1 // SPDX-License-Identifier: GPL-2.0
3 * Virtio driver for the paravirtualized IOMMU
5 * Copyright (C) 2019 Arm Limited
8 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
10 #include <linux/delay.h>
11 #include <linux/dma-map-ops.h>
12 #include <linux/freezer.h>
13 #include <linux/interval_tree.h>
14 #include <linux/iommu.h>
15 #include <linux/module.h>
16 #include <linux/of_platform.h>
17 #include <linux/pci.h>
18 #include <linux/virtio.h>
19 #include <linux/virtio_config.h>
20 #include <linux/virtio_ids.h>
21 #include <linux/wait.h>
23 #include <uapi/linux/virtio_iommu.h>
25 #include "dma-iommu.h"
27 #define MSI_IOVA_BASE 0x8000000
28 #define MSI_IOVA_LENGTH 0x100000
30 #define VIOMMU_REQUEST_VQ 0
31 #define VIOMMU_EVENT_VQ 1
32 #define VIOMMU_NR_VQS 2
35 struct iommu_device iommu;
37 struct virtio_device *vdev;
39 struct ida domain_ids;
41 struct virtqueue *vqs[VIOMMU_NR_VQS];
42 spinlock_t request_lock;
43 struct list_head requests;
46 /* Device configuration */
47 struct iommu_domain_geometry geometry;
51 /* Supported MAP flags */
56 struct viommu_mapping {
58 struct interval_tree_node iova;
62 struct viommu_domain {
63 struct iommu_domain domain;
64 struct viommu_dev *viommu;
65 struct mutex mutex; /* protects viommu pointer */
69 spinlock_t mappings_lock;
70 struct rb_root_cached mappings;
72 unsigned long nr_endpoints;
76 struct viommu_endpoint {
78 struct viommu_dev *viommu;
79 struct viommu_domain *vdomain;
80 struct list_head resv_regions;
83 struct viommu_request {
84 struct list_head list;
86 unsigned int write_offset;
91 #define VIOMMU_FAULT_RESV_MASK 0xffffff00
96 struct virtio_iommu_fault fault;
100 #define to_viommu_domain(domain) \
101 container_of(domain, struct viommu_domain, domain)
103 static int viommu_get_req_errno(void *buf, size_t len)
105 struct virtio_iommu_req_tail *tail = buf + len - sizeof(*tail);
107 switch (tail->status) {
108 case VIRTIO_IOMMU_S_OK:
110 case VIRTIO_IOMMU_S_UNSUPP:
112 case VIRTIO_IOMMU_S_INVAL:
114 case VIRTIO_IOMMU_S_RANGE:
116 case VIRTIO_IOMMU_S_NOENT:
118 case VIRTIO_IOMMU_S_FAULT:
120 case VIRTIO_IOMMU_S_NOMEM:
122 case VIRTIO_IOMMU_S_IOERR:
123 case VIRTIO_IOMMU_S_DEVERR:
129 static void viommu_set_req_status(void *buf, size_t len, int status)
131 struct virtio_iommu_req_tail *tail = buf + len - sizeof(*tail);
133 tail->status = status;
136 static off_t viommu_get_write_desc_offset(struct viommu_dev *viommu,
137 struct virtio_iommu_req_head *req,
140 size_t tail_size = sizeof(struct virtio_iommu_req_tail);
142 if (req->type == VIRTIO_IOMMU_T_PROBE)
143 return len - viommu->probe_size - tail_size;
145 return len - tail_size;
149 * __viommu_sync_req - Complete all in-flight requests
151 * Wait for all added requests to complete. When this function returns, all
152 * requests that were in-flight at the time of the call have completed.
154 static int __viommu_sync_req(struct viommu_dev *viommu)
158 struct viommu_request *req;
159 struct virtqueue *vq = viommu->vqs[VIOMMU_REQUEST_VQ];
161 assert_spin_locked(&viommu->request_lock);
165 while (!list_empty(&viommu->requests)) {
167 req = virtqueue_get_buf(vq, &len);
172 viommu_set_req_status(req->buf, req->len,
173 VIRTIO_IOMMU_S_IOERR);
175 write_len = req->len - req->write_offset;
176 if (req->writeback && len == write_len)
177 memcpy(req->writeback, req->buf + req->write_offset,
180 list_del(&req->list);
187 static int viommu_sync_req(struct viommu_dev *viommu)
192 spin_lock_irqsave(&viommu->request_lock, flags);
193 ret = __viommu_sync_req(viommu);
195 dev_dbg(viommu->dev, "could not sync requests (%d)\n", ret);
196 spin_unlock_irqrestore(&viommu->request_lock, flags);
202 * __viommu_add_request - Add one request to the queue
203 * @buf: pointer to the request buffer
204 * @len: length of the request buffer
205 * @writeback: copy data back to the buffer when the request completes.
207 * Add a request to the queue. Only synchronize the queue if it's already full.
208 * Otherwise don't kick the queue nor wait for requests to complete.
210 * When @writeback is true, data written by the device, including the request
211 * status, is copied into @buf after the request completes. This is unsafe if
212 * the caller allocates @buf on stack and drops the lock between add_req() and
215 * Return 0 if the request was successfully added to the queue.
217 static int __viommu_add_req(struct viommu_dev *viommu, void *buf, size_t len,
222 struct viommu_request *req;
223 struct scatterlist top_sg, bottom_sg;
224 struct scatterlist *sg[2] = { &top_sg, &bottom_sg };
225 struct virtqueue *vq = viommu->vqs[VIOMMU_REQUEST_VQ];
227 assert_spin_locked(&viommu->request_lock);
229 write_offset = viommu_get_write_desc_offset(viommu, buf, len);
230 if (write_offset <= 0)
233 req = kzalloc(sizeof(*req) + len, GFP_ATOMIC);
239 req->writeback = buf + write_offset;
240 req->write_offset = write_offset;
242 memcpy(&req->buf, buf, write_offset);
244 sg_init_one(&top_sg, req->buf, write_offset);
245 sg_init_one(&bottom_sg, req->buf + write_offset, len - write_offset);
247 ret = virtqueue_add_sgs(vq, sg, 1, 1, req, GFP_ATOMIC);
248 if (ret == -ENOSPC) {
249 /* If the queue is full, sync and retry */
250 if (!__viommu_sync_req(viommu))
251 ret = virtqueue_add_sgs(vq, sg, 1, 1, req, GFP_ATOMIC);
256 list_add_tail(&req->list, &viommu->requests);
264 static int viommu_add_req(struct viommu_dev *viommu, void *buf, size_t len)
269 spin_lock_irqsave(&viommu->request_lock, flags);
270 ret = __viommu_add_req(viommu, buf, len, false);
272 dev_dbg(viommu->dev, "could not add request: %d\n", ret);
273 spin_unlock_irqrestore(&viommu->request_lock, flags);
279 * Send a request and wait for it to complete. Return the request status (as an
282 static int viommu_send_req_sync(struct viommu_dev *viommu, void *buf,
288 spin_lock_irqsave(&viommu->request_lock, flags);
290 ret = __viommu_add_req(viommu, buf, len, true);
292 dev_dbg(viommu->dev, "could not add request (%d)\n", ret);
296 ret = __viommu_sync_req(viommu);
298 dev_dbg(viommu->dev, "could not sync requests (%d)\n", ret);
299 /* Fall-through (get the actual request status) */
302 ret = viommu_get_req_errno(buf, len);
304 spin_unlock_irqrestore(&viommu->request_lock, flags);
309 * viommu_add_mapping - add a mapping to the internal tree
311 * On success, return the new mapping. Otherwise return NULL.
313 static int viommu_add_mapping(struct viommu_domain *vdomain, u64 iova, u64 end,
314 phys_addr_t paddr, u32 flags)
316 unsigned long irqflags;
317 struct viommu_mapping *mapping;
319 mapping = kzalloc(sizeof(*mapping), GFP_ATOMIC);
323 mapping->paddr = paddr;
324 mapping->iova.start = iova;
325 mapping->iova.last = end;
326 mapping->flags = flags;
328 spin_lock_irqsave(&vdomain->mappings_lock, irqflags);
329 interval_tree_insert(&mapping->iova, &vdomain->mappings);
330 spin_unlock_irqrestore(&vdomain->mappings_lock, irqflags);
336 * viommu_del_mappings - remove mappings from the internal tree
338 * @vdomain: the domain
339 * @iova: start of the range
340 * @end: end of the range
342 * On success, returns the number of unmapped bytes
344 static size_t viommu_del_mappings(struct viommu_domain *vdomain,
349 struct viommu_mapping *mapping = NULL;
350 struct interval_tree_node *node, *next;
352 spin_lock_irqsave(&vdomain->mappings_lock, flags);
353 next = interval_tree_iter_first(&vdomain->mappings, iova, end);
356 mapping = container_of(node, struct viommu_mapping, iova);
357 next = interval_tree_iter_next(node, iova, end);
359 /* Trying to split a mapping? */
360 if (mapping->iova.start < iova)
364 * Virtio-iommu doesn't allow UNMAP to split a mapping created
365 * with a single MAP request, so remove the full mapping.
367 unmapped += mapping->iova.last - mapping->iova.start + 1;
369 interval_tree_remove(node, &vdomain->mappings);
372 spin_unlock_irqrestore(&vdomain->mappings_lock, flags);
378 * Fill the domain with identity mappings, skipping the device's reserved
381 static int viommu_domain_map_identity(struct viommu_endpoint *vdev,
382 struct viommu_domain *vdomain)
385 struct iommu_resv_region *resv;
386 u64 iova = vdomain->domain.geometry.aperture_start;
387 u64 limit = vdomain->domain.geometry.aperture_end;
388 u32 flags = VIRTIO_IOMMU_MAP_F_READ | VIRTIO_IOMMU_MAP_F_WRITE;
389 unsigned long granule = 1UL << __ffs(vdomain->domain.pgsize_bitmap);
391 iova = ALIGN(iova, granule);
392 limit = ALIGN_DOWN(limit + 1, granule) - 1;
394 list_for_each_entry(resv, &vdev->resv_regions, list) {
395 u64 resv_start = ALIGN_DOWN(resv->start, granule);
396 u64 resv_end = ALIGN(resv->start + resv->length, granule) - 1;
398 if (resv_end < iova || resv_start > limit)
402 if (resv_start > iova) {
403 ret = viommu_add_mapping(vdomain, iova, resv_start - 1,
404 (phys_addr_t)iova, flags);
409 if (resv_end >= limit)
415 ret = viommu_add_mapping(vdomain, iova, limit, (phys_addr_t)iova,
422 viommu_del_mappings(vdomain, 0, iova);
427 * viommu_replay_mappings - re-send MAP requests
429 * When reattaching a domain that was previously detached from all endpoints,
430 * mappings were deleted from the device. Re-create the mappings available in
433 static int viommu_replay_mappings(struct viommu_domain *vdomain)
437 struct viommu_mapping *mapping;
438 struct interval_tree_node *node;
439 struct virtio_iommu_req_map map;
441 spin_lock_irqsave(&vdomain->mappings_lock, flags);
442 node = interval_tree_iter_first(&vdomain->mappings, 0, -1UL);
444 mapping = container_of(node, struct viommu_mapping, iova);
445 map = (struct virtio_iommu_req_map) {
446 .head.type = VIRTIO_IOMMU_T_MAP,
447 .domain = cpu_to_le32(vdomain->id),
448 .virt_start = cpu_to_le64(mapping->iova.start),
449 .virt_end = cpu_to_le64(mapping->iova.last),
450 .phys_start = cpu_to_le64(mapping->paddr),
451 .flags = cpu_to_le32(mapping->flags),
454 ret = viommu_send_req_sync(vdomain->viommu, &map, sizeof(map));
458 node = interval_tree_iter_next(node, 0, -1UL);
460 spin_unlock_irqrestore(&vdomain->mappings_lock, flags);
465 static int viommu_add_resv_mem(struct viommu_endpoint *vdev,
466 struct virtio_iommu_probe_resv_mem *mem,
471 phys_addr_t start, end;
472 struct iommu_resv_region *region = NULL, *next;
473 unsigned long prot = IOMMU_WRITE | IOMMU_NOEXEC | IOMMU_MMIO;
475 start = start64 = le64_to_cpu(mem->start);
476 end = end64 = le64_to_cpu(mem->end);
477 size = end64 - start64 + 1;
479 /* Catch any overflow, including the unlikely end64 - start64 + 1 = 0 */
480 if (start != start64 || end != end64 || size < end64 - start64)
483 if (len < sizeof(*mem))
486 switch (mem->subtype) {
488 dev_warn(vdev->dev, "unknown resv mem subtype 0x%x\n",
491 case VIRTIO_IOMMU_RESV_MEM_T_RESERVED:
492 region = iommu_alloc_resv_region(start, size, 0,
496 case VIRTIO_IOMMU_RESV_MEM_T_MSI:
497 region = iommu_alloc_resv_region(start, size, prot,
505 /* Keep the list sorted */
506 list_for_each_entry(next, &vdev->resv_regions, list) {
507 if (next->start > region->start)
510 list_add_tail(®ion->list, &next->list);
514 static int viommu_probe_endpoint(struct viommu_dev *viommu, struct device *dev)
520 struct virtio_iommu_req_probe *probe;
521 struct virtio_iommu_probe_property *prop;
522 struct iommu_fwspec *fwspec = dev_iommu_fwspec_get(dev);
523 struct viommu_endpoint *vdev = dev_iommu_priv_get(dev);
525 if (!fwspec->num_ids)
528 probe_len = sizeof(*probe) + viommu->probe_size +
529 sizeof(struct virtio_iommu_req_tail);
530 probe = kzalloc(probe_len, GFP_KERNEL);
534 probe->head.type = VIRTIO_IOMMU_T_PROBE;
536 * For now, assume that properties of an endpoint that outputs multiple
537 * IDs are consistent. Only probe the first one.
539 probe->endpoint = cpu_to_le32(fwspec->ids[0]);
541 ret = viommu_send_req_sync(viommu, probe, probe_len);
545 prop = (void *)probe->properties;
546 type = le16_to_cpu(prop->type) & VIRTIO_IOMMU_PROBE_T_MASK;
548 while (type != VIRTIO_IOMMU_PROBE_T_NONE &&
549 cur < viommu->probe_size) {
550 len = le16_to_cpu(prop->length) + sizeof(*prop);
553 case VIRTIO_IOMMU_PROBE_T_RESV_MEM:
554 ret = viommu_add_resv_mem(vdev, (void *)prop, len);
557 dev_err(dev, "unknown viommu prop 0x%x\n", type);
561 dev_err(dev, "failed to parse viommu prop 0x%x\n", type);
564 if (cur >= viommu->probe_size)
567 prop = (void *)probe->properties + cur;
568 type = le16_to_cpu(prop->type) & VIRTIO_IOMMU_PROBE_T_MASK;
576 static int viommu_fault_handler(struct viommu_dev *viommu,
577 struct virtio_iommu_fault *fault)
581 u8 reason = fault->reason;
582 u32 flags = le32_to_cpu(fault->flags);
583 u32 endpoint = le32_to_cpu(fault->endpoint);
584 u64 address = le64_to_cpu(fault->address);
587 case VIRTIO_IOMMU_FAULT_R_DOMAIN:
588 reason_str = "domain";
590 case VIRTIO_IOMMU_FAULT_R_MAPPING:
593 case VIRTIO_IOMMU_FAULT_R_UNKNOWN:
595 reason_str = "unknown";
599 /* TODO: find EP by ID and report_iommu_fault */
600 if (flags & VIRTIO_IOMMU_FAULT_F_ADDRESS)
601 dev_err_ratelimited(viommu->dev, "%s fault from EP %u at %#llx [%s%s%s]\n",
602 reason_str, endpoint, address,
603 flags & VIRTIO_IOMMU_FAULT_F_READ ? "R" : "",
604 flags & VIRTIO_IOMMU_FAULT_F_WRITE ? "W" : "",
605 flags & VIRTIO_IOMMU_FAULT_F_EXEC ? "X" : "");
607 dev_err_ratelimited(viommu->dev, "%s fault from EP %u\n",
608 reason_str, endpoint);
612 static void viommu_event_handler(struct virtqueue *vq)
616 struct scatterlist sg[1];
617 struct viommu_event *evt;
618 struct viommu_dev *viommu = vq->vdev->priv;
620 while ((evt = virtqueue_get_buf(vq, &len)) != NULL) {
621 if (len > sizeof(*evt)) {
623 "invalid event buffer (len %u != %zu)\n",
625 } else if (!(evt->head & VIOMMU_FAULT_RESV_MASK)) {
626 viommu_fault_handler(viommu, &evt->fault);
629 sg_init_one(sg, evt, sizeof(*evt));
630 ret = virtqueue_add_inbuf(vq, sg, 1, evt, GFP_ATOMIC);
632 dev_err(viommu->dev, "could not add event buffer\n");
640 static struct iommu_domain *viommu_domain_alloc(unsigned type)
642 struct viommu_domain *vdomain;
644 if (type != IOMMU_DOMAIN_UNMANAGED &&
645 type != IOMMU_DOMAIN_DMA &&
646 type != IOMMU_DOMAIN_IDENTITY)
649 vdomain = kzalloc(sizeof(*vdomain), GFP_KERNEL);
653 mutex_init(&vdomain->mutex);
654 spin_lock_init(&vdomain->mappings_lock);
655 vdomain->mappings = RB_ROOT_CACHED;
657 return &vdomain->domain;
660 static int viommu_domain_finalise(struct viommu_endpoint *vdev,
661 struct iommu_domain *domain)
664 unsigned long viommu_page_size;
665 struct viommu_dev *viommu = vdev->viommu;
666 struct viommu_domain *vdomain = to_viommu_domain(domain);
668 viommu_page_size = 1UL << __ffs(viommu->pgsize_bitmap);
669 if (viommu_page_size > PAGE_SIZE) {
671 "granule 0x%lx larger than system page size 0x%lx\n",
672 viommu_page_size, PAGE_SIZE);
676 ret = ida_alloc_range(&viommu->domain_ids, viommu->first_domain,
677 viommu->last_domain, GFP_KERNEL);
681 vdomain->id = (unsigned int)ret;
683 domain->pgsize_bitmap = viommu->pgsize_bitmap;
684 domain->geometry = viommu->geometry;
686 vdomain->map_flags = viommu->map_flags;
687 vdomain->viommu = viommu;
689 if (domain->type == IOMMU_DOMAIN_IDENTITY) {
690 if (virtio_has_feature(viommu->vdev,
691 VIRTIO_IOMMU_F_BYPASS_CONFIG)) {
692 vdomain->bypass = true;
696 ret = viommu_domain_map_identity(vdev, vdomain);
698 ida_free(&viommu->domain_ids, vdomain->id);
699 vdomain->viommu = NULL;
707 static void viommu_domain_free(struct iommu_domain *domain)
709 struct viommu_domain *vdomain = to_viommu_domain(domain);
711 /* Free all remaining mappings */
712 viommu_del_mappings(vdomain, 0, ULLONG_MAX);
715 ida_free(&vdomain->viommu->domain_ids, vdomain->id);
720 static int viommu_attach_dev(struct iommu_domain *domain, struct device *dev)
724 struct virtio_iommu_req_attach req;
725 struct iommu_fwspec *fwspec = dev_iommu_fwspec_get(dev);
726 struct viommu_endpoint *vdev = dev_iommu_priv_get(dev);
727 struct viommu_domain *vdomain = to_viommu_domain(domain);
729 mutex_lock(&vdomain->mutex);
730 if (!vdomain->viommu) {
732 * Properly initialize the domain now that we know which viommu
735 ret = viommu_domain_finalise(vdev, domain);
736 } else if (vdomain->viommu != vdev->viommu) {
737 dev_err(dev, "cannot attach to foreign vIOMMU\n");
740 mutex_unlock(&vdomain->mutex);
746 * In the virtio-iommu device, when attaching the endpoint to a new
747 * domain, it is detached from the old one and, if as a result the
748 * old domain isn't attached to any endpoint, all mappings are removed
749 * from the old domain and it is freed.
751 * In the driver the old domain still exists, and its mappings will be
752 * recreated if it gets reattached to an endpoint. Otherwise it will be
755 * vdev->vdomain is protected by group->mutex
758 vdev->vdomain->nr_endpoints--;
760 req = (struct virtio_iommu_req_attach) {
761 .head.type = VIRTIO_IOMMU_T_ATTACH,
762 .domain = cpu_to_le32(vdomain->id),
766 req.flags |= cpu_to_le32(VIRTIO_IOMMU_ATTACH_F_BYPASS);
768 for (i = 0; i < fwspec->num_ids; i++) {
769 req.endpoint = cpu_to_le32(fwspec->ids[i]);
771 ret = viommu_send_req_sync(vdomain->viommu, &req, sizeof(req));
776 if (!vdomain->nr_endpoints) {
778 * This endpoint is the first to be attached to the domain.
779 * Replay existing mappings (e.g. SW MSI).
781 ret = viommu_replay_mappings(vdomain);
786 vdomain->nr_endpoints++;
787 vdev->vdomain = vdomain;
792 static int viommu_map_pages(struct iommu_domain *domain, unsigned long iova,
793 phys_addr_t paddr, size_t pgsize, size_t pgcount,
794 int prot, gfp_t gfp, size_t *mapped)
798 size_t size = pgsize * pgcount;
799 u64 end = iova + size - 1;
800 struct virtio_iommu_req_map map;
801 struct viommu_domain *vdomain = to_viommu_domain(domain);
803 flags = (prot & IOMMU_READ ? VIRTIO_IOMMU_MAP_F_READ : 0) |
804 (prot & IOMMU_WRITE ? VIRTIO_IOMMU_MAP_F_WRITE : 0) |
805 (prot & IOMMU_MMIO ? VIRTIO_IOMMU_MAP_F_MMIO : 0);
807 if (flags & ~vdomain->map_flags)
810 ret = viommu_add_mapping(vdomain, iova, end, paddr, flags);
814 map = (struct virtio_iommu_req_map) {
815 .head.type = VIRTIO_IOMMU_T_MAP,
816 .domain = cpu_to_le32(vdomain->id),
817 .virt_start = cpu_to_le64(iova),
818 .phys_start = cpu_to_le64(paddr),
819 .virt_end = cpu_to_le64(end),
820 .flags = cpu_to_le32(flags),
823 if (!vdomain->nr_endpoints)
826 ret = viommu_send_req_sync(vdomain->viommu, &map, sizeof(map));
828 viommu_del_mappings(vdomain, iova, end);
835 static size_t viommu_unmap_pages(struct iommu_domain *domain, unsigned long iova,
836 size_t pgsize, size_t pgcount,
837 struct iommu_iotlb_gather *gather)
841 struct virtio_iommu_req_unmap unmap;
842 struct viommu_domain *vdomain = to_viommu_domain(domain);
843 size_t size = pgsize * pgcount;
845 unmapped = viommu_del_mappings(vdomain, iova, iova + size - 1);
849 /* Device already removed all mappings after detach. */
850 if (!vdomain->nr_endpoints)
853 unmap = (struct virtio_iommu_req_unmap) {
854 .head.type = VIRTIO_IOMMU_T_UNMAP,
855 .domain = cpu_to_le32(vdomain->id),
856 .virt_start = cpu_to_le64(iova),
857 .virt_end = cpu_to_le64(iova + unmapped - 1),
860 ret = viommu_add_req(vdomain->viommu, &unmap, sizeof(unmap));
861 return ret ? 0 : unmapped;
864 static phys_addr_t viommu_iova_to_phys(struct iommu_domain *domain,
869 struct viommu_mapping *mapping;
870 struct interval_tree_node *node;
871 struct viommu_domain *vdomain = to_viommu_domain(domain);
873 spin_lock_irqsave(&vdomain->mappings_lock, flags);
874 node = interval_tree_iter_first(&vdomain->mappings, iova, iova);
876 mapping = container_of(node, struct viommu_mapping, iova);
877 paddr = mapping->paddr + (iova - mapping->iova.start);
879 spin_unlock_irqrestore(&vdomain->mappings_lock, flags);
884 static void viommu_iotlb_sync(struct iommu_domain *domain,
885 struct iommu_iotlb_gather *gather)
887 struct viommu_domain *vdomain = to_viommu_domain(domain);
889 viommu_sync_req(vdomain->viommu);
892 static void viommu_get_resv_regions(struct device *dev, struct list_head *head)
894 struct iommu_resv_region *entry, *new_entry, *msi = NULL;
895 struct viommu_endpoint *vdev = dev_iommu_priv_get(dev);
896 int prot = IOMMU_WRITE | IOMMU_NOEXEC | IOMMU_MMIO;
898 list_for_each_entry(entry, &vdev->resv_regions, list) {
899 if (entry->type == IOMMU_RESV_MSI)
902 new_entry = kmemdup(entry, sizeof(*entry), GFP_KERNEL);
905 list_add_tail(&new_entry->list, head);
909 * If the device didn't register any bypass MSI window, add a
910 * software-mapped region.
913 msi = iommu_alloc_resv_region(MSI_IOVA_BASE, MSI_IOVA_LENGTH,
914 prot, IOMMU_RESV_SW_MSI,
919 list_add_tail(&msi->list, head);
922 iommu_dma_get_resv_regions(dev, head);
925 static struct iommu_ops viommu_ops;
926 static struct virtio_driver virtio_iommu_drv;
928 static int viommu_match_node(struct device *dev, const void *data)
930 return device_match_fwnode(dev->parent, data);
933 static struct viommu_dev *viommu_get_by_fwnode(struct fwnode_handle *fwnode)
935 struct device *dev = driver_find_device(&virtio_iommu_drv.driver, NULL,
936 fwnode, viommu_match_node);
939 return dev ? dev_to_virtio(dev)->priv : NULL;
942 static struct iommu_device *viommu_probe_device(struct device *dev)
945 struct viommu_endpoint *vdev;
946 struct viommu_dev *viommu = NULL;
947 struct iommu_fwspec *fwspec = dev_iommu_fwspec_get(dev);
949 if (!fwspec || fwspec->ops != &viommu_ops)
950 return ERR_PTR(-ENODEV);
952 viommu = viommu_get_by_fwnode(fwspec->iommu_fwnode);
954 return ERR_PTR(-ENODEV);
956 vdev = kzalloc(sizeof(*vdev), GFP_KERNEL);
958 return ERR_PTR(-ENOMEM);
961 vdev->viommu = viommu;
962 INIT_LIST_HEAD(&vdev->resv_regions);
963 dev_iommu_priv_set(dev, vdev);
965 if (viommu->probe_size) {
966 /* Get additional information for this endpoint */
967 ret = viommu_probe_endpoint(viommu, dev);
972 return &viommu->iommu;
975 iommu_put_resv_regions(dev, &vdev->resv_regions);
981 static void viommu_probe_finalize(struct device *dev)
983 #ifndef CONFIG_ARCH_HAS_SETUP_DMA_OPS
984 /* First clear the DMA ops in case we're switching from a DMA domain */
985 set_dma_ops(dev, NULL);
986 iommu_setup_dma_ops(dev, 0, U64_MAX);
990 static void viommu_release_device(struct device *dev)
992 struct viommu_endpoint *vdev = dev_iommu_priv_get(dev);
994 iommu_put_resv_regions(dev, &vdev->resv_regions);
998 static struct iommu_group *viommu_device_group(struct device *dev)
1000 if (dev_is_pci(dev))
1001 return pci_device_group(dev);
1003 return generic_device_group(dev);
1006 static int viommu_of_xlate(struct device *dev, struct of_phandle_args *args)
1008 return iommu_fwspec_add_ids(dev, args->args, 1);
1011 static bool viommu_capable(struct device *dev, enum iommu_cap cap)
1014 case IOMMU_CAP_CACHE_COHERENCY:
1021 static struct iommu_ops viommu_ops = {
1022 .capable = viommu_capable,
1023 .domain_alloc = viommu_domain_alloc,
1024 .probe_device = viommu_probe_device,
1025 .probe_finalize = viommu_probe_finalize,
1026 .release_device = viommu_release_device,
1027 .device_group = viommu_device_group,
1028 .get_resv_regions = viommu_get_resv_regions,
1029 .of_xlate = viommu_of_xlate,
1030 .owner = THIS_MODULE,
1031 .default_domain_ops = &(const struct iommu_domain_ops) {
1032 .attach_dev = viommu_attach_dev,
1033 .map_pages = viommu_map_pages,
1034 .unmap_pages = viommu_unmap_pages,
1035 .iova_to_phys = viommu_iova_to_phys,
1036 .iotlb_sync = viommu_iotlb_sync,
1037 .free = viommu_domain_free,
1041 static int viommu_init_vqs(struct viommu_dev *viommu)
1043 struct virtio_device *vdev = dev_to_virtio(viommu->dev);
1044 const char *names[] = { "request", "event" };
1045 vq_callback_t *callbacks[] = {
1046 NULL, /* No async requests */
1047 viommu_event_handler,
1050 return virtio_find_vqs(vdev, VIOMMU_NR_VQS, viommu->vqs, callbacks,
1054 static int viommu_fill_evtq(struct viommu_dev *viommu)
1057 struct scatterlist sg[1];
1058 struct viommu_event *evts;
1059 struct virtqueue *vq = viommu->vqs[VIOMMU_EVENT_VQ];
1060 size_t nr_evts = vq->num_free;
1062 viommu->evts = evts = devm_kmalloc_array(viommu->dev, nr_evts,
1063 sizeof(*evts), GFP_KERNEL);
1067 for (i = 0; i < nr_evts; i++) {
1068 sg_init_one(sg, &evts[i], sizeof(*evts));
1069 ret = virtqueue_add_inbuf(vq, sg, 1, &evts[i], GFP_KERNEL);
1077 static int viommu_probe(struct virtio_device *vdev)
1079 struct device *parent_dev = vdev->dev.parent;
1080 struct viommu_dev *viommu = NULL;
1081 struct device *dev = &vdev->dev;
1082 u64 input_start = 0;
1083 u64 input_end = -1UL;
1086 if (!virtio_has_feature(vdev, VIRTIO_F_VERSION_1) ||
1087 !virtio_has_feature(vdev, VIRTIO_IOMMU_F_MAP_UNMAP))
1090 viommu = devm_kzalloc(dev, sizeof(*viommu), GFP_KERNEL);
1094 spin_lock_init(&viommu->request_lock);
1095 ida_init(&viommu->domain_ids);
1097 viommu->vdev = vdev;
1098 INIT_LIST_HEAD(&viommu->requests);
1100 ret = viommu_init_vqs(viommu);
1104 virtio_cread_le(vdev, struct virtio_iommu_config, page_size_mask,
1105 &viommu->pgsize_bitmap);
1107 if (!viommu->pgsize_bitmap) {
1112 viommu->map_flags = VIRTIO_IOMMU_MAP_F_READ | VIRTIO_IOMMU_MAP_F_WRITE;
1113 viommu->last_domain = ~0U;
1115 /* Optional features */
1116 virtio_cread_le_feature(vdev, VIRTIO_IOMMU_F_INPUT_RANGE,
1117 struct virtio_iommu_config, input_range.start,
1120 virtio_cread_le_feature(vdev, VIRTIO_IOMMU_F_INPUT_RANGE,
1121 struct virtio_iommu_config, input_range.end,
1124 virtio_cread_le_feature(vdev, VIRTIO_IOMMU_F_DOMAIN_RANGE,
1125 struct virtio_iommu_config, domain_range.start,
1126 &viommu->first_domain);
1128 virtio_cread_le_feature(vdev, VIRTIO_IOMMU_F_DOMAIN_RANGE,
1129 struct virtio_iommu_config, domain_range.end,
1130 &viommu->last_domain);
1132 virtio_cread_le_feature(vdev, VIRTIO_IOMMU_F_PROBE,
1133 struct virtio_iommu_config, probe_size,
1134 &viommu->probe_size);
1136 viommu->geometry = (struct iommu_domain_geometry) {
1137 .aperture_start = input_start,
1138 .aperture_end = input_end,
1139 .force_aperture = true,
1142 if (virtio_has_feature(vdev, VIRTIO_IOMMU_F_MMIO))
1143 viommu->map_flags |= VIRTIO_IOMMU_MAP_F_MMIO;
1145 viommu_ops.pgsize_bitmap = viommu->pgsize_bitmap;
1147 virtio_device_ready(vdev);
1149 /* Populate the event queue with buffers */
1150 ret = viommu_fill_evtq(viommu);
1154 ret = iommu_device_sysfs_add(&viommu->iommu, dev, NULL, "%s",
1155 virtio_bus_name(vdev));
1159 iommu_device_register(&viommu->iommu, &viommu_ops, parent_dev);
1161 vdev->priv = viommu;
1163 dev_info(dev, "input address: %u bits\n",
1164 order_base_2(viommu->geometry.aperture_end));
1165 dev_info(dev, "page mask: %#llx\n", viommu->pgsize_bitmap);
1170 vdev->config->del_vqs(vdev);
1175 static void viommu_remove(struct virtio_device *vdev)
1177 struct viommu_dev *viommu = vdev->priv;
1179 iommu_device_sysfs_remove(&viommu->iommu);
1180 iommu_device_unregister(&viommu->iommu);
1182 /* Stop all virtqueues */
1183 virtio_reset_device(vdev);
1184 vdev->config->del_vqs(vdev);
1186 dev_info(&vdev->dev, "device removed\n");
1189 static void viommu_config_changed(struct virtio_device *vdev)
1191 dev_warn(&vdev->dev, "config changed\n");
1194 static unsigned int features[] = {
1195 VIRTIO_IOMMU_F_MAP_UNMAP,
1196 VIRTIO_IOMMU_F_INPUT_RANGE,
1197 VIRTIO_IOMMU_F_DOMAIN_RANGE,
1198 VIRTIO_IOMMU_F_PROBE,
1199 VIRTIO_IOMMU_F_MMIO,
1200 VIRTIO_IOMMU_F_BYPASS_CONFIG,
1203 static struct virtio_device_id id_table[] = {
1204 { VIRTIO_ID_IOMMU, VIRTIO_DEV_ANY_ID },
1207 MODULE_DEVICE_TABLE(virtio, id_table);
1209 static struct virtio_driver virtio_iommu_drv = {
1210 .driver.name = KBUILD_MODNAME,
1211 .driver.owner = THIS_MODULE,
1212 .id_table = id_table,
1213 .feature_table = features,
1214 .feature_table_size = ARRAY_SIZE(features),
1215 .probe = viommu_probe,
1216 .remove = viommu_remove,
1217 .config_changed = viommu_config_changed,
1220 module_virtio_driver(virtio_iommu_drv);
1222 MODULE_DESCRIPTION("Virtio IOMMU driver");
1223 MODULE_AUTHOR("Jean-Philippe Brucker <jean-philippe.brucker@arm.com>");
1224 MODULE_LICENSE("GPL v2");