drivers/iommu/amd/iommu_v2.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright (C) 2010-2012 Advanced Micro Devices, Inc.
   4  * Author: Joerg Roedel <jroedel@suse.de>
   5  */
   6
   7 #define pr_fmt(fmt)     "AMD-Vi: " fmt
   8
   9 #include <linux/mmu_notifier.h>
  10 #include <linux/amd-iommu.h>
  11 #include <linux/mm_types.h>
  12 #include <linux/profile.h>
  13 #include <linux/module.h>
  14 #include <linux/sched.h>
  15 #include <linux/sched/mm.h>
  16 #include <linux/wait.h>
  17 #include <linux/pci.h>
  18 #include <linux/gfp.h>
  19
  20 #include "amd_iommu.h"
  21
  22 MODULE_LICENSE("GPL v2");
  23 MODULE_AUTHOR("Joerg Roedel <jroedel@suse.de>");
  24
  25 #define MAX_DEVICES             0x10000
  26 #define PRI_QUEUE_SIZE          512
  27
  28 struct pri_queue {
  29         atomic_t inflight;
  30         bool finish;
  31         int status;
  32 };
  33
  34 struct pasid_state {
  35         struct list_head list;                  /* For global state-list */
  36         atomic_t count;                         /* Reference count */
  37         unsigned mmu_notifier_count;            /* Counting nested mmu_notifier
  38                                                    calls */
  39         struct mm_struct *mm;                   /* mm_struct for the faults */
  40         struct mmu_notifier mn;                 /* mmu_notifier handle */
  41         struct pri_queue pri[PRI_QUEUE_SIZE];   /* PRI tag states */
  42         struct device_state *device_state;      /* Link to our device_state */
  43         int pasid;                              /* PASID index */
  44         bool invalid;                           /* Used during setup and
  45                                                    teardown of the pasid */
  46         spinlock_t lock;                        /* Protect pri_queues and
  47                                                    mmu_notifer_count */
  48         wait_queue_head_t wq;                   /* To wait for count == 0 */
  49 };
  50
  51 struct device_state {
  52         struct list_head list;
  53         u16 devid;
  54         atomic_t count;
  55         struct pci_dev *pdev;
  56         struct pasid_state **states;
  57         struct iommu_domain *domain;
  58         int pasid_levels;
  59         int max_pasids;
  60         amd_iommu_invalid_ppr_cb inv_ppr_cb;
  61         amd_iommu_invalidate_ctx inv_ctx_cb;
  62         spinlock_t lock;
  63         wait_queue_head_t wq;
  64 };
  65
  66 struct fault {
  67         struct work_struct work;
  68         struct device_state *dev_state;
  69         struct pasid_state *state;
  70         struct mm_struct *mm;
  71         u64 address;
  72         u16 devid;
  73         u16 pasid;
  74         u16 tag;
  75         u16 finish;
  76         u16 flags;
  77 };
  78
  79 static LIST_HEAD(state_list);
  80 static spinlock_t state_lock;
  81
  82 static struct workqueue_struct *iommu_wq;
  83
  84 static void free_pasid_states(struct device_state *dev_state);
  85
  86 static u16 device_id(struct pci_dev *pdev)
  87 {
  88         u16 devid;
  89
  90         devid = pdev->bus->number;
  91         devid = (devid << 8) | pdev->devfn;
  92
  93         return devid;
  94 }
  95
  96 static struct device_state *__get_device_state(u16 devid)
  97 {
  98         struct device_state *dev_state;
  99
 100         list_for_each_entry(dev_state, &state_list, list) {
 101                 if (dev_state->devid == devid)
 102                         return dev_state;
 103         }
 104
 105         return NULL;
 106 }
 107
 108 static struct device_state *get_device_state(u16 devid)
 109 {
 110         struct device_state *dev_state;
 111         unsigned long flags;
 112
 113         spin_lock_irqsave(&state_lock, flags);
 114         dev_state = __get_device_state(devid);
 115         if (dev_state != NULL)
 116                 atomic_inc(&dev_state->count);
 117         spin_unlock_irqrestore(&state_lock, flags);
 118
 119         return dev_state;
 120 }
 121
 122 static void free_device_state(struct device_state *dev_state)
 123 {
 124         struct iommu_group *group;
 125
 126         /*
 127          * First detach device from domain - No more PRI requests will arrive
 128          * from that device after it is unbound from the IOMMUv2 domain.
 129          */
 130         group = iommu_group_get(&dev_state->pdev->dev);
 131         if (WARN_ON(!group))
 132                 return;
 133
 134         iommu_detach_group(dev_state->domain, group);
 135
 136         iommu_group_put(group);
 137
 138         /* Everything is down now, free the IOMMUv2 domain */
 139         iommu_domain_free(dev_state->domain);
 140
 141         /* Finally get rid of the device-state */
 142         kfree(dev_state);
 143 }
 144
 145 static void put_device_state(struct device_state *dev_state)
 146 {
 147         if (atomic_dec_and_test(&dev_state->count))
 148                 wake_up(&dev_state->wq);
 149 }
 150
 151 /* Must be called under dev_state->lock */
 152 static struct pasid_state **__get_pasid_state_ptr(struct device_state *dev_state,
 153                                                   int pasid, bool alloc)
 154 {
 155         struct pasid_state **root, **ptr;
 156         int level, index;
 157
 158         level = dev_state->pasid_levels;
 159         root  = dev_state->states;
 160
 161         while (true) {
 162
 163                 index = (pasid >> (9 * level)) & 0x1ff;
 164                 ptr   = &root[index];
 165
 166                 if (level == 0)
 167                         break;
 168
 169                 if (*ptr == NULL) {
 170                         if (!alloc)
 171                                 return NULL;
 172
 173                         *ptr = (void *)get_zeroed_page(GFP_ATOMIC);
 174                         if (*ptr == NULL)
 175                                 return NULL;
 176                 }
 177
 178                 root   = (struct pasid_state **)*ptr;
 179                 level -= 1;
 180         }
 181
 182         return ptr;
 183 }
 184
 185 static int set_pasid_state(struct device_state *dev_state,
 186                            struct pasid_state *pasid_state,
 187                            int pasid)
 188 {
 189         struct pasid_state **ptr;
 190         unsigned long flags;
 191         int ret;
 192
 193         spin_lock_irqsave(&dev_state->lock, flags);
 194         ptr = __get_pasid_state_ptr(dev_state, pasid, true);
 195
 196         ret = -ENOMEM;
 197         if (ptr == NULL)
 198                 goto out_unlock;
 199
 200         ret = -ENOMEM;
 201         if (*ptr != NULL)
 202                 goto out_unlock;
 203
 204         *ptr = pasid_state;
 205
 206         ret = 0;
 207
 208 out_unlock:
 209         spin_unlock_irqrestore(&dev_state->lock, flags);
 210
 211         return ret;
 212 }
 213
 214 static void clear_pasid_state(struct device_state *dev_state, int pasid)
 215 {
 216         struct pasid_state **ptr;
 217         unsigned long flags;
 218
 219         spin_lock_irqsave(&dev_state->lock, flags);
 220         ptr = __get_pasid_state_ptr(dev_state, pasid, true);
 221
 222         if (ptr == NULL)
 223                 goto out_unlock;
 224
 225         *ptr = NULL;
 226
 227 out_unlock:
 228         spin_unlock_irqrestore(&dev_state->lock, flags);
 229 }
 230
 231 static struct pasid_state *get_pasid_state(struct device_state *dev_state,
 232                                            int pasid)
 233 {
 234         struct pasid_state **ptr, *ret = NULL;
 235         unsigned long flags;
 236
 237         spin_lock_irqsave(&dev_state->lock, flags);
 238         ptr = __get_pasid_state_ptr(dev_state, pasid, false);
 239
 240         if (ptr == NULL)
 241                 goto out_unlock;
 242
 243         ret = *ptr;
 244         if (ret)
 245                 atomic_inc(&ret->count);
 246
 247 out_unlock:
 248         spin_unlock_irqrestore(&dev_state->lock, flags);
 249
 250         return ret;
 251 }
 252
 253 static void free_pasid_state(struct pasid_state *pasid_state)
 254 {
 255         kfree(pasid_state);
 256 }
 257
 258 static void put_pasid_state(struct pasid_state *pasid_state)
 259 {
 260         if (atomic_dec_and_test(&pasid_state->count))
 261                 wake_up(&pasid_state->wq);
 262 }
 263
 264 static void put_pasid_state_wait(struct pasid_state *pasid_state)
 265 {
 266         atomic_dec(&pasid_state->count);
 267         wait_event(pasid_state->wq, !atomic_read(&pasid_state->count));
 268         free_pasid_state(pasid_state);
 269 }
 270
 271 static void unbind_pasid(struct pasid_state *pasid_state)
 272 {
 273         struct iommu_domain *domain;
 274
 275         domain = pasid_state->device_state->domain;
 276
 277         /*
 278          * Mark pasid_state as invalid, no more faults will we added to the
 279          * work queue after this is visible everywhere.
 280          */
 281         pasid_state->invalid = true;
 282
 283         /* Make sure this is visible */
 284         smp_wmb();
 285
 286         /* After this the device/pasid can't access the mm anymore */
 287         amd_iommu_domain_clear_gcr3(domain, pasid_state->pasid);
 288
 289         /* Make sure no more pending faults are in the queue */
 290         flush_workqueue(iommu_wq);
 291 }
 292
 293 static void free_pasid_states_level1(struct pasid_state **tbl)
 294 {
 295         int i;
 296
 297         for (i = 0; i < 512; ++i) {
 298                 if (tbl[i] == NULL)
 299                         continue;
 300
 301                 free_page((unsigned long)tbl[i]);
 302         }
 303 }
 304
 305 static void free_pasid_states_level2(struct pasid_state **tbl)
 306 {
 307         struct pasid_state **ptr;
 308         int i;
 309
 310         for (i = 0; i < 512; ++i) {
 311                 if (tbl[i] == NULL)
 312                         continue;
 313
 314                 ptr = (struct pasid_state **)tbl[i];
 315                 free_pasid_states_level1(ptr);
 316         }
 317 }
 318
 319 static void free_pasid_states(struct device_state *dev_state)
 320 {
 321         struct pasid_state *pasid_state;
 322         int i;
 323
 324         for (i = 0; i < dev_state->max_pasids; ++i) {
 325                 pasid_state = get_pasid_state(dev_state, i);
 326                 if (pasid_state == NULL)
 327                         continue;
 328
 329                 put_pasid_state(pasid_state);
 330
 331                 /*
 332                  * This will call the mn_release function and
 333                  * unbind the PASID
 334                  */
 335                 mmu_notifier_unregister(&pasid_state->mn, pasid_state->mm);
 336
 337                 put_pasid_state_wait(pasid_state); /* Reference taken in
 338                                                       amd_iommu_bind_pasid */
 339
 340                 /* Drop reference taken in amd_iommu_bind_pasid */
 341                 put_device_state(dev_state);
 342         }
 343
 344         if (dev_state->pasid_levels == 2)
 345                 free_pasid_states_level2(dev_state->states);
 346         else if (dev_state->pasid_levels == 1)
 347                 free_pasid_states_level1(dev_state->states);
 348         else
 349                 BUG_ON(dev_state->pasid_levels != 0);
 350
 351         free_page((unsigned long)dev_state->states);
 352 }
 353
 354 static struct pasid_state *mn_to_state(struct mmu_notifier *mn)
 355 {
 356         return container_of(mn, struct pasid_state, mn);
 357 }
 358
 359 static void mn_invalidate_range(struct mmu_notifier *mn,
 360                                 struct mm_struct *mm,
 361                                 unsigned long start, unsigned long end)
 362 {
 363         struct pasid_state *pasid_state;
 364         struct device_state *dev_state;
 365
 366         pasid_state = mn_to_state(mn);
 367         dev_state   = pasid_state->device_state;
 368
 369         if ((start ^ (end - 1)) < PAGE_SIZE)
 370                 amd_iommu_flush_page(dev_state->domain, pasid_state->pasid,
 371                                      start);
 372         else
 373                 amd_iommu_flush_tlb(dev_state->domain, pasid_state->pasid);
 374 }
 375
 376 static void mn_release(struct mmu_notifier *mn, struct mm_struct *mm)
 377 {
 378         struct pasid_state *pasid_state;
 379         struct device_state *dev_state;
 380         bool run_inv_ctx_cb;
 381
 382         might_sleep();
 383
 384         pasid_state    = mn_to_state(mn);
 385         dev_state      = pasid_state->device_state;
 386         run_inv_ctx_cb = !pasid_state->invalid;
 387
 388         if (run_inv_ctx_cb && dev_state->inv_ctx_cb)
 389                 dev_state->inv_ctx_cb(dev_state->pdev, pasid_state->pasid);
 390
 391         unbind_pasid(pasid_state);
 392 }
 393
 394 static const struct mmu_notifier_ops iommu_mn = {
 395         .release                = mn_release,
 396         .invalidate_range       = mn_invalidate_range,
 397 };
 398
 399 static void set_pri_tag_status(struct pasid_state *pasid_state,
 400                                u16 tag, int status)
 401 {
 402         unsigned long flags;
 403
 404         spin_lock_irqsave(&pasid_state->lock, flags);
 405         pasid_state->pri[tag].status = status;
 406         spin_unlock_irqrestore(&pasid_state->lock, flags);
 407 }
 408
 409 static void finish_pri_tag(struct device_state *dev_state,
 410                            struct pasid_state *pasid_state,
 411                            u16 tag)
 412 {
 413         unsigned long flags;
 414
 415         spin_lock_irqsave(&pasid_state->lock, flags);
 416         if (atomic_dec_and_test(&pasid_state->pri[tag].inflight) &&
 417             pasid_state->pri[tag].finish) {
 418                 amd_iommu_complete_ppr(dev_state->pdev, pasid_state->pasid,
 419                                        pasid_state->pri[tag].status, tag);
 420                 pasid_state->pri[tag].finish = false;
 421                 pasid_state->pri[tag].status = PPR_SUCCESS;
 422         }
 423         spin_unlock_irqrestore(&pasid_state->lock, flags);
 424 }
 425
 426 static void handle_fault_error(struct fault *fault)
 427 {
 428         int status;
 429
 430         if (!fault->dev_state->inv_ppr_cb) {
 431                 set_pri_tag_status(fault->state, fault->tag, PPR_INVALID);
 432                 return;
 433         }
 434
 435         status = fault->dev_state->inv_ppr_cb(fault->dev_state->pdev,
 436                                               fault->pasid,
 437                                               fault->address,
 438                                               fault->flags);
 439         switch (status) {
 440         case AMD_IOMMU_INV_PRI_RSP_SUCCESS:
 441                 set_pri_tag_status(fault->state, fault->tag, PPR_SUCCESS);
 442                 break;
 443         case AMD_IOMMU_INV_PRI_RSP_INVALID:
 444                 set_pri_tag_status(fault->state, fault->tag, PPR_INVALID);
 445                 break;
 446         case AMD_IOMMU_INV_PRI_RSP_FAIL:
 447                 set_pri_tag_status(fault->state, fault->tag, PPR_FAILURE);
 448                 break;
 449         default:
 450                 BUG();
 451         }
 452 }
 453
 454 static bool access_error(struct vm_area_struct *vma, struct fault *fault)
 455 {
 456         unsigned long requested = 0;
 457
 458         if (fault->flags & PPR_FAULT_EXEC)
 459                 requested |= VM_EXEC;
 460
 461         if (fault->flags & PPR_FAULT_READ)
 462                 requested |= VM_READ;
 463
 464         if (fault->flags & PPR_FAULT_WRITE)
 465                 requested |= VM_WRITE;
 466
 467         return (requested & ~vma->vm_flags) != 0;
 468 }
 469
 470 static void do_fault(struct work_struct *work)
 471 {
 472         struct fault *fault = container_of(work, struct fault, work);
 473         struct vm_area_struct *vma;
 474         vm_fault_t ret = VM_FAULT_ERROR;
 475         unsigned int flags = 0;
 476         struct mm_struct *mm;
 477         u64 address;
 478
 479         mm = fault->state->mm;
 480         address = fault->address;
 481
 482         if (fault->flags & PPR_FAULT_USER)
 483                 flags |= FAULT_FLAG_USER;
 484         if (fault->flags & PPR_FAULT_WRITE)
 485                 flags |= FAULT_FLAG_WRITE;
 486         flags |= FAULT_FLAG_REMOTE;
 487
 488         mmap_read_lock(mm);
 489         vma = find_extend_vma(mm, address);
 490         if (!vma || address < vma->vm_start)
 491                 /* failed to get a vma in the right range */
 492                 goto out;
 493
 494         /* Check if we have the right permissions on the vma */
 495         if (access_error(vma, fault))
 496                 goto out;
 497
 498         ret = handle_mm_fault(vma, address, flags);
 499 out:
 500         mmap_read_unlock(mm);
 501
 502         if (ret & VM_FAULT_ERROR)
 503                 /* failed to service fault */
 504                 handle_fault_error(fault);
 505
 506         finish_pri_tag(fault->dev_state, fault->state, fault->tag);
 507
 508         put_pasid_state(fault->state);
 509
 510         kfree(fault);
 511 }
 512
 513 static int ppr_notifier(struct notifier_block *nb, unsigned long e, void *data)
 514 {
 515         struct amd_iommu_fault *iommu_fault;
 516         struct pasid_state *pasid_state;
 517         struct device_state *dev_state;
 518         struct pci_dev *pdev = NULL;
 519         unsigned long flags;
 520         struct fault *fault;
 521         bool finish;
 522         u16 tag, devid;
 523         int ret;
 524
 525         iommu_fault = data;
 526         tag         = iommu_fault->tag & 0x1ff;
 527         finish      = (iommu_fault->tag >> 9) & 1;
 528
 529         devid = iommu_fault->device_id;
 530         pdev = pci_get_domain_bus_and_slot(0, PCI_BUS_NUM(devid),
 531                                            devid & 0xff);
 532         if (!pdev)
 533                 return -ENODEV;
 534
 535         ret = NOTIFY_DONE;
 536
 537         /* In kdump kernel pci dev is not initialized yet -> send INVALID */
 538         if (amd_iommu_is_attach_deferred(NULL, &pdev->dev)) {
 539                 amd_iommu_complete_ppr(pdev, iommu_fault->pasid,
 540                                        PPR_INVALID, tag);
 541                 goto out;
 542         }
 543
 544         dev_state = get_device_state(iommu_fault->device_id);
 545         if (dev_state == NULL)
 546                 goto out;
 547
 548         pasid_state = get_pasid_state(dev_state, iommu_fault->pasid);
 549         if (pasid_state == NULL || pasid_state->invalid) {
 550                 /* We know the device but not the PASID -> send INVALID */
 551                 amd_iommu_complete_ppr(dev_state->pdev, iommu_fault->pasid,
 552                                        PPR_INVALID, tag);
 553                 goto out_drop_state;
 554         }
 555
 556         spin_lock_irqsave(&pasid_state->lock, flags);
 557         atomic_inc(&pasid_state->pri[tag].inflight);
 558         if (finish)
 559                 pasid_state->pri[tag].finish = true;
 560         spin_unlock_irqrestore(&pasid_state->lock, flags);
 561
 562         fault = kzalloc(sizeof(*fault), GFP_ATOMIC);
 563         if (fault == NULL) {
 564                 /* We are OOM - send success and let the device re-fault */
 565                 finish_pri_tag(dev_state, pasid_state, tag);
 566                 goto out_drop_state;
 567         }
 568
 569         fault->dev_state = dev_state;
 570         fault->address   = iommu_fault->address;
 571         fault->state     = pasid_state;
 572         fault->tag       = tag;
 573         fault->finish    = finish;
 574         fault->pasid     = iommu_fault->pasid;
 575         fault->flags     = iommu_fault->flags;
 576         INIT_WORK(&fault->work, do_fault);
 577
 578         queue_work(iommu_wq, &fault->work);
 579
 580         ret = NOTIFY_OK;
 581
 582 out_drop_state:
 583
 584         if (ret != NOTIFY_OK && pasid_state)
 585                 put_pasid_state(pasid_state);
 586
 587         put_device_state(dev_state);
 588
 589 out:
 590         return ret;
 591 }
 592
 593 static struct notifier_block ppr_nb = {
 594         .notifier_call = ppr_notifier,
 595 };
 596
 597 int amd_iommu_bind_pasid(struct pci_dev *pdev, int pasid,
 598                          struct task_struct *task)
 599 {
 600         struct pasid_state *pasid_state;
 601         struct device_state *dev_state;
 602         struct mm_struct *mm;
 603         u16 devid;
 604         int ret;
 605
 606         might_sleep();
 607
 608         if (!amd_iommu_v2_supported())
 609                 return -ENODEV;
 610
 611         devid     = device_id(pdev);
 612         dev_state = get_device_state(devid);
 613
 614         if (dev_state == NULL)
 615                 return -EINVAL;
 616
 617         ret = -EINVAL;
 618         if (pasid < 0 || pasid >= dev_state->max_pasids)
 619                 goto out;
 620
 621         ret = -ENOMEM;
 622         pasid_state = kzalloc(sizeof(*pasid_state), GFP_KERNEL);
 623         if (pasid_state == NULL)
 624                 goto out;
 625
 626
 627         atomic_set(&pasid_state->count, 1);
 628         init_waitqueue_head(&pasid_state->wq);
 629         spin_lock_init(&pasid_state->lock);
 630
 631         mm                        = get_task_mm(task);
 632         pasid_state->mm           = mm;
 633         pasid_state->device_state = dev_state;
 634         pasid_state->pasid        = pasid;
 635         pasid_state->invalid      = true; /* Mark as valid only if we are
 636                                              done with setting up the pasid */
 637         pasid_state->mn.ops       = &iommu_mn;
 638
 639         if (pasid_state->mm == NULL)
 640                 goto out_free;
 641
 642         mmu_notifier_register(&pasid_state->mn, mm);
 643
 644         ret = set_pasid_state(dev_state, pasid_state, pasid);
 645         if (ret)
 646                 goto out_unregister;
 647
 648         ret = amd_iommu_domain_set_gcr3(dev_state->domain, pasid,
 649                                         __pa(pasid_state->mm->pgd));
 650         if (ret)
 651                 goto out_clear_state;
 652
 653         /* Now we are ready to handle faults */
 654         pasid_state->invalid = false;
 655
 656         /*
 657          * Drop the reference to the mm_struct here. We rely on the
 658          * mmu_notifier release call-back to inform us when the mm
 659          * is going away.
 660          */
 661         mmput(mm);
 662
 663         return 0;
 664
 665 out_clear_state:
 666         clear_pasid_state(dev_state, pasid);
 667
 668 out_unregister:
 669         mmu_notifier_unregister(&pasid_state->mn, mm);
 670         mmput(mm);
 671
 672 out_free:
 673         free_pasid_state(pasid_state);
 674
 675 out:
 676         put_device_state(dev_state);
 677
 678         return ret;
 679 }
 680 EXPORT_SYMBOL(amd_iommu_bind_pasid);
 681
 682 void amd_iommu_unbind_pasid(struct pci_dev *pdev, int pasid)
 683 {
 684         struct pasid_state *pasid_state;
 685         struct device_state *dev_state;
 686         u16 devid;
 687
 688         might_sleep();
 689
 690         if (!amd_iommu_v2_supported())
 691                 return;
 692
 693         devid = device_id(pdev);
 694         dev_state = get_device_state(devid);
 695         if (dev_state == NULL)
 696                 return;
 697
 698         if (pasid < 0 || pasid >= dev_state->max_pasids)
 699                 goto out;
 700
 701         pasid_state = get_pasid_state(dev_state, pasid);
 702         if (pasid_state == NULL)
 703                 goto out;
 704         /*
 705          * Drop reference taken here. We are safe because we still hold
 706          * the reference taken in the amd_iommu_bind_pasid function.
 707          */
 708         put_pasid_state(pasid_state);
 709
 710         /* Clear the pasid state so that the pasid can be re-used */
 711         clear_pasid_state(dev_state, pasid_state->pasid);
 712
 713         /*
 714          * Call mmu_notifier_unregister to drop our reference
 715          * to pasid_state->mm
 716          */
 717         mmu_notifier_unregister(&pasid_state->mn, pasid_state->mm);
 718
 719         put_pasid_state_wait(pasid_state); /* Reference taken in
 720                                               amd_iommu_bind_pasid */
 721 out:
 722         /* Drop reference taken in this function */
 723         put_device_state(dev_state);
 724
 725         /* Drop reference taken in amd_iommu_bind_pasid */
 726         put_device_state(dev_state);
 727 }
 728 EXPORT_SYMBOL(amd_iommu_unbind_pasid);
 729
 730 int amd_iommu_init_device(struct pci_dev *pdev, int pasids)
 731 {
 732         struct device_state *dev_state;
 733         struct iommu_group *group;
 734         unsigned long flags;
 735         int ret, tmp;
 736         u16 devid;
 737
 738         might_sleep();
 739
 740         if (!amd_iommu_v2_supported())
 741                 return -ENODEV;
 742
 743         if (pasids <= 0 || pasids > (PASID_MASK + 1))
 744                 return -EINVAL;
 745
 746         devid = device_id(pdev);
 747
 748         dev_state = kzalloc(sizeof(*dev_state), GFP_KERNEL);
 749         if (dev_state == NULL)
 750                 return -ENOMEM;
 751
 752         spin_lock_init(&dev_state->lock);
 753         init_waitqueue_head(&dev_state->wq);
 754         dev_state->pdev  = pdev;
 755         dev_state->devid = devid;
 756
 757         tmp = pasids;
 758         for (dev_state->pasid_levels = 0; (tmp - 1) & ~0x1ff; tmp >>= 9)
 759                 dev_state->pasid_levels += 1;
 760
 761         atomic_set(&dev_state->count, 1);
 762         dev_state->max_pasids = pasids;
 763
 764         ret = -ENOMEM;
 765         dev_state->states = (void *)get_zeroed_page(GFP_KERNEL);
 766         if (dev_state->states == NULL)
 767                 goto out_free_dev_state;
 768
 769         dev_state->domain = iommu_domain_alloc(&pci_bus_type);
 770         if (dev_state->domain == NULL)
 771                 goto out_free_states;
 772
 773         amd_iommu_domain_direct_map(dev_state->domain);
 774
 775         ret = amd_iommu_domain_enable_v2(dev_state->domain, pasids);
 776         if (ret)
 777                 goto out_free_domain;
 778
 779         group = iommu_group_get(&pdev->dev);
 780         if (!group) {
 781                 ret = -EINVAL;
 782                 goto out_free_domain;
 783         }
 784
 785         ret = iommu_attach_group(dev_state->domain, group);
 786         if (ret != 0)
 787                 goto out_drop_group;
 788
 789         iommu_group_put(group);
 790
 791         spin_lock_irqsave(&state_lock, flags);
 792
 793         if (__get_device_state(devid) != NULL) {
 794                 spin_unlock_irqrestore(&state_lock, flags);
 795                 ret = -EBUSY;
 796                 goto out_free_domain;
 797         }
 798
 799         list_add_tail(&dev_state->list, &state_list);
 800
 801         spin_unlock_irqrestore(&state_lock, flags);
 802
 803         return 0;
 804
 805 out_drop_group:
 806         iommu_group_put(group);
 807
 808 out_free_domain:
 809         iommu_domain_free(dev_state->domain);
 810
 811 out_free_states:
 812         free_page((unsigned long)dev_state->states);
 813
 814 out_free_dev_state:
 815         kfree(dev_state);
 816
 817         return ret;
 818 }
 819 EXPORT_SYMBOL(amd_iommu_init_device);
 820
 821 void amd_iommu_free_device(struct pci_dev *pdev)
 822 {
 823         struct device_state *dev_state;
 824         unsigned long flags;
 825         u16 devid;
 826
 827         if (!amd_iommu_v2_supported())
 828                 return;
 829
 830         devid = device_id(pdev);
 831
 832         spin_lock_irqsave(&state_lock, flags);
 833
 834         dev_state = __get_device_state(devid);
 835         if (dev_state == NULL) {
 836                 spin_unlock_irqrestore(&state_lock, flags);
 837                 return;
 838         }
 839
 840         list_del(&dev_state->list);
 841
 842         spin_unlock_irqrestore(&state_lock, flags);
 843
 844         /* Get rid of any remaining pasid states */
 845         free_pasid_states(dev_state);
 846
 847         put_device_state(dev_state);
 848         /*
 849          * Wait until the last reference is dropped before freeing
 850          * the device state.
 851          */
 852         wait_event(dev_state->wq, !atomic_read(&dev_state->count));
 853         free_device_state(dev_state);
 854 }
 855 EXPORT_SYMBOL(amd_iommu_free_device);
 856
 857 int amd_iommu_set_invalid_ppr_cb(struct pci_dev *pdev,
 858                                  amd_iommu_invalid_ppr_cb cb)
 859 {
 860         struct device_state *dev_state;
 861         unsigned long flags;
 862         u16 devid;
 863         int ret;
 864
 865         if (!amd_iommu_v2_supported())
 866                 return -ENODEV;
 867
 868         devid = device_id(pdev);
 869
 870         spin_lock_irqsave(&state_lock, flags);
 871
 872         ret = -EINVAL;
 873         dev_state = __get_device_state(devid);
 874         if (dev_state == NULL)
 875                 goto out_unlock;
 876
 877         dev_state->inv_ppr_cb = cb;
 878
 879         ret = 0;
 880
 881 out_unlock:
 882         spin_unlock_irqrestore(&state_lock, flags);
 883
 884         return ret;
 885 }
 886 EXPORT_SYMBOL(amd_iommu_set_invalid_ppr_cb);
 887
 888 int amd_iommu_set_invalidate_ctx_cb(struct pci_dev *pdev,
 889                                     amd_iommu_invalidate_ctx cb)
 890 {
 891         struct device_state *dev_state;
 892         unsigned long flags;
 893         u16 devid;
 894         int ret;
 895
 896         if (!amd_iommu_v2_supported())
 897                 return -ENODEV;
 898
 899         devid = device_id(pdev);
 900
 901         spin_lock_irqsave(&state_lock, flags);
 902
 903         ret = -EINVAL;
 904         dev_state = __get_device_state(devid);
 905         if (dev_state == NULL)
 906                 goto out_unlock;
 907
 908         dev_state->inv_ctx_cb = cb;
 909
 910         ret = 0;
 911
 912 out_unlock:
 913         spin_unlock_irqrestore(&state_lock, flags);
 914
 915         return ret;
 916 }
 917 EXPORT_SYMBOL(amd_iommu_set_invalidate_ctx_cb);
 918
 919 static int __init amd_iommu_v2_init(void)
 920 {
 921         int ret;
 922
 923         pr_info("AMD IOMMUv2 driver by Joerg Roedel <jroedel@suse.de>\n");
 924
 925         if (!amd_iommu_v2_supported()) {
 926                 pr_info("AMD IOMMUv2 functionality not available on this system\n");
 927                 /*
 928                  * Load anyway to provide the symbols to other modules
 929                  * which may use AMD IOMMUv2 optionally.
 930                  */
 931                 return 0;
 932         }
 933
 934         spin_lock_init(&state_lock);
 935
 936         ret = -ENOMEM;
 937         iommu_wq = alloc_workqueue("amd_iommu_v2", WQ_MEM_RECLAIM, 0);
 938         if (iommu_wq == NULL)
 939                 goto out;
 940
 941         amd_iommu_register_ppr_notifier(&ppr_nb);
 942
 943         return 0;
 944
 945 out:
 946         return ret;
 947 }
 948
 949 static void __exit amd_iommu_v2_exit(void)
 950 {
 951         struct device_state *dev_state;
 952         int i;
 953
 954         if (!amd_iommu_v2_supported())
 955                 return;
 956
 957         amd_iommu_unregister_ppr_notifier(&ppr_nb);
 958
 959         flush_workqueue(iommu_wq);
 960
 961         /*
 962          * The loop below might call flush_workqueue(), so call
 963          * destroy_workqueue() after it
 964          */
 965         for (i = 0; i < MAX_DEVICES; ++i) {
 966                 dev_state = get_device_state(i);
 967
 968                 if (dev_state == NULL)
 969                         continue;
 970
 971                 WARN_ON_ONCE(1);
 972
 973                 put_device_state(dev_state);
 974                 amd_iommu_free_device(dev_state->pdev);
 975         }
 976
 977         destroy_workqueue(iommu_wq);
 978 }
 979
 980 module_init(amd_iommu_v2_init);
 981 module_exit(amd_iommu_v2_exit);