drivers/vhost/vdpa.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2018-2020 Intel Corporation.
   4  * Copyright (C) 2020 Red Hat, Inc.
   5  *
   6  * Author: Tiwei Bie <tiwei.bie@intel.com>
   7  *         Jason Wang <jasowang@redhat.com>
   8  *
   9  * Thanks Michael S. Tsirkin for the valuable comments and
  10  * suggestions.  And thanks to Cunming Liang and Zhihong Wang for all
  11  * their supports.
  12  */
  13
  14 #include <linux/kernel.h>
  15 #include <linux/module.h>
  16 #include <linux/cdev.h>
  17 #include <linux/device.h>
  18 #include <linux/mm.h>
  19 #include <linux/iommu.h>
  20 #include <linux/uuid.h>
  21 #include <linux/vdpa.h>
  22 #include <linux/nospec.h>
  23 #include <linux/vhost.h>
  24 #include <linux/virtio_net.h>
  25
  26 #include "vhost.h"
  27
  28 enum {
  29         VHOST_VDPA_BACKEND_FEATURES =
  30         (1ULL << VHOST_BACKEND_F_IOTLB_MSG_V2) |
  31         (1ULL << VHOST_BACKEND_F_IOTLB_BATCH),
  32 };
  33
  34 #define VHOST_VDPA_DEV_MAX (1U << MINORBITS)
  35
  36 struct vhost_vdpa {
  37         struct vhost_dev vdev;
  38         struct iommu_domain *domain;
  39         struct vhost_virtqueue *vqs;
  40         struct completion completion;
  41         struct vdpa_device *vdpa;
  42         struct device dev;
  43         struct cdev cdev;
  44         atomic_t opened;
  45         int nvqs;
  46         int virtio_id;
  47         int minor;
  48         struct eventfd_ctx *config_ctx;
  49         int in_batch;
  50         struct vdpa_iova_range range;
  51 };
  52
  53 static DEFINE_IDA(vhost_vdpa_ida);
  54
  55 static dev_t vhost_vdpa_major;
  56
  57 static void handle_vq_kick(struct vhost_work *work)
  58 {
  59         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
  60                                                   poll.work);
  61         struct vhost_vdpa *v = container_of(vq->dev, struct vhost_vdpa, vdev);
  62         const struct vdpa_config_ops *ops = v->vdpa->config;
  63
  64         ops->kick_vq(v->vdpa, vq - v->vqs);
  65 }
  66
  67 static irqreturn_t vhost_vdpa_virtqueue_cb(void *private)
  68 {
  69         struct vhost_virtqueue *vq = private;
  70         struct eventfd_ctx *call_ctx = vq->call_ctx.ctx;
  71
  72         if (call_ctx)
  73                 eventfd_signal(call_ctx, 1);
  74
  75         return IRQ_HANDLED;
  76 }
  77
  78 static irqreturn_t vhost_vdpa_config_cb(void *private)
  79 {
  80         struct vhost_vdpa *v = private;
  81         struct eventfd_ctx *config_ctx = v->config_ctx;
  82
  83         if (config_ctx)
  84                 eventfd_signal(config_ctx, 1);
  85
  86         return IRQ_HANDLED;
  87 }
  88
  89 static void vhost_vdpa_setup_vq_irq(struct vhost_vdpa *v, u16 qid)
  90 {
  91         struct vhost_virtqueue *vq = &v->vqs[qid];
  92         const struct vdpa_config_ops *ops = v->vdpa->config;
  93         struct vdpa_device *vdpa = v->vdpa;
  94         int ret, irq;
  95
  96         if (!ops->get_vq_irq)
  97                 return;
  98
  99         irq = ops->get_vq_irq(vdpa, qid);
 100         irq_bypass_unregister_producer(&vq->call_ctx.producer);
 101         if (!vq->call_ctx.ctx || irq < 0)
 102                 return;
 103
 104         vq->call_ctx.producer.token = vq->call_ctx.ctx;
 105         vq->call_ctx.producer.irq = irq;
 106         ret = irq_bypass_register_producer(&vq->call_ctx.producer);
 107         if (unlikely(ret))
 108                 dev_info(&v->dev, "vq %u, irq bypass producer (token %p) registration fails, ret =  %d\n",
 109                          qid, vq->call_ctx.producer.token, ret);
 110 }
 111
 112 static void vhost_vdpa_unsetup_vq_irq(struct vhost_vdpa *v, u16 qid)
 113 {
 114         struct vhost_virtqueue *vq = &v->vqs[qid];
 115
 116         irq_bypass_unregister_producer(&vq->call_ctx.producer);
 117 }
 118
 119 static void vhost_vdpa_reset(struct vhost_vdpa *v)
 120 {
 121         struct vdpa_device *vdpa = v->vdpa;
 122
 123         vdpa_reset(vdpa);
 124         v->in_batch = 0;
 125 }
 126
 127 static long vhost_vdpa_get_device_id(struct vhost_vdpa *v, u8 __user *argp)
 128 {
 129         struct vdpa_device *vdpa = v->vdpa;
 130         const struct vdpa_config_ops *ops = vdpa->config;
 131         u32 device_id;
 132
 133         device_id = ops->get_device_id(vdpa);
 134
 135         if (copy_to_user(argp, &device_id, sizeof(device_id)))
 136                 return -EFAULT;
 137
 138         return 0;
 139 }
 140
 141 static long vhost_vdpa_get_status(struct vhost_vdpa *v, u8 __user *statusp)
 142 {
 143         struct vdpa_device *vdpa = v->vdpa;
 144         const struct vdpa_config_ops *ops = vdpa->config;
 145         u8 status;
 146
 147         status = ops->get_status(vdpa);
 148
 149         if (copy_to_user(statusp, &status, sizeof(status)))
 150                 return -EFAULT;
 151
 152         return 0;
 153 }
 154
 155 static long vhost_vdpa_set_status(struct vhost_vdpa *v, u8 __user *statusp)
 156 {
 157         struct vdpa_device *vdpa = v->vdpa;
 158         const struct vdpa_config_ops *ops = vdpa->config;
 159         u8 status, status_old;
 160         int nvqs = v->nvqs;
 161         u16 i;
 162
 163         if (copy_from_user(&status, statusp, sizeof(status)))
 164                 return -EFAULT;
 165
 166         status_old = ops->get_status(vdpa);
 167
 168         /*
 169          * Userspace shouldn't remove status bits unless reset the
 170          * status to 0.
 171          */
 172         if (status != 0 && (ops->get_status(vdpa) & ~status) != 0)
 173                 return -EINVAL;
 174
 175         ops->set_status(vdpa, status);
 176
 177         if ((status & VIRTIO_CONFIG_S_DRIVER_OK) && !(status_old & VIRTIO_CONFIG_S_DRIVER_OK))
 178                 for (i = 0; i < nvqs; i++)
 179                         vhost_vdpa_setup_vq_irq(v, i);
 180
 181         if ((status_old & VIRTIO_CONFIG_S_DRIVER_OK) && !(status & VIRTIO_CONFIG_S_DRIVER_OK))
 182                 for (i = 0; i < nvqs; i++)
 183                         vhost_vdpa_unsetup_vq_irq(v, i);
 184
 185         return 0;
 186 }
 187
 188 static int vhost_vdpa_config_validate(struct vhost_vdpa *v,
 189                                       struct vhost_vdpa_config *c)
 190 {
 191         long size = 0;
 192
 193         switch (v->virtio_id) {
 194         case VIRTIO_ID_NET:
 195                 size = sizeof(struct virtio_net_config);
 196                 break;
 197         }
 198
 199         if (c->len == 0)
 200                 return -EINVAL;
 201
 202         if (c->len > size - c->off)
 203                 return -E2BIG;
 204
 205         return 0;
 206 }
 207
 208 static long vhost_vdpa_get_config(struct vhost_vdpa *v,
 209                                   struct vhost_vdpa_config __user *c)
 210 {
 211         struct vdpa_device *vdpa = v->vdpa;
 212         struct vhost_vdpa_config config;
 213         unsigned long size = offsetof(struct vhost_vdpa_config, buf);
 214         u8 *buf;
 215
 216         if (copy_from_user(&config, c, size))
 217                 return -EFAULT;
 218         if (vhost_vdpa_config_validate(v, &config))
 219                 return -EINVAL;
 220         buf = kvzalloc(config.len, GFP_KERNEL);
 221         if (!buf)
 222                 return -ENOMEM;
 223
 224         vdpa_get_config(vdpa, config.off, buf, config.len);
 225
 226         if (copy_to_user(c->buf, buf, config.len)) {
 227                 kvfree(buf);
 228                 return -EFAULT;
 229         }
 230
 231         kvfree(buf);
 232         return 0;
 233 }
 234
 235 static long vhost_vdpa_set_config(struct vhost_vdpa *v,
 236                                   struct vhost_vdpa_config __user *c)
 237 {
 238         struct vdpa_device *vdpa = v->vdpa;
 239         const struct vdpa_config_ops *ops = vdpa->config;
 240         struct vhost_vdpa_config config;
 241         unsigned long size = offsetof(struct vhost_vdpa_config, buf);
 242         u8 *buf;
 243
 244         if (copy_from_user(&config, c, size))
 245                 return -EFAULT;
 246         if (vhost_vdpa_config_validate(v, &config))
 247                 return -EINVAL;
 248
 249         buf = vmemdup_user(c->buf, config.len);
 250         if (IS_ERR(buf))
 251                 return PTR_ERR(buf);
 252
 253         ops->set_config(vdpa, config.off, buf, config.len);
 254
 255         kvfree(buf);
 256         return 0;
 257 }
 258
 259 static long vhost_vdpa_get_features(struct vhost_vdpa *v, u64 __user *featurep)
 260 {
 261         struct vdpa_device *vdpa = v->vdpa;
 262         const struct vdpa_config_ops *ops = vdpa->config;
 263         u64 features;
 264
 265         features = ops->get_features(vdpa);
 266
 267         if (copy_to_user(featurep, &features, sizeof(features)))
 268                 return -EFAULT;
 269
 270         return 0;
 271 }
 272
 273 static long vhost_vdpa_set_features(struct vhost_vdpa *v, u64 __user *featurep)
 274 {
 275         struct vdpa_device *vdpa = v->vdpa;
 276         const struct vdpa_config_ops *ops = vdpa->config;
 277         u64 features;
 278
 279         /*
 280          * It's not allowed to change the features after they have
 281          * been negotiated.
 282          */
 283         if (ops->get_status(vdpa) & VIRTIO_CONFIG_S_FEATURES_OK)
 284                 return -EBUSY;
 285
 286         if (copy_from_user(&features, featurep, sizeof(features)))
 287                 return -EFAULT;
 288
 289         if (vdpa_set_features(vdpa, features))
 290                 return -EINVAL;
 291
 292         return 0;
 293 }
 294
 295 static long vhost_vdpa_get_vring_num(struct vhost_vdpa *v, u16 __user *argp)
 296 {
 297         struct vdpa_device *vdpa = v->vdpa;
 298         const struct vdpa_config_ops *ops = vdpa->config;
 299         u16 num;
 300
 301         num = ops->get_vq_num_max(vdpa);
 302
 303         if (copy_to_user(argp, &num, sizeof(num)))
 304                 return -EFAULT;
 305
 306         return 0;
 307 }
 308
 309 static void vhost_vdpa_config_put(struct vhost_vdpa *v)
 310 {
 311         if (v->config_ctx) {
 312                 eventfd_ctx_put(v->config_ctx);
 313                 v->config_ctx = NULL;
 314         }
 315 }
 316
 317 static long vhost_vdpa_set_config_call(struct vhost_vdpa *v, u32 __user *argp)
 318 {
 319         struct vdpa_callback cb;
 320         int fd;
 321         struct eventfd_ctx *ctx;
 322
 323         cb.callback = vhost_vdpa_config_cb;
 324         cb.private = v->vdpa;
 325         if (copy_from_user(&fd, argp, sizeof(fd)))
 326                 return  -EFAULT;
 327
 328         ctx = fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(fd);
 329         swap(ctx, v->config_ctx);
 330
 331         if (!IS_ERR_OR_NULL(ctx))
 332                 eventfd_ctx_put(ctx);
 333
 334         if (IS_ERR(v->config_ctx)) {
 335                 long ret = PTR_ERR(v->config_ctx);
 336
 337                 v->config_ctx = NULL;
 338                 return ret;
 339         }
 340
 341         v->vdpa->config->set_config_cb(v->vdpa, &cb);
 342
 343         return 0;
 344 }
 345
 346 static long vhost_vdpa_get_iova_range(struct vhost_vdpa *v, u32 __user *argp)
 347 {
 348         struct vhost_vdpa_iova_range range = {
 349                 .first = v->range.first,
 350                 .last = v->range.last,
 351         };
 352
 353         if (copy_to_user(argp, &range, sizeof(range)))
 354                 return -EFAULT;
 355         return 0;
 356 }
 357
 358 static long vhost_vdpa_vring_ioctl(struct vhost_vdpa *v, unsigned int cmd,
 359                                    void __user *argp)
 360 {
 361         struct vdpa_device *vdpa = v->vdpa;
 362         const struct vdpa_config_ops *ops = vdpa->config;
 363         struct vdpa_vq_state vq_state;
 364         struct vdpa_callback cb;
 365         struct vhost_virtqueue *vq;
 366         struct vhost_vring_state s;
 367         u32 idx;
 368         long r;
 369
 370         r = get_user(idx, (u32 __user *)argp);
 371         if (r < 0)
 372                 return r;
 373
 374         if (idx >= v->nvqs)
 375                 return -ENOBUFS;
 376
 377         idx = array_index_nospec(idx, v->nvqs);
 378         vq = &v->vqs[idx];
 379
 380         switch (cmd) {
 381         case VHOST_VDPA_SET_VRING_ENABLE:
 382                 if (copy_from_user(&s, argp, sizeof(s)))
 383                         return -EFAULT;
 384                 ops->set_vq_ready(vdpa, idx, s.num);
 385                 return 0;
 386         case VHOST_GET_VRING_BASE:
 387                 r = ops->get_vq_state(v->vdpa, idx, &vq_state);
 388                 if (r)
 389                         return r;
 390
 391                 vq->last_avail_idx = vq_state.avail_index;
 392                 break;
 393         }
 394
 395         r = vhost_vring_ioctl(&v->vdev, cmd, argp);
 396         if (r)
 397                 return r;
 398
 399         switch (cmd) {
 400         case VHOST_SET_VRING_ADDR:
 401                 if (ops->set_vq_address(vdpa, idx,
 402                                         (u64)(uintptr_t)vq->desc,
 403                                         (u64)(uintptr_t)vq->avail,
 404                                         (u64)(uintptr_t)vq->used))
 405                         r = -EINVAL;
 406                 break;
 407
 408         case VHOST_SET_VRING_BASE:
 409                 vq_state.avail_index = vq->last_avail_idx;
 410                 if (ops->set_vq_state(vdpa, idx, &vq_state))
 411                         r = -EINVAL;
 412                 break;
 413
 414         case VHOST_SET_VRING_CALL:
 415                 if (vq->call_ctx.ctx) {
 416                         cb.callback = vhost_vdpa_virtqueue_cb;
 417                         cb.private = vq;
 418                 } else {
 419                         cb.callback = NULL;
 420                         cb.private = NULL;
 421                 }
 422                 ops->set_vq_cb(vdpa, idx, &cb);
 423                 vhost_vdpa_setup_vq_irq(v, idx);
 424                 break;
 425
 426         case VHOST_SET_VRING_NUM:
 427                 ops->set_vq_num(vdpa, idx, vq->num);
 428                 break;
 429         }
 430
 431         return r;
 432 }
 433
 434 static long vhost_vdpa_unlocked_ioctl(struct file *filep,
 435                                       unsigned int cmd, unsigned long arg)
 436 {
 437         struct vhost_vdpa *v = filep->private_data;
 438         struct vhost_dev *d = &v->vdev;
 439         void __user *argp = (void __user *)arg;
 440         u64 __user *featurep = argp;
 441         u64 features;
 442         long r = 0;
 443
 444         if (cmd == VHOST_SET_BACKEND_FEATURES) {
 445                 if (copy_from_user(&features, featurep, sizeof(features)))
 446                         return -EFAULT;
 447                 if (features & ~VHOST_VDPA_BACKEND_FEATURES)
 448                         return -EOPNOTSUPP;
 449                 vhost_set_backend_features(&v->vdev, features);
 450                 return 0;
 451         }
 452
 453         mutex_lock(&d->mutex);
 454
 455         switch (cmd) {
 456         case VHOST_VDPA_GET_DEVICE_ID:
 457                 r = vhost_vdpa_get_device_id(v, argp);
 458                 break;
 459         case VHOST_VDPA_GET_STATUS:
 460                 r = vhost_vdpa_get_status(v, argp);
 461                 break;
 462         case VHOST_VDPA_SET_STATUS:
 463                 r = vhost_vdpa_set_status(v, argp);
 464                 break;
 465         case VHOST_VDPA_GET_CONFIG:
 466                 r = vhost_vdpa_get_config(v, argp);
 467                 break;
 468         case VHOST_VDPA_SET_CONFIG:
 469                 r = vhost_vdpa_set_config(v, argp);
 470                 break;
 471         case VHOST_GET_FEATURES:
 472                 r = vhost_vdpa_get_features(v, argp);
 473                 break;
 474         case VHOST_SET_FEATURES:
 475                 r = vhost_vdpa_set_features(v, argp);
 476                 break;
 477         case VHOST_VDPA_GET_VRING_NUM:
 478                 r = vhost_vdpa_get_vring_num(v, argp);
 479                 break;
 480         case VHOST_SET_LOG_BASE:
 481         case VHOST_SET_LOG_FD:
 482                 r = -ENOIOCTLCMD;
 483                 break;
 484         case VHOST_VDPA_SET_CONFIG_CALL:
 485                 r = vhost_vdpa_set_config_call(v, argp);
 486                 break;
 487         case VHOST_GET_BACKEND_FEATURES:
 488                 features = VHOST_VDPA_BACKEND_FEATURES;
 489                 if (copy_to_user(featurep, &features, sizeof(features)))
 490                         r = -EFAULT;
 491                 break;
 492         case VHOST_VDPA_GET_IOVA_RANGE:
 493                 r = vhost_vdpa_get_iova_range(v, argp);
 494                 break;
 495         default:
 496                 r = vhost_dev_ioctl(&v->vdev, cmd, argp);
 497                 if (r == -ENOIOCTLCMD)
 498                         r = vhost_vdpa_vring_ioctl(v, cmd, argp);
 499                 break;
 500         }
 501
 502         mutex_unlock(&d->mutex);
 503         return r;
 504 }
 505
 506 static void vhost_vdpa_iotlb_unmap(struct vhost_vdpa *v, u64 start, u64 last)
 507 {
 508         struct vhost_dev *dev = &v->vdev;
 509         struct vhost_iotlb *iotlb = dev->iotlb;
 510         struct vhost_iotlb_map *map;
 511         struct page *page;
 512         unsigned long pfn, pinned;
 513
 514         while ((map = vhost_iotlb_itree_first(iotlb, start, last)) != NULL) {
 515                 pinned = map->size >> PAGE_SHIFT;
 516                 for (pfn = map->addr >> PAGE_SHIFT;
 517                      pinned > 0; pfn++, pinned--) {
 518                         page = pfn_to_page(pfn);
 519                         if (map->perm & VHOST_ACCESS_WO)
 520                                 set_page_dirty_lock(page);
 521                         unpin_user_page(page);
 522                 }
 523                 atomic64_sub(map->size >> PAGE_SHIFT, &dev->mm->pinned_vm);
 524                 vhost_iotlb_map_free(iotlb, map);
 525         }
 526 }
 527
 528 static void vhost_vdpa_iotlb_free(struct vhost_vdpa *v)
 529 {
 530         struct vhost_dev *dev = &v->vdev;
 531
 532         vhost_vdpa_iotlb_unmap(v, 0ULL, 0ULL - 1);
 533         kfree(dev->iotlb);
 534         dev->iotlb = NULL;
 535 }
 536
 537 static int perm_to_iommu_flags(u32 perm)
 538 {
 539         int flags = 0;
 540
 541         switch (perm) {
 542         case VHOST_ACCESS_WO:
 543                 flags |= IOMMU_WRITE;
 544                 break;
 545         case VHOST_ACCESS_RO:
 546                 flags |= IOMMU_READ;
 547                 break;
 548         case VHOST_ACCESS_RW:
 549                 flags |= (IOMMU_WRITE | IOMMU_READ);
 550                 break;
 551         default:
 552                 WARN(1, "invalidate vhost IOTLB permission\n");
 553                 break;
 554         }
 555
 556         return flags | IOMMU_CACHE;
 557 }
 558
 559 static int vhost_vdpa_map(struct vhost_vdpa *v,
 560                           u64 iova, u64 size, u64 pa, u32 perm)
 561 {
 562         struct vhost_dev *dev = &v->vdev;
 563         struct vdpa_device *vdpa = v->vdpa;
 564         const struct vdpa_config_ops *ops = vdpa->config;
 565         int r = 0;
 566
 567         r = vhost_iotlb_add_range(dev->iotlb, iova, iova + size - 1,
 568                                   pa, perm);
 569         if (r)
 570                 return r;
 571
 572         if (ops->dma_map) {
 573                 r = ops->dma_map(vdpa, iova, size, pa, perm);
 574         } else if (ops->set_map) {
 575                 if (!v->in_batch)
 576                         r = ops->set_map(vdpa, dev->iotlb);
 577         } else {
 578                 r = iommu_map(v->domain, iova, pa, size,
 579                               perm_to_iommu_flags(perm));
 580         }
 581
 582         if (r)
 583                 vhost_iotlb_del_range(dev->iotlb, iova, iova + size - 1);
 584         else
 585                 atomic64_add(size >> PAGE_SHIFT, &dev->mm->pinned_vm);
 586
 587         return r;
 588 }
 589
 590 static void vhost_vdpa_unmap(struct vhost_vdpa *v, u64 iova, u64 size)
 591 {
 592         struct vhost_dev *dev = &v->vdev;
 593         struct vdpa_device *vdpa = v->vdpa;
 594         const struct vdpa_config_ops *ops = vdpa->config;
 595
 596         vhost_vdpa_iotlb_unmap(v, iova, iova + size - 1);
 597
 598         if (ops->dma_map) {
 599                 ops->dma_unmap(vdpa, iova, size);
 600         } else if (ops->set_map) {
 601                 if (!v->in_batch)
 602                         ops->set_map(vdpa, dev->iotlb);
 603         } else {
 604                 iommu_unmap(v->domain, iova, size);
 605         }
 606 }
 607
 608 static int vhost_vdpa_process_iotlb_update(struct vhost_vdpa *v,
 609                                            struct vhost_iotlb_msg *msg)
 610 {
 611         struct vhost_dev *dev = &v->vdev;
 612         struct vhost_iotlb *iotlb = dev->iotlb;
 613         struct page **page_list;
 614         unsigned long list_size = PAGE_SIZE / sizeof(struct page *);
 615         unsigned int gup_flags = FOLL_LONGTERM;
 616         unsigned long npages, cur_base, map_pfn, last_pfn = 0;
 617         unsigned long lock_limit, sz2pin, nchunks, i;
 618         u64 iova = msg->iova;
 619         long pinned;
 620         int ret = 0;
 621
 622         if (msg->iova < v->range.first ||
 623             msg->iova + msg->size - 1 > v->range.last)
 624                 return -EINVAL;
 625
 626         if (vhost_iotlb_itree_first(iotlb, msg->iova,
 627                                     msg->iova + msg->size - 1))
 628                 return -EEXIST;
 629
 630         /* Limit the use of memory for bookkeeping */
 631         page_list = (struct page **) __get_free_page(GFP_KERNEL);
 632         if (!page_list)
 633                 return -ENOMEM;
 634
 635         if (msg->perm & VHOST_ACCESS_WO)
 636                 gup_flags |= FOLL_WRITE;
 637
 638         npages = PAGE_ALIGN(msg->size + (iova & ~PAGE_MASK)) >> PAGE_SHIFT;
 639         if (!npages) {
 640                 ret = -EINVAL;
 641                 goto free;
 642         }
 643
 644         mmap_read_lock(dev->mm);
 645
 646         lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
 647         if (npages + atomic64_read(&dev->mm->pinned_vm) > lock_limit) {
 648                 ret = -ENOMEM;
 649                 goto unlock;
 650         }
 651
 652         cur_base = msg->uaddr & PAGE_MASK;
 653         iova &= PAGE_MASK;
 654         nchunks = 0;
 655
 656         while (npages) {
 657                 sz2pin = min_t(unsigned long, npages, list_size);
 658                 pinned = pin_user_pages(cur_base, sz2pin,
 659                                         gup_flags, page_list, NULL);
 660                 if (sz2pin != pinned) {
 661                         if (pinned < 0) {
 662                                 ret = pinned;
 663                         } else {
 664                                 unpin_user_pages(page_list, pinned);
 665                                 ret = -ENOMEM;
 666                         }
 667                         goto out;
 668                 }
 669                 nchunks++;
 670
 671                 if (!last_pfn)
 672                         map_pfn = page_to_pfn(page_list[0]);
 673
 674                 for (i = 0; i < pinned; i++) {
 675                         unsigned long this_pfn = page_to_pfn(page_list[i]);
 676                         u64 csize;
 677
 678                         if (last_pfn && (this_pfn != last_pfn + 1)) {
 679                                 /* Pin a contiguous chunk of memory */
 680                                 csize = (last_pfn - map_pfn + 1) << PAGE_SHIFT;
 681                                 ret = vhost_vdpa_map(v, iova, csize,
 682                                                      map_pfn << PAGE_SHIFT,
 683                                                      msg->perm);
 684                                 if (ret) {
 685                                         /*
 686                                          * Unpin the pages that are left unmapped
 687                                          * from this point on in the current
 688                                          * page_list. The remaining outstanding
 689                                          * ones which may stride across several
 690                                          * chunks will be covered in the common
 691                                          * error path subsequently.
 692                                          */
 693                                         unpin_user_pages(&page_list[i],
 694                                                          pinned - i);
 695                                         goto out;
 696                                 }
 697
 698                                 map_pfn = this_pfn;
 699                                 iova += csize;
 700                                 nchunks = 0;
 701                         }
 702
 703                         last_pfn = this_pfn;
 704                 }
 705
 706                 cur_base += pinned << PAGE_SHIFT;
 707                 npages -= pinned;
 708         }
 709
 710         /* Pin the rest chunk */
 711         ret = vhost_vdpa_map(v, iova, (last_pfn - map_pfn + 1) << PAGE_SHIFT,
 712                              map_pfn << PAGE_SHIFT, msg->perm);
 713 out:
 714         if (ret) {
 715                 if (nchunks) {
 716                         unsigned long pfn;
 717
 718                         /*
 719                          * Unpin the outstanding pages which are yet to be
 720                          * mapped but haven't due to vdpa_map() or
 721                          * pin_user_pages() failure.
 722                          *
 723                          * Mapped pages are accounted in vdpa_map(), hence
 724                          * the corresponding unpinning will be handled by
 725                          * vdpa_unmap().
 726                          */
 727                         WARN_ON(!last_pfn);
 728                         for (pfn = map_pfn; pfn <= last_pfn; pfn++)
 729                                 unpin_user_page(pfn_to_page(pfn));
 730                 }
 731                 vhost_vdpa_unmap(v, msg->iova, msg->size);
 732         }
 733 unlock:
 734         mmap_read_unlock(dev->mm);
 735 free:
 736         free_page((unsigned long)page_list);
 737         return ret;
 738 }
 739
 740 static int vhost_vdpa_process_iotlb_msg(struct vhost_dev *dev,
 741                                         struct vhost_iotlb_msg *msg)
 742 {
 743         struct vhost_vdpa *v = container_of(dev, struct vhost_vdpa, vdev);
 744         struct vdpa_device *vdpa = v->vdpa;
 745         const struct vdpa_config_ops *ops = vdpa->config;
 746         int r = 0;
 747
 748         mutex_lock(&dev->mutex);
 749
 750         r = vhost_dev_check_owner(dev);
 751         if (r)
 752                 goto unlock;
 753
 754         switch (msg->type) {
 755         case VHOST_IOTLB_UPDATE:
 756                 r = vhost_vdpa_process_iotlb_update(v, msg);
 757                 break;
 758         case VHOST_IOTLB_INVALIDATE:
 759                 vhost_vdpa_unmap(v, msg->iova, msg->size);
 760                 break;
 761         case VHOST_IOTLB_BATCH_BEGIN:
 762                 v->in_batch = true;
 763                 break;
 764         case VHOST_IOTLB_BATCH_END:
 765                 if (v->in_batch && ops->set_map)
 766                         ops->set_map(vdpa, dev->iotlb);
 767                 v->in_batch = false;
 768                 break;
 769         default:
 770                 r = -EINVAL;
 771                 break;
 772         }
 773 unlock:
 774         mutex_unlock(&dev->mutex);
 775
 776         return r;
 777 }
 778
 779 static ssize_t vhost_vdpa_chr_write_iter(struct kiocb *iocb,
 780                                          struct iov_iter *from)
 781 {
 782         struct file *file = iocb->ki_filp;
 783         struct vhost_vdpa *v = file->private_data;
 784         struct vhost_dev *dev = &v->vdev;
 785
 786         return vhost_chr_write_iter(dev, from);
 787 }
 788
 789 static int vhost_vdpa_alloc_domain(struct vhost_vdpa *v)
 790 {
 791         struct vdpa_device *vdpa = v->vdpa;
 792         const struct vdpa_config_ops *ops = vdpa->config;
 793         struct device *dma_dev = vdpa_get_dma_dev(vdpa);
 794         struct bus_type *bus;
 795         int ret;
 796
 797         /* Device want to do DMA by itself */
 798         if (ops->set_map || ops->dma_map)
 799                 return 0;
 800
 801         bus = dma_dev->bus;
 802         if (!bus)
 803                 return -EFAULT;
 804
 805         if (!iommu_capable(bus, IOMMU_CAP_CACHE_COHERENCY))
 806                 return -ENOTSUPP;
 807
 808         v->domain = iommu_domain_alloc(bus);
 809         if (!v->domain)
 810                 return -EIO;
 811
 812         ret = iommu_attach_device(v->domain, dma_dev);
 813         if (ret)
 814                 goto err_attach;
 815
 816         return 0;
 817
 818 err_attach:
 819         iommu_domain_free(v->domain);
 820         return ret;
 821 }
 822
 823 static void vhost_vdpa_free_domain(struct vhost_vdpa *v)
 824 {
 825         struct vdpa_device *vdpa = v->vdpa;
 826         struct device *dma_dev = vdpa_get_dma_dev(vdpa);
 827
 828         if (v->domain) {
 829                 iommu_detach_device(v->domain, dma_dev);
 830                 iommu_domain_free(v->domain);
 831         }
 832
 833         v->domain = NULL;
 834 }
 835
 836 static void vhost_vdpa_set_iova_range(struct vhost_vdpa *v)
 837 {
 838         struct vdpa_iova_range *range = &v->range;
 839         struct vdpa_device *vdpa = v->vdpa;
 840         const struct vdpa_config_ops *ops = vdpa->config;
 841
 842         if (ops->get_iova_range) {
 843                 *range = ops->get_iova_range(vdpa);
 844         } else if (v->domain && v->domain->geometry.force_aperture) {
 845                 range->first = v->domain->geometry.aperture_start;
 846                 range->last = v->domain->geometry.aperture_end;
 847         } else {
 848                 range->first = 0;
 849                 range->last = ULLONG_MAX;
 850         }
 851 }
 852
 853 static int vhost_vdpa_open(struct inode *inode, struct file *filep)
 854 {
 855         struct vhost_vdpa *v;
 856         struct vhost_dev *dev;
 857         struct vhost_virtqueue **vqs;
 858         int nvqs, i, r, opened;
 859
 860         v = container_of(inode->i_cdev, struct vhost_vdpa, cdev);
 861
 862         opened = atomic_cmpxchg(&v->opened, 0, 1);
 863         if (opened)
 864                 return -EBUSY;
 865
 866         nvqs = v->nvqs;
 867         vhost_vdpa_reset(v);
 868
 869         vqs = kmalloc_array(nvqs, sizeof(*vqs), GFP_KERNEL);
 870         if (!vqs) {
 871                 r = -ENOMEM;
 872                 goto err;
 873         }
 874
 875         dev = &v->vdev;
 876         for (i = 0; i < nvqs; i++) {
 877                 vqs[i] = &v->vqs[i];
 878                 vqs[i]->handle_kick = handle_vq_kick;
 879         }
 880         vhost_dev_init(dev, vqs, nvqs, 0, 0, 0, false,
 881                        vhost_vdpa_process_iotlb_msg);
 882
 883         dev->iotlb = vhost_iotlb_alloc(0, 0);
 884         if (!dev->iotlb) {
 885                 r = -ENOMEM;
 886                 goto err_init_iotlb;
 887         }
 888
 889         r = vhost_vdpa_alloc_domain(v);
 890         if (r)
 891                 goto err_init_iotlb;
 892
 893         vhost_vdpa_set_iova_range(v);
 894
 895         filep->private_data = v;
 896
 897         return 0;
 898
 899 err_init_iotlb:
 900         vhost_dev_cleanup(&v->vdev);
 901         kfree(vqs);
 902 err:
 903         atomic_dec(&v->opened);
 904         return r;
 905 }
 906
 907 static void vhost_vdpa_clean_irq(struct vhost_vdpa *v)
 908 {
 909         int i;
 910
 911         for (i = 0; i < v->nvqs; i++)
 912                 vhost_vdpa_unsetup_vq_irq(v, i);
 913 }
 914
 915 static int vhost_vdpa_release(struct inode *inode, struct file *filep)
 916 {
 917         struct vhost_vdpa *v = filep->private_data;
 918         struct vhost_dev *d = &v->vdev;
 919
 920         mutex_lock(&d->mutex);
 921         filep->private_data = NULL;
 922         vhost_vdpa_reset(v);
 923         vhost_dev_stop(&v->vdev);
 924         vhost_vdpa_iotlb_free(v);
 925         vhost_vdpa_free_domain(v);
 926         vhost_vdpa_config_put(v);
 927         vhost_vdpa_clean_irq(v);
 928         vhost_dev_cleanup(&v->vdev);
 929         kfree(v->vdev.vqs);
 930         mutex_unlock(&d->mutex);
 931
 932         atomic_dec(&v->opened);
 933         complete(&v->completion);
 934
 935         return 0;
 936 }
 937
 938 #ifdef CONFIG_MMU
 939 static vm_fault_t vhost_vdpa_fault(struct vm_fault *vmf)
 940 {
 941         struct vhost_vdpa *v = vmf->vma->vm_file->private_data;
 942         struct vdpa_device *vdpa = v->vdpa;
 943         const struct vdpa_config_ops *ops = vdpa->config;
 944         struct vdpa_notification_area notify;
 945         struct vm_area_struct *vma = vmf->vma;
 946         u16 index = vma->vm_pgoff;
 947
 948         notify = ops->get_vq_notification(vdpa, index);
 949
 950         vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
 951         if (remap_pfn_range(vma, vmf->address & PAGE_MASK,
 952                             notify.addr >> PAGE_SHIFT, PAGE_SIZE,
 953                             vma->vm_page_prot))
 954                 return VM_FAULT_SIGBUS;
 955
 956         return VM_FAULT_NOPAGE;
 957 }
 958
 959 static const struct vm_operations_struct vhost_vdpa_vm_ops = {
 960         .fault = vhost_vdpa_fault,
 961 };
 962
 963 static int vhost_vdpa_mmap(struct file *file, struct vm_area_struct *vma)
 964 {
 965         struct vhost_vdpa *v = vma->vm_file->private_data;
 966         struct vdpa_device *vdpa = v->vdpa;
 967         const struct vdpa_config_ops *ops = vdpa->config;
 968         struct vdpa_notification_area notify;
 969         unsigned long index = vma->vm_pgoff;
 970
 971         if (vma->vm_end - vma->vm_start != PAGE_SIZE)
 972                 return -EINVAL;
 973         if ((vma->vm_flags & VM_SHARED) == 0)
 974                 return -EINVAL;
 975         if (vma->vm_flags & VM_READ)
 976                 return -EINVAL;
 977         if (index > 65535)
 978                 return -EINVAL;
 979         if (!ops->get_vq_notification)
 980                 return -ENOTSUPP;
 981
 982         /* To be safe and easily modelled by userspace, We only
 983          * support the doorbell which sits on the page boundary and
 984          * does not share the page with other registers.
 985          */
 986         notify = ops->get_vq_notification(vdpa, index);
 987         if (notify.addr & (PAGE_SIZE - 1))
 988                 return -EINVAL;
 989         if (vma->vm_end - vma->vm_start != notify.size)
 990                 return -ENOTSUPP;
 991
 992         vma->vm_ops = &vhost_vdpa_vm_ops;
 993         return 0;
 994 }
 995 #endif /* CONFIG_MMU */
 996
 997 static const struct file_operations vhost_vdpa_fops = {
 998         .owner          = THIS_MODULE,
 999         .open           = vhost_vdpa_open,
1000         .release        = vhost_vdpa_release,
1001         .write_iter     = vhost_vdpa_chr_write_iter,
1002         .unlocked_ioctl = vhost_vdpa_unlocked_ioctl,
1003 #ifdef CONFIG_MMU
1004         .mmap           = vhost_vdpa_mmap,
1005 #endif /* CONFIG_MMU */
1006         .compat_ioctl   = compat_ptr_ioctl,
1007 };
1008
1009 static void vhost_vdpa_release_dev(struct device *device)
1010 {
1011         struct vhost_vdpa *v =
1012                container_of(device, struct vhost_vdpa, dev);
1013
1014         ida_simple_remove(&vhost_vdpa_ida, v->minor);
1015         kfree(v->vqs);
1016         kfree(v);
1017 }
1018
1019 static int vhost_vdpa_probe(struct vdpa_device *vdpa)
1020 {
1021         const struct vdpa_config_ops *ops = vdpa->config;
1022         struct vhost_vdpa *v;
1023         int minor;
1024         int r;
1025
1026         /* Currently, we only accept the network devices. */
1027         if (ops->get_device_id(vdpa) != VIRTIO_ID_NET)
1028                 return -ENOTSUPP;
1029
1030         v = kzalloc(sizeof(*v), GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1031         if (!v)
1032                 return -ENOMEM;
1033
1034         minor = ida_simple_get(&vhost_vdpa_ida, 0,
1035                                VHOST_VDPA_DEV_MAX, GFP_KERNEL);
1036         if (minor < 0) {
1037                 kfree(v);
1038                 return minor;
1039         }
1040
1041         atomic_set(&v->opened, 0);
1042         v->minor = minor;
1043         v->vdpa = vdpa;
1044         v->nvqs = vdpa->nvqs;
1045         v->virtio_id = ops->get_device_id(vdpa);
1046
1047         device_initialize(&v->dev);
1048         v->dev.release = vhost_vdpa_release_dev;
1049         v->dev.parent = &vdpa->dev;
1050         v->dev.devt = MKDEV(MAJOR(vhost_vdpa_major), minor);
1051         v->vqs = kmalloc_array(v->nvqs, sizeof(struct vhost_virtqueue),
1052                                GFP_KERNEL);
1053         if (!v->vqs) {
1054                 r = -ENOMEM;
1055                 goto err;
1056         }
1057
1058         r = dev_set_name(&v->dev, "vhost-vdpa-%u", minor);
1059         if (r)
1060                 goto err;
1061
1062         cdev_init(&v->cdev, &vhost_vdpa_fops);
1063         v->cdev.owner = THIS_MODULE;
1064
1065         r = cdev_device_add(&v->cdev, &v->dev);
1066         if (r)
1067                 goto err;
1068
1069         init_completion(&v->completion);
1070         vdpa_set_drvdata(vdpa, v);
1071
1072         return 0;
1073
1074 err:
1075         put_device(&v->dev);
1076         return r;
1077 }
1078
1079 static void vhost_vdpa_remove(struct vdpa_device *vdpa)
1080 {
1081         struct vhost_vdpa *v = vdpa_get_drvdata(vdpa);
1082         int opened;
1083
1084         cdev_device_del(&v->cdev, &v->dev);
1085
1086         do {
1087                 opened = atomic_cmpxchg(&v->opened, 0, 1);
1088                 if (!opened)
1089                         break;
1090                 wait_for_completion(&v->completion);
1091         } while (1);
1092
1093         put_device(&v->dev);
1094 }
1095
1096 static struct vdpa_driver vhost_vdpa_driver = {
1097         .driver = {
1098                 .name   = "vhost_vdpa",
1099         },
1100         .probe  = vhost_vdpa_probe,
1101         .remove = vhost_vdpa_remove,
1102 };
1103
1104 static int __init vhost_vdpa_init(void)
1105 {
1106         int r;
1107
1108         r = alloc_chrdev_region(&vhost_vdpa_major, 0, VHOST_VDPA_DEV_MAX,
1109                                 "vhost-vdpa");
1110         if (r)
1111                 goto err_alloc_chrdev;
1112
1113         r = vdpa_register_driver(&vhost_vdpa_driver);
1114         if (r)
1115                 goto err_vdpa_register_driver;
1116
1117         return 0;
1118
1119 err_vdpa_register_driver:
1120         unregister_chrdev_region(vhost_vdpa_major, VHOST_VDPA_DEV_MAX);
1121 err_alloc_chrdev:
1122         return r;
1123 }
1124 module_init(vhost_vdpa_init);
1125
1126 static void __exit vhost_vdpa_exit(void)
1127 {
1128         vdpa_unregister_driver(&vhost_vdpa_driver);
1129         unregister_chrdev_region(vhost_vdpa_major, VHOST_VDPA_DEV_MAX);
1130 }
1131 module_exit(vhost_vdpa_exit);
1132
1133 MODULE_VERSION("0.0.1");
1134 MODULE_LICENSE("GPL v2");
1135 MODULE_AUTHOR("Intel Corporation");
1136 MODULE_DESCRIPTION("vDPA-based vhost backend for virtio");