drivers/iommu/virtio-iommu.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Virtio driver for the paravirtualized IOMMU
   4  *
   5  * Copyright (C) 2019 Arm Limited
   6  */
   7
   8 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
   9
  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.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>
  22
  23 #include <uapi/linux/virtio_iommu.h>
  24
  25 #include "dma-iommu.h"
  26
  27 #define MSI_IOVA_BASE                   0x8000000
  28 #define MSI_IOVA_LENGTH                 0x100000
  29
  30 #define VIOMMU_REQUEST_VQ               0
  31 #define VIOMMU_EVENT_VQ                 1
  32 #define VIOMMU_NR_VQS                   2
  33
  34 struct viommu_dev {
  35         struct iommu_device             iommu;
  36         struct device                   *dev;
  37         struct virtio_device            *vdev;
  38
  39         struct ida                      domain_ids;
  40
  41         struct virtqueue                *vqs[VIOMMU_NR_VQS];
  42         spinlock_t                      request_lock;
  43         struct list_head                requests;
  44         void                            *evts;
  45
  46         /* Device configuration */
  47         struct iommu_domain_geometry    geometry;
  48         u64                             pgsize_bitmap;
  49         u32                             first_domain;
  50         u32                             last_domain;
  51         /* Supported MAP flags */
  52         u32                             map_flags;
  53         u32                             probe_size;
  54 };
  55
  56 struct viommu_mapping {
  57         phys_addr_t                     paddr;
  58         struct interval_tree_node       iova;
  59         u32                             flags;
  60 };
  61
  62 struct viommu_domain {
  63         struct iommu_domain             domain;
  64         struct viommu_dev               *viommu;
  65         struct mutex                    mutex; /* protects viommu pointer */
  66         unsigned int                    id;
  67         u32                             map_flags;
  68
  69         spinlock_t                      mappings_lock;
  70         struct rb_root_cached           mappings;
  71
  72         unsigned long                   nr_endpoints;
  73         bool                            bypass;
  74 };
  75
  76 struct viommu_endpoint {
  77         struct device                   *dev;
  78         struct viommu_dev               *viommu;
  79         struct viommu_domain            *vdomain;
  80         struct list_head                resv_regions;
  81 };
  82
  83 struct viommu_request {
  84         struct list_head                list;
  85         void                            *writeback;
  86         unsigned int                    write_offset;
  87         unsigned int                    len;
  88         char                            buf[] __counted_by(len);
  89 };
  90
  91 #define VIOMMU_FAULT_RESV_MASK          0xffffff00
  92
  93 struct viommu_event {
  94         union {
  95                 u32                     head;
  96                 struct virtio_iommu_fault fault;
  97         };
  98 };
  99
 100 #define to_viommu_domain(domain)        \
 101         container_of(domain, struct viommu_domain, domain)
 102
 103 static int viommu_get_req_errno(void *buf, size_t len)
 104 {
 105         struct virtio_iommu_req_tail *tail = buf + len - sizeof(*tail);
 106
 107         switch (tail->status) {
 108         case VIRTIO_IOMMU_S_OK:
 109                 return 0;
 110         case VIRTIO_IOMMU_S_UNSUPP:
 111                 return -ENOSYS;
 112         case VIRTIO_IOMMU_S_INVAL:
 113                 return -EINVAL;
 114         case VIRTIO_IOMMU_S_RANGE:
 115                 return -ERANGE;
 116         case VIRTIO_IOMMU_S_NOENT:
 117                 return -ENOENT;
 118         case VIRTIO_IOMMU_S_FAULT:
 119                 return -EFAULT;
 120         case VIRTIO_IOMMU_S_NOMEM:
 121                 return -ENOMEM;
 122         case VIRTIO_IOMMU_S_IOERR:
 123         case VIRTIO_IOMMU_S_DEVERR:
 124         default:
 125                 return -EIO;
 126         }
 127 }
 128
 129 static void viommu_set_req_status(void *buf, size_t len, int status)
 130 {
 131         struct virtio_iommu_req_tail *tail = buf + len - sizeof(*tail);
 132
 133         tail->status = status;
 134 }
 135
 136 static off_t viommu_get_write_desc_offset(struct viommu_dev *viommu,
 137                                           struct virtio_iommu_req_head *req,
 138                                           size_t len)
 139 {
 140         size_t tail_size = sizeof(struct virtio_iommu_req_tail);
 141
 142         if (req->type == VIRTIO_IOMMU_T_PROBE)
 143                 return len - viommu->probe_size - tail_size;
 144
 145         return len - tail_size;
 146 }
 147
 148 /*
 149  * __viommu_sync_req - Complete all in-flight requests
 150  *
 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.
 153  */
 154 static int __viommu_sync_req(struct viommu_dev *viommu)
 155 {
 156         unsigned int len;
 157         size_t write_len;
 158         struct viommu_request *req;
 159         struct virtqueue *vq = viommu->vqs[VIOMMU_REQUEST_VQ];
 160
 161         assert_spin_locked(&viommu->request_lock);
 162
 163         virtqueue_kick(vq);
 164
 165         while (!list_empty(&viommu->requests)) {
 166                 len = 0;
 167                 req = virtqueue_get_buf(vq, &len);
 168                 if (!req)
 169                         continue;
 170
 171                 if (!len)
 172                         viommu_set_req_status(req->buf, req->len,
 173                                               VIRTIO_IOMMU_S_IOERR);
 174
 175                 write_len = req->len - req->write_offset;
 176                 if (req->writeback && len == write_len)
 177                         memcpy(req->writeback, req->buf + req->write_offset,
 178                                write_len);
 179
 180                 list_del(&req->list);
 181                 kfree(req);
 182         }
 183
 184         return 0;
 185 }
 186
 187 static int viommu_sync_req(struct viommu_dev *viommu)
 188 {
 189         int ret;
 190         unsigned long flags;
 191
 192         spin_lock_irqsave(&viommu->request_lock, flags);
 193         ret = __viommu_sync_req(viommu);
 194         if (ret)
 195                 dev_dbg(viommu->dev, "could not sync requests (%d)\n", ret);
 196         spin_unlock_irqrestore(&viommu->request_lock, flags);
 197
 198         return ret;
 199 }
 200
 201 /*
 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.
 206  *
 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.
 209  *
 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
 213  * sync_req().
 214  *
 215  * Return 0 if the request was successfully added to the queue.
 216  */
 217 static int __viommu_add_req(struct viommu_dev *viommu, void *buf, size_t len,
 218                             bool writeback)
 219 {
 220         int ret;
 221         off_t write_offset;
 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];
 226
 227         assert_spin_locked(&viommu->request_lock);
 228
 229         write_offset = viommu_get_write_desc_offset(viommu, buf, len);
 230         if (write_offset <= 0)
 231                 return -EINVAL;
 232
 233         req = kzalloc(struct_size(req, buf, len), GFP_ATOMIC);
 234         if (!req)
 235                 return -ENOMEM;
 236
 237         req->len = len;
 238         if (writeback) {
 239                 req->writeback = buf + write_offset;
 240                 req->write_offset = write_offset;
 241         }
 242         memcpy(&req->buf, buf, write_offset);
 243
 244         sg_init_one(&top_sg, req->buf, write_offset);
 245         sg_init_one(&bottom_sg, req->buf + write_offset, len - write_offset);
 246
 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);
 252         }
 253         if (ret)
 254                 goto err_free;
 255
 256         list_add_tail(&req->list, &viommu->requests);
 257         return 0;
 258
 259 err_free:
 260         kfree(req);
 261         return ret;
 262 }
 263
 264 static int viommu_add_req(struct viommu_dev *viommu, void *buf, size_t len)
 265 {
 266         int ret;
 267         unsigned long flags;
 268
 269         spin_lock_irqsave(&viommu->request_lock, flags);
 270         ret = __viommu_add_req(viommu, buf, len, false);
 271         if (ret)
 272                 dev_dbg(viommu->dev, "could not add request: %d\n", ret);
 273         spin_unlock_irqrestore(&viommu->request_lock, flags);
 274
 275         return ret;
 276 }
 277
 278 /*
 279  * Send a request and wait for it to complete. Return the request status (as an
 280  * errno)
 281  */
 282 static int viommu_send_req_sync(struct viommu_dev *viommu, void *buf,
 283                                 size_t len)
 284 {
 285         int ret;
 286         unsigned long flags;
 287
 288         spin_lock_irqsave(&viommu->request_lock, flags);
 289
 290         ret = __viommu_add_req(viommu, buf, len, true);
 291         if (ret) {
 292                 dev_dbg(viommu->dev, "could not add request (%d)\n", ret);
 293                 goto out_unlock;
 294         }
 295
 296         ret = __viommu_sync_req(viommu);
 297         if (ret) {
 298                 dev_dbg(viommu->dev, "could not sync requests (%d)\n", ret);
 299                 /* Fall-through (get the actual request status) */
 300         }
 301
 302         ret = viommu_get_req_errno(buf, len);
 303 out_unlock:
 304         spin_unlock_irqrestore(&viommu->request_lock, flags);
 305         return ret;
 306 }
 307
 308 /*
 309  * viommu_add_mapping - add a mapping to the internal tree
 310  *
 311  * On success, return the new mapping. Otherwise return NULL.
 312  */
 313 static int viommu_add_mapping(struct viommu_domain *vdomain, u64 iova, u64 end,
 314                               phys_addr_t paddr, u32 flags)
 315 {
 316         unsigned long irqflags;
 317         struct viommu_mapping *mapping;
 318
 319         mapping = kzalloc(sizeof(*mapping), GFP_ATOMIC);
 320         if (!mapping)
 321                 return -ENOMEM;
 322
 323         mapping->paddr          = paddr;
 324         mapping->iova.start     = iova;
 325         mapping->iova.last      = end;
 326         mapping->flags          = flags;
 327
 328         spin_lock_irqsave(&vdomain->mappings_lock, irqflags);
 329         interval_tree_insert(&mapping->iova, &vdomain->mappings);
 330         spin_unlock_irqrestore(&vdomain->mappings_lock, irqflags);
 331
 332         return 0;
 333 }
 334
 335 /*
 336  * viommu_del_mappings - remove mappings from the internal tree
 337  *
 338  * @vdomain: the domain
 339  * @iova: start of the range
 340  * @end: end of the range
 341  *
 342  * On success, returns the number of unmapped bytes
 343  */
 344 static size_t viommu_del_mappings(struct viommu_domain *vdomain,
 345                                   u64 iova, u64 end)
 346 {
 347         size_t unmapped = 0;
 348         unsigned long flags;
 349         struct viommu_mapping *mapping = NULL;
 350         struct interval_tree_node *node, *next;
 351
 352         spin_lock_irqsave(&vdomain->mappings_lock, flags);
 353         next = interval_tree_iter_first(&vdomain->mappings, iova, end);
 354         while (next) {
 355                 node = next;
 356                 mapping = container_of(node, struct viommu_mapping, iova);
 357                 next = interval_tree_iter_next(node, iova, end);
 358
 359                 /* Trying to split a mapping? */
 360                 if (mapping->iova.start < iova)
 361                         break;
 362
 363                 /*
 364                  * Virtio-iommu doesn't allow UNMAP to split a mapping created
 365                  * with a single MAP request, so remove the full mapping.
 366                  */
 367                 unmapped += mapping->iova.last - mapping->iova.start + 1;
 368
 369                 interval_tree_remove(node, &vdomain->mappings);
 370                 kfree(mapping);
 371         }
 372         spin_unlock_irqrestore(&vdomain->mappings_lock, flags);
 373
 374         return unmapped;
 375 }
 376
 377 /*
 378  * Fill the domain with identity mappings, skipping the device's reserved
 379  * regions.
 380  */
 381 static int viommu_domain_map_identity(struct viommu_endpoint *vdev,
 382                                       struct viommu_domain *vdomain)
 383 {
 384         int ret;
 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);
 390
 391         iova = ALIGN(iova, granule);
 392         limit = ALIGN_DOWN(limit + 1, granule) - 1;
 393
 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;
 397
 398                 if (resv_end < iova || resv_start > limit)
 399                         /* No overlap */
 400                         continue;
 401
 402                 if (resv_start > iova) {
 403                         ret = viommu_add_mapping(vdomain, iova, resv_start - 1,
 404                                                  (phys_addr_t)iova, flags);
 405                         if (ret)
 406                                 goto err_unmap;
 407                 }
 408
 409                 if (resv_end >= limit)
 410                         return 0;
 411
 412                 iova = resv_end + 1;
 413         }
 414
 415         ret = viommu_add_mapping(vdomain, iova, limit, (phys_addr_t)iova,
 416                                  flags);
 417         if (ret)
 418                 goto err_unmap;
 419         return 0;
 420
 421 err_unmap:
 422         viommu_del_mappings(vdomain, 0, iova);
 423         return ret;
 424 }
 425
 426 /*
 427  * viommu_replay_mappings - re-send MAP requests
 428  *
 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
 431  * the internal tree.
 432  */
 433 static int viommu_replay_mappings(struct viommu_domain *vdomain)
 434 {
 435         int ret = 0;
 436         unsigned long flags;
 437         struct viommu_mapping *mapping;
 438         struct interval_tree_node *node;
 439         struct virtio_iommu_req_map map;
 440
 441         spin_lock_irqsave(&vdomain->mappings_lock, flags);
 442         node = interval_tree_iter_first(&vdomain->mappings, 0, -1UL);
 443         while (node) {
 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),
 452                 };
 453
 454                 ret = viommu_send_req_sync(vdomain->viommu, &map, sizeof(map));
 455                 if (ret)
 456                         break;
 457
 458                 node = interval_tree_iter_next(node, 0, -1UL);
 459         }
 460         spin_unlock_irqrestore(&vdomain->mappings_lock, flags);
 461
 462         return ret;
 463 }
 464
 465 static int viommu_add_resv_mem(struct viommu_endpoint *vdev,
 466                                struct virtio_iommu_probe_resv_mem *mem,
 467                                size_t len)
 468 {
 469         size_t size;
 470         u64 start64, end64;
 471         phys_addr_t start, end;
 472         struct iommu_resv_region *region = NULL, *next;
 473         unsigned long prot = IOMMU_WRITE | IOMMU_NOEXEC | IOMMU_MMIO;
 474
 475         start = start64 = le64_to_cpu(mem->start);
 476         end = end64 = le64_to_cpu(mem->end);
 477         size = end64 - start64 + 1;
 478
 479         /* Catch any overflow, including the unlikely end64 - start64 + 1 = 0 */
 480         if (start != start64 || end != end64 || size < end64 - start64)
 481                 return -EOVERFLOW;
 482
 483         if (len < sizeof(*mem))
 484                 return -EINVAL;
 485
 486         switch (mem->subtype) {
 487         default:
 488                 dev_warn(vdev->dev, "unknown resv mem subtype 0x%x\n",
 489                          mem->subtype);
 490                 fallthrough;
 491         case VIRTIO_IOMMU_RESV_MEM_T_RESERVED:
 492                 region = iommu_alloc_resv_region(start, size, 0,
 493                                                  IOMMU_RESV_RESERVED,
 494                                                  GFP_KERNEL);
 495                 break;
 496         case VIRTIO_IOMMU_RESV_MEM_T_MSI:
 497                 region = iommu_alloc_resv_region(start, size, prot,
 498                                                  IOMMU_RESV_MSI,
 499                                                  GFP_KERNEL);
 500                 break;
 501         }
 502         if (!region)
 503                 return -ENOMEM;
 504
 505         /* Keep the list sorted */
 506         list_for_each_entry(next, &vdev->resv_regions, list) {
 507                 if (next->start > region->start)
 508                         break;
 509         }
 510         list_add_tail(&region->list, &next->list);
 511         return 0;
 512 }
 513
 514 static int viommu_probe_endpoint(struct viommu_dev *viommu, struct device *dev)
 515 {
 516         int ret;
 517         u16 type, len;
 518         size_t cur = 0;
 519         size_t probe_len;
 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);
 524
 525         if (!fwspec->num_ids)
 526                 return -EINVAL;
 527
 528         probe_len = sizeof(*probe) + viommu->probe_size +
 529                     sizeof(struct virtio_iommu_req_tail);
 530         probe = kzalloc(probe_len, GFP_KERNEL);
 531         if (!probe)
 532                 return -ENOMEM;
 533
 534         probe->head.type = VIRTIO_IOMMU_T_PROBE;
 535         /*
 536          * For now, assume that properties of an endpoint that outputs multiple
 537          * IDs are consistent. Only probe the first one.
 538          */
 539         probe->endpoint = cpu_to_le32(fwspec->ids[0]);
 540
 541         ret = viommu_send_req_sync(viommu, probe, probe_len);
 542         if (ret)
 543                 goto out_free;
 544
 545         prop = (void *)probe->properties;
 546         type = le16_to_cpu(prop->type) & VIRTIO_IOMMU_PROBE_T_MASK;
 547
 548         while (type != VIRTIO_IOMMU_PROBE_T_NONE &&
 549                cur < viommu->probe_size) {
 550                 len = le16_to_cpu(prop->length) + sizeof(*prop);
 551
 552                 switch (type) {
 553                 case VIRTIO_IOMMU_PROBE_T_RESV_MEM:
 554                         ret = viommu_add_resv_mem(vdev, (void *)prop, len);
 555                         break;
 556                 default:
 557                         dev_err(dev, "unknown viommu prop 0x%x\n", type);
 558                 }
 559
 560                 if (ret)
 561                         dev_err(dev, "failed to parse viommu prop 0x%x\n", type);
 562
 563                 cur += len;
 564                 if (cur >= viommu->probe_size)
 565                         break;
 566
 567                 prop = (void *)probe->properties + cur;
 568                 type = le16_to_cpu(prop->type) & VIRTIO_IOMMU_PROBE_T_MASK;
 569         }
 570
 571 out_free:
 572         kfree(probe);
 573         return ret;
 574 }
 575
 576 static int viommu_fault_handler(struct viommu_dev *viommu,
 577                                 struct virtio_iommu_fault *fault)
 578 {
 579         char *reason_str;
 580
 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);
 585
 586         switch (reason) {
 587         case VIRTIO_IOMMU_FAULT_R_DOMAIN:
 588                 reason_str = "domain";
 589                 break;
 590         case VIRTIO_IOMMU_FAULT_R_MAPPING:
 591                 reason_str = "page";
 592                 break;
 593         case VIRTIO_IOMMU_FAULT_R_UNKNOWN:
 594         default:
 595                 reason_str = "unknown";
 596                 break;
 597         }
 598
 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" : "");
 606         else
 607                 dev_err_ratelimited(viommu->dev, "%s fault from EP %u\n",
 608                                     reason_str, endpoint);
 609         return 0;
 610 }
 611
 612 static void viommu_event_handler(struct virtqueue *vq)
 613 {
 614         int ret;
 615         unsigned int len;
 616         struct scatterlist sg[1];
 617         struct viommu_event *evt;
 618         struct viommu_dev *viommu = vq->vdev->priv;
 619
 620         while ((evt = virtqueue_get_buf(vq, &len)) != NULL) {
 621                 if (len > sizeof(*evt)) {
 622                         dev_err(viommu->dev,
 623                                 "invalid event buffer (len %u != %zu)\n",
 624                                 len, sizeof(*evt));
 625                 } else if (!(evt->head & VIOMMU_FAULT_RESV_MASK)) {
 626                         viommu_fault_handler(viommu, &evt->fault);
 627                 }
 628
 629                 sg_init_one(sg, evt, sizeof(*evt));
 630                 ret = virtqueue_add_inbuf(vq, sg, 1, evt, GFP_ATOMIC);
 631                 if (ret)
 632                         dev_err(viommu->dev, "could not add event buffer\n");
 633         }
 634
 635         virtqueue_kick(vq);
 636 }
 637
 638 /* IOMMU API */
 639
 640 static struct iommu_domain *viommu_domain_alloc(unsigned type)
 641 {
 642         struct viommu_domain *vdomain;
 643
 644         if (type != IOMMU_DOMAIN_UNMANAGED &&
 645             type != IOMMU_DOMAIN_DMA &&
 646             type != IOMMU_DOMAIN_IDENTITY)
 647                 return NULL;
 648
 649         vdomain = kzalloc(sizeof(*vdomain), GFP_KERNEL);
 650         if (!vdomain)
 651                 return NULL;
 652
 653         mutex_init(&vdomain->mutex);
 654         spin_lock_init(&vdomain->mappings_lock);
 655         vdomain->mappings = RB_ROOT_CACHED;
 656
 657         return &vdomain->domain;
 658 }
 659
 660 static int viommu_domain_finalise(struct viommu_endpoint *vdev,
 661                                   struct iommu_domain *domain)
 662 {
 663         int ret;
 664         unsigned long viommu_page_size;
 665         struct viommu_dev *viommu = vdev->viommu;
 666         struct viommu_domain *vdomain = to_viommu_domain(domain);
 667
 668         viommu_page_size = 1UL << __ffs(viommu->pgsize_bitmap);
 669         if (viommu_page_size > PAGE_SIZE) {
 670                 dev_err(vdev->dev,
 671                         "granule 0x%lx larger than system page size 0x%lx\n",
 672                         viommu_page_size, PAGE_SIZE);
 673                 return -ENODEV;
 674         }
 675
 676         ret = ida_alloc_range(&viommu->domain_ids, viommu->first_domain,
 677                               viommu->last_domain, GFP_KERNEL);
 678         if (ret < 0)
 679                 return ret;
 680
 681         vdomain->id             = (unsigned int)ret;
 682
 683         domain->pgsize_bitmap   = viommu->pgsize_bitmap;
 684         domain->geometry        = viommu->geometry;
 685
 686         vdomain->map_flags      = viommu->map_flags;
 687         vdomain->viommu         = viommu;
 688
 689         if (domain->type == IOMMU_DOMAIN_IDENTITY) {
 690                 if (virtio_has_feature(viommu->vdev,
 691                                        VIRTIO_IOMMU_F_BYPASS_CONFIG)) {
 692                         vdomain->bypass = true;
 693                         return 0;
 694                 }
 695
 696                 ret = viommu_domain_map_identity(vdev, vdomain);
 697                 if (ret) {
 698                         ida_free(&viommu->domain_ids, vdomain->id);
 699                         vdomain->viommu = NULL;
 700                         return ret;
 701                 }
 702         }
 703
 704         return 0;
 705 }
 706
 707 static void viommu_domain_free(struct iommu_domain *domain)
 708 {
 709         struct viommu_domain *vdomain = to_viommu_domain(domain);
 710
 711         /* Free all remaining mappings */
 712         viommu_del_mappings(vdomain, 0, ULLONG_MAX);
 713
 714         if (vdomain->viommu)
 715                 ida_free(&vdomain->viommu->domain_ids, vdomain->id);
 716
 717         kfree(vdomain);
 718 }
 719
 720 static int viommu_attach_dev(struct iommu_domain *domain, struct device *dev)
 721 {
 722         int i;
 723         int ret = 0;
 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);
 728
 729         mutex_lock(&vdomain->mutex);
 730         if (!vdomain->viommu) {
 731                 /*
 732                  * Properly initialize the domain now that we know which viommu
 733                  * owns it.
 734                  */
 735                 ret = viommu_domain_finalise(vdev, domain);
 736         } else if (vdomain->viommu != vdev->viommu) {
 737                 ret = -EINVAL;
 738         }
 739         mutex_unlock(&vdomain->mutex);
 740
 741         if (ret)
 742                 return ret;
 743
 744         /*
 745          * In the virtio-iommu device, when attaching the endpoint to a new
 746          * domain, it is detached from the old one and, if as a result the
 747          * old domain isn't attached to any endpoint, all mappings are removed
 748          * from the old domain and it is freed.
 749          *
 750          * In the driver the old domain still exists, and its mappings will be
 751          * recreated if it gets reattached to an endpoint. Otherwise it will be
 752          * freed explicitly.
 753          *
 754          * vdev->vdomain is protected by group->mutex
 755          */
 756         if (vdev->vdomain)
 757                 vdev->vdomain->nr_endpoints--;
 758
 759         req = (struct virtio_iommu_req_attach) {
 760                 .head.type      = VIRTIO_IOMMU_T_ATTACH,
 761                 .domain         = cpu_to_le32(vdomain->id),
 762         };
 763
 764         if (vdomain->bypass)
 765                 req.flags |= cpu_to_le32(VIRTIO_IOMMU_ATTACH_F_BYPASS);
 766
 767         for (i = 0; i < fwspec->num_ids; i++) {
 768                 req.endpoint = cpu_to_le32(fwspec->ids[i]);
 769
 770                 ret = viommu_send_req_sync(vdomain->viommu, &req, sizeof(req));
 771                 if (ret)
 772                         return ret;
 773         }
 774
 775         if (!vdomain->nr_endpoints) {
 776                 /*
 777                  * This endpoint is the first to be attached to the domain.
 778                  * Replay existing mappings (e.g. SW MSI).
 779                  */
 780                 ret = viommu_replay_mappings(vdomain);
 781                 if (ret)
 782                         return ret;
 783         }
 784
 785         vdomain->nr_endpoints++;
 786         vdev->vdomain = vdomain;
 787
 788         return 0;
 789 }
 790
 791 static void viommu_detach_dev(struct viommu_endpoint *vdev)
 792 {
 793         int i;
 794         struct virtio_iommu_req_detach req;
 795         struct viommu_domain *vdomain = vdev->vdomain;
 796         struct iommu_fwspec *fwspec = dev_iommu_fwspec_get(vdev->dev);
 797
 798         if (!vdomain)
 799                 return;
 800
 801         req = (struct virtio_iommu_req_detach) {
 802                 .head.type      = VIRTIO_IOMMU_T_DETACH,
 803                 .domain         = cpu_to_le32(vdomain->id),
 804         };
 805
 806         for (i = 0; i < fwspec->num_ids; i++) {
 807                 req.endpoint = cpu_to_le32(fwspec->ids[i]);
 808                 WARN_ON(viommu_send_req_sync(vdev->viommu, &req, sizeof(req)));
 809         }
 810         vdomain->nr_endpoints--;
 811         vdev->vdomain = NULL;
 812 }
 813
 814 static int viommu_map_pages(struct iommu_domain *domain, unsigned long iova,
 815                             phys_addr_t paddr, size_t pgsize, size_t pgcount,
 816                             int prot, gfp_t gfp, size_t *mapped)
 817 {
 818         int ret;
 819         u32 flags;
 820         size_t size = pgsize * pgcount;
 821         u64 end = iova + size - 1;
 822         struct virtio_iommu_req_map map;
 823         struct viommu_domain *vdomain = to_viommu_domain(domain);
 824
 825         flags = (prot & IOMMU_READ ? VIRTIO_IOMMU_MAP_F_READ : 0) |
 826                 (prot & IOMMU_WRITE ? VIRTIO_IOMMU_MAP_F_WRITE : 0) |
 827                 (prot & IOMMU_MMIO ? VIRTIO_IOMMU_MAP_F_MMIO : 0);
 828
 829         if (flags & ~vdomain->map_flags)
 830                 return -EINVAL;
 831
 832         ret = viommu_add_mapping(vdomain, iova, end, paddr, flags);
 833         if (ret)
 834                 return ret;
 835
 836         if (vdomain->nr_endpoints) {
 837                 map = (struct virtio_iommu_req_map) {
 838                         .head.type      = VIRTIO_IOMMU_T_MAP,
 839                         .domain         = cpu_to_le32(vdomain->id),
 840                         .virt_start     = cpu_to_le64(iova),
 841                         .phys_start     = cpu_to_le64(paddr),
 842                         .virt_end       = cpu_to_le64(end),
 843                         .flags          = cpu_to_le32(flags),
 844                 };
 845
 846                 ret = viommu_add_req(vdomain->viommu, &map, sizeof(map));
 847                 if (ret) {
 848                         viommu_del_mappings(vdomain, iova, end);
 849                         return ret;
 850                 }
 851         }
 852         if (mapped)
 853                 *mapped = size;
 854
 855         return 0;
 856 }
 857
 858 static size_t viommu_unmap_pages(struct iommu_domain *domain, unsigned long iova,
 859                                  size_t pgsize, size_t pgcount,
 860                                  struct iommu_iotlb_gather *gather)
 861 {
 862         int ret = 0;
 863         size_t unmapped;
 864         struct virtio_iommu_req_unmap unmap;
 865         struct viommu_domain *vdomain = to_viommu_domain(domain);
 866         size_t size = pgsize * pgcount;
 867
 868         unmapped = viommu_del_mappings(vdomain, iova, iova + size - 1);
 869         if (unmapped < size)
 870                 return 0;
 871
 872         /* Device already removed all mappings after detach. */
 873         if (!vdomain->nr_endpoints)
 874                 return unmapped;
 875
 876         unmap = (struct virtio_iommu_req_unmap) {
 877                 .head.type      = VIRTIO_IOMMU_T_UNMAP,
 878                 .domain         = cpu_to_le32(vdomain->id),
 879                 .virt_start     = cpu_to_le64(iova),
 880                 .virt_end       = cpu_to_le64(iova + unmapped - 1),
 881         };
 882
 883         ret = viommu_add_req(vdomain->viommu, &unmap, sizeof(unmap));
 884         return ret ? 0 : unmapped;
 885 }
 886
 887 static phys_addr_t viommu_iova_to_phys(struct iommu_domain *domain,
 888                                        dma_addr_t iova)
 889 {
 890         u64 paddr = 0;
 891         unsigned long flags;
 892         struct viommu_mapping *mapping;
 893         struct interval_tree_node *node;
 894         struct viommu_domain *vdomain = to_viommu_domain(domain);
 895
 896         spin_lock_irqsave(&vdomain->mappings_lock, flags);
 897         node = interval_tree_iter_first(&vdomain->mappings, iova, iova);
 898         if (node) {
 899                 mapping = container_of(node, struct viommu_mapping, iova);
 900                 paddr = mapping->paddr + (iova - mapping->iova.start);
 901         }
 902         spin_unlock_irqrestore(&vdomain->mappings_lock, flags);
 903
 904         return paddr;
 905 }
 906
 907 static void viommu_iotlb_sync(struct iommu_domain *domain,
 908                               struct iommu_iotlb_gather *gather)
 909 {
 910         struct viommu_domain *vdomain = to_viommu_domain(domain);
 911
 912         viommu_sync_req(vdomain->viommu);
 913 }
 914
 915 static int viommu_iotlb_sync_map(struct iommu_domain *domain,
 916                                  unsigned long iova, size_t size)
 917 {
 918         struct viommu_domain *vdomain = to_viommu_domain(domain);
 919
 920         /*
 921          * May be called before the viommu is initialized including
 922          * while creating direct mapping
 923          */
 924         if (!vdomain->nr_endpoints)
 925                 return 0;
 926         return viommu_sync_req(vdomain->viommu);
 927 }
 928
 929 static void viommu_flush_iotlb_all(struct iommu_domain *domain)
 930 {
 931         struct viommu_domain *vdomain = to_viommu_domain(domain);
 932
 933         /*
 934          * May be called before the viommu is initialized including
 935          * while creating direct mapping
 936          */
 937         if (!vdomain->nr_endpoints)
 938                 return;
 939         viommu_sync_req(vdomain->viommu);
 940 }
 941
 942 static void viommu_get_resv_regions(struct device *dev, struct list_head *head)
 943 {
 944         struct iommu_resv_region *entry, *new_entry, *msi = NULL;
 945         struct viommu_endpoint *vdev = dev_iommu_priv_get(dev);
 946         int prot = IOMMU_WRITE | IOMMU_NOEXEC | IOMMU_MMIO;
 947
 948         list_for_each_entry(entry, &vdev->resv_regions, list) {
 949                 if (entry->type == IOMMU_RESV_MSI)
 950                         msi = entry;
 951
 952                 new_entry = kmemdup(entry, sizeof(*entry), GFP_KERNEL);
 953                 if (!new_entry)
 954                         return;
 955                 list_add_tail(&new_entry->list, head);
 956         }
 957
 958         /*
 959          * If the device didn't register any bypass MSI window, add a
 960          * software-mapped region.
 961          */
 962         if (!msi) {
 963                 msi = iommu_alloc_resv_region(MSI_IOVA_BASE, MSI_IOVA_LENGTH,
 964                                               prot, IOMMU_RESV_SW_MSI,
 965                                               GFP_KERNEL);
 966                 if (!msi)
 967                         return;
 968
 969                 list_add_tail(&msi->list, head);
 970         }
 971
 972         iommu_dma_get_resv_regions(dev, head);
 973 }
 974
 975 static struct iommu_ops viommu_ops;
 976 static struct virtio_driver virtio_iommu_drv;
 977
 978 static int viommu_match_node(struct device *dev, const void *data)
 979 {
 980         return device_match_fwnode(dev->parent, data);
 981 }
 982
 983 static struct viommu_dev *viommu_get_by_fwnode(struct fwnode_handle *fwnode)
 984 {
 985         struct device *dev = driver_find_device(&virtio_iommu_drv.driver, NULL,
 986                                                 fwnode, viommu_match_node);
 987         put_device(dev);
 988
 989         return dev ? dev_to_virtio(dev)->priv : NULL;
 990 }
 991
 992 static struct iommu_device *viommu_probe_device(struct device *dev)
 993 {
 994         int ret;
 995         struct viommu_endpoint *vdev;
 996         struct viommu_dev *viommu = NULL;
 997         struct iommu_fwspec *fwspec = dev_iommu_fwspec_get(dev);
 998
 999         viommu = viommu_get_by_fwnode(fwspec->iommu_fwnode);
1000         if (!viommu)
1001                 return ERR_PTR(-ENODEV);
1002
1003         vdev = kzalloc(sizeof(*vdev), GFP_KERNEL);
1004         if (!vdev)
1005                 return ERR_PTR(-ENOMEM);
1006
1007         vdev->dev = dev;
1008         vdev->viommu = viommu;
1009         INIT_LIST_HEAD(&vdev->resv_regions);
1010         dev_iommu_priv_set(dev, vdev);
1011
1012         if (viommu->probe_size) {
1013                 /* Get additional information for this endpoint */
1014                 ret = viommu_probe_endpoint(viommu, dev);
1015                 if (ret)
1016                         goto err_free_dev;
1017         }
1018
1019         return &viommu->iommu;
1020
1021 err_free_dev:
1022         iommu_put_resv_regions(dev, &vdev->resv_regions);
1023         kfree(vdev);
1024
1025         return ERR_PTR(ret);
1026 }
1027
1028 static void viommu_probe_finalize(struct device *dev)
1029 {
1030 #ifndef CONFIG_ARCH_HAS_SETUP_DMA_OPS
1031         /* First clear the DMA ops in case we're switching from a DMA domain */
1032         set_dma_ops(dev, NULL);
1033         iommu_setup_dma_ops(dev, 0, U64_MAX);
1034 #endif
1035 }
1036
1037 static void viommu_release_device(struct device *dev)
1038 {
1039         struct viommu_endpoint *vdev = dev_iommu_priv_get(dev);
1040
1041         viommu_detach_dev(vdev);
1042         iommu_put_resv_regions(dev, &vdev->resv_regions);
1043         kfree(vdev);
1044 }
1045
1046 static struct iommu_group *viommu_device_group(struct device *dev)
1047 {
1048         if (dev_is_pci(dev))
1049                 return pci_device_group(dev);
1050         else
1051                 return generic_device_group(dev);
1052 }
1053
1054 static int viommu_of_xlate(struct device *dev, struct of_phandle_args *args)
1055 {
1056         return iommu_fwspec_add_ids(dev, args->args, 1);
1057 }
1058
1059 static bool viommu_capable(struct device *dev, enum iommu_cap cap)
1060 {
1061         switch (cap) {
1062         case IOMMU_CAP_CACHE_COHERENCY:
1063                 return true;
1064         case IOMMU_CAP_DEFERRED_FLUSH:
1065                 return true;
1066         default:
1067                 return false;
1068         }
1069 }
1070
1071 static struct iommu_ops viommu_ops = {
1072         .capable                = viommu_capable,
1073         .domain_alloc           = viommu_domain_alloc,
1074         .probe_device           = viommu_probe_device,
1075         .probe_finalize         = viommu_probe_finalize,
1076         .release_device         = viommu_release_device,
1077         .device_group           = viommu_device_group,
1078         .get_resv_regions       = viommu_get_resv_regions,
1079         .of_xlate               = viommu_of_xlate,
1080         .owner                  = THIS_MODULE,
1081         .default_domain_ops = &(const struct iommu_domain_ops) {
1082                 .attach_dev             = viommu_attach_dev,
1083                 .map_pages              = viommu_map_pages,
1084                 .unmap_pages            = viommu_unmap_pages,
1085                 .iova_to_phys           = viommu_iova_to_phys,
1086                 .flush_iotlb_all        = viommu_flush_iotlb_all,
1087                 .iotlb_sync             = viommu_iotlb_sync,
1088                 .iotlb_sync_map         = viommu_iotlb_sync_map,
1089                 .free                   = viommu_domain_free,
1090         }
1091 };
1092
1093 static int viommu_init_vqs(struct viommu_dev *viommu)
1094 {
1095         struct virtio_device *vdev = dev_to_virtio(viommu->dev);
1096         const char *names[] = { "request", "event" };
1097         vq_callback_t *callbacks[] = {
1098                 NULL, /* No async requests */
1099                 viommu_event_handler,
1100         };
1101
1102         return virtio_find_vqs(vdev, VIOMMU_NR_VQS, viommu->vqs, callbacks,
1103                                names, NULL);
1104 }
1105
1106 static int viommu_fill_evtq(struct viommu_dev *viommu)
1107 {
1108         int i, ret;
1109         struct scatterlist sg[1];
1110         struct viommu_event *evts;
1111         struct virtqueue *vq = viommu->vqs[VIOMMU_EVENT_VQ];
1112         size_t nr_evts = vq->num_free;
1113
1114         viommu->evts = evts = devm_kmalloc_array(viommu->dev, nr_evts,
1115                                                  sizeof(*evts), GFP_KERNEL);
1116         if (!evts)
1117                 return -ENOMEM;
1118
1119         for (i = 0; i < nr_evts; i++) {
1120                 sg_init_one(sg, &evts[i], sizeof(*evts));
1121                 ret = virtqueue_add_inbuf(vq, sg, 1, &evts[i], GFP_KERNEL);
1122                 if (ret)
1123                         return ret;
1124         }
1125
1126         return 0;
1127 }
1128
1129 static int viommu_probe(struct virtio_device *vdev)
1130 {
1131         struct device *parent_dev = vdev->dev.parent;
1132         struct viommu_dev *viommu = NULL;
1133         struct device *dev = &vdev->dev;
1134         u64 input_start = 0;
1135         u64 input_end = -1UL;
1136         int ret;
1137
1138         if (!virtio_has_feature(vdev, VIRTIO_F_VERSION_1) ||
1139             !virtio_has_feature(vdev, VIRTIO_IOMMU_F_MAP_UNMAP))
1140                 return -ENODEV;
1141
1142         viommu = devm_kzalloc(dev, sizeof(*viommu), GFP_KERNEL);
1143         if (!viommu)
1144                 return -ENOMEM;
1145
1146         spin_lock_init(&viommu->request_lock);
1147         ida_init(&viommu->domain_ids);
1148         viommu->dev = dev;
1149         viommu->vdev = vdev;
1150         INIT_LIST_HEAD(&viommu->requests);
1151
1152         ret = viommu_init_vqs(viommu);
1153         if (ret)
1154                 return ret;
1155
1156         virtio_cread_le(vdev, struct virtio_iommu_config, page_size_mask,
1157                         &viommu->pgsize_bitmap);
1158
1159         if (!viommu->pgsize_bitmap) {
1160                 ret = -EINVAL;
1161                 goto err_free_vqs;
1162         }
1163
1164         viommu->map_flags = VIRTIO_IOMMU_MAP_F_READ | VIRTIO_IOMMU_MAP_F_WRITE;
1165         viommu->last_domain = ~0U;
1166
1167         /* Optional features */
1168         virtio_cread_le_feature(vdev, VIRTIO_IOMMU_F_INPUT_RANGE,
1169                                 struct virtio_iommu_config, input_range.start,
1170                                 &input_start);
1171
1172         virtio_cread_le_feature(vdev, VIRTIO_IOMMU_F_INPUT_RANGE,
1173                                 struct virtio_iommu_config, input_range.end,
1174                                 &input_end);
1175
1176         virtio_cread_le_feature(vdev, VIRTIO_IOMMU_F_DOMAIN_RANGE,
1177                                 struct virtio_iommu_config, domain_range.start,
1178                                 &viommu->first_domain);
1179
1180         virtio_cread_le_feature(vdev, VIRTIO_IOMMU_F_DOMAIN_RANGE,
1181                                 struct virtio_iommu_config, domain_range.end,
1182                                 &viommu->last_domain);
1183
1184         virtio_cread_le_feature(vdev, VIRTIO_IOMMU_F_PROBE,
1185                                 struct virtio_iommu_config, probe_size,
1186                                 &viommu->probe_size);
1187
1188         viommu->geometry = (struct iommu_domain_geometry) {
1189                 .aperture_start = input_start,
1190                 .aperture_end   = input_end,
1191                 .force_aperture = true,
1192         };
1193
1194         if (virtio_has_feature(vdev, VIRTIO_IOMMU_F_MMIO))
1195                 viommu->map_flags |= VIRTIO_IOMMU_MAP_F_MMIO;
1196
1197         viommu_ops.pgsize_bitmap = viommu->pgsize_bitmap;
1198
1199         virtio_device_ready(vdev);
1200
1201         /* Populate the event queue with buffers */
1202         ret = viommu_fill_evtq(viommu);
1203         if (ret)
1204                 goto err_free_vqs;
1205
1206         ret = iommu_device_sysfs_add(&viommu->iommu, dev, NULL, "%s",
1207                                      virtio_bus_name(vdev));
1208         if (ret)
1209                 goto err_free_vqs;
1210
1211         iommu_device_register(&viommu->iommu, &viommu_ops, parent_dev);
1212
1213         vdev->priv = viommu;
1214
1215         dev_info(dev, "input address: %u bits\n",
1216                  order_base_2(viommu->geometry.aperture_end));
1217         dev_info(dev, "page mask: %#llx\n", viommu->pgsize_bitmap);
1218
1219         return 0;
1220
1221 err_free_vqs:
1222         vdev->config->del_vqs(vdev);
1223
1224         return ret;
1225 }
1226
1227 static void viommu_remove(struct virtio_device *vdev)
1228 {
1229         struct viommu_dev *viommu = vdev->priv;
1230
1231         iommu_device_sysfs_remove(&viommu->iommu);
1232         iommu_device_unregister(&viommu->iommu);
1233
1234         /* Stop all virtqueues */
1235         virtio_reset_device(vdev);
1236         vdev->config->del_vqs(vdev);
1237
1238         dev_info(&vdev->dev, "device removed\n");
1239 }
1240
1241 static void viommu_config_changed(struct virtio_device *vdev)
1242 {
1243         dev_warn(&vdev->dev, "config changed\n");
1244 }
1245
1246 static unsigned int features[] = {
1247         VIRTIO_IOMMU_F_MAP_UNMAP,
1248         VIRTIO_IOMMU_F_INPUT_RANGE,
1249         VIRTIO_IOMMU_F_DOMAIN_RANGE,
1250         VIRTIO_IOMMU_F_PROBE,
1251         VIRTIO_IOMMU_F_MMIO,
1252         VIRTIO_IOMMU_F_BYPASS_CONFIG,
1253 };
1254
1255 static struct virtio_device_id id_table[] = {
1256         { VIRTIO_ID_IOMMU, VIRTIO_DEV_ANY_ID },
1257         { 0 },
1258 };
1259 MODULE_DEVICE_TABLE(virtio, id_table);
1260
1261 static struct virtio_driver virtio_iommu_drv = {
1262         .driver.name            = KBUILD_MODNAME,
1263         .driver.owner           = THIS_MODULE,
1264         .id_table               = id_table,
1265         .feature_table          = features,
1266         .feature_table_size     = ARRAY_SIZE(features),
1267         .probe                  = viommu_probe,
1268         .remove                 = viommu_remove,
1269         .config_changed         = viommu_config_changed,
1270 };
1271
1272 module_virtio_driver(virtio_iommu_drv);
1273
1274 MODULE_DESCRIPTION("Virtio IOMMU driver");
1275 MODULE_AUTHOR("Jean-Philippe Brucker <jean-philippe.brucker@arm.com>");
1276 MODULE_LICENSE("GPL v2");