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         if (irq < 0)
 101                 return;
 102
 103         irq_bypass_unregister_producer(&vq->call_ctx.producer);
 104         if (!vq->call_ctx.ctx)
 105                 return;
 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         if (unlikely(ret))
 111                 dev_info(&v->dev, "vq %u, irq bypass producer (token %p) registration fails, ret =  %d\n",
 112                          qid, vq->call_ctx.producer.token, ret);
 113 }
 114
 115 static void vhost_vdpa_unsetup_vq_irq(struct vhost_vdpa *v, u16 qid)
 116 {
 117         struct vhost_virtqueue *vq = &v->vqs[qid];
 118
 119         irq_bypass_unregister_producer(&vq->call_ctx.producer);
 120 }
 121
 122 static int vhost_vdpa_reset(struct vhost_vdpa *v)
 123 {
 124         struct vdpa_device *vdpa = v->vdpa;
 125
 126         v->in_batch = 0;
 127
 128         return vdpa_reset(vdpa);
 129 }
 130
 131 static long vhost_vdpa_get_device_id(struct vhost_vdpa *v, u8 __user *argp)
 132 {
 133         struct vdpa_device *vdpa = v->vdpa;
 134         const struct vdpa_config_ops *ops = vdpa->config;
 135         u32 device_id;
 136
 137         device_id = ops->get_device_id(vdpa);
 138
 139         if (copy_to_user(argp, &device_id, sizeof(device_id)))
 140                 return -EFAULT;
 141
 142         return 0;
 143 }
 144
 145 static long vhost_vdpa_get_status(struct vhost_vdpa *v, u8 __user *statusp)
 146 {
 147         struct vdpa_device *vdpa = v->vdpa;
 148         const struct vdpa_config_ops *ops = vdpa->config;
 149         u8 status;
 150
 151         status = ops->get_status(vdpa);
 152
 153         if (copy_to_user(statusp, &status, sizeof(status)))
 154                 return -EFAULT;
 155
 156         return 0;
 157 }
 158
 159 static long vhost_vdpa_set_status(struct vhost_vdpa *v, u8 __user *statusp)
 160 {
 161         struct vdpa_device *vdpa = v->vdpa;
 162         const struct vdpa_config_ops *ops = vdpa->config;
 163         u8 status, status_old;
 164         int ret, nvqs = v->nvqs;
 165         u16 i;
 166
 167         if (copy_from_user(&status, statusp, sizeof(status)))
 168                 return -EFAULT;
 169
 170         status_old = ops->get_status(vdpa);
 171
 172         /*
 173          * Userspace shouldn't remove status bits unless reset the
 174          * status to 0.
 175          */
 176         if (status != 0 && (ops->get_status(vdpa) & ~status) != 0)
 177                 return -EINVAL;
 178
 179         if ((status_old & VIRTIO_CONFIG_S_DRIVER_OK) && !(status & VIRTIO_CONFIG_S_DRIVER_OK))
 180                 for (i = 0; i < nvqs; i++)
 181                         vhost_vdpa_unsetup_vq_irq(v, i);
 182
 183         if (status == 0) {
 184                 ret = ops->reset(vdpa);
 185                 if (ret)
 186                         return ret;
 187         } else
 188                 ops->set_status(vdpa, status);
 189
 190         if ((status & VIRTIO_CONFIG_S_DRIVER_OK) && !(status_old & VIRTIO_CONFIG_S_DRIVER_OK))
 191                 for (i = 0; i < nvqs; i++)
 192                         vhost_vdpa_setup_vq_irq(v, i);
 193
 194         return 0;
 195 }
 196
 197 static int vhost_vdpa_config_validate(struct vhost_vdpa *v,
 198                                       struct vhost_vdpa_config *c)
 199 {
 200         struct vdpa_device *vdpa = v->vdpa;
 201         long size = vdpa->config->get_config_size(vdpa);
 202
 203         if (c->len == 0 || c->off > size)
 204                 return -EINVAL;
 205
 206         if (c->len > size - c->off)
 207                 return -E2BIG;
 208
 209         return 0;
 210 }
 211
 212 static long vhost_vdpa_get_config(struct vhost_vdpa *v,
 213                                   struct vhost_vdpa_config __user *c)
 214 {
 215         struct vdpa_device *vdpa = v->vdpa;
 216         struct vhost_vdpa_config config;
 217         unsigned long size = offsetof(struct vhost_vdpa_config, buf);
 218         u8 *buf;
 219
 220         if (copy_from_user(&config, c, size))
 221                 return -EFAULT;
 222         if (vhost_vdpa_config_validate(v, &config))
 223                 return -EINVAL;
 224         buf = kvzalloc(config.len, GFP_KERNEL);
 225         if (!buf)
 226                 return -ENOMEM;
 227
 228         vdpa_get_config(vdpa, config.off, buf, config.len);
 229
 230         if (copy_to_user(c->buf, buf, config.len)) {
 231                 kvfree(buf);
 232                 return -EFAULT;
 233         }
 234
 235         kvfree(buf);
 236         return 0;
 237 }
 238
 239 static long vhost_vdpa_set_config(struct vhost_vdpa *v,
 240                                   struct vhost_vdpa_config __user *c)
 241 {
 242         struct vdpa_device *vdpa = v->vdpa;
 243         const struct vdpa_config_ops *ops = vdpa->config;
 244         struct vhost_vdpa_config config;
 245         unsigned long size = offsetof(struct vhost_vdpa_config, buf);
 246         u8 *buf;
 247
 248         if (copy_from_user(&config, c, size))
 249                 return -EFAULT;
 250         if (vhost_vdpa_config_validate(v, &config))
 251                 return -EINVAL;
 252
 253         buf = vmemdup_user(c->buf, config.len);
 254         if (IS_ERR(buf))
 255                 return PTR_ERR(buf);
 256
 257         ops->set_config(vdpa, config.off, buf, config.len);
 258
 259         kvfree(buf);
 260         return 0;
 261 }
 262
 263 static long vhost_vdpa_get_features(struct vhost_vdpa *v, u64 __user *featurep)
 264 {
 265         struct vdpa_device *vdpa = v->vdpa;
 266         const struct vdpa_config_ops *ops = vdpa->config;
 267         u64 features;
 268
 269         features = ops->get_features(vdpa);
 270
 271         if (copy_to_user(featurep, &features, sizeof(features)))
 272                 return -EFAULT;
 273
 274         return 0;
 275 }
 276
 277 static long vhost_vdpa_set_features(struct vhost_vdpa *v, u64 __user *featurep)
 278 {
 279         struct vdpa_device *vdpa = v->vdpa;
 280         const struct vdpa_config_ops *ops = vdpa->config;
 281         u64 features;
 282
 283         /*
 284          * It's not allowed to change the features after they have
 285          * been negotiated.
 286          */
 287         if (ops->get_status(vdpa) & VIRTIO_CONFIG_S_FEATURES_OK)
 288                 return -EBUSY;
 289
 290         if (copy_from_user(&features, featurep, sizeof(features)))
 291                 return -EFAULT;
 292
 293         if (vdpa_set_features(vdpa, features))
 294                 return -EINVAL;
 295
 296         return 0;
 297 }
 298
 299 static long vhost_vdpa_get_vring_num(struct vhost_vdpa *v, u16 __user *argp)
 300 {
 301         struct vdpa_device *vdpa = v->vdpa;
 302         const struct vdpa_config_ops *ops = vdpa->config;
 303         u16 num;
 304
 305         num = ops->get_vq_num_max(vdpa);
 306
 307         if (copy_to_user(argp, &num, sizeof(num)))
 308                 return -EFAULT;
 309
 310         return 0;
 311 }
 312
 313 static void vhost_vdpa_config_put(struct vhost_vdpa *v)
 314 {
 315         if (v->config_ctx) {
 316                 eventfd_ctx_put(v->config_ctx);
 317                 v->config_ctx = NULL;
 318         }
 319 }
 320
 321 static long vhost_vdpa_set_config_call(struct vhost_vdpa *v, u32 __user *argp)
 322 {
 323         struct vdpa_callback cb;
 324         int fd;
 325         struct eventfd_ctx *ctx;
 326
 327         cb.callback = vhost_vdpa_config_cb;
 328         cb.private = v;
 329         if (copy_from_user(&fd, argp, sizeof(fd)))
 330                 return  -EFAULT;
 331
 332         ctx = fd == VHOST_FILE_UNBIND ? NULL : eventfd_ctx_fdget(fd);
 333         swap(ctx, v->config_ctx);
 334
 335         if (!IS_ERR_OR_NULL(ctx))
 336                 eventfd_ctx_put(ctx);
 337
 338         if (IS_ERR(v->config_ctx)) {
 339                 long ret = PTR_ERR(v->config_ctx);
 340
 341                 v->config_ctx = NULL;
 342                 return ret;
 343         }
 344
 345         v->vdpa->config->set_config_cb(v->vdpa, &cb);
 346
 347         return 0;
 348 }
 349
 350 static long vhost_vdpa_get_iova_range(struct vhost_vdpa *v, u32 __user *argp)
 351 {
 352         struct vhost_vdpa_iova_range range = {
 353                 .first = v->range.first,
 354                 .last = v->range.last,
 355         };
 356
 357         if (copy_to_user(argp, &range, sizeof(range)))
 358                 return -EFAULT;
 359         return 0;
 360 }
 361
 362 static long vhost_vdpa_vring_ioctl(struct vhost_vdpa *v, unsigned int cmd,
 363                                    void __user *argp)
 364 {
 365         struct vdpa_device *vdpa = v->vdpa;
 366         const struct vdpa_config_ops *ops = vdpa->config;
 367         struct vdpa_vq_state vq_state;
 368         struct vdpa_callback cb;
 369         struct vhost_virtqueue *vq;
 370         struct vhost_vring_state s;
 371         u32 idx;
 372         long r;
 373
 374         r = get_user(idx, (u32 __user *)argp);
 375         if (r < 0)
 376                 return r;
 377
 378         if (idx >= v->nvqs)
 379                 return -ENOBUFS;
 380
 381         idx = array_index_nospec(idx, v->nvqs);
 382         vq = &v->vqs[idx];
 383
 384         switch (cmd) {
 385         case VHOST_VDPA_SET_VRING_ENABLE:
 386                 if (copy_from_user(&s, argp, sizeof(s)))
 387                         return -EFAULT;
 388                 ops->set_vq_ready(vdpa, idx, s.num);
 389                 return 0;
 390         case VHOST_GET_VRING_BASE:
 391                 r = ops->get_vq_state(v->vdpa, idx, &vq_state);
 392                 if (r)
 393                         return r;
 394
 395                 vq->last_avail_idx = vq_state.split.avail_index;
 396                 break;
 397         }
 398
 399         r = vhost_vring_ioctl(&v->vdev, cmd, argp);
 400         if (r)
 401                 return r;
 402
 403         switch (cmd) {
 404         case VHOST_SET_VRING_ADDR:
 405                 if (ops->set_vq_address(vdpa, idx,
 406                                         (u64)(uintptr_t)vq->desc,
 407                                         (u64)(uintptr_t)vq->avail,
 408                                         (u64)(uintptr_t)vq->used))
 409                         r = -EINVAL;
 410                 break;
 411
 412         case VHOST_SET_VRING_BASE:
 413                 vq_state.split.avail_index = vq->last_avail_idx;
 414                 if (ops->set_vq_state(vdpa, idx, &vq_state))
 415                         r = -EINVAL;
 416                 break;
 417
 418         case VHOST_SET_VRING_CALL:
 419                 if (vq->call_ctx.ctx) {
 420                         cb.callback = vhost_vdpa_virtqueue_cb;
 421                         cb.private = vq;
 422                 } else {
 423                         cb.callback = NULL;
 424                         cb.private = NULL;
 425                 }
 426                 ops->set_vq_cb(vdpa, idx, &cb);
 427                 vhost_vdpa_setup_vq_irq(v, idx);
 428                 break;
 429
 430         case VHOST_SET_VRING_NUM:
 431                 ops->set_vq_num(vdpa, idx, vq->num);
 432                 break;
 433         }
 434
 435         return r;
 436 }
 437
 438 static long vhost_vdpa_unlocked_ioctl(struct file *filep,
 439                                       unsigned int cmd, unsigned long arg)
 440 {
 441         struct vhost_vdpa *v = filep->private_data;
 442         struct vhost_dev *d = &v->vdev;
 443         void __user *argp = (void __user *)arg;
 444         u64 __user *featurep = argp;
 445         u64 features;
 446         long r = 0;
 447
 448         if (cmd == VHOST_SET_BACKEND_FEATURES) {
 449                 if (copy_from_user(&features, featurep, sizeof(features)))
 450                         return -EFAULT;
 451                 if (features & ~VHOST_VDPA_BACKEND_FEATURES)
 452                         return -EOPNOTSUPP;
 453                 vhost_set_backend_features(&v->vdev, features);
 454                 return 0;
 455         }
 456
 457         mutex_lock(&d->mutex);
 458
 459         switch (cmd) {
 460         case VHOST_VDPA_GET_DEVICE_ID:
 461                 r = vhost_vdpa_get_device_id(v, argp);
 462                 break;
 463         case VHOST_VDPA_GET_STATUS:
 464                 r = vhost_vdpa_get_status(v, argp);
 465                 break;
 466         case VHOST_VDPA_SET_STATUS:
 467                 r = vhost_vdpa_set_status(v, argp);
 468                 break;
 469         case VHOST_VDPA_GET_CONFIG:
 470                 r = vhost_vdpa_get_config(v, argp);
 471                 break;
 472         case VHOST_VDPA_SET_CONFIG:
 473                 r = vhost_vdpa_set_config(v, argp);
 474                 break;
 475         case VHOST_GET_FEATURES:
 476                 r = vhost_vdpa_get_features(v, argp);
 477                 break;
 478         case VHOST_SET_FEATURES:
 479                 r = vhost_vdpa_set_features(v, argp);
 480                 break;
 481         case VHOST_VDPA_GET_VRING_NUM:
 482                 r = vhost_vdpa_get_vring_num(v, argp);
 483                 break;
 484         case VHOST_SET_LOG_BASE:
 485         case VHOST_SET_LOG_FD:
 486                 r = -ENOIOCTLCMD;
 487                 break;
 488         case VHOST_VDPA_SET_CONFIG_CALL:
 489                 r = vhost_vdpa_set_config_call(v, argp);
 490                 break;
 491         case VHOST_GET_BACKEND_FEATURES:
 492                 features = VHOST_VDPA_BACKEND_FEATURES;
 493                 if (copy_to_user(featurep, &features, sizeof(features)))
 494                         r = -EFAULT;
 495                 break;
 496         case VHOST_VDPA_GET_IOVA_RANGE:
 497                 r = vhost_vdpa_get_iova_range(v, argp);
 498                 break;
 499         default:
 500                 r = vhost_dev_ioctl(&v->vdev, cmd, argp);
 501                 if (r == -ENOIOCTLCMD)
 502                         r = vhost_vdpa_vring_ioctl(v, cmd, argp);
 503                 break;
 504         }
 505
 506         mutex_unlock(&d->mutex);
 507         return r;
 508 }
 509
 510 static void vhost_vdpa_pa_unmap(struct vhost_vdpa *v, u64 start, u64 last)
 511 {
 512         struct vhost_dev *dev = &v->vdev;
 513         struct vhost_iotlb *iotlb = dev->iotlb;
 514         struct vhost_iotlb_map *map;
 515         struct page *page;
 516         unsigned long pfn, pinned;
 517
 518         while ((map = vhost_iotlb_itree_first(iotlb, start, last)) != NULL) {
 519                 pinned = PFN_DOWN(map->size);
 520                 for (pfn = PFN_DOWN(map->addr);
 521                      pinned > 0; pfn++, pinned--) {
 522                         page = pfn_to_page(pfn);
 523                         if (map->perm & VHOST_ACCESS_WO)
 524                                 set_page_dirty_lock(page);
 525                         unpin_user_page(page);
 526                 }
 527                 atomic64_sub(PFN_DOWN(map->size), &dev->mm->pinned_vm);
 528                 vhost_iotlb_map_free(iotlb, map);
 529         }
 530 }
 531
 532 static void vhost_vdpa_va_unmap(struct vhost_vdpa *v, u64 start, u64 last)
 533 {
 534         struct vhost_dev *dev = &v->vdev;
 535         struct vhost_iotlb *iotlb = dev->iotlb;
 536         struct vhost_iotlb_map *map;
 537         struct vdpa_map_file *map_file;
 538
 539         while ((map = vhost_iotlb_itree_first(iotlb, start, last)) != NULL) {
 540                 map_file = (struct vdpa_map_file *)map->opaque;
 541                 fput(map_file->file);
 542                 kfree(map_file);
 543                 vhost_iotlb_map_free(iotlb, map);
 544         }
 545 }
 546
 547 static void vhost_vdpa_iotlb_unmap(struct vhost_vdpa *v, u64 start, u64 last)
 548 {
 549         struct vdpa_device *vdpa = v->vdpa;
 550
 551         if (vdpa->use_va)
 552                 return vhost_vdpa_va_unmap(v, start, last);
 553
 554         return vhost_vdpa_pa_unmap(v, start, last);
 555 }
 556
 557 static void vhost_vdpa_iotlb_free(struct vhost_vdpa *v)
 558 {
 559         struct vhost_dev *dev = &v->vdev;
 560
 561         vhost_vdpa_iotlb_unmap(v, 0ULL, 0ULL - 1);
 562         kfree(dev->iotlb);
 563         dev->iotlb = NULL;
 564 }
 565
 566 static int perm_to_iommu_flags(u32 perm)
 567 {
 568         int flags = 0;
 569
 570         switch (perm) {
 571         case VHOST_ACCESS_WO:
 572                 flags |= IOMMU_WRITE;
 573                 break;
 574         case VHOST_ACCESS_RO:
 575                 flags |= IOMMU_READ;
 576                 break;
 577         case VHOST_ACCESS_RW:
 578                 flags |= (IOMMU_WRITE | IOMMU_READ);
 579                 break;
 580         default:
 581                 WARN(1, "invalidate vhost IOTLB permission\n");
 582                 break;
 583         }
 584
 585         return flags | IOMMU_CACHE;
 586 }
 587
 588 static int vhost_vdpa_map(struct vhost_vdpa *v, u64 iova,
 589                           u64 size, u64 pa, u32 perm, void *opaque)
 590 {
 591         struct vhost_dev *dev = &v->vdev;
 592         struct vdpa_device *vdpa = v->vdpa;
 593         const struct vdpa_config_ops *ops = vdpa->config;
 594         int r = 0;
 595
 596         r = vhost_iotlb_add_range_ctx(dev->iotlb, iova, iova + size - 1,
 597                                       pa, perm, opaque);
 598         if (r)
 599                 return r;
 600
 601         if (ops->dma_map) {
 602                 r = ops->dma_map(vdpa, iova, size, pa, perm, opaque);
 603         } else if (ops->set_map) {
 604                 if (!v->in_batch)
 605                         r = ops->set_map(vdpa, dev->iotlb);
 606         } else {
 607                 r = iommu_map(v->domain, iova, pa, size,
 608                               perm_to_iommu_flags(perm));
 609         }
 610         if (r) {
 611                 vhost_iotlb_del_range(dev->iotlb, iova, iova + size - 1);
 612                 return r;
 613         }
 614
 615         if (!vdpa->use_va)
 616                 atomic64_add(PFN_DOWN(size), &dev->mm->pinned_vm);
 617
 618         return 0;
 619 }
 620
 621 static void vhost_vdpa_unmap(struct vhost_vdpa *v, u64 iova, u64 size)
 622 {
 623         struct vhost_dev *dev = &v->vdev;
 624         struct vdpa_device *vdpa = v->vdpa;
 625         const struct vdpa_config_ops *ops = vdpa->config;
 626
 627         vhost_vdpa_iotlb_unmap(v, iova, iova + size - 1);
 628
 629         if (ops->dma_map) {
 630                 ops->dma_unmap(vdpa, iova, size);
 631         } else if (ops->set_map) {
 632                 if (!v->in_batch)
 633                         ops->set_map(vdpa, dev->iotlb);
 634         } else {
 635                 iommu_unmap(v->domain, iova, size);
 636         }
 637 }
 638
 639 static int vhost_vdpa_va_map(struct vhost_vdpa *v,
 640                              u64 iova, u64 size, u64 uaddr, u32 perm)
 641 {
 642         struct vhost_dev *dev = &v->vdev;
 643         u64 offset, map_size, map_iova = iova;
 644         struct vdpa_map_file *map_file;
 645         struct vm_area_struct *vma;
 646         int ret = 0;
 647
 648         mmap_read_lock(dev->mm);
 649
 650         while (size) {
 651                 vma = find_vma(dev->mm, uaddr);
 652                 if (!vma) {
 653                         ret = -EINVAL;
 654                         break;
 655                 }
 656                 map_size = min(size, vma->vm_end - uaddr);
 657                 if (!(vma->vm_file && (vma->vm_flags & VM_SHARED) &&
 658                         !(vma->vm_flags & (VM_IO | VM_PFNMAP))))
 659                         goto next;
 660
 661                 map_file = kzalloc(sizeof(*map_file), GFP_KERNEL);
 662                 if (!map_file) {
 663                         ret = -ENOMEM;
 664                         break;
 665                 }
 666                 offset = (vma->vm_pgoff << PAGE_SHIFT) + uaddr - vma->vm_start;
 667                 map_file->offset = offset;
 668                 map_file->file = get_file(vma->vm_file);
 669                 ret = vhost_vdpa_map(v, map_iova, map_size, uaddr,
 670                                      perm, map_file);
 671                 if (ret) {
 672                         fput(map_file->file);
 673                         kfree(map_file);
 674                         break;
 675                 }
 676 next:
 677                 size -= map_size;
 678                 uaddr += map_size;
 679                 map_iova += map_size;
 680         }
 681         if (ret)
 682                 vhost_vdpa_unmap(v, iova, map_iova - iova);
 683
 684         mmap_read_unlock(dev->mm);
 685
 686         return ret;
 687 }
 688
 689 static int vhost_vdpa_pa_map(struct vhost_vdpa *v,
 690                              u64 iova, u64 size, u64 uaddr, u32 perm)
 691 {
 692         struct vhost_dev *dev = &v->vdev;
 693         struct page **page_list;
 694         unsigned long list_size = PAGE_SIZE / sizeof(struct page *);
 695         unsigned int gup_flags = FOLL_LONGTERM;
 696         unsigned long npages, cur_base, map_pfn, last_pfn = 0;
 697         unsigned long lock_limit, sz2pin, nchunks, i;
 698         u64 start = iova;
 699         long pinned;
 700         int ret = 0;
 701
 702         /* Limit the use of memory for bookkeeping */
 703         page_list = (struct page **) __get_free_page(GFP_KERNEL);
 704         if (!page_list)
 705                 return -ENOMEM;
 706
 707         if (perm & VHOST_ACCESS_WO)
 708                 gup_flags |= FOLL_WRITE;
 709
 710         npages = PFN_UP(size + (iova & ~PAGE_MASK));
 711         if (!npages) {
 712                 ret = -EINVAL;
 713                 goto free;
 714         }
 715
 716         mmap_read_lock(dev->mm);
 717
 718         lock_limit = PFN_DOWN(rlimit(RLIMIT_MEMLOCK));
 719         if (npages + atomic64_read(&dev->mm->pinned_vm) > lock_limit) {
 720                 ret = -ENOMEM;
 721                 goto unlock;
 722         }
 723
 724         cur_base = uaddr & PAGE_MASK;
 725         iova &= PAGE_MASK;
 726         nchunks = 0;
 727
 728         while (npages) {
 729                 sz2pin = min_t(unsigned long, npages, list_size);
 730                 pinned = pin_user_pages(cur_base, sz2pin,
 731                                         gup_flags, page_list, NULL);
 732                 if (sz2pin != pinned) {
 733                         if (pinned < 0) {
 734                                 ret = pinned;
 735                         } else {
 736                                 unpin_user_pages(page_list, pinned);
 737                                 ret = -ENOMEM;
 738                         }
 739                         goto out;
 740                 }
 741                 nchunks++;
 742
 743                 if (!last_pfn)
 744                         map_pfn = page_to_pfn(page_list[0]);
 745
 746                 for (i = 0; i < pinned; i++) {
 747                         unsigned long this_pfn = page_to_pfn(page_list[i]);
 748                         u64 csize;
 749
 750                         if (last_pfn && (this_pfn != last_pfn + 1)) {
 751                                 /* Pin a contiguous chunk of memory */
 752                                 csize = PFN_PHYS(last_pfn - map_pfn + 1);
 753                                 ret = vhost_vdpa_map(v, iova, csize,
 754                                                      PFN_PHYS(map_pfn),
 755                                                      perm, NULL);
 756                                 if (ret) {
 757                                         /*
 758                                          * Unpin the pages that are left unmapped
 759                                          * from this point on in the current
 760                                          * page_list. The remaining outstanding
 761                                          * ones which may stride across several
 762                                          * chunks will be covered in the common
 763                                          * error path subsequently.
 764                                          */
 765                                         unpin_user_pages(&page_list[i],
 766                                                          pinned - i);
 767                                         goto out;
 768                                 }
 769
 770                                 map_pfn = this_pfn;
 771                                 iova += csize;
 772                                 nchunks = 0;
 773                         }
 774
 775                         last_pfn = this_pfn;
 776                 }
 777
 778                 cur_base += PFN_PHYS(pinned);
 779                 npages -= pinned;
 780         }
 781
 782         /* Pin the rest chunk */
 783         ret = vhost_vdpa_map(v, iova, PFN_PHYS(last_pfn - map_pfn + 1),
 784                              PFN_PHYS(map_pfn), perm, NULL);
 785 out:
 786         if (ret) {
 787                 if (nchunks) {
 788                         unsigned long pfn;
 789
 790                         /*
 791                          * Unpin the outstanding pages which are yet to be
 792                          * mapped but haven't due to vdpa_map() or
 793                          * pin_user_pages() failure.
 794                          *
 795                          * Mapped pages are accounted in vdpa_map(), hence
 796                          * the corresponding unpinning will be handled by
 797                          * vdpa_unmap().
 798                          */
 799                         WARN_ON(!last_pfn);
 800                         for (pfn = map_pfn; pfn <= last_pfn; pfn++)
 801                                 unpin_user_page(pfn_to_page(pfn));
 802                 }
 803                 vhost_vdpa_unmap(v, start, size);
 804         }
 805 unlock:
 806         mmap_read_unlock(dev->mm);
 807 free:
 808         free_page((unsigned long)page_list);
 809         return ret;
 810
 811 }
 812
 813 static int vhost_vdpa_process_iotlb_update(struct vhost_vdpa *v,
 814                                            struct vhost_iotlb_msg *msg)
 815 {
 816         struct vhost_dev *dev = &v->vdev;
 817         struct vdpa_device *vdpa = v->vdpa;
 818         struct vhost_iotlb *iotlb = dev->iotlb;
 819
 820         if (msg->iova < v->range.first || !msg->size ||
 821             msg->iova > U64_MAX - msg->size + 1 ||
 822             msg->iova + msg->size - 1 > v->range.last)
 823                 return -EINVAL;
 824
 825         if (vhost_iotlb_itree_first(iotlb, msg->iova,
 826                                     msg->iova + msg->size - 1))
 827                 return -EEXIST;
 828
 829         if (vdpa->use_va)
 830                 return vhost_vdpa_va_map(v, msg->iova, msg->size,
 831                                          msg->uaddr, msg->perm);
 832
 833         return vhost_vdpa_pa_map(v, msg->iova, msg->size, msg->uaddr,
 834                                  msg->perm);
 835 }
 836
 837 static int vhost_vdpa_process_iotlb_msg(struct vhost_dev *dev,
 838                                         struct vhost_iotlb_msg *msg)
 839 {
 840         struct vhost_vdpa *v = container_of(dev, struct vhost_vdpa, vdev);
 841         struct vdpa_device *vdpa = v->vdpa;
 842         const struct vdpa_config_ops *ops = vdpa->config;
 843         int r = 0;
 844
 845         mutex_lock(&dev->mutex);
 846
 847         r = vhost_dev_check_owner(dev);
 848         if (r)
 849                 goto unlock;
 850
 851         switch (msg->type) {
 852         case VHOST_IOTLB_UPDATE:
 853                 r = vhost_vdpa_process_iotlb_update(v, msg);
 854                 break;
 855         case VHOST_IOTLB_INVALIDATE:
 856                 vhost_vdpa_unmap(v, msg->iova, msg->size);
 857                 break;
 858         case VHOST_IOTLB_BATCH_BEGIN:
 859                 v->in_batch = true;
 860                 break;
 861         case VHOST_IOTLB_BATCH_END:
 862                 if (v->in_batch && ops->set_map)
 863                         ops->set_map(vdpa, dev->iotlb);
 864                 v->in_batch = false;
 865                 break;
 866         default:
 867                 r = -EINVAL;
 868                 break;
 869         }
 870 unlock:
 871         mutex_unlock(&dev->mutex);
 872
 873         return r;
 874 }
 875
 876 static ssize_t vhost_vdpa_chr_write_iter(struct kiocb *iocb,
 877                                          struct iov_iter *from)
 878 {
 879         struct file *file = iocb->ki_filp;
 880         struct vhost_vdpa *v = file->private_data;
 881         struct vhost_dev *dev = &v->vdev;
 882
 883         return vhost_chr_write_iter(dev, from);
 884 }
 885
 886 static int vhost_vdpa_alloc_domain(struct vhost_vdpa *v)
 887 {
 888         struct vdpa_device *vdpa = v->vdpa;
 889         const struct vdpa_config_ops *ops = vdpa->config;
 890         struct device *dma_dev = vdpa_get_dma_dev(vdpa);
 891         struct bus_type *bus;
 892         int ret;
 893
 894         /* Device want to do DMA by itself */
 895         if (ops->set_map || ops->dma_map)
 896                 return 0;
 897
 898         bus = dma_dev->bus;
 899         if (!bus)
 900                 return -EFAULT;
 901
 902         if (!iommu_capable(bus, IOMMU_CAP_CACHE_COHERENCY))
 903                 return -ENOTSUPP;
 904
 905         v->domain = iommu_domain_alloc(bus);
 906         if (!v->domain)
 907                 return -EIO;
 908
 909         ret = iommu_attach_device(v->domain, dma_dev);
 910         if (ret)
 911                 goto err_attach;
 912
 913         return 0;
 914
 915 err_attach:
 916         iommu_domain_free(v->domain);
 917         return ret;
 918 }
 919
 920 static void vhost_vdpa_free_domain(struct vhost_vdpa *v)
 921 {
 922         struct vdpa_device *vdpa = v->vdpa;
 923         struct device *dma_dev = vdpa_get_dma_dev(vdpa);
 924
 925         if (v->domain) {
 926                 iommu_detach_device(v->domain, dma_dev);
 927                 iommu_domain_free(v->domain);
 928         }
 929
 930         v->domain = NULL;
 931 }
 932
 933 static void vhost_vdpa_set_iova_range(struct vhost_vdpa *v)
 934 {
 935         struct vdpa_iova_range *range = &v->range;
 936         struct vdpa_device *vdpa = v->vdpa;
 937         const struct vdpa_config_ops *ops = vdpa->config;
 938
 939         if (ops->get_iova_range) {
 940                 *range = ops->get_iova_range(vdpa);
 941         } else if (v->domain && v->domain->geometry.force_aperture) {
 942                 range->first = v->domain->geometry.aperture_start;
 943                 range->last = v->domain->geometry.aperture_end;
 944         } else {
 945                 range->first = 0;
 946                 range->last = ULLONG_MAX;
 947         }
 948 }
 949
 950 static int vhost_vdpa_open(struct inode *inode, struct file *filep)
 951 {
 952         struct vhost_vdpa *v;
 953         struct vhost_dev *dev;
 954         struct vhost_virtqueue **vqs;
 955         int nvqs, i, r, opened;
 956
 957         v = container_of(inode->i_cdev, struct vhost_vdpa, cdev);
 958
 959         opened = atomic_cmpxchg(&v->opened, 0, 1);
 960         if (opened)
 961                 return -EBUSY;
 962
 963         nvqs = v->nvqs;
 964         r = vhost_vdpa_reset(v);
 965         if (r)
 966                 goto err;
 967
 968         vqs = kmalloc_array(nvqs, sizeof(*vqs), GFP_KERNEL);
 969         if (!vqs) {
 970                 r = -ENOMEM;
 971                 goto err;
 972         }
 973
 974         dev = &v->vdev;
 975         for (i = 0; i < nvqs; i++) {
 976                 vqs[i] = &v->vqs[i];
 977                 vqs[i]->handle_kick = handle_vq_kick;
 978         }
 979         vhost_dev_init(dev, vqs, nvqs, 0, 0, 0, false,
 980                        vhost_vdpa_process_iotlb_msg);
 981
 982         dev->iotlb = vhost_iotlb_alloc(0, 0);
 983         if (!dev->iotlb) {
 984                 r = -ENOMEM;
 985                 goto err_init_iotlb;
 986         }
 987
 988         r = vhost_vdpa_alloc_domain(v);
 989         if (r)
 990                 goto err_init_iotlb;
 991
 992         vhost_vdpa_set_iova_range(v);
 993
 994         filep->private_data = v;
 995
 996         return 0;
 997
 998 err_init_iotlb:
 999         vhost_dev_cleanup(&v->vdev);
1000         kfree(vqs);
1001 err:
1002         atomic_dec(&v->opened);
1003         return r;
1004 }
1005
1006 static void vhost_vdpa_clean_irq(struct vhost_vdpa *v)
1007 {
1008         int i;
1009
1010         for (i = 0; i < v->nvqs; i++)
1011                 vhost_vdpa_unsetup_vq_irq(v, i);
1012 }
1013
1014 static int vhost_vdpa_release(struct inode *inode, struct file *filep)
1015 {
1016         struct vhost_vdpa *v = filep->private_data;
1017         struct vhost_dev *d = &v->vdev;
1018
1019         mutex_lock(&d->mutex);
1020         filep->private_data = NULL;
1021         vhost_vdpa_reset(v);
1022         vhost_dev_stop(&v->vdev);
1023         vhost_vdpa_iotlb_free(v);
1024         vhost_vdpa_free_domain(v);
1025         vhost_vdpa_config_put(v);
1026         vhost_vdpa_clean_irq(v);
1027         vhost_dev_cleanup(&v->vdev);
1028         kfree(v->vdev.vqs);
1029         mutex_unlock(&d->mutex);
1030
1031         atomic_dec(&v->opened);
1032         complete(&v->completion);
1033
1034         return 0;
1035 }
1036
1037 #ifdef CONFIG_MMU
1038 static vm_fault_t vhost_vdpa_fault(struct vm_fault *vmf)
1039 {
1040         struct vhost_vdpa *v = vmf->vma->vm_file->private_data;
1041         struct vdpa_device *vdpa = v->vdpa;
1042         const struct vdpa_config_ops *ops = vdpa->config;
1043         struct vdpa_notification_area notify;
1044         struct vm_area_struct *vma = vmf->vma;
1045         u16 index = vma->vm_pgoff;
1046
1047         notify = ops->get_vq_notification(vdpa, index);
1048
1049         vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1050         if (remap_pfn_range(vma, vmf->address & PAGE_MASK,
1051                             PFN_DOWN(notify.addr), PAGE_SIZE,
1052                             vma->vm_page_prot))
1053                 return VM_FAULT_SIGBUS;
1054
1055         return VM_FAULT_NOPAGE;
1056 }
1057
1058 static const struct vm_operations_struct vhost_vdpa_vm_ops = {
1059         .fault = vhost_vdpa_fault,
1060 };
1061
1062 static int vhost_vdpa_mmap(struct file *file, struct vm_area_struct *vma)
1063 {
1064         struct vhost_vdpa *v = vma->vm_file->private_data;
1065         struct vdpa_device *vdpa = v->vdpa;
1066         const struct vdpa_config_ops *ops = vdpa->config;
1067         struct vdpa_notification_area notify;
1068         unsigned long index = vma->vm_pgoff;
1069
1070         if (vma->vm_end - vma->vm_start != PAGE_SIZE)
1071                 return -EINVAL;
1072         if ((vma->vm_flags & VM_SHARED) == 0)
1073                 return -EINVAL;
1074         if (vma->vm_flags & VM_READ)
1075                 return -EINVAL;
1076         if (index > 65535)
1077                 return -EINVAL;
1078         if (!ops->get_vq_notification)
1079                 return -ENOTSUPP;
1080
1081         /* To be safe and easily modelled by userspace, We only
1082          * support the doorbell which sits on the page boundary and
1083          * does not share the page with other registers.
1084          */
1085         notify = ops->get_vq_notification(vdpa, index);
1086         if (notify.addr & (PAGE_SIZE - 1))
1087                 return -EINVAL;
1088         if (vma->vm_end - vma->vm_start != notify.size)
1089                 return -ENOTSUPP;
1090
1091         vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
1092         vma->vm_ops = &vhost_vdpa_vm_ops;
1093         return 0;
1094 }
1095 #endif /* CONFIG_MMU */
1096
1097 static const struct file_operations vhost_vdpa_fops = {
1098         .owner          = THIS_MODULE,
1099         .open           = vhost_vdpa_open,
1100         .release        = vhost_vdpa_release,
1101         .write_iter     = vhost_vdpa_chr_write_iter,
1102         .unlocked_ioctl = vhost_vdpa_unlocked_ioctl,
1103 #ifdef CONFIG_MMU
1104         .mmap           = vhost_vdpa_mmap,
1105 #endif /* CONFIG_MMU */
1106         .compat_ioctl   = compat_ptr_ioctl,
1107 };
1108
1109 static void vhost_vdpa_release_dev(struct device *device)
1110 {
1111         struct vhost_vdpa *v =
1112                container_of(device, struct vhost_vdpa, dev);
1113
1114         ida_simple_remove(&vhost_vdpa_ida, v->minor);
1115         kfree(v->vqs);
1116         kfree(v);
1117 }
1118
1119 static int vhost_vdpa_probe(struct vdpa_device *vdpa)
1120 {
1121         const struct vdpa_config_ops *ops = vdpa->config;
1122         struct vhost_vdpa *v;
1123         int minor;
1124         int r;
1125
1126         v = kzalloc(sizeof(*v), GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1127         if (!v)
1128                 return -ENOMEM;
1129
1130         minor = ida_simple_get(&vhost_vdpa_ida, 0,
1131                                VHOST_VDPA_DEV_MAX, GFP_KERNEL);
1132         if (minor < 0) {
1133                 kfree(v);
1134                 return minor;
1135         }
1136
1137         atomic_set(&v->opened, 0);
1138         v->minor = minor;
1139         v->vdpa = vdpa;
1140         v->nvqs = vdpa->nvqs;
1141         v->virtio_id = ops->get_device_id(vdpa);
1142
1143         device_initialize(&v->dev);
1144         v->dev.release = vhost_vdpa_release_dev;
1145         v->dev.parent = &vdpa->dev;
1146         v->dev.devt = MKDEV(MAJOR(vhost_vdpa_major), minor);
1147         v->vqs = kmalloc_array(v->nvqs, sizeof(struct vhost_virtqueue),
1148                                GFP_KERNEL);
1149         if (!v->vqs) {
1150                 r = -ENOMEM;
1151                 goto err;
1152         }
1153
1154         r = dev_set_name(&v->dev, "vhost-vdpa-%u", minor);
1155         if (r)
1156                 goto err;
1157
1158         cdev_init(&v->cdev, &vhost_vdpa_fops);
1159         v->cdev.owner = THIS_MODULE;
1160
1161         r = cdev_device_add(&v->cdev, &v->dev);
1162         if (r)
1163                 goto err;
1164
1165         init_completion(&v->completion);
1166         vdpa_set_drvdata(vdpa, v);
1167
1168         return 0;
1169
1170 err:
1171         put_device(&v->dev);
1172         return r;
1173 }
1174
1175 static void vhost_vdpa_remove(struct vdpa_device *vdpa)
1176 {
1177         struct vhost_vdpa *v = vdpa_get_drvdata(vdpa);
1178         int opened;
1179
1180         cdev_device_del(&v->cdev, &v->dev);
1181
1182         do {
1183                 opened = atomic_cmpxchg(&v->opened, 0, 1);
1184                 if (!opened)
1185                         break;
1186                 wait_for_completion(&v->completion);
1187         } while (1);
1188
1189         put_device(&v->dev);
1190 }
1191
1192 static struct vdpa_driver vhost_vdpa_driver = {
1193         .driver = {
1194                 .name   = "vhost_vdpa",
1195         },
1196         .probe  = vhost_vdpa_probe,
1197         .remove = vhost_vdpa_remove,
1198 };
1199
1200 static int __init vhost_vdpa_init(void)
1201 {
1202         int r;
1203
1204         r = alloc_chrdev_region(&vhost_vdpa_major, 0, VHOST_VDPA_DEV_MAX,
1205                                 "vhost-vdpa");
1206         if (r)
1207                 goto err_alloc_chrdev;
1208
1209         r = vdpa_register_driver(&vhost_vdpa_driver);
1210         if (r)
1211                 goto err_vdpa_register_driver;
1212
1213         return 0;
1214
1215 err_vdpa_register_driver:
1216         unregister_chrdev_region(vhost_vdpa_major, VHOST_VDPA_DEV_MAX);
1217 err_alloc_chrdev:
1218         return r;
1219 }
1220 module_init(vhost_vdpa_init);
1221
1222 static void __exit vhost_vdpa_exit(void)
1223 {
1224         vdpa_unregister_driver(&vhost_vdpa_driver);
1225         unregister_chrdev_region(vhost_vdpa_major, VHOST_VDPA_DEV_MAX);
1226 }
1227 module_exit(vhost_vdpa_exit);
1228
1229 MODULE_VERSION("0.0.1");
1230 MODULE_LICENSE("GPL v2");
1231 MODULE_AUTHOR("Intel Corporation");
1232 MODULE_DESCRIPTION("vDPA-based vhost backend for virtio");