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/slab.h>
  20 #include <linux/iommu.h>
  21 #include <linux/uuid.h>
  22 #include <linux/vdpa.h>
  23 #include <linux/nospec.h>
  24 #include <linux/vhost.h>
  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         struct vdpa_device *vdpa = v->vdpa;
 192         long size = vdpa->config->get_config_size(vdpa);
 193
 194         if (c->len == 0)
 195                 return -EINVAL;
 196
 197         if (c->len > size - c->off)
 198                 return -E2BIG;
 199
 200         return 0;
 201 }
 202
 203 static long vhost_vdpa_get_config(struct vhost_vdpa *v,
 204                                   struct vhost_vdpa_config __user *c)
 205 {
 206         struct vdpa_device *vdpa = v->vdpa;
 207         struct vhost_vdpa_config config;
 208         unsigned long size = offsetof(struct vhost_vdpa_config, buf);
 209         u8 *buf;
 210
 211         if (copy_from_user(&config, c, size))
 212                 return -EFAULT;
 213         if (vhost_vdpa_config_validate(v, &config))
 214                 return -EINVAL;
 215         buf = kvzalloc(config.len, GFP_KERNEL);
 216         if (!buf)
 217                 return -ENOMEM;
 218
 219         vdpa_get_config(vdpa, config.off, buf, config.len);
 220
 221         if (copy_to_user(c->buf, buf, config.len)) {
 222                 kvfree(buf);
 223                 return -EFAULT;
 224         }
 225
 226         kvfree(buf);
 227         return 0;
 228 }
 229
 230 static long vhost_vdpa_set_config(struct vhost_vdpa *v,
 231                                   struct vhost_vdpa_config __user *c)
 232 {
 233         struct vdpa_device *vdpa = v->vdpa;
 234         const struct vdpa_config_ops *ops = vdpa->config;
 235         struct vhost_vdpa_config config;
 236         unsigned long size = offsetof(struct vhost_vdpa_config, buf);
 237         u8 *buf;
 238
 239         if (copy_from_user(&config, c, size))
 240                 return -EFAULT;
 241         if (vhost_vdpa_config_validate(v, &config))
 242                 return -EINVAL;
 243
 244         buf = vmemdup_user(c->buf, config.len);
 245         if (IS_ERR(buf))
 246                 return PTR_ERR(buf);
 247
 248         ops->set_config(vdpa, config.off, buf, config.len);
 249
 250         kvfree(buf);
 251         return 0;
 252 }
 253
 254 static long vhost_vdpa_get_features(struct vhost_vdpa *v, u64 __user *featurep)
 255 {
 256         struct vdpa_device *vdpa = v->vdpa;
 257         const struct vdpa_config_ops *ops = vdpa->config;
 258         u64 features;
 259
 260         features = ops->get_features(vdpa);
 261
 262         if (copy_to_user(featurep, &features, sizeof(features)))
 263                 return -EFAULT;
 264
 265         return 0;
 266 }
 267
 268 static long vhost_vdpa_set_features(struct vhost_vdpa *v, u64 __user *featurep)
 269 {
 270         struct vdpa_device *vdpa = v->vdpa;
 271         const struct vdpa_config_ops *ops = vdpa->config;
 272         u64 features;
 273
 274         /*
 275          * It's not allowed to change the features after they have
 276          * been negotiated.
 277          */
 278         if (ops->get_status(vdpa) & VIRTIO_CONFIG_S_FEATURES_OK)
 279                 return -EBUSY;
 280
 281         if (copy_from_user(&features, featurep, sizeof(features)))
 282                 return -EFAULT;
 283
 284         if (vdpa_set_features(vdpa, features))
 285                 return -EINVAL;
 286
 287         return 0;
 288 }
 289
 290 static long vhost_vdpa_get_vring_num(struct vhost_vdpa *v, u16 __user *argp)
 291 {
 292         struct vdpa_device *vdpa = v->vdpa;
 293         const struct vdpa_config_ops *ops = vdpa->config;
 294         u16 num;
 295
 296         num = ops->get_vq_num_max(vdpa);
 297
 298         if (copy_to_user(argp, &num, sizeof(num)))
 299                 return -EFAULT;
 300
 301         return 0;
 302 }
 303
 304 static void vhost_vdpa_config_put(struct vhost_vdpa *v)
 305 {
 306         if (v->config_ctx) {
 307                 eventfd_ctx_put(v->config_ctx);
 308                 v->config_ctx = NULL;
 309         }
 310 }
 311
 312 static long vhost_vdpa_set_config_call(struct vhost_vdpa *v, u32 __user *argp)
 313 {
 314         struct vdpa_callback cb;
 315         int fd;
 316         struct eventfd_ctx *ctx;
 317
 318         cb.callback = vhost_vdpa_config_cb;
 319         cb.private = v->vdpa;
 320         if (copy_from_user(&fd, argp, sizeof(fd)))
 321                 return  -EFAULT;
 322
 323         ctx = fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(fd);
 324         swap(ctx, v->config_ctx);
 325
 326         if (!IS_ERR_OR_NULL(ctx))
 327                 eventfd_ctx_put(ctx);
 328
 329         if (IS_ERR(v->config_ctx)) {
 330                 long ret = PTR_ERR(v->config_ctx);
 331
 332                 v->config_ctx = NULL;
 333                 return ret;
 334         }
 335
 336         v->vdpa->config->set_config_cb(v->vdpa, &cb);
 337
 338         return 0;
 339 }
 340
 341 static long vhost_vdpa_get_iova_range(struct vhost_vdpa *v, u32 __user *argp)
 342 {
 343         struct vhost_vdpa_iova_range range = {
 344                 .first = v->range.first,
 345                 .last = v->range.last,
 346         };
 347
 348         if (copy_to_user(argp, &range, sizeof(range)))
 349                 return -EFAULT;
 350         return 0;
 351 }
 352
 353 static long vhost_vdpa_vring_ioctl(struct vhost_vdpa *v, unsigned int cmd,
 354                                    void __user *argp)
 355 {
 356         struct vdpa_device *vdpa = v->vdpa;
 357         const struct vdpa_config_ops *ops = vdpa->config;
 358         struct vdpa_vq_state vq_state;
 359         struct vdpa_callback cb;
 360         struct vhost_virtqueue *vq;
 361         struct vhost_vring_state s;
 362         u32 idx;
 363         long r;
 364
 365         r = get_user(idx, (u32 __user *)argp);
 366         if (r < 0)
 367                 return r;
 368
 369         if (idx >= v->nvqs)
 370                 return -ENOBUFS;
 371
 372         idx = array_index_nospec(idx, v->nvqs);
 373         vq = &v->vqs[idx];
 374
 375         switch (cmd) {
 376         case VHOST_VDPA_SET_VRING_ENABLE:
 377                 if (copy_from_user(&s, argp, sizeof(s)))
 378                         return -EFAULT;
 379                 ops->set_vq_ready(vdpa, idx, s.num);
 380                 return 0;
 381         case VHOST_GET_VRING_BASE:
 382                 r = ops->get_vq_state(v->vdpa, idx, &vq_state);
 383                 if (r)
 384                         return r;
 385
 386                 vq->last_avail_idx = vq_state.split.avail_index;
 387                 break;
 388         }
 389
 390         r = vhost_vring_ioctl(&v->vdev, cmd, argp);
 391         if (r)
 392                 return r;
 393
 394         switch (cmd) {
 395         case VHOST_SET_VRING_ADDR:
 396                 if (ops->set_vq_address(vdpa, idx,
 397                                         (u64)(uintptr_t)vq->desc,
 398                                         (u64)(uintptr_t)vq->avail,
 399                                         (u64)(uintptr_t)vq->used))
 400                         r = -EINVAL;
 401                 break;
 402
 403         case VHOST_SET_VRING_BASE:
 404                 vq_state.split.avail_index = vq->last_avail_idx;
 405                 if (ops->set_vq_state(vdpa, idx, &vq_state))
 406                         r = -EINVAL;
 407                 break;
 408
 409         case VHOST_SET_VRING_CALL:
 410                 if (vq->call_ctx.ctx) {
 411                         cb.callback = vhost_vdpa_virtqueue_cb;
 412                         cb.private = vq;
 413                 } else {
 414                         cb.callback = NULL;
 415                         cb.private = NULL;
 416                 }
 417                 ops->set_vq_cb(vdpa, idx, &cb);
 418                 vhost_vdpa_setup_vq_irq(v, idx);
 419                 break;
 420
 421         case VHOST_SET_VRING_NUM:
 422                 ops->set_vq_num(vdpa, idx, vq->num);
 423                 break;
 424         }
 425
 426         return r;
 427 }
 428
 429 static long vhost_vdpa_unlocked_ioctl(struct file *filep,
 430                                       unsigned int cmd, unsigned long arg)
 431 {
 432         struct vhost_vdpa *v = filep->private_data;
 433         struct vhost_dev *d = &v->vdev;
 434         void __user *argp = (void __user *)arg;
 435         u64 __user *featurep = argp;
 436         u64 features;
 437         long r = 0;
 438
 439         if (cmd == VHOST_SET_BACKEND_FEATURES) {
 440                 if (copy_from_user(&features, featurep, sizeof(features)))
 441                         return -EFAULT;
 442                 if (features & ~VHOST_VDPA_BACKEND_FEATURES)
 443                         return -EOPNOTSUPP;
 444                 vhost_set_backend_features(&v->vdev, features);
 445                 return 0;
 446         }
 447
 448         mutex_lock(&d->mutex);
 449
 450         switch (cmd) {
 451         case VHOST_VDPA_GET_DEVICE_ID:
 452                 r = vhost_vdpa_get_device_id(v, argp);
 453                 break;
 454         case VHOST_VDPA_GET_STATUS:
 455                 r = vhost_vdpa_get_status(v, argp);
 456                 break;
 457         case VHOST_VDPA_SET_STATUS:
 458                 r = vhost_vdpa_set_status(v, argp);
 459                 break;
 460         case VHOST_VDPA_GET_CONFIG:
 461                 r = vhost_vdpa_get_config(v, argp);
 462                 break;
 463         case VHOST_VDPA_SET_CONFIG:
 464                 r = vhost_vdpa_set_config(v, argp);
 465                 break;
 466         case VHOST_GET_FEATURES:
 467                 r = vhost_vdpa_get_features(v, argp);
 468                 break;
 469         case VHOST_SET_FEATURES:
 470                 r = vhost_vdpa_set_features(v, argp);
 471                 break;
 472         case VHOST_VDPA_GET_VRING_NUM:
 473                 r = vhost_vdpa_get_vring_num(v, argp);
 474                 break;
 475         case VHOST_SET_LOG_BASE:
 476         case VHOST_SET_LOG_FD:
 477                 r = -ENOIOCTLCMD;
 478                 break;
 479         case VHOST_VDPA_SET_CONFIG_CALL:
 480                 r = vhost_vdpa_set_config_call(v, argp);
 481                 break;
 482         case VHOST_GET_BACKEND_FEATURES:
 483                 features = VHOST_VDPA_BACKEND_FEATURES;
 484                 if (copy_to_user(featurep, &features, sizeof(features)))
 485                         r = -EFAULT;
 486                 break;
 487         case VHOST_VDPA_GET_IOVA_RANGE:
 488                 r = vhost_vdpa_get_iova_range(v, argp);
 489                 break;
 490         default:
 491                 r = vhost_dev_ioctl(&v->vdev, cmd, argp);
 492                 if (r == -ENOIOCTLCMD)
 493                         r = vhost_vdpa_vring_ioctl(v, cmd, argp);
 494                 break;
 495         }
 496
 497         mutex_unlock(&d->mutex);
 498         return r;
 499 }
 500
 501 static void vhost_vdpa_iotlb_unmap(struct vhost_vdpa *v, u64 start, u64 last)
 502 {
 503         struct vhost_dev *dev = &v->vdev;
 504         struct vhost_iotlb *iotlb = dev->iotlb;
 505         struct vhost_iotlb_map *map;
 506         struct page *page;
 507         unsigned long pfn, pinned;
 508
 509         while ((map = vhost_iotlb_itree_first(iotlb, start, last)) != NULL) {
 510                 pinned = map->size >> PAGE_SHIFT;
 511                 for (pfn = map->addr >> PAGE_SHIFT;
 512                      pinned > 0; pfn++, pinned--) {
 513                         page = pfn_to_page(pfn);
 514                         if (map->perm & VHOST_ACCESS_WO)
 515                                 set_page_dirty_lock(page);
 516                         unpin_user_page(page);
 517                 }
 518                 atomic64_sub(map->size >> PAGE_SHIFT, &dev->mm->pinned_vm);
 519                 vhost_iotlb_map_free(iotlb, map);
 520         }
 521 }
 522
 523 static void vhost_vdpa_iotlb_free(struct vhost_vdpa *v)
 524 {
 525         struct vhost_dev *dev = &v->vdev;
 526
 527         vhost_vdpa_iotlb_unmap(v, 0ULL, 0ULL - 1);
 528         kfree(dev->iotlb);
 529         dev->iotlb = NULL;
 530 }
 531
 532 static int perm_to_iommu_flags(u32 perm)
 533 {
 534         int flags = 0;
 535
 536         switch (perm) {
 537         case VHOST_ACCESS_WO:
 538                 flags |= IOMMU_WRITE;
 539                 break;
 540         case VHOST_ACCESS_RO:
 541                 flags |= IOMMU_READ;
 542                 break;
 543         case VHOST_ACCESS_RW:
 544                 flags |= (IOMMU_WRITE | IOMMU_READ);
 545                 break;
 546         default:
 547                 WARN(1, "invalidate vhost IOTLB permission\n");
 548                 break;
 549         }
 550
 551         return flags | IOMMU_CACHE;
 552 }
 553
 554 static int vhost_vdpa_map(struct vhost_vdpa *v,
 555                           u64 iova, u64 size, u64 pa, u32 perm)
 556 {
 557         struct vhost_dev *dev = &v->vdev;
 558         struct vdpa_device *vdpa = v->vdpa;
 559         const struct vdpa_config_ops *ops = vdpa->config;
 560         int r = 0;
 561
 562         r = vhost_iotlb_add_range(dev->iotlb, iova, iova + size - 1,
 563                                   pa, perm);
 564         if (r)
 565                 return r;
 566
 567         if (ops->dma_map) {
 568                 r = ops->dma_map(vdpa, iova, size, pa, perm);
 569         } else if (ops->set_map) {
 570                 if (!v->in_batch)
 571                         r = ops->set_map(vdpa, dev->iotlb);
 572         } else {
 573                 r = iommu_map(v->domain, iova, pa, size,
 574                               perm_to_iommu_flags(perm));
 575         }
 576
 577         if (r)
 578                 vhost_iotlb_del_range(dev->iotlb, iova, iova + size - 1);
 579         else
 580                 atomic64_add(size >> PAGE_SHIFT, &dev->mm->pinned_vm);
 581
 582         return r;
 583 }
 584
 585 static void vhost_vdpa_unmap(struct vhost_vdpa *v, u64 iova, u64 size)
 586 {
 587         struct vhost_dev *dev = &v->vdev;
 588         struct vdpa_device *vdpa = v->vdpa;
 589         const struct vdpa_config_ops *ops = vdpa->config;
 590
 591         vhost_vdpa_iotlb_unmap(v, iova, iova + size - 1);
 592
 593         if (ops->dma_map) {
 594                 ops->dma_unmap(vdpa, iova, size);
 595         } else if (ops->set_map) {
 596                 if (!v->in_batch)
 597                         ops->set_map(vdpa, dev->iotlb);
 598         } else {
 599                 iommu_unmap(v->domain, iova, size);
 600         }
 601 }
 602
 603 static int vhost_vdpa_process_iotlb_update(struct vhost_vdpa *v,
 604                                            struct vhost_iotlb_msg *msg)
 605 {
 606         struct vhost_dev *dev = &v->vdev;
 607         struct vhost_iotlb *iotlb = dev->iotlb;
 608         struct page **page_list;
 609         unsigned long list_size = PAGE_SIZE / sizeof(struct page *);
 610         unsigned int gup_flags = FOLL_LONGTERM;
 611         unsigned long npages, cur_base, map_pfn, last_pfn = 0;
 612         unsigned long lock_limit, sz2pin, nchunks, i;
 613         u64 iova = msg->iova;
 614         long pinned;
 615         int ret = 0;
 616
 617         if (msg->iova < v->range.first || !msg->size ||
 618             msg->iova > U64_MAX - msg->size + 1 ||
 619             msg->iova + msg->size - 1 > v->range.last)
 620                 return -EINVAL;
 621
 622         if (vhost_iotlb_itree_first(iotlb, msg->iova,
 623                                     msg->iova + msg->size - 1))
 624                 return -EEXIST;
 625
 626         /* Limit the use of memory for bookkeeping */
 627         page_list = (struct page **) __get_free_page(GFP_KERNEL);
 628         if (!page_list)
 629                 return -ENOMEM;
 630
 631         if (msg->perm & VHOST_ACCESS_WO)
 632                 gup_flags |= FOLL_WRITE;
 633
 634         npages = PAGE_ALIGN(msg->size + (iova & ~PAGE_MASK)) >> PAGE_SHIFT;
 635         if (!npages) {
 636                 ret = -EINVAL;
 637                 goto free;
 638         }
 639
 640         mmap_read_lock(dev->mm);
 641
 642         lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
 643         if (npages + atomic64_read(&dev->mm->pinned_vm) > lock_limit) {
 644                 ret = -ENOMEM;
 645                 goto unlock;
 646         }
 647
 648         cur_base = msg->uaddr & PAGE_MASK;
 649         iova &= PAGE_MASK;
 650         nchunks = 0;
 651
 652         while (npages) {
 653                 sz2pin = min_t(unsigned long, npages, list_size);
 654                 pinned = pin_user_pages(cur_base, sz2pin,
 655                                         gup_flags, page_list, NULL);
 656                 if (sz2pin != pinned) {
 657                         if (pinned < 0) {
 658                                 ret = pinned;
 659                         } else {
 660                                 unpin_user_pages(page_list, pinned);
 661                                 ret = -ENOMEM;
 662                         }
 663                         goto out;
 664                 }
 665                 nchunks++;
 666
 667                 if (!last_pfn)
 668                         map_pfn = page_to_pfn(page_list[0]);
 669
 670                 for (i = 0; i < pinned; i++) {
 671                         unsigned long this_pfn = page_to_pfn(page_list[i]);
 672                         u64 csize;
 673
 674                         if (last_pfn && (this_pfn != last_pfn + 1)) {
 675                                 /* Pin a contiguous chunk of memory */
 676                                 csize = (last_pfn - map_pfn + 1) << PAGE_SHIFT;
 677                                 ret = vhost_vdpa_map(v, iova, csize,
 678                                                      map_pfn << PAGE_SHIFT,
 679                                                      msg->perm);
 680                                 if (ret) {
 681                                         /*
 682                                          * Unpin the pages that are left unmapped
 683                                          * from this point on in the current
 684                                          * page_list. The remaining outstanding
 685                                          * ones which may stride across several
 686                                          * chunks will be covered in the common
 687                                          * error path subsequently.
 688                                          */
 689                                         unpin_user_pages(&page_list[i],
 690                                                          pinned - i);
 691                                         goto out;
 692                                 }
 693
 694                                 map_pfn = this_pfn;
 695                                 iova += csize;
 696                                 nchunks = 0;
 697                         }
 698
 699                         last_pfn = this_pfn;
 700                 }
 701
 702                 cur_base += pinned << PAGE_SHIFT;
 703                 npages -= pinned;
 704         }
 705
 706         /* Pin the rest chunk */
 707         ret = vhost_vdpa_map(v, iova, (last_pfn - map_pfn + 1) << PAGE_SHIFT,
 708                              map_pfn << PAGE_SHIFT, msg->perm);
 709 out:
 710         if (ret) {
 711                 if (nchunks) {
 712                         unsigned long pfn;
 713
 714                         /*
 715                          * Unpin the outstanding pages which are yet to be
 716                          * mapped but haven't due to vdpa_map() or
 717                          * pin_user_pages() failure.
 718                          *
 719                          * Mapped pages are accounted in vdpa_map(), hence
 720                          * the corresponding unpinning will be handled by
 721                          * vdpa_unmap().
 722                          */
 723                         WARN_ON(!last_pfn);
 724                         for (pfn = map_pfn; pfn <= last_pfn; pfn++)
 725                                 unpin_user_page(pfn_to_page(pfn));
 726                 }
 727                 vhost_vdpa_unmap(v, msg->iova, msg->size);
 728         }
 729 unlock:
 730         mmap_read_unlock(dev->mm);
 731 free:
 732         free_page((unsigned long)page_list);
 733         return ret;
 734 }
 735
 736 static int vhost_vdpa_process_iotlb_msg(struct vhost_dev *dev,
 737                                         struct vhost_iotlb_msg *msg)
 738 {
 739         struct vhost_vdpa *v = container_of(dev, struct vhost_vdpa, vdev);
 740         struct vdpa_device *vdpa = v->vdpa;
 741         const struct vdpa_config_ops *ops = vdpa->config;
 742         int r = 0;
 743
 744         mutex_lock(&dev->mutex);
 745
 746         r = vhost_dev_check_owner(dev);
 747         if (r)
 748                 goto unlock;
 749
 750         switch (msg->type) {
 751         case VHOST_IOTLB_UPDATE:
 752                 r = vhost_vdpa_process_iotlb_update(v, msg);
 753                 break;
 754         case VHOST_IOTLB_INVALIDATE:
 755                 vhost_vdpa_unmap(v, msg->iova, msg->size);
 756                 break;
 757         case VHOST_IOTLB_BATCH_BEGIN:
 758                 v->in_batch = true;
 759                 break;
 760         case VHOST_IOTLB_BATCH_END:
 761                 if (v->in_batch && ops->set_map)
 762                         ops->set_map(vdpa, dev->iotlb);
 763                 v->in_batch = false;
 764                 break;
 765         default:
 766                 r = -EINVAL;
 767                 break;
 768         }
 769 unlock:
 770         mutex_unlock(&dev->mutex);
 771
 772         return r;
 773 }
 774
 775 static ssize_t vhost_vdpa_chr_write_iter(struct kiocb *iocb,
 776                                          struct iov_iter *from)
 777 {
 778         struct file *file = iocb->ki_filp;
 779         struct vhost_vdpa *v = file->private_data;
 780         struct vhost_dev *dev = &v->vdev;
 781
 782         return vhost_chr_write_iter(dev, from);
 783 }
 784
 785 static int vhost_vdpa_alloc_domain(struct vhost_vdpa *v)
 786 {
 787         struct vdpa_device *vdpa = v->vdpa;
 788         const struct vdpa_config_ops *ops = vdpa->config;
 789         struct device *dma_dev = vdpa_get_dma_dev(vdpa);
 790         struct bus_type *bus;
 791         int ret;
 792
 793         /* Device want to do DMA by itself */
 794         if (ops->set_map || ops->dma_map)
 795                 return 0;
 796
 797         bus = dma_dev->bus;
 798         if (!bus)
 799                 return -EFAULT;
 800
 801         if (!iommu_capable(bus, IOMMU_CAP_CACHE_COHERENCY))
 802                 return -ENOTSUPP;
 803
 804         v->domain = iommu_domain_alloc(bus);
 805         if (!v->domain)
 806                 return -EIO;
 807
 808         ret = iommu_attach_device(v->domain, dma_dev);
 809         if (ret)
 810                 goto err_attach;
 811
 812         return 0;
 813
 814 err_attach:
 815         iommu_domain_free(v->domain);
 816         return ret;
 817 }
 818
 819 static void vhost_vdpa_free_domain(struct vhost_vdpa *v)
 820 {
 821         struct vdpa_device *vdpa = v->vdpa;
 822         struct device *dma_dev = vdpa_get_dma_dev(vdpa);
 823
 824         if (v->domain) {
 825                 iommu_detach_device(v->domain, dma_dev);
 826                 iommu_domain_free(v->domain);
 827         }
 828
 829         v->domain = NULL;
 830 }
 831
 832 static void vhost_vdpa_set_iova_range(struct vhost_vdpa *v)
 833 {
 834         struct vdpa_iova_range *range = &v->range;
 835         struct vdpa_device *vdpa = v->vdpa;
 836         const struct vdpa_config_ops *ops = vdpa->config;
 837
 838         if (ops->get_iova_range) {
 839                 *range = ops->get_iova_range(vdpa);
 840         } else if (v->domain && v->domain->geometry.force_aperture) {
 841                 range->first = v->domain->geometry.aperture_start;
 842                 range->last = v->domain->geometry.aperture_end;
 843         } else {
 844                 range->first = 0;
 845                 range->last = ULLONG_MAX;
 846         }
 847 }
 848
 849 static int vhost_vdpa_open(struct inode *inode, struct file *filep)
 850 {
 851         struct vhost_vdpa *v;
 852         struct vhost_dev *dev;
 853         struct vhost_virtqueue **vqs;
 854         int nvqs, i, r, opened;
 855
 856         v = container_of(inode->i_cdev, struct vhost_vdpa, cdev);
 857
 858         opened = atomic_cmpxchg(&v->opened, 0, 1);
 859         if (opened)
 860                 return -EBUSY;
 861
 862         nvqs = v->nvqs;
 863         vhost_vdpa_reset(v);
 864
 865         vqs = kmalloc_array(nvqs, sizeof(*vqs), GFP_KERNEL);
 866         if (!vqs) {
 867                 r = -ENOMEM;
 868                 goto err;
 869         }
 870
 871         dev = &v->vdev;
 872         for (i = 0; i < nvqs; i++) {
 873                 vqs[i] = &v->vqs[i];
 874                 vqs[i]->handle_kick = handle_vq_kick;
 875         }
 876         vhost_dev_init(dev, vqs, nvqs, 0, 0, 0, false,
 877                        vhost_vdpa_process_iotlb_msg);
 878
 879         dev->iotlb = vhost_iotlb_alloc(0, 0);
 880         if (!dev->iotlb) {
 881                 r = -ENOMEM;
 882                 goto err_init_iotlb;
 883         }
 884
 885         r = vhost_vdpa_alloc_domain(v);
 886         if (r)
 887                 goto err_init_iotlb;
 888
 889         vhost_vdpa_set_iova_range(v);
 890
 891         filep->private_data = v;
 892
 893         return 0;
 894
 895 err_init_iotlb:
 896         vhost_dev_cleanup(&v->vdev);
 897         kfree(vqs);
 898 err:
 899         atomic_dec(&v->opened);
 900         return r;
 901 }
 902
 903 static void vhost_vdpa_clean_irq(struct vhost_vdpa *v)
 904 {
 905         int i;
 906
 907         for (i = 0; i < v->nvqs; i++)
 908                 vhost_vdpa_unsetup_vq_irq(v, i);
 909 }
 910
 911 static int vhost_vdpa_release(struct inode *inode, struct file *filep)
 912 {
 913         struct vhost_vdpa *v = filep->private_data;
 914         struct vhost_dev *d = &v->vdev;
 915
 916         mutex_lock(&d->mutex);
 917         filep->private_data = NULL;
 918         vhost_vdpa_reset(v);
 919         vhost_dev_stop(&v->vdev);
 920         vhost_vdpa_iotlb_free(v);
 921         vhost_vdpa_free_domain(v);
 922         vhost_vdpa_config_put(v);
 923         vhost_vdpa_clean_irq(v);
 924         vhost_dev_cleanup(&v->vdev);
 925         kfree(v->vdev.vqs);
 926         mutex_unlock(&d->mutex);
 927
 928         atomic_dec(&v->opened);
 929         complete(&v->completion);
 930
 931         return 0;
 932 }
 933
 934 #ifdef CONFIG_MMU
 935 static vm_fault_t vhost_vdpa_fault(struct vm_fault *vmf)
 936 {
 937         struct vhost_vdpa *v = vmf->vma->vm_file->private_data;
 938         struct vdpa_device *vdpa = v->vdpa;
 939         const struct vdpa_config_ops *ops = vdpa->config;
 940         struct vdpa_notification_area notify;
 941         struct vm_area_struct *vma = vmf->vma;
 942         u16 index = vma->vm_pgoff;
 943
 944         notify = ops->get_vq_notification(vdpa, index);
 945
 946         vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
 947         if (remap_pfn_range(vma, vmf->address & PAGE_MASK,
 948                             notify.addr >> PAGE_SHIFT, PAGE_SIZE,
 949                             vma->vm_page_prot))
 950                 return VM_FAULT_SIGBUS;
 951
 952         return VM_FAULT_NOPAGE;
 953 }
 954
 955 static const struct vm_operations_struct vhost_vdpa_vm_ops = {
 956         .fault = vhost_vdpa_fault,
 957 };
 958
 959 static int vhost_vdpa_mmap(struct file *file, struct vm_area_struct *vma)
 960 {
 961         struct vhost_vdpa *v = vma->vm_file->private_data;
 962         struct vdpa_device *vdpa = v->vdpa;
 963         const struct vdpa_config_ops *ops = vdpa->config;
 964         struct vdpa_notification_area notify;
 965         unsigned long index = vma->vm_pgoff;
 966
 967         if (vma->vm_end - vma->vm_start != PAGE_SIZE)
 968                 return -EINVAL;
 969         if ((vma->vm_flags & VM_SHARED) == 0)
 970                 return -EINVAL;
 971         if (vma->vm_flags & VM_READ)
 972                 return -EINVAL;
 973         if (index > 65535)
 974                 return -EINVAL;
 975         if (!ops->get_vq_notification)
 976                 return -ENOTSUPP;
 977
 978         /* To be safe and easily modelled by userspace, We only
 979          * support the doorbell which sits on the page boundary and
 980          * does not share the page with other registers.
 981          */
 982         notify = ops->get_vq_notification(vdpa, index);
 983         if (notify.addr & (PAGE_SIZE - 1))
 984                 return -EINVAL;
 985         if (vma->vm_end - vma->vm_start != notify.size)
 986                 return -ENOTSUPP;
 987
 988         vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
 989         vma->vm_ops = &vhost_vdpa_vm_ops;
 990         return 0;
 991 }
 992 #endif /* CONFIG_MMU */
 993
 994 static const struct file_operations vhost_vdpa_fops = {
 995         .owner          = THIS_MODULE,
 996         .open           = vhost_vdpa_open,
 997         .release        = vhost_vdpa_release,
 998         .write_iter     = vhost_vdpa_chr_write_iter,
 999         .unlocked_ioctl = vhost_vdpa_unlocked_ioctl,
1000 #ifdef CONFIG_MMU
1001         .mmap           = vhost_vdpa_mmap,
1002 #endif /* CONFIG_MMU */
1003         .compat_ioctl   = compat_ptr_ioctl,
1004 };
1005
1006 static void vhost_vdpa_release_dev(struct device *device)
1007 {
1008         struct vhost_vdpa *v =
1009                container_of(device, struct vhost_vdpa, dev);
1010
1011         ida_simple_remove(&vhost_vdpa_ida, v->minor);
1012         kfree(v->vqs);
1013         kfree(v);
1014 }
1015
1016 static int vhost_vdpa_probe(struct vdpa_device *vdpa)
1017 {
1018         const struct vdpa_config_ops *ops = vdpa->config;
1019         struct vhost_vdpa *v;
1020         int minor;
1021         int r;
1022
1023         v = kzalloc(sizeof(*v), GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1024         if (!v)
1025                 return -ENOMEM;
1026
1027         minor = ida_simple_get(&vhost_vdpa_ida, 0,
1028                                VHOST_VDPA_DEV_MAX, GFP_KERNEL);
1029         if (minor < 0) {
1030                 kfree(v);
1031                 return minor;
1032         }
1033
1034         atomic_set(&v->opened, 0);
1035         v->minor = minor;
1036         v->vdpa = vdpa;
1037         v->nvqs = vdpa->nvqs;
1038         v->virtio_id = ops->get_device_id(vdpa);
1039
1040         device_initialize(&v->dev);
1041         v->dev.release = vhost_vdpa_release_dev;
1042         v->dev.parent = &vdpa->dev;
1043         v->dev.devt = MKDEV(MAJOR(vhost_vdpa_major), minor);
1044         v->vqs = kmalloc_array(v->nvqs, sizeof(struct vhost_virtqueue),
1045                                GFP_KERNEL);
1046         if (!v->vqs) {
1047                 r = -ENOMEM;
1048                 goto err;
1049         }
1050
1051         r = dev_set_name(&v->dev, "vhost-vdpa-%u", minor);
1052         if (r)
1053                 goto err;
1054
1055         cdev_init(&v->cdev, &vhost_vdpa_fops);
1056         v->cdev.owner = THIS_MODULE;
1057
1058         r = cdev_device_add(&v->cdev, &v->dev);
1059         if (r)
1060                 goto err;
1061
1062         init_completion(&v->completion);
1063         vdpa_set_drvdata(vdpa, v);
1064
1065         return 0;
1066
1067 err:
1068         put_device(&v->dev);
1069         return r;
1070 }
1071
1072 static void vhost_vdpa_remove(struct vdpa_device *vdpa)
1073 {
1074         struct vhost_vdpa *v = vdpa_get_drvdata(vdpa);
1075         int opened;
1076
1077         cdev_device_del(&v->cdev, &v->dev);
1078
1079         do {
1080                 opened = atomic_cmpxchg(&v->opened, 0, 1);
1081                 if (!opened)
1082                         break;
1083                 wait_for_completion(&v->completion);
1084         } while (1);
1085
1086         put_device(&v->dev);
1087 }
1088
1089 static struct vdpa_driver vhost_vdpa_driver = {
1090         .driver = {
1091                 .name   = "vhost_vdpa",
1092         },
1093         .probe  = vhost_vdpa_probe,
1094         .remove = vhost_vdpa_remove,
1095 };
1096
1097 static int __init vhost_vdpa_init(void)
1098 {
1099         int r;
1100
1101         r = alloc_chrdev_region(&vhost_vdpa_major, 0, VHOST_VDPA_DEV_MAX,
1102                                 "vhost-vdpa");
1103         if (r)
1104                 goto err_alloc_chrdev;
1105
1106         r = vdpa_register_driver(&vhost_vdpa_driver);
1107         if (r)
1108                 goto err_vdpa_register_driver;
1109
1110         return 0;
1111
1112 err_vdpa_register_driver:
1113         unregister_chrdev_region(vhost_vdpa_major, VHOST_VDPA_DEV_MAX);
1114 err_alloc_chrdev:
1115         return r;
1116 }
1117 module_init(vhost_vdpa_init);
1118
1119 static void __exit vhost_vdpa_exit(void)
1120 {
1121         vdpa_unregister_driver(&vhost_vdpa_driver);
1122         unregister_chrdev_region(vhost_vdpa_major, VHOST_VDPA_DEV_MAX);
1123 }
1124 module_exit(vhost_vdpa_exit);
1125
1126 MODULE_VERSION("0.0.1");
1127 MODULE_LICENSE("GPL v2");
1128 MODULE_AUTHOR("Intel Corporation");
1129 MODULE_DESCRIPTION("vDPA-based vhost backend for virtio");