback cb;
508 struct vhost_virtqueue *vq;
509 struct vhost_vring_state s;
513 r = get_user(idx, (u32 __user *)argp);
520 idx = array_index_nospec(idx, v->nvqs);
524 case VHOST_VDPA_SET_VRING_ENABLE:
525 if (copy_from_user(&s, argp, sizeof(s)))
527 ops->set_vq_ready(vdpa, idx, s.num);
529 case VHOST_VDPA_GET_VRING_GROUP:
530 if (!ops->get_vq_group)
533 s.num = ops->get_vq_group(vdpa, idx);
534 if (s.num >= vdpa->ngroups)
536 else if (copy_to_user(argp, &s, sizeof(s)))
539 case VHOST_VDPA_SET_GROUP_ASID:
540 if (copy_from_user(&s, argp, sizeof(s)))
542 if (s.num >= vdpa->nas)
544 if (!ops->set_group_asid)
546 return ops->set_group_asid(vdpa, idx, s.num);
547 case VHOST_GET_VRING_BASE:
548 r = ops->get_vq_state(v->vdpa, idx, &vq_state);
552 vq->last_avail_idx = vq_state.split.avail_index;
556 r = vhost_vring_ioctl(&v->vdev, cmd, argp);
561 case VHOST_SET_VRING_ADDR:
562 if (ops->set_vq_address(vdpa, idx,
563 (u64)(uintptr_t)vq->desc,
564 (u64)(uintptr_t)vq->avail,
565 (u64)(uintptr_t)vq->used))
569 case VHOST_SET_VRING_BASE:
570 vq_state.split.avail_index = vq->last_avail_idx;
571 if (ops->set_vq_state(vdpa, idx, &vq_state))
575 case VHOST_SET_VRING_CALL:
576 if (vq->call_ctx.ctx) {
577 cb.callback = vhost_vdpa_virtqueue_cb;
583 ops->set_vq_cb(vdpa, idx, &cb);
584 vhost_vdpa_setup_vq_irq(v, idx);
587 case VHOST_SET_VRING_NUM:
588 ops->set_vq_num(vdpa, idx, vq->num);
595 static long vhost_vdpa_unlocked_ioctl(struct file *filep,
596 unsigned int cmd, unsigned long arg)
598 struct vhost_vdpa *v = filep->private_data;
599 struct vhost_dev *d = &v->vdev;
600 void __user *argp = (void __user *)arg;
601 u64 __user *featurep = argp;
605 if (cmd == VHOST_SET_BACKEND_FEATURES) {
606 if (copy_from_user(&features, featurep, sizeof(features)))
608 if (features & ~(VHOST_VDPA_BACKEND_FEATURES |
609 BIT_ULL(VHOST_BACKEND_F_SUSPEND)))
611 if ((features & BIT_ULL(VHOST_BACKEND_F_SUSPEND)) &&
612 !vhost_vdpa_can_suspend(v))
614 vhost_set_backend_features(&v->vdev, features);
618 mutex_lock(&d->mutex);
621 case VHOST_VDPA_GET_DEVICE_ID:
622 r = vhost_vdpa_get_device_id(v, argp);
624 case VHOST_VDPA_GET_STATUS:
625 r = vhost_vdpa_get_status(v, argp);
627 case VHOST_VDPA_SET_STATUS:
628 r = vhost_vdpa_set_status(v, argp);
630 case VHOST_VDPA_GET_CONFIG:
631 r = vhost_vdpa_get_config(v, argp);
633 case VHOST_VDPA_SET_CONFIG:
634 r = vhost_vdpa_set_config(v, argp);
636 case VHOST_GET_FEATURES:
637 r = vhost_vdpa_get_features(v, argp);
639 case VHOST_SET_FEATURES:
640 r = vhost_vdpa_set_features(v, argp);
642 case VHOST_VDPA_GET_VRING_NUM:
643 r = vhost_vdpa_get_vring_num(v, argp);
645 case VHOST_VDPA_GET_GROUP_NUM:
646 if (copy_to_user(argp, &v->vdpa->ngroups,
647 sizeof(v->vdpa->ngroups)))
650 case VHOST_VDPA_GET_AS_NUM:
651 if (copy_to_user(argp, &v->vdpa->nas, sizeof(v->vdpa->nas)))
654 case VHOST_SET_LOG_BASE:
655 case VHOST_SET_LOG_FD:
658 case VHOST_VDPA_SET_CONFIG_CALL:
659 r = vhost_vdpa_set_config_call(v, argp);
661 case VHOST_GET_BACKEND_FEATURES:
662 features = VHOST_VDPA_BACKEND_FEATURES;
663 if (vhost_vdpa_can_suspend(v))
664 features |= BIT_ULL(VHOST_BACKEND_F_SUSPEND);
665 if (copy_to_user(featurep, &features, sizeof(features)))
668 case VHOST_VDPA_GET_IOVA_RANGE:
669 r = vhost_vdpa_get_iova_range(v, argp);
671 case VHOST_VDPA_GET_CONFIG_SIZE:
672 r = vhost_vdpa_get_config_size(v, argp);
674 case VHOST_VDPA_GET_VQS_COUNT:
675 r = vhost_vdpa_get_vqs_count(v, argp);
677 case VHOST_VDPA_SUSPEND:
678 r = vhost_vdpa_suspend(v);
681 r = vhost_dev_ioctl(&v->vdev, cmd, argp);
682 if (r == -ENOIOCTLCMD)
683 r = vhost_vdpa_vring_ioctl(v, cmd, argp);
687 mutex_unlock(&d->mutex);
690 static void vhost_vdpa_general_unmap(struct vhost_vdpa *v,
691 struct vhost_iotlb_map *map, u32 asid)
693 struct vdpa_device *vdpa = v->vdpa;
694 const struct vdpa_config_ops *ops = vdpa->config;
696 ops->dma_unmap(vdpa, asid, map->start, map->size);
697 } else if (ops->set_map == NULL) {
698 iommu_unmap(v->domain, map->start, map->size);
702 static void vhost_vdpa_pa_unmap(struct vhost_vdpa *v, struct vhost_iotlb *iotlb,
703 u64 start, u64 last, u32 asid)
705 struct vhost_dev *dev = &v->vdev;
706 struct vhost_iotlb_map *map;
708 unsigned long pfn, pinned;
710 while ((map = vhost_iotlb_itree_first(iotlb, start, last)) != NULL) {
711 pinned = PFN_DOWN(map->size);
712 for (pfn = PFN_DOWN(map->addr);
713 pinned > 0; pfn++, pinned--) {
714 page = pfn_to_page(pfn);
715 if (map->perm & VHOST_ACCESS_WO)
716 set_page_dirty_lock(page);
717 unpin_user_page(page);
719 atomic64_sub(PFN_DOWN(map->size), &dev->mm->pinned_vm);
720 vhost_vdpa_general_unmap(v, map, asid);
721 vhost_iotlb_map_free(iotlb, map);
725 static void vhost_vdpa_va_unmap(struct vhost_vdpa *v, struct vhost_iotlb *iotlb,
726 u64 start, u64 last, u32 asid)
728 struct vhost_iotlb_map *map;
729 struct vdpa_map_file *map_file;
731 while ((map = vhost_iotlb_itree_first(iotlb, start, last)) != NULL) {
732 map_file = (struct vdpa_map_file *)map->opaque;
733 fput(map_file->file);
735 vhost_vdpa_general_unmap(v, map, asid);
736 vhost_iotlb_map_free(iotlb, map);
740 static void vhost_vdpa_iotlb_unmap(struct vhost_vdpa *v,
741 struct vhost_iotlb *iotlb, u64 start,
744 struct vdpa_device *vdpa = v->vdpa;
747 return vhost_vdpa_va_unmap(v, iotlb, start, last, asid);
749 return vhost_vdpa_pa_unmap(v, iotlb, start, last, asid);
752 static int perm_to_iommu_flags(u32 perm)
757 case VHOST_ACCESS_WO:
758 flags |= IOMMU_WRITE;
760 case VHOST_ACCESS_RO:
763 case VHOST_ACCESS_RW:
764 flags |= (IOMMU_WRITE | IOMMU_READ);
767 WARN(1, "invalidate vhost IOTLB permission\n");
771 return flags | IOMMU_CACHE;
774 static int vhost_vdpa_map(struct vhost_vdpa *v, struct vhost_iotlb *iotlb,
775 u64 iova, u64 size, u64 pa, u32 perm, void *opaque)
777 struct vhost_dev *dev = &v->vdev;
778 struct vdpa_device *vdpa = v->vdpa;
779 const struct vdpa_config_ops *ops = vdpa->config;
780 u32 asid = iotlb_to_asid(iotlb);
783 r = vhost_iotlb_add_range_ctx(iotlb, iova, iova + size - 1,
789 r = ops->dma_map(vdpa, asid, iova, size, pa, perm, opaque);
790 } else if (ops->set_map) {
792 r = ops->set_map(vdpa, asid, iotlb);
794 r = iommu_map(v->domain, iova, pa, size,
795 perm_to_iommu_flags(perm));
798 vhost_iotlb_del_range(iotlb, iova, iova + size - 1);
803 atomic64_add(PFN_DOWN(size), &dev->mm->pinned_vm);
808 static void vhost_vdpa_unmap(struct vhost_vdpa *v,
809 struct vhost_iotlb *iotlb,
812 struct vdpa_device *vdpa = v->vdpa;
813 const struct vdpa_config_ops *ops = vdpa->config;
814 u32 asid = iotlb_to_asid(iotlb);
816 vhost_vdpa_iotlb_unmap(v, iotlb, iova, iova + size - 1, asid);
820 ops->set_map(vdpa, asid, iotlb);
822 /* If we are in the middle of batch processing, delay the free
823 * of AS until BATCH_END.
825 if (!v->in_batch && !iotlb->nmaps)
826 vhost_vdpa_remove_as(v, asid);
829 static int vhost_vdpa_va_map(struct vhost_vdpa *v,
830 struct vhost_iotlb *iotlb,
831 u64 iova, u64 size, u64 uaddr, u32 perm)
833 struct vhost_dev *dev = &v->vdev;
834 u64 offset, map_size, map_iova = iova;
835 struct vdpa_map_file *map_file;
836 struct vm_area_struct *vma;
839 mmap_read_lock(dev->mm);
842 vma = find_vma(dev->mm, uaddr);
847 map_size = min(size, vma->vm_end - uaddr);
848 if (!(vma->vm_file && (vma->vm_flags & VM_SHARED) &&
849 !(vma->vm_flags & (VM_IO | VM_PFNMAP))))
852 map_file = kzalloc(sizeof(*map_file), GFP_KERNEL);
857 offset = (vma->vm_pgoff << PAGE_SHIFT) + uaddr - vma->vm_start;
858 map_file->offset = offset;
859 map_file->file = get_file(vma->vm_file);
860 ret = vhost_vdpa_map(v, iotlb, map_iova, map_size, uaddr,
863 fput(map_file->file);
870 map_iova += map_size;
873 vhost_vdpa_unmap(v, iotlb, iova, map_iova - iova);
875 mmap_read_unlock(dev->mm);
880 static int vhost_vdpa_pa_map(struct vhost_vdpa *v,
881 struct vhost_iotlb *iotlb,
882 u64 iova, u64 size, u64 uaddr, u32 perm)
884 struct vhost_dev *dev = &v->vdev;
885 struct page **page_list;
886 unsigned long list_size = PAGE_SIZE / sizeof(struct page *);
887 unsigned int gup_flags = FOLL_LONGTERM;
888 unsigned long npages, cur_base, map_pfn, last_pfn = 0;
889 unsigned long lock_limit, sz2pin, nchunks, i;
894 /* Limit the use of memory for bookkeeping */
895 page_list = (struct page **) __get_free_page(GFP_KERNEL);
899 if (perm & VHOST_ACCESS_WO)
900 gup_flags |= FOLL_WRITE;
902 npages = PFN_UP(size + (iova & ~PAGE_MASK));
908 mmap_read_lock(dev->mm);
910 lock_limit = PFN_DOWN(rlimit(RLIMIT_MEMLOCK));
911 if (npages + atomic64_read(&dev->mm->pinned_vm) > lock_limit) {
916 cur_base = uaddr & PAGE_MASK;
921 sz2pin = min_t(unsigned long, npages, list_size);
922 pinned = pin_user_pages(cur_base, sz2pin,
923 gup_flags, page_list, NULL);
924 if (sz2pin != pinned) {
928 unpin_user_pages(page_list, pinned);
936 map_pfn = page_to_pfn(page_list[0]);
938 for (i = 0; i < pinned; i++) {
939 unsigned long this_pfn = page_to_pfn(page_list[i]);
942 if (last_pfn && (this_pfn != last_pfn + 1)) {
943 /* Pin a contiguous chunk of memory */
944 csize = PFN_PHYS(last_pfn - map_pfn + 1);
945 ret = vhost_vdpa_map(v, iotlb, iova, csize,
950 * Unpin the pages that are left unmapped
951 * from this point on in the current
952 * page_list. The remaining outstanding
953 * ones which may stride across several
954 * chunks will be covered in the common
955 * error path subsequently.
957 unpin_user_pages(&page_list[i],
970 cur_base += PFN_PHYS(pinned);
974 /* Pin the rest chunk */
975 ret = vhost_vdpa_map(v, iotlb, iova, PFN_PHYS(last_pfn - map_pfn + 1),
976 PFN_PHYS(map_pfn), perm, NULL);
983 * Unpin the outstanding pages which are yet to be
984 * mapped but haven't due to vdpa_map() or
985 * pin_user_pages() failure.
987 * Mapped pages are accounted in vdpa_map(), hence
988 * the corresponding unpinning will be handled by
992 for (pfn = map_pfn; pfn <= last_pfn; pfn++)
993 unpin_user_page(pfn_to_page(pfn));
995 vhost_vdpa_unmap(v, iotlb, start, size);
998 mmap_read_unlock(dev->mm);
1000 free_page((unsigned long)page_list);
1005 static int vhost_vdpa_process_iotlb_update(struct vhost_vdpa *v,
1006 struct vhost_iotlb *iotlb,
1007 struct vhost_iotlb_msg *msg)
1009 struct vdpa_device *vdpa = v->vdpa;
1011 if (msg->iova < v->range.first || !msg->size ||
1012 msg->iova > U64_MAX - msg->size + 1 ||
1013 msg->iova + msg->size - 1 > v->range.last)
1016 if (vhost_iotlb_itree_first(iotlb, msg->iova,
1017 msg->iova + msg->size - 1))
1021 return vhost_vdpa_va_map(v, iotlb, msg->iova, msg->size,
1022 msg->uaddr, msg->perm);
1024 return vhost_vdpa_pa_map(v, iotlb, msg->iova, msg->size, msg->uaddr,
1028 static int vhost_vdpa_process_iotlb_msg(struct vhost_dev *dev, u32 asid,
1029 struct vhost_iotlb_msg *msg)
1031 struct vhost_vdpa *v = container_of(dev, struct vhost_vdpa, vdev);
1032 struct vdpa_device *vdpa = v->vdpa;
1033 const struct vdpa_config_ops *ops = vdpa->config;
1034 struct vhost_iotlb *iotlb = NULL;
1035 struct vhost_vdpa_as *as = NULL;
1038 mutex_lock(&dev->mutex);
1040 r = vhost_dev_check_owner(dev);
1044 if (msg->type == VHOST_IOTLB_UPDATE ||
1045 msg->type == VHOST_IOTLB_BATCH_BEGIN) {
1046 as = vhost_vdpa_find_alloc_as(v, asid);
1048 dev_err(&v->dev, "can't find and alloc asid %d\n",
1055 iotlb = asid_to_iotlb(v, asid);
1057 if ((v->in_batch && v->batch_asid != asid) || !iotlb) {
1058 if (v->in_batch && v->batch_asid != asid) {
1059 dev_info(&v->dev, "batch id %d asid %d\n",
1060 v->batch_asid, asid);
1063 dev_err(&v->dev, "no iotlb for asid %d\n", asid);
1068 switch (msg->type) {
1069 case VHOST_IOTLB_UPDATE:
1070 r = vhost_vdpa_process_iotlb_update(v, iotlb, msg);
1072 case VHOST_IOTLB_INVALIDATE:
1073 vhost_vdpa_unmap(v, iotlb, msg->iova, msg->size);
1075 case VHOST_IOTLB_BATCH_BEGIN:
1076 v->batch_asid = asid;
1079 case VHOST_IOTLB_BATCH_END:
1080 if (v->in_batch && ops->set_map)
1081 ops->set_map(vdpa, asid, iotlb);
1082 v->in_batch = false;
1084 vhost_vdpa_remove_as(v, asid);
1091 mutex_unlock(&dev->mutex);
1096 static ssize_t vhost_vdpa_chr_write_iter(struct kiocb *iocb,
1097 struct iov_iter *from)
1099 struct file *file = iocb->ki_filp;
1100 struct vhost_vdpa *v = file->private_data;
1101 struct vhost_dev *dev = &v->vdev;
1103 return vhost_chr_write_iter(dev, from);
1106 static int vhost_vdpa_alloc_domain(struct vhost_vdpa *v)
1108 struct vdpa_device *vdpa = v->vdpa;
1109 const struct vdpa_config_ops *ops = vdpa->config;
1110 struct device *dma_dev = vdpa_get_dma_dev(vdpa);
1111 struct bus_type *bus;
1114 /* Device want to do DMA by itself */
1115 if (ops->set_map || ops->dma_map)
1122 if (!device_iommu_capable(dma_dev, IOMMU_CAP_CACHE_COHERENCY))
1125 v->domain = iommu_domain_alloc(bus);
1129 ret = iommu_attach_device(v->domain, dma_dev);
1136 iommu_domain_free(v->domain);
1141 static void vhost_vdpa_free_domain(struct vhost_vdpa *v)
1143 struct vdpa_device *vdpa = v->vdpa;
1144 struct device *dma_dev = vdpa_get_dma_dev(vdpa);
1147 iommu_detach_device(v->domain, dma_dev);
1148 iommu_domain_free(v->domain);
1154 static void vhost_vdpa_set_iova_range(struct vhost_vdpa *v)
1156 struct vdpa_iova_range *range = &v->range;
1157 struct vdpa_device *vdpa = v->vdpa;
1158 const struct vdpa_config_ops *ops = vdpa->config;
1160 if (ops->get_iova_range) {
1161 *range = ops->get_iova_range(vdpa);
1162 } else if (v->domain && v->domain->geometry.force_aperture) {
1163 range->first = v->domain->geometry.aperture_start;
1164 range->last = v->domain->geometry.aperture_end;
1167 range->last = ULLONG_MAX;
1171 static void vhost_vdpa_cleanup(struct vhost_vdpa *v)
1173 struct vhost_vdpa_as *as;
1176 for (asid = 0; asid < v->vdpa->nas; asid++) {
1177 as = asid_to_as(v, asid);
1179 vhost_vdpa_remove_as(v, asid);
1182 vhost_vdpa_free_domain(v);
1183 vhost_dev_cleanup(&v->vdev);
1187 static int vhost_vdpa_open(struct inode *inode, struct file *filep)
1189 struct vhost_vdpa *v;
1190 struct vhost_dev *dev;
1191 struct vhost_virtqueue **vqs;
1195 v = container_of(inode->i_cdev, struct vhost_vdpa, cdev);
1197 opened = atomic_cmpxchg(&v->opened, 0, 1);
1202 r = vhost_vdpa_reset(v);
1206 vqs = kmalloc_array(nvqs, sizeof(*vqs), GFP_KERNEL);
1213 for (i = 0; i < nvqs; i++) {
1214 vqs[i] = &v->vqs[i];
1215 vqs[i]->handle_kick = handle_vq_kick;
1217 vhost_dev_init(dev, vqs, nvqs, 0, 0, 0, false,
1218 vhost_vdpa_process_iotlb_msg);
1220 r = vhost_vdpa_alloc_domain(v);
1222 goto err_alloc_domain;
1224 vhost_vdpa_set_iova_range(v);
1226 filep->private_data = v;
1231 vhost_vdpa_cleanup(v);
1233 atomic_dec(&v->opened);
1237 static void vhost_vdpa_clean_irq(struct vhost_vdpa *v)
1241 for (i = 0; i < v->nvqs; i++)
1242 vhost_vdpa_unsetup_vq_irq(v, i);
1245 static int vhost_vdpa_release(struct inode *inode, struct file *filep)
1247 struct vhost_vdpa *v = filep->private_data;
1248 struct vhost_dev *d = &v->vdev;
1250 mutex_lock(&d->mutex);
1251 filep->private_data = NULL;
1252 vhost_vdpa_clean_irq(v);
1253 vhost_vdpa_reset(v);
1254 vhost_dev_stop(&v->vdev);
1255 vhost_vdpa_config_put(v);
1256 vhost_vdpa_cleanup(v);
1257 mutex_unlock(&d->mutex);
1259 atomic_dec(&v->opened);
1260 complete(&v->completion);
1266 static vm_fault_t vhost_vdpa_fault(struct vm_fault *vmf)
1268 struct vhost_vdpa *v = vmf->vma->vm_file->private_data;
1269 struct vdpa_device *vdpa = v->vdpa;
1270 const struct vdpa_config_ops *ops = vdpa->config;
1271 struct vdpa_notification_area notify;
1272 struct vm_area_struct *vma = vmf->vma;
1273 u16 index = vma->vm_pgoff;
1275 notify = ops->get_vq_notification(vdpa, index);
1277 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1278 if (remap_pfn_range(vma, vmf->address & PAGE_MASK,
1279 PFN_DOWN(notify.addr), PAGE_SIZE,
1281 return VM_FAULT_SIGBUS;
1283 return VM_FAULT_NOPAGE;
1286 static const struct vm_operations_struct vhost_vdpa_vm_ops = {
1287 .fault = vhost_vdpa_fault,
1290 static int vhost_vdpa_mmap(struct file *file, struct vm_area_struct *vma)
1292 struct vhost_vdpa *v = vma->vm_file->private_data;
1293 struct vdpa_device *vdpa = v->vdpa;
1294 const struct vdpa_config_ops *ops = vdpa->config;
1295 struct vdpa_notification_area notify;
1296 unsigned long index = vma->vm_pgoff;
1298 if (vma->vm_end - vma->vm_start != PAGE_SIZE)
1300 if ((vma->vm_flags & VM_SHARED) == 0)
1302 if (vma->vm_flags & VM_READ)
1306 if (!ops->get_vq_notification)
1309 /* To be safe and easily modelled by userspace, We only
1310 * support the doorbell which sits on the page boundary and
1311 * does not share the page with other registers.
1313 notify = ops->get_vq_notification(vdpa, index);
1314 if (notify.addr & (PAGE_SIZE - 1))
1316 if (vma->vm_end - vma->vm_start != notify.size)
1319 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
1320 vma->vm_ops = &vhost_vdpa_vm_ops;
1323 #endif /* CONFIG_MMU */
1325 static const struct file_operations vhost_vdpa_fops = {
1326 .owner = THIS_MODULE,
1327 .open = vhost_vdpa_open,
1328 .release = vhost_vdpa_release,
1329 .write_iter = vhost_vdpa_chr_write_iter,
1330 .unlocked_ioctl = vhost_vdpa_unlocked_ioctl,
1332 .mmap = vhost_vdpa_mmap,
1333 #endif /* CONFIG_MMU */
1334 .compat_ioctl = compat_ptr_ioctl,
1337 static void vhost_vdpa_release_dev(struct device *device)
1339 struct vhost_vdpa *v =
1340 container_of(device, struct vhost_vdpa, dev);
1342 ida_simple_remove(&vhost_vdpa_ida, v->minor);
1347 static int vhost_vdpa_probe(struct vdpa_device *vdpa)
1349 const struct vdpa_config_ops *ops = vdpa->config;
1350 struct vhost_vdpa *v;
1354 /* We can't support platform IOMMU device with more than 1
1357 if (!ops->set_map && !ops->dma_map &&
1358 (vdpa->ngroups > 1 || vdpa->nas > 1))
1361 v = kzalloc(sizeof(*v), GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1365 minor = ida_simple_get(&vhost_vdpa_ida, 0,
1366 VHOST_VDPA_DEV_MAX, GFP_KERNEL);
1372 atomic_set(&v->opened, 0);
1375 v->nvqs = vdpa->nvqs;
1376 v->virtio_id = ops->get_device_id(vdpa);
1378 device_initialize(&v->dev);
1379 v->dev.release = vhost_vdpa_release_dev;
1380 v->dev.parent = &vdpa->dev;
1381 v->dev.devt = MKDEV(MAJOR(vhost_vdpa_major), minor);
1382 v->vqs = kmalloc_array(v->nvqs, sizeof(struct vhost_virtqueue),
1389 r = dev_set_name(&v->dev, "vhost-vdpa-%u", minor);
1393 cdev_init(&v->cdev, &vhost_vdpa_fops);
1394 v->cdev.owner = THIS_MODULE;
1396 r = cdev_device_add(&v->cdev, &v->dev);
1400 init_completion(&v->completion);
1401 vdpa_set_drvdata(vdpa, v);
1403 for (i = 0; i < VHOST_VDPA_IOTLB_BUCKETS; i++)
1404 INIT_HLIST_HEAD(&v->as[i]);
1409 put_device(&v->dev);
1410 ida_simple_remove(&vhost_vdpa_ida, v->minor);
1414 static void vhost_vdpa_remove(struct vdpa_device *vdpa)
1416 struct vhost_vdpa *v = vdpa_get_drvdata(vdpa);
1419 cdev_device_del(&v->cdev, &v->dev);
1422 opened = atomic_cmpxchg(&v->opened, 0, 1);
1425 wait_for_completion(&v->completion);
1428 put_device(&v->dev);
1431 static struct vdpa_driver vhost_vdpa_driver = {
1433 .name = "vhost_vdpa",
1435 .probe = vhost_vdpa_probe,
1436 .remove = vhost_vdpa_remove,
1439 static int __init vhost_vdpa_init(void)
1443 r = alloc_chrdev_region(&vhost_vdpa_major, 0, VHOST_VDPA_DEV_MAX,
1446 goto err_alloc_chrdev;
1448 r = vdpa_register_driver(&vhost_vdpa_driver);
1450 goto err_vdpa_register_driver;
1454 err_vdpa_register_driver:
1455 unregister_chrdev_region(vhost_vdpa_major, VHOST_VDPA_DEV_MAX);
1459 module_init(vhost_vdpa_init);
1461 static void __exit vhost_vdpa_exit(void)
1463 vdpa_unregister_driver(&vhost_vdpa_driver);
1464 unregister_chrdev_region(vhost_vdpa_major, VHOST_VDPA_DEV_MAX);
1466 module_exit(vhost_vdpa_exit);
1468 MODULE_VERSION("0.0.1");
1469 MODULE_LICENSE("GPL v2");
1470 MODULE_AUTHOR("Intel Corporation");
1471 MODULE_DESCRIPTION("vDPA-based vhost backend for virtio");