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