Back to home page

OSCL-LXR
 
 

     

0001 // SPDX-License-Identifier: GPL-2.0
0002 /*
0003  * Virtio driver for the paravirtualized IOMMU
0004  *
0005  * Copyright (C) 2019 Arm Limited
0006  */
0007 
0008 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
0009 
0010 #include <linux/amba/bus.h>
0011 #include <linux/delay.h>
0012 #include <linux/dma-iommu.h>
0013 #include <linux/dma-map-ops.h>
0014 #include <linux/freezer.h>
0015 #include <linux/interval_tree.h>
0016 #include <linux/iommu.h>
0017 #include <linux/module.h>
0018 #include <linux/of_platform.h>
0019 #include <linux/pci.h>
0020 #include <linux/platform_device.h>
0021 #include <linux/virtio.h>
0022 #include <linux/virtio_config.h>
0023 #include <linux/virtio_ids.h>
0024 #include <linux/wait.h>
0025 
0026 #include <uapi/linux/virtio_iommu.h>
0027 
0028 #define MSI_IOVA_BASE           0x8000000
0029 #define MSI_IOVA_LENGTH         0x100000
0030 
0031 #define VIOMMU_REQUEST_VQ       0
0032 #define VIOMMU_EVENT_VQ         1
0033 #define VIOMMU_NR_VQS           2
0034 
0035 struct viommu_dev {
0036     struct iommu_device     iommu;
0037     struct device           *dev;
0038     struct virtio_device        *vdev;
0039 
0040     struct ida          domain_ids;
0041 
0042     struct virtqueue        *vqs[VIOMMU_NR_VQS];
0043     spinlock_t          request_lock;
0044     struct list_head        requests;
0045     void                *evts;
0046 
0047     /* Device configuration */
0048     struct iommu_domain_geometry    geometry;
0049     u64             pgsize_bitmap;
0050     u32             first_domain;
0051     u32             last_domain;
0052     /* Supported MAP flags */
0053     u32             map_flags;
0054     u32             probe_size;
0055 };
0056 
0057 struct viommu_mapping {
0058     phys_addr_t         paddr;
0059     struct interval_tree_node   iova;
0060     u32             flags;
0061 };
0062 
0063 struct viommu_domain {
0064     struct iommu_domain     domain;
0065     struct viommu_dev       *viommu;
0066     struct mutex            mutex; /* protects viommu pointer */
0067     unsigned int            id;
0068     u32             map_flags;
0069 
0070     spinlock_t          mappings_lock;
0071     struct rb_root_cached       mappings;
0072 
0073     unsigned long           nr_endpoints;
0074     bool                bypass;
0075 };
0076 
0077 struct viommu_endpoint {
0078     struct device           *dev;
0079     struct viommu_dev       *viommu;
0080     struct viommu_domain        *vdomain;
0081     struct list_head        resv_regions;
0082 };
0083 
0084 struct viommu_request {
0085     struct list_head        list;
0086     void                *writeback;
0087     unsigned int            write_offset;
0088     unsigned int            len;
0089     char                buf[];
0090 };
0091 
0092 #define VIOMMU_FAULT_RESV_MASK      0xffffff00
0093 
0094 struct viommu_event {
0095     union {
0096         u32         head;
0097         struct virtio_iommu_fault fault;
0098     };
0099 };
0100 
0101 #define to_viommu_domain(domain)    \
0102     container_of(domain, struct viommu_domain, domain)
0103 
0104 static int viommu_get_req_errno(void *buf, size_t len)
0105 {
0106     struct virtio_iommu_req_tail *tail = buf + len - sizeof(*tail);
0107 
0108     switch (tail->status) {
0109     case VIRTIO_IOMMU_S_OK:
0110         return 0;
0111     case VIRTIO_IOMMU_S_UNSUPP:
0112         return -ENOSYS;
0113     case VIRTIO_IOMMU_S_INVAL:
0114         return -EINVAL;
0115     case VIRTIO_IOMMU_S_RANGE:
0116         return -ERANGE;
0117     case VIRTIO_IOMMU_S_NOENT:
0118         return -ENOENT;
0119     case VIRTIO_IOMMU_S_FAULT:
0120         return -EFAULT;
0121     case VIRTIO_IOMMU_S_NOMEM:
0122         return -ENOMEM;
0123     case VIRTIO_IOMMU_S_IOERR:
0124     case VIRTIO_IOMMU_S_DEVERR:
0125     default:
0126         return -EIO;
0127     }
0128 }
0129 
0130 static void viommu_set_req_status(void *buf, size_t len, int status)
0131 {
0132     struct virtio_iommu_req_tail *tail = buf + len - sizeof(*tail);
0133 
0134     tail->status = status;
0135 }
0136 
0137 static off_t viommu_get_write_desc_offset(struct viommu_dev *viommu,
0138                       struct virtio_iommu_req_head *req,
0139                       size_t len)
0140 {
0141     size_t tail_size = sizeof(struct virtio_iommu_req_tail);
0142 
0143     if (req->type == VIRTIO_IOMMU_T_PROBE)
0144         return len - viommu->probe_size - tail_size;
0145 
0146     return len - tail_size;
0147 }
0148 
0149 /*
0150  * __viommu_sync_req - Complete all in-flight requests
0151  *
0152  * Wait for all added requests to complete. When this function returns, all
0153  * requests that were in-flight at the time of the call have completed.
0154  */
0155 static int __viommu_sync_req(struct viommu_dev *viommu)
0156 {
0157     unsigned int len;
0158     size_t write_len;
0159     struct viommu_request *req;
0160     struct virtqueue *vq = viommu->vqs[VIOMMU_REQUEST_VQ];
0161 
0162     assert_spin_locked(&viommu->request_lock);
0163 
0164     virtqueue_kick(vq);
0165 
0166     while (!list_empty(&viommu->requests)) {
0167         len = 0;
0168         req = virtqueue_get_buf(vq, &len);
0169         if (!req)
0170             continue;
0171 
0172         if (!len)
0173             viommu_set_req_status(req->buf, req->len,
0174                           VIRTIO_IOMMU_S_IOERR);
0175 
0176         write_len = req->len - req->write_offset;
0177         if (req->writeback && len == write_len)
0178             memcpy(req->writeback, req->buf + req->write_offset,
0179                    write_len);
0180 
0181         list_del(&req->list);
0182         kfree(req);
0183     }
0184 
0185     return 0;
0186 }
0187 
0188 static int viommu_sync_req(struct viommu_dev *viommu)
0189 {
0190     int ret;
0191     unsigned long flags;
0192 
0193     spin_lock_irqsave(&viommu->request_lock, flags);
0194     ret = __viommu_sync_req(viommu);
0195     if (ret)
0196         dev_dbg(viommu->dev, "could not sync requests (%d)\n", ret);
0197     spin_unlock_irqrestore(&viommu->request_lock, flags);
0198 
0199     return ret;
0200 }
0201 
0202 /*
0203  * __viommu_add_request - Add one request to the queue
0204  * @buf: pointer to the request buffer
0205  * @len: length of the request buffer
0206  * @writeback: copy data back to the buffer when the request completes.
0207  *
0208  * Add a request to the queue. Only synchronize the queue if it's already full.
0209  * Otherwise don't kick the queue nor wait for requests to complete.
0210  *
0211  * When @writeback is true, data written by the device, including the request
0212  * status, is copied into @buf after the request completes. This is unsafe if
0213  * the caller allocates @buf on stack and drops the lock between add_req() and
0214  * sync_req().
0215  *
0216  * Return 0 if the request was successfully added to the queue.
0217  */
0218 static int __viommu_add_req(struct viommu_dev *viommu, void *buf, size_t len,
0219                 bool writeback)
0220 {
0221     int ret;
0222     off_t write_offset;
0223     struct viommu_request *req;
0224     struct scatterlist top_sg, bottom_sg;
0225     struct scatterlist *sg[2] = { &top_sg, &bottom_sg };
0226     struct virtqueue *vq = viommu->vqs[VIOMMU_REQUEST_VQ];
0227 
0228     assert_spin_locked(&viommu->request_lock);
0229 
0230     write_offset = viommu_get_write_desc_offset(viommu, buf, len);
0231     if (write_offset <= 0)
0232         return -EINVAL;
0233 
0234     req = kzalloc(sizeof(*req) + len, GFP_ATOMIC);
0235     if (!req)
0236         return -ENOMEM;
0237 
0238     req->len = len;
0239     if (writeback) {
0240         req->writeback = buf + write_offset;
0241         req->write_offset = write_offset;
0242     }
0243     memcpy(&req->buf, buf, write_offset);
0244 
0245     sg_init_one(&top_sg, req->buf, write_offset);
0246     sg_init_one(&bottom_sg, req->buf + write_offset, len - write_offset);
0247 
0248     ret = virtqueue_add_sgs(vq, sg, 1, 1, req, GFP_ATOMIC);
0249     if (ret == -ENOSPC) {
0250         /* If the queue is full, sync and retry */
0251         if (!__viommu_sync_req(viommu))
0252             ret = virtqueue_add_sgs(vq, sg, 1, 1, req, GFP_ATOMIC);
0253     }
0254     if (ret)
0255         goto err_free;
0256 
0257     list_add_tail(&req->list, &viommu->requests);
0258     return 0;
0259 
0260 err_free:
0261     kfree(req);
0262     return ret;
0263 }
0264 
0265 static int viommu_add_req(struct viommu_dev *viommu, void *buf, size_t len)
0266 {
0267     int ret;
0268     unsigned long flags;
0269 
0270     spin_lock_irqsave(&viommu->request_lock, flags);
0271     ret = __viommu_add_req(viommu, buf, len, false);
0272     if (ret)
0273         dev_dbg(viommu->dev, "could not add request: %d\n", ret);
0274     spin_unlock_irqrestore(&viommu->request_lock, flags);
0275 
0276     return ret;
0277 }
0278 
0279 /*
0280  * Send a request and wait for it to complete. Return the request status (as an
0281  * errno)
0282  */
0283 static int viommu_send_req_sync(struct viommu_dev *viommu, void *buf,
0284                 size_t len)
0285 {
0286     int ret;
0287     unsigned long flags;
0288 
0289     spin_lock_irqsave(&viommu->request_lock, flags);
0290 
0291     ret = __viommu_add_req(viommu, buf, len, true);
0292     if (ret) {
0293         dev_dbg(viommu->dev, "could not add request (%d)\n", ret);
0294         goto out_unlock;
0295     }
0296 
0297     ret = __viommu_sync_req(viommu);
0298     if (ret) {
0299         dev_dbg(viommu->dev, "could not sync requests (%d)\n", ret);
0300         /* Fall-through (get the actual request status) */
0301     }
0302 
0303     ret = viommu_get_req_errno(buf, len);
0304 out_unlock:
0305     spin_unlock_irqrestore(&viommu->request_lock, flags);
0306     return ret;
0307 }
0308 
0309 /*
0310  * viommu_add_mapping - add a mapping to the internal tree
0311  *
0312  * On success, return the new mapping. Otherwise return NULL.
0313  */
0314 static int viommu_add_mapping(struct viommu_domain *vdomain, u64 iova, u64 end,
0315                   phys_addr_t paddr, u32 flags)
0316 {
0317     unsigned long irqflags;
0318     struct viommu_mapping *mapping;
0319 
0320     mapping = kzalloc(sizeof(*mapping), GFP_ATOMIC);
0321     if (!mapping)
0322         return -ENOMEM;
0323 
0324     mapping->paddr      = paddr;
0325     mapping->iova.start = iova;
0326     mapping->iova.last  = end;
0327     mapping->flags      = flags;
0328 
0329     spin_lock_irqsave(&vdomain->mappings_lock, irqflags);
0330     interval_tree_insert(&mapping->iova, &vdomain->mappings);
0331     spin_unlock_irqrestore(&vdomain->mappings_lock, irqflags);
0332 
0333     return 0;
0334 }
0335 
0336 /*
0337  * viommu_del_mappings - remove mappings from the internal tree
0338  *
0339  * @vdomain: the domain
0340  * @iova: start of the range
0341  * @end: end of the range
0342  *
0343  * On success, returns the number of unmapped bytes
0344  */
0345 static size_t viommu_del_mappings(struct viommu_domain *vdomain,
0346                   u64 iova, u64 end)
0347 {
0348     size_t unmapped = 0;
0349     unsigned long flags;
0350     struct viommu_mapping *mapping = NULL;
0351     struct interval_tree_node *node, *next;
0352 
0353     spin_lock_irqsave(&vdomain->mappings_lock, flags);
0354     next = interval_tree_iter_first(&vdomain->mappings, iova, end);
0355     while (next) {
0356         node = next;
0357         mapping = container_of(node, struct viommu_mapping, iova);
0358         next = interval_tree_iter_next(node, iova, end);
0359 
0360         /* Trying to split a mapping? */
0361         if (mapping->iova.start < iova)
0362             break;
0363 
0364         /*
0365          * Virtio-iommu doesn't allow UNMAP to split a mapping created
0366          * with a single MAP request, so remove the full mapping.
0367          */
0368         unmapped += mapping->iova.last - mapping->iova.start + 1;
0369 
0370         interval_tree_remove(node, &vdomain->mappings);
0371         kfree(mapping);
0372     }
0373     spin_unlock_irqrestore(&vdomain->mappings_lock, flags);
0374 
0375     return unmapped;
0376 }
0377 
0378 /*
0379  * Fill the domain with identity mappings, skipping the device's reserved
0380  * regions.
0381  */
0382 static int viommu_domain_map_identity(struct viommu_endpoint *vdev,
0383                       struct viommu_domain *vdomain)
0384 {
0385     int ret;
0386     struct iommu_resv_region *resv;
0387     u64 iova = vdomain->domain.geometry.aperture_start;
0388     u64 limit = vdomain->domain.geometry.aperture_end;
0389     u32 flags = VIRTIO_IOMMU_MAP_F_READ | VIRTIO_IOMMU_MAP_F_WRITE;
0390     unsigned long granule = 1UL << __ffs(vdomain->domain.pgsize_bitmap);
0391 
0392     iova = ALIGN(iova, granule);
0393     limit = ALIGN_DOWN(limit + 1, granule) - 1;
0394 
0395     list_for_each_entry(resv, &vdev->resv_regions, list) {
0396         u64 resv_start = ALIGN_DOWN(resv->start, granule);
0397         u64 resv_end = ALIGN(resv->start + resv->length, granule) - 1;
0398 
0399         if (resv_end < iova || resv_start > limit)
0400             /* No overlap */
0401             continue;
0402 
0403         if (resv_start > iova) {
0404             ret = viommu_add_mapping(vdomain, iova, resv_start - 1,
0405                          (phys_addr_t)iova, flags);
0406             if (ret)
0407                 goto err_unmap;
0408         }
0409 
0410         if (resv_end >= limit)
0411             return 0;
0412 
0413         iova = resv_end + 1;
0414     }
0415 
0416     ret = viommu_add_mapping(vdomain, iova, limit, (phys_addr_t)iova,
0417                  flags);
0418     if (ret)
0419         goto err_unmap;
0420     return 0;
0421 
0422 err_unmap:
0423     viommu_del_mappings(vdomain, 0, iova);
0424     return ret;
0425 }
0426 
0427 /*
0428  * viommu_replay_mappings - re-send MAP requests
0429  *
0430  * When reattaching a domain that was previously detached from all endpoints,
0431  * mappings were deleted from the device. Re-create the mappings available in
0432  * the internal tree.
0433  */
0434 static int viommu_replay_mappings(struct viommu_domain *vdomain)
0435 {
0436     int ret = 0;
0437     unsigned long flags;
0438     struct viommu_mapping *mapping;
0439     struct interval_tree_node *node;
0440     struct virtio_iommu_req_map map;
0441 
0442     spin_lock_irqsave(&vdomain->mappings_lock, flags);
0443     node = interval_tree_iter_first(&vdomain->mappings, 0, -1UL);
0444     while (node) {
0445         mapping = container_of(node, struct viommu_mapping, iova);
0446         map = (struct virtio_iommu_req_map) {
0447             .head.type  = VIRTIO_IOMMU_T_MAP,
0448             .domain     = cpu_to_le32(vdomain->id),
0449             .virt_start = cpu_to_le64(mapping->iova.start),
0450             .virt_end   = cpu_to_le64(mapping->iova.last),
0451             .phys_start = cpu_to_le64(mapping->paddr),
0452             .flags      = cpu_to_le32(mapping->flags),
0453         };
0454 
0455         ret = viommu_send_req_sync(vdomain->viommu, &map, sizeof(map));
0456         if (ret)
0457             break;
0458 
0459         node = interval_tree_iter_next(node, 0, -1UL);
0460     }
0461     spin_unlock_irqrestore(&vdomain->mappings_lock, flags);
0462 
0463     return ret;
0464 }
0465 
0466 static int viommu_add_resv_mem(struct viommu_endpoint *vdev,
0467                    struct virtio_iommu_probe_resv_mem *mem,
0468                    size_t len)
0469 {
0470     size_t size;
0471     u64 start64, end64;
0472     phys_addr_t start, end;
0473     struct iommu_resv_region *region = NULL, *next;
0474     unsigned long prot = IOMMU_WRITE | IOMMU_NOEXEC | IOMMU_MMIO;
0475 
0476     start = start64 = le64_to_cpu(mem->start);
0477     end = end64 = le64_to_cpu(mem->end);
0478     size = end64 - start64 + 1;
0479 
0480     /* Catch any overflow, including the unlikely end64 - start64 + 1 = 0 */
0481     if (start != start64 || end != end64 || size < end64 - start64)
0482         return -EOVERFLOW;
0483 
0484     if (len < sizeof(*mem))
0485         return -EINVAL;
0486 
0487     switch (mem->subtype) {
0488     default:
0489         dev_warn(vdev->dev, "unknown resv mem subtype 0x%x\n",
0490              mem->subtype);
0491         fallthrough;
0492     case VIRTIO_IOMMU_RESV_MEM_T_RESERVED:
0493         region = iommu_alloc_resv_region(start, size, 0,
0494                          IOMMU_RESV_RESERVED);
0495         break;
0496     case VIRTIO_IOMMU_RESV_MEM_T_MSI:
0497         region = iommu_alloc_resv_region(start, size, prot,
0498                          IOMMU_RESV_MSI);
0499         break;
0500     }
0501     if (!region)
0502         return -ENOMEM;
0503 
0504     /* Keep the list sorted */
0505     list_for_each_entry(next, &vdev->resv_regions, list) {
0506         if (next->start > region->start)
0507             break;
0508     }
0509     list_add_tail(&region->list, &next->list);
0510     return 0;
0511 }
0512 
0513 static int viommu_probe_endpoint(struct viommu_dev *viommu, struct device *dev)
0514 {
0515     int ret;
0516     u16 type, len;
0517     size_t cur = 0;
0518     size_t probe_len;
0519     struct virtio_iommu_req_probe *probe;
0520     struct virtio_iommu_probe_property *prop;
0521     struct iommu_fwspec *fwspec = dev_iommu_fwspec_get(dev);
0522     struct viommu_endpoint *vdev = dev_iommu_priv_get(dev);
0523 
0524     if (!fwspec->num_ids)
0525         return -EINVAL;
0526 
0527     probe_len = sizeof(*probe) + viommu->probe_size +
0528             sizeof(struct virtio_iommu_req_tail);
0529     probe = kzalloc(probe_len, GFP_KERNEL);
0530     if (!probe)
0531         return -ENOMEM;
0532 
0533     probe->head.type = VIRTIO_IOMMU_T_PROBE;
0534     /*
0535      * For now, assume that properties of an endpoint that outputs multiple
0536      * IDs are consistent. Only probe the first one.
0537      */
0538     probe->endpoint = cpu_to_le32(fwspec->ids[0]);
0539 
0540     ret = viommu_send_req_sync(viommu, probe, probe_len);
0541     if (ret)
0542         goto out_free;
0543 
0544     prop = (void *)probe->properties;
0545     type = le16_to_cpu(prop->type) & VIRTIO_IOMMU_PROBE_T_MASK;
0546 
0547     while (type != VIRTIO_IOMMU_PROBE_T_NONE &&
0548            cur < viommu->probe_size) {
0549         len = le16_to_cpu(prop->length) + sizeof(*prop);
0550 
0551         switch (type) {
0552         case VIRTIO_IOMMU_PROBE_T_RESV_MEM:
0553             ret = viommu_add_resv_mem(vdev, (void *)prop, len);
0554             break;
0555         default:
0556             dev_err(dev, "unknown viommu prop 0x%x\n", type);
0557         }
0558 
0559         if (ret)
0560             dev_err(dev, "failed to parse viommu prop 0x%x\n", type);
0561 
0562         cur += len;
0563         if (cur >= viommu->probe_size)
0564             break;
0565 
0566         prop = (void *)probe->properties + cur;
0567         type = le16_to_cpu(prop->type) & VIRTIO_IOMMU_PROBE_T_MASK;
0568     }
0569 
0570 out_free:
0571     kfree(probe);
0572     return ret;
0573 }
0574 
0575 static int viommu_fault_handler(struct viommu_dev *viommu,
0576                 struct virtio_iommu_fault *fault)
0577 {
0578     char *reason_str;
0579 
0580     u8 reason   = fault->reason;
0581     u32 flags   = le32_to_cpu(fault->flags);
0582     u32 endpoint    = le32_to_cpu(fault->endpoint);
0583     u64 address = le64_to_cpu(fault->address);
0584 
0585     switch (reason) {
0586     case VIRTIO_IOMMU_FAULT_R_DOMAIN:
0587         reason_str = "domain";
0588         break;
0589     case VIRTIO_IOMMU_FAULT_R_MAPPING:
0590         reason_str = "page";
0591         break;
0592     case VIRTIO_IOMMU_FAULT_R_UNKNOWN:
0593     default:
0594         reason_str = "unknown";
0595         break;
0596     }
0597 
0598     /* TODO: find EP by ID and report_iommu_fault */
0599     if (flags & VIRTIO_IOMMU_FAULT_F_ADDRESS)
0600         dev_err_ratelimited(viommu->dev, "%s fault from EP %u at %#llx [%s%s%s]\n",
0601                     reason_str, endpoint, address,
0602                     flags & VIRTIO_IOMMU_FAULT_F_READ ? "R" : "",
0603                     flags & VIRTIO_IOMMU_FAULT_F_WRITE ? "W" : "",
0604                     flags & VIRTIO_IOMMU_FAULT_F_EXEC ? "X" : "");
0605     else
0606         dev_err_ratelimited(viommu->dev, "%s fault from EP %u\n",
0607                     reason_str, endpoint);
0608     return 0;
0609 }
0610 
0611 static void viommu_event_handler(struct virtqueue *vq)
0612 {
0613     int ret;
0614     unsigned int len;
0615     struct scatterlist sg[1];
0616     struct viommu_event *evt;
0617     struct viommu_dev *viommu = vq->vdev->priv;
0618 
0619     while ((evt = virtqueue_get_buf(vq, &len)) != NULL) {
0620         if (len > sizeof(*evt)) {
0621             dev_err(viommu->dev,
0622                 "invalid event buffer (len %u != %zu)\n",
0623                 len, sizeof(*evt));
0624         } else if (!(evt->head & VIOMMU_FAULT_RESV_MASK)) {
0625             viommu_fault_handler(viommu, &evt->fault);
0626         }
0627 
0628         sg_init_one(sg, evt, sizeof(*evt));
0629         ret = virtqueue_add_inbuf(vq, sg, 1, evt, GFP_ATOMIC);
0630         if (ret)
0631             dev_err(viommu->dev, "could not add event buffer\n");
0632     }
0633 
0634     virtqueue_kick(vq);
0635 }
0636 
0637 /* IOMMU API */
0638 
0639 static struct iommu_domain *viommu_domain_alloc(unsigned type)
0640 {
0641     struct viommu_domain *vdomain;
0642 
0643     if (type != IOMMU_DOMAIN_UNMANAGED &&
0644         type != IOMMU_DOMAIN_DMA &&
0645         type != IOMMU_DOMAIN_IDENTITY)
0646         return NULL;
0647 
0648     vdomain = kzalloc(sizeof(*vdomain), GFP_KERNEL);
0649     if (!vdomain)
0650         return NULL;
0651 
0652     mutex_init(&vdomain->mutex);
0653     spin_lock_init(&vdomain->mappings_lock);
0654     vdomain->mappings = RB_ROOT_CACHED;
0655 
0656     return &vdomain->domain;
0657 }
0658 
0659 static int viommu_domain_finalise(struct viommu_endpoint *vdev,
0660                   struct iommu_domain *domain)
0661 {
0662     int ret;
0663     unsigned long viommu_page_size;
0664     struct viommu_dev *viommu = vdev->viommu;
0665     struct viommu_domain *vdomain = to_viommu_domain(domain);
0666 
0667     viommu_page_size = 1UL << __ffs(viommu->pgsize_bitmap);
0668     if (viommu_page_size > PAGE_SIZE) {
0669         dev_err(vdev->dev,
0670             "granule 0x%lx larger than system page size 0x%lx\n",
0671             viommu_page_size, PAGE_SIZE);
0672         return -EINVAL;
0673     }
0674 
0675     ret = ida_alloc_range(&viommu->domain_ids, viommu->first_domain,
0676                   viommu->last_domain, GFP_KERNEL);
0677     if (ret < 0)
0678         return ret;
0679 
0680     vdomain->id     = (unsigned int)ret;
0681 
0682     domain->pgsize_bitmap   = viommu->pgsize_bitmap;
0683     domain->geometry    = viommu->geometry;
0684 
0685     vdomain->map_flags  = viommu->map_flags;
0686     vdomain->viommu     = viommu;
0687 
0688     if (domain->type == IOMMU_DOMAIN_IDENTITY) {
0689         if (virtio_has_feature(viommu->vdev,
0690                        VIRTIO_IOMMU_F_BYPASS_CONFIG)) {
0691             vdomain->bypass = true;
0692             return 0;
0693         }
0694 
0695         ret = viommu_domain_map_identity(vdev, vdomain);
0696         if (ret) {
0697             ida_free(&viommu->domain_ids, vdomain->id);
0698             vdomain->viommu = NULL;
0699             return -EOPNOTSUPP;
0700         }
0701     }
0702 
0703     return 0;
0704 }
0705 
0706 static void viommu_domain_free(struct iommu_domain *domain)
0707 {
0708     struct viommu_domain *vdomain = to_viommu_domain(domain);
0709 
0710     /* Free all remaining mappings */
0711     viommu_del_mappings(vdomain, 0, ULLONG_MAX);
0712 
0713     if (vdomain->viommu)
0714         ida_free(&vdomain->viommu->domain_ids, vdomain->id);
0715 
0716     kfree(vdomain);
0717 }
0718 
0719 static int viommu_attach_dev(struct iommu_domain *domain, struct device *dev)
0720 {
0721     int i;
0722     int ret = 0;
0723     struct virtio_iommu_req_attach req;
0724     struct iommu_fwspec *fwspec = dev_iommu_fwspec_get(dev);
0725     struct viommu_endpoint *vdev = dev_iommu_priv_get(dev);
0726     struct viommu_domain *vdomain = to_viommu_domain(domain);
0727 
0728     mutex_lock(&vdomain->mutex);
0729     if (!vdomain->viommu) {
0730         /*
0731          * Properly initialize the domain now that we know which viommu
0732          * owns it.
0733          */
0734         ret = viommu_domain_finalise(vdev, domain);
0735     } else if (vdomain->viommu != vdev->viommu) {
0736         dev_err(dev, "cannot attach to foreign vIOMMU\n");
0737         ret = -EXDEV;
0738     }
0739     mutex_unlock(&vdomain->mutex);
0740 
0741     if (ret)
0742         return ret;
0743 
0744     /*
0745      * In the virtio-iommu device, when attaching the endpoint to a new
0746      * domain, it is detached from the old one and, if as a result the
0747      * old domain isn't attached to any endpoint, all mappings are removed
0748      * from the old domain and it is freed.
0749      *
0750      * In the driver the old domain still exists, and its mappings will be
0751      * recreated if it gets reattached to an endpoint. Otherwise it will be
0752      * freed explicitly.
0753      *
0754      * vdev->vdomain is protected by group->mutex
0755      */
0756     if (vdev->vdomain)
0757         vdev->vdomain->nr_endpoints--;
0758 
0759     req = (struct virtio_iommu_req_attach) {
0760         .head.type  = VIRTIO_IOMMU_T_ATTACH,
0761         .domain     = cpu_to_le32(vdomain->id),
0762     };
0763 
0764     if (vdomain->bypass)
0765         req.flags |= cpu_to_le32(VIRTIO_IOMMU_ATTACH_F_BYPASS);
0766 
0767     for (i = 0; i < fwspec->num_ids; i++) {
0768         req.endpoint = cpu_to_le32(fwspec->ids[i]);
0769 
0770         ret = viommu_send_req_sync(vdomain->viommu, &req, sizeof(req));
0771         if (ret)
0772             return ret;
0773     }
0774 
0775     if (!vdomain->nr_endpoints) {
0776         /*
0777          * This endpoint is the first to be attached to the domain.
0778          * Replay existing mappings (e.g. SW MSI).
0779          */
0780         ret = viommu_replay_mappings(vdomain);
0781         if (ret)
0782             return ret;
0783     }
0784 
0785     vdomain->nr_endpoints++;
0786     vdev->vdomain = vdomain;
0787 
0788     return 0;
0789 }
0790 
0791 static int viommu_map_pages(struct iommu_domain *domain, unsigned long iova,
0792                 phys_addr_t paddr, size_t pgsize, size_t pgcount,
0793                 int prot, gfp_t gfp, size_t *mapped)
0794 {
0795     int ret;
0796     u32 flags;
0797     size_t size = pgsize * pgcount;
0798     u64 end = iova + size - 1;
0799     struct virtio_iommu_req_map map;
0800     struct viommu_domain *vdomain = to_viommu_domain(domain);
0801 
0802     flags = (prot & IOMMU_READ ? VIRTIO_IOMMU_MAP_F_READ : 0) |
0803         (prot & IOMMU_WRITE ? VIRTIO_IOMMU_MAP_F_WRITE : 0) |
0804         (prot & IOMMU_MMIO ? VIRTIO_IOMMU_MAP_F_MMIO : 0);
0805 
0806     if (flags & ~vdomain->map_flags)
0807         return -EINVAL;
0808 
0809     ret = viommu_add_mapping(vdomain, iova, end, paddr, flags);
0810     if (ret)
0811         return ret;
0812 
0813     map = (struct virtio_iommu_req_map) {
0814         .head.type  = VIRTIO_IOMMU_T_MAP,
0815         .domain     = cpu_to_le32(vdomain->id),
0816         .virt_start = cpu_to_le64(iova),
0817         .phys_start = cpu_to_le64(paddr),
0818         .virt_end   = cpu_to_le64(end),
0819         .flags      = cpu_to_le32(flags),
0820     };
0821 
0822     if (!vdomain->nr_endpoints)
0823         return 0;
0824 
0825     ret = viommu_send_req_sync(vdomain->viommu, &map, sizeof(map));
0826     if (ret)
0827         viommu_del_mappings(vdomain, iova, end);
0828     else if (mapped)
0829         *mapped = size;
0830 
0831     return ret;
0832 }
0833 
0834 static size_t viommu_unmap_pages(struct iommu_domain *domain, unsigned long iova,
0835                  size_t pgsize, size_t pgcount,
0836                  struct iommu_iotlb_gather *gather)
0837 {
0838     int ret = 0;
0839     size_t unmapped;
0840     struct virtio_iommu_req_unmap unmap;
0841     struct viommu_domain *vdomain = to_viommu_domain(domain);
0842     size_t size = pgsize * pgcount;
0843 
0844     unmapped = viommu_del_mappings(vdomain, iova, iova + size - 1);
0845     if (unmapped < size)
0846         return 0;
0847 
0848     /* Device already removed all mappings after detach. */
0849     if (!vdomain->nr_endpoints)
0850         return unmapped;
0851 
0852     unmap = (struct virtio_iommu_req_unmap) {
0853         .head.type  = VIRTIO_IOMMU_T_UNMAP,
0854         .domain     = cpu_to_le32(vdomain->id),
0855         .virt_start = cpu_to_le64(iova),
0856         .virt_end   = cpu_to_le64(iova + unmapped - 1),
0857     };
0858 
0859     ret = viommu_add_req(vdomain->viommu, &unmap, sizeof(unmap));
0860     return ret ? 0 : unmapped;
0861 }
0862 
0863 static phys_addr_t viommu_iova_to_phys(struct iommu_domain *domain,
0864                        dma_addr_t iova)
0865 {
0866     u64 paddr = 0;
0867     unsigned long flags;
0868     struct viommu_mapping *mapping;
0869     struct interval_tree_node *node;
0870     struct viommu_domain *vdomain = to_viommu_domain(domain);
0871 
0872     spin_lock_irqsave(&vdomain->mappings_lock, flags);
0873     node = interval_tree_iter_first(&vdomain->mappings, iova, iova);
0874     if (node) {
0875         mapping = container_of(node, struct viommu_mapping, iova);
0876         paddr = mapping->paddr + (iova - mapping->iova.start);
0877     }
0878     spin_unlock_irqrestore(&vdomain->mappings_lock, flags);
0879 
0880     return paddr;
0881 }
0882 
0883 static void viommu_iotlb_sync(struct iommu_domain *domain,
0884                   struct iommu_iotlb_gather *gather)
0885 {
0886     struct viommu_domain *vdomain = to_viommu_domain(domain);
0887 
0888     viommu_sync_req(vdomain->viommu);
0889 }
0890 
0891 static void viommu_get_resv_regions(struct device *dev, struct list_head *head)
0892 {
0893     struct iommu_resv_region *entry, *new_entry, *msi = NULL;
0894     struct viommu_endpoint *vdev = dev_iommu_priv_get(dev);
0895     int prot = IOMMU_WRITE | IOMMU_NOEXEC | IOMMU_MMIO;
0896 
0897     list_for_each_entry(entry, &vdev->resv_regions, list) {
0898         if (entry->type == IOMMU_RESV_MSI)
0899             msi = entry;
0900 
0901         new_entry = kmemdup(entry, sizeof(*entry), GFP_KERNEL);
0902         if (!new_entry)
0903             return;
0904         list_add_tail(&new_entry->list, head);
0905     }
0906 
0907     /*
0908      * If the device didn't register any bypass MSI window, add a
0909      * software-mapped region.
0910      */
0911     if (!msi) {
0912         msi = iommu_alloc_resv_region(MSI_IOVA_BASE, MSI_IOVA_LENGTH,
0913                           prot, IOMMU_RESV_SW_MSI);
0914         if (!msi)
0915             return;
0916 
0917         list_add_tail(&msi->list, head);
0918     }
0919 
0920     iommu_dma_get_resv_regions(dev, head);
0921 }
0922 
0923 static struct iommu_ops viommu_ops;
0924 static struct virtio_driver virtio_iommu_drv;
0925 
0926 static int viommu_match_node(struct device *dev, const void *data)
0927 {
0928     return dev->parent->fwnode == data;
0929 }
0930 
0931 static struct viommu_dev *viommu_get_by_fwnode(struct fwnode_handle *fwnode)
0932 {
0933     struct device *dev = driver_find_device(&virtio_iommu_drv.driver, NULL,
0934                         fwnode, viommu_match_node);
0935     put_device(dev);
0936 
0937     return dev ? dev_to_virtio(dev)->priv : NULL;
0938 }
0939 
0940 static struct iommu_device *viommu_probe_device(struct device *dev)
0941 {
0942     int ret;
0943     struct viommu_endpoint *vdev;
0944     struct viommu_dev *viommu = NULL;
0945     struct iommu_fwspec *fwspec = dev_iommu_fwspec_get(dev);
0946 
0947     if (!fwspec || fwspec->ops != &viommu_ops)
0948         return ERR_PTR(-ENODEV);
0949 
0950     viommu = viommu_get_by_fwnode(fwspec->iommu_fwnode);
0951     if (!viommu)
0952         return ERR_PTR(-ENODEV);
0953 
0954     vdev = kzalloc(sizeof(*vdev), GFP_KERNEL);
0955     if (!vdev)
0956         return ERR_PTR(-ENOMEM);
0957 
0958     vdev->dev = dev;
0959     vdev->viommu = viommu;
0960     INIT_LIST_HEAD(&vdev->resv_regions);
0961     dev_iommu_priv_set(dev, vdev);
0962 
0963     if (viommu->probe_size) {
0964         /* Get additional information for this endpoint */
0965         ret = viommu_probe_endpoint(viommu, dev);
0966         if (ret)
0967             goto err_free_dev;
0968     }
0969 
0970     return &viommu->iommu;
0971 
0972 err_free_dev:
0973     iommu_put_resv_regions(dev, &vdev->resv_regions);
0974     kfree(vdev);
0975 
0976     return ERR_PTR(ret);
0977 }
0978 
0979 static void viommu_probe_finalize(struct device *dev)
0980 {
0981 #ifndef CONFIG_ARCH_HAS_SETUP_DMA_OPS
0982     /* First clear the DMA ops in case we're switching from a DMA domain */
0983     set_dma_ops(dev, NULL);
0984     iommu_setup_dma_ops(dev, 0, U64_MAX);
0985 #endif
0986 }
0987 
0988 static void viommu_release_device(struct device *dev)
0989 {
0990     struct viommu_endpoint *vdev = dev_iommu_priv_get(dev);
0991 
0992     iommu_put_resv_regions(dev, &vdev->resv_regions);
0993     kfree(vdev);
0994 }
0995 
0996 static struct iommu_group *viommu_device_group(struct device *dev)
0997 {
0998     if (dev_is_pci(dev))
0999         return pci_device_group(dev);
1000     else
1001         return generic_device_group(dev);
1002 }
1003 
1004 static int viommu_of_xlate(struct device *dev, struct of_phandle_args *args)
1005 {
1006     return iommu_fwspec_add_ids(dev, args->args, 1);
1007 }
1008 
1009 static bool viommu_capable(enum iommu_cap cap)
1010 {
1011     switch (cap) {
1012     case IOMMU_CAP_CACHE_COHERENCY:
1013         return true;
1014     default:
1015         return false;
1016     }
1017 }
1018 
1019 static struct iommu_ops viommu_ops = {
1020     .capable        = viommu_capable,
1021     .domain_alloc       = viommu_domain_alloc,
1022     .probe_device       = viommu_probe_device,
1023     .probe_finalize     = viommu_probe_finalize,
1024     .release_device     = viommu_release_device,
1025     .device_group       = viommu_device_group,
1026     .get_resv_regions   = viommu_get_resv_regions,
1027     .of_xlate       = viommu_of_xlate,
1028     .owner          = THIS_MODULE,
1029     .default_domain_ops = &(const struct iommu_domain_ops) {
1030         .attach_dev     = viommu_attach_dev,
1031         .map_pages      = viommu_map_pages,
1032         .unmap_pages        = viommu_unmap_pages,
1033         .iova_to_phys       = viommu_iova_to_phys,
1034         .iotlb_sync     = viommu_iotlb_sync,
1035         .free           = viommu_domain_free,
1036     }
1037 };
1038 
1039 static int viommu_init_vqs(struct viommu_dev *viommu)
1040 {
1041     struct virtio_device *vdev = dev_to_virtio(viommu->dev);
1042     const char *names[] = { "request", "event" };
1043     vq_callback_t *callbacks[] = {
1044         NULL, /* No async requests */
1045         viommu_event_handler,
1046     };
1047 
1048     return virtio_find_vqs(vdev, VIOMMU_NR_VQS, viommu->vqs, callbacks,
1049                    names, NULL);
1050 }
1051 
1052 static int viommu_fill_evtq(struct viommu_dev *viommu)
1053 {
1054     int i, ret;
1055     struct scatterlist sg[1];
1056     struct viommu_event *evts;
1057     struct virtqueue *vq = viommu->vqs[VIOMMU_EVENT_VQ];
1058     size_t nr_evts = vq->num_free;
1059 
1060     viommu->evts = evts = devm_kmalloc_array(viommu->dev, nr_evts,
1061                          sizeof(*evts), GFP_KERNEL);
1062     if (!evts)
1063         return -ENOMEM;
1064 
1065     for (i = 0; i < nr_evts; i++) {
1066         sg_init_one(sg, &evts[i], sizeof(*evts));
1067         ret = virtqueue_add_inbuf(vq, sg, 1, &evts[i], GFP_KERNEL);
1068         if (ret)
1069             return ret;
1070     }
1071 
1072     return 0;
1073 }
1074 
1075 static int viommu_probe(struct virtio_device *vdev)
1076 {
1077     struct device *parent_dev = vdev->dev.parent;
1078     struct viommu_dev *viommu = NULL;
1079     struct device *dev = &vdev->dev;
1080     u64 input_start = 0;
1081     u64 input_end = -1UL;
1082     int ret;
1083 
1084     if (!virtio_has_feature(vdev, VIRTIO_F_VERSION_1) ||
1085         !virtio_has_feature(vdev, VIRTIO_IOMMU_F_MAP_UNMAP))
1086         return -ENODEV;
1087 
1088     viommu = devm_kzalloc(dev, sizeof(*viommu), GFP_KERNEL);
1089     if (!viommu)
1090         return -ENOMEM;
1091 
1092     spin_lock_init(&viommu->request_lock);
1093     ida_init(&viommu->domain_ids);
1094     viommu->dev = dev;
1095     viommu->vdev = vdev;
1096     INIT_LIST_HEAD(&viommu->requests);
1097 
1098     ret = viommu_init_vqs(viommu);
1099     if (ret)
1100         return ret;
1101 
1102     virtio_cread_le(vdev, struct virtio_iommu_config, page_size_mask,
1103             &viommu->pgsize_bitmap);
1104 
1105     if (!viommu->pgsize_bitmap) {
1106         ret = -EINVAL;
1107         goto err_free_vqs;
1108     }
1109 
1110     viommu->map_flags = VIRTIO_IOMMU_MAP_F_READ | VIRTIO_IOMMU_MAP_F_WRITE;
1111     viommu->last_domain = ~0U;
1112 
1113     /* Optional features */
1114     virtio_cread_le_feature(vdev, VIRTIO_IOMMU_F_INPUT_RANGE,
1115                 struct virtio_iommu_config, input_range.start,
1116                 &input_start);
1117 
1118     virtio_cread_le_feature(vdev, VIRTIO_IOMMU_F_INPUT_RANGE,
1119                 struct virtio_iommu_config, input_range.end,
1120                 &input_end);
1121 
1122     virtio_cread_le_feature(vdev, VIRTIO_IOMMU_F_DOMAIN_RANGE,
1123                 struct virtio_iommu_config, domain_range.start,
1124                 &viommu->first_domain);
1125 
1126     virtio_cread_le_feature(vdev, VIRTIO_IOMMU_F_DOMAIN_RANGE,
1127                 struct virtio_iommu_config, domain_range.end,
1128                 &viommu->last_domain);
1129 
1130     virtio_cread_le_feature(vdev, VIRTIO_IOMMU_F_PROBE,
1131                 struct virtio_iommu_config, probe_size,
1132                 &viommu->probe_size);
1133 
1134     viommu->geometry = (struct iommu_domain_geometry) {
1135         .aperture_start = input_start,
1136         .aperture_end   = input_end,
1137         .force_aperture = true,
1138     };
1139 
1140     if (virtio_has_feature(vdev, VIRTIO_IOMMU_F_MMIO))
1141         viommu->map_flags |= VIRTIO_IOMMU_MAP_F_MMIO;
1142 
1143     viommu_ops.pgsize_bitmap = viommu->pgsize_bitmap;
1144 
1145     virtio_device_ready(vdev);
1146 
1147     /* Populate the event queue with buffers */
1148     ret = viommu_fill_evtq(viommu);
1149     if (ret)
1150         goto err_free_vqs;
1151 
1152     ret = iommu_device_sysfs_add(&viommu->iommu, dev, NULL, "%s",
1153                      virtio_bus_name(vdev));
1154     if (ret)
1155         goto err_free_vqs;
1156 
1157     iommu_device_register(&viommu->iommu, &viommu_ops, parent_dev);
1158 
1159 #ifdef CONFIG_PCI
1160     if (pci_bus_type.iommu_ops != &viommu_ops) {
1161         ret = bus_set_iommu(&pci_bus_type, &viommu_ops);
1162         if (ret)
1163             goto err_unregister;
1164     }
1165 #endif
1166 #ifdef CONFIG_ARM_AMBA
1167     if (amba_bustype.iommu_ops != &viommu_ops) {
1168         ret = bus_set_iommu(&amba_bustype, &viommu_ops);
1169         if (ret)
1170             goto err_unregister;
1171     }
1172 #endif
1173     if (platform_bus_type.iommu_ops != &viommu_ops) {
1174         ret = bus_set_iommu(&platform_bus_type, &viommu_ops);
1175         if (ret)
1176             goto err_unregister;
1177     }
1178 
1179     vdev->priv = viommu;
1180 
1181     dev_info(dev, "input address: %u bits\n",
1182          order_base_2(viommu->geometry.aperture_end));
1183     dev_info(dev, "page mask: %#llx\n", viommu->pgsize_bitmap);
1184 
1185     return 0;
1186 
1187 err_unregister:
1188     iommu_device_sysfs_remove(&viommu->iommu);
1189     iommu_device_unregister(&viommu->iommu);
1190 err_free_vqs:
1191     vdev->config->del_vqs(vdev);
1192 
1193     return ret;
1194 }
1195 
1196 static void viommu_remove(struct virtio_device *vdev)
1197 {
1198     struct viommu_dev *viommu = vdev->priv;
1199 
1200     iommu_device_sysfs_remove(&viommu->iommu);
1201     iommu_device_unregister(&viommu->iommu);
1202 
1203     /* Stop all virtqueues */
1204     virtio_reset_device(vdev);
1205     vdev->config->del_vqs(vdev);
1206 
1207     dev_info(&vdev->dev, "device removed\n");
1208 }
1209 
1210 static void viommu_config_changed(struct virtio_device *vdev)
1211 {
1212     dev_warn(&vdev->dev, "config changed\n");
1213 }
1214 
1215 static unsigned int features[] = {
1216     VIRTIO_IOMMU_F_MAP_UNMAP,
1217     VIRTIO_IOMMU_F_INPUT_RANGE,
1218     VIRTIO_IOMMU_F_DOMAIN_RANGE,
1219     VIRTIO_IOMMU_F_PROBE,
1220     VIRTIO_IOMMU_F_MMIO,
1221     VIRTIO_IOMMU_F_BYPASS_CONFIG,
1222 };
1223 
1224 static struct virtio_device_id id_table[] = {
1225     { VIRTIO_ID_IOMMU, VIRTIO_DEV_ANY_ID },
1226     { 0 },
1227 };
1228 MODULE_DEVICE_TABLE(virtio, id_table);
1229 
1230 static struct virtio_driver virtio_iommu_drv = {
1231     .driver.name        = KBUILD_MODNAME,
1232     .driver.owner       = THIS_MODULE,
1233     .id_table       = id_table,
1234     .feature_table      = features,
1235     .feature_table_size = ARRAY_SIZE(features),
1236     .probe          = viommu_probe,
1237     .remove         = viommu_remove,
1238     .config_changed     = viommu_config_changed,
1239 };
1240 
1241 module_virtio_driver(virtio_iommu_drv);
1242 
1243 MODULE_DESCRIPTION("Virtio IOMMU driver");
1244 MODULE_AUTHOR("Jean-Philippe Brucker <jean-philippe.brucker@arm.com>");
1245 MODULE_LICENSE("GPL v2");
[ Source navigation ] [ Diff markup ] [ Identifier search ] [ general search ]

This page was automatically generated by the 2.1.0 LXR engine.

The LXR team
Valid CSS 2.1! Valid HTML 4.01!