drivers/iommu/amd_iommu_v2.c

   1 /*
   2  * Copyright (C) 2010-2012 Advanced Micro Devices, Inc.
   3  * Author: Joerg Roedel <jroedel@suse.de>
   4  *
   5  * This program is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 as published
   7  * by the Free Software Foundation.
   8  *
   9  * This program is distributed in the hope that it will be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12  * GNU General Public License for more details.
  13  *
  14  * You should have received a copy of the GNU General Public License
  15  * along with this program; if not, write to the Free Software
  16  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307 USA
  17  */
  18
  19 #include <linux/mmu_notifier.h>
  20 #include <linux/amd-iommu.h>
  21 #include <linux/mm_types.h>
  22 #include <linux/profile.h>
  23 #include <linux/module.h>
  24 #include <linux/sched.h>
  25 #include <linux/sched/mm.h>
  26 #include <linux/iommu.h>
  27 #include <linux/wait.h>
  28 #include <linux/pci.h>
  29 #include <linux/gfp.h>
  30
  31 #include "amd_iommu_types.h"
  32 #include "amd_iommu_proto.h"
  33
  34 MODULE_LICENSE("GPL v2");
  35 MODULE_AUTHOR("Joerg Roedel <jroedel@suse.de>");
  36
  37 #define MAX_DEVICES             0x10000
  38 #define PRI_QUEUE_SIZE          512
  39
  40 struct pri_queue {
  41         atomic_t inflight;
  42         bool finish;
  43         int status;
  44 };
  45
  46 struct pasid_state {
  47         struct list_head list;                  /* For global state-list */
  48         atomic_t count;                         /* Reference count */
  49         unsigned mmu_notifier_count;            /* Counting nested mmu_notifier
  50                                                    calls */
  51         struct mm_struct *mm;                   /* mm_struct for the faults */
  52         struct mmu_notifier mn;                 /* mmu_notifier handle */
  53         struct pri_queue pri[PRI_QUEUE_SIZE];   /* PRI tag states */
  54         struct device_state *device_state;      /* Link to our device_state */
  55         int pasid;                              /* PASID index */
  56         bool invalid;                           /* Used during setup and
  57                                                    teardown of the pasid */
  58         spinlock_t lock;                        /* Protect pri_queues and
  59                                                    mmu_notifer_count */
  60         wait_queue_head_t wq;                   /* To wait for count == 0 */
  61 };
  62
  63 struct device_state {
  64         struct list_head list;
  65         u16 devid;
  66         atomic_t count;
  67         struct pci_dev *pdev;
  68         struct pasid_state **states;
  69         struct iommu_domain *domain;
  70         int pasid_levels;
  71         int max_pasids;
  72         amd_iommu_invalid_ppr_cb inv_ppr_cb;
  73         amd_iommu_invalidate_ctx inv_ctx_cb;
  74         spinlock_t lock;
  75         wait_queue_head_t wq;
  76 };
  77
  78 struct fault {
  79         struct work_struct work;
  80         struct device_state *dev_state;
  81         struct pasid_state *state;
  82         struct mm_struct *mm;
  83         u64 address;
  84         u16 devid;
  85         u16 pasid;
  86         u16 tag;
  87         u16 finish;
  88         u16 flags;
  89 };
  90
  91 static LIST_HEAD(state_list);
  92 static spinlock_t state_lock;
  93
  94 static struct workqueue_struct *iommu_wq;
  95
  96 static void free_pasid_states(struct device_state *dev_state);
  97
  98 static u16 device_id(struct pci_dev *pdev)
  99 {
 100         u16 devid;
 101
 102         devid = pdev->bus->number;
 103         devid = (devid << 8) | pdev->devfn;
 104
 105         return devid;
 106 }
 107
 108 static struct device_state *__get_device_state(u16 devid)
 109 {
 110         struct device_state *dev_state;
 111
 112         list_for_each_entry(dev_state, &state_list, list) {
 113                 if (dev_state->devid == devid)
 114                         return dev_state;
 115         }
 116
 117         return NULL;
 118 }
 119
 120 static struct device_state *get_device_state(u16 devid)
 121 {
 122         struct device_state *dev_state;
 123         unsigned long flags;
 124
 125         spin_lock_irqsave(&state_lock, flags);
 126         dev_state = __get_device_state(devid);
 127         if (dev_state != NULL)
 128                 atomic_inc(&dev_state->count);
 129         spin_unlock_irqrestore(&state_lock, flags);
 130
 131         return dev_state;
 132 }
 133
 134 static void free_device_state(struct device_state *dev_state)
 135 {
 136         struct iommu_group *group;
 137
 138         /*
 139          * First detach device from domain - No more PRI requests will arrive
 140          * from that device after it is unbound from the IOMMUv2 domain.
 141          */
 142         group = iommu_group_get(&dev_state->pdev->dev);
 143         if (WARN_ON(!group))
 144                 return;
 145
 146         iommu_detach_group(dev_state->domain, group);
 147
 148         iommu_group_put(group);
 149
 150         /* Everything is down now, free the IOMMUv2 domain */
 151         iommu_domain_free(dev_state->domain);
 152
 153         /* Finally get rid of the device-state */
 154         kfree(dev_state);
 155 }
 156
 157 static void put_device_state(struct device_state *dev_state)
 158 {
 159         if (atomic_dec_and_test(&dev_state->count))
 160                 wake_up(&dev_state->wq);
 161 }
 162
 163 /* Must be called under dev_state->lock */
 164 static struct pasid_state **__get_pasid_state_ptr(struct device_state *dev_state,
 165                                                   int pasid, bool alloc)
 166 {
 167         struct pasid_state **root, **ptr;
 168         int level, index;
 169
 170         level = dev_state->pasid_levels;
 171         root  = dev_state->states;
 172
 173         while (true) {
 174
 175                 index = (pasid >> (9 * level)) & 0x1ff;
 176                 ptr   = &root[index];
 177
 178                 if (level == 0)
 179                         break;
 180
 181                 if (*ptr == NULL) {
 182                         if (!alloc)
 183                                 return NULL;
 184
 185                         *ptr = (void *)get_zeroed_page(GFP_ATOMIC);
 186                         if (*ptr == NULL)
 187                                 return NULL;
 188                 }
 189
 190                 root   = (struct pasid_state **)*ptr;
 191                 level -= 1;
 192         }
 193
 194         return ptr;
 195 }
 196
 197 static int set_pasid_state(struct device_state *dev_state,
 198                            struct pasid_state *pasid_state,
 199                            int pasid)
 200 {
 201         struct pasid_state **ptr;
 202         unsigned long flags;
 203         int ret;
 204
 205         spin_lock_irqsave(&dev_state->lock, flags);
 206         ptr = __get_pasid_state_ptr(dev_state, pasid, true);
 207
 208         ret = -ENOMEM;
 209         if (ptr == NULL)
 210                 goto out_unlock;
 211
 212         ret = -ENOMEM;
 213         if (*ptr != NULL)
 214                 goto out_unlock;
 215
 216         *ptr = pasid_state;
 217
 218         ret = 0;
 219
 220 out_unlock:
 221         spin_unlock_irqrestore(&dev_state->lock, flags);
 222
 223         return ret;
 224 }
 225
 226 static void clear_pasid_state(struct device_state *dev_state, int pasid)
 227 {
 228         struct pasid_state **ptr;
 229         unsigned long flags;
 230
 231         spin_lock_irqsave(&dev_state->lock, flags);
 232         ptr = __get_pasid_state_ptr(dev_state, pasid, true);
 233
 234         if (ptr == NULL)
 235                 goto out_unlock;
 236
 237         *ptr = NULL;
 238
 239 out_unlock:
 240         spin_unlock_irqrestore(&dev_state->lock, flags);
 241 }
 242
 243 static struct pasid_state *get_pasid_state(struct device_state *dev_state,
 244                                            int pasid)
 245 {
 246         struct pasid_state **ptr, *ret = NULL;
 247         unsigned long flags;
 248
 249         spin_lock_irqsave(&dev_state->lock, flags);
 250         ptr = __get_pasid_state_ptr(dev_state, pasid, false);
 251
 252         if (ptr == NULL)
 253                 goto out_unlock;
 254
 255         ret = *ptr;
 256         if (ret)
 257                 atomic_inc(&ret->count);
 258
 259 out_unlock:
 260         spin_unlock_irqrestore(&dev_state->lock, flags);
 261
 262         return ret;
 263 }
 264
 265 static void free_pasid_state(struct pasid_state *pasid_state)
 266 {
 267         kfree(pasid_state);
 268 }
 269
 270 static void put_pasid_state(struct pasid_state *pasid_state)
 271 {
 272         if (atomic_dec_and_test(&pasid_state->count))
 273                 wake_up(&pasid_state->wq);
 274 }
 275
 276 static void put_pasid_state_wait(struct pasid_state *pasid_state)
 277 {
 278         atomic_dec(&pasid_state->count);
 279         wait_event(pasid_state->wq, !atomic_read(&pasid_state->count));
 280         free_pasid_state(pasid_state);
 281 }
 282
 283 static void unbind_pasid(struct pasid_state *pasid_state)
 284 {
 285         struct iommu_domain *domain;
 286
 287         domain = pasid_state->device_state->domain;
 288
 289         /*
 290          * Mark pasid_state as invalid, no more faults will we added to the
 291          * work queue after this is visible everywhere.
 292          */
 293         pasid_state->invalid = true;
 294
 295         /* Make sure this is visible */
 296         smp_wmb();
 297
 298         /* After this the device/pasid can't access the mm anymore */
 299         amd_iommu_domain_clear_gcr3(domain, pasid_state->pasid);
 300
 301         /* Make sure no more pending faults are in the queue */
 302         flush_workqueue(iommu_wq);
 303 }
 304
 305 static void free_pasid_states_level1(struct pasid_state **tbl)
 306 {
 307         int i;
 308
 309         for (i = 0; i < 512; ++i) {
 310                 if (tbl[i] == NULL)
 311                         continue;
 312
 313                 free_page((unsigned long)tbl[i]);
 314         }
 315 }
 316
 317 static void free_pasid_states_level2(struct pasid_state **tbl)
 318 {
 319         struct pasid_state **ptr;
 320         int i;
 321
 322         for (i = 0; i < 512; ++i) {
 323                 if (tbl[i] == NULL)
 324                         continue;
 325
 326                 ptr = (struct pasid_state **)tbl[i];
 327                 free_pasid_states_level1(ptr);
 328         }
 329 }
 330
 331 static void free_pasid_states(struct device_state *dev_state)
 332 {
 333         struct pasid_state *pasid_state;
 334         int i;
 335
 336         for (i = 0; i < dev_state->max_pasids; ++i) {
 337                 pasid_state = get_pasid_state(dev_state, i);
 338                 if (pasid_state == NULL)
 339                         continue;
 340
 341                 put_pasid_state(pasid_state);
 342
 343                 /*
 344                  * This will call the mn_release function and
 345                  * unbind the PASID
 346                  */
 347                 mmu_notifier_unregister(&pasid_state->mn, pasid_state->mm);
 348
 349                 put_pasid_state_wait(pasid_state); /* Reference taken in
 350                                                       amd_iommu_bind_pasid */
 351
 352                 /* Drop reference taken in amd_iommu_bind_pasid */
 353                 put_device_state(dev_state);
 354         }
 355
 356         if (dev_state->pasid_levels == 2)
 357                 free_pasid_states_level2(dev_state->states);
 358         else if (dev_state->pasid_levels == 1)
 359                 free_pasid_states_level1(dev_state->states);
 360         else
 361                 BUG_ON(dev_state->pasid_levels != 0);
 362
 363         free_page((unsigned long)dev_state->states);
 364 }
 365
 366 static struct pasid_state *mn_to_state(struct mmu_notifier *mn)
 367 {
 368         return container_of(mn, struct pasid_state, mn);
 369 }
 370
 371 static void __mn_flush_page(struct mmu_notifier *mn,
 372                             unsigned long address)
 373 {
 374         struct pasid_state *pasid_state;
 375         struct device_state *dev_state;
 376
 377         pasid_state = mn_to_state(mn);
 378         dev_state   = pasid_state->device_state;
 379
 380         amd_iommu_flush_page(dev_state->domain, pasid_state->pasid, address);
 381 }
 382
 383 static int mn_clear_flush_young(struct mmu_notifier *mn,
 384                                 struct mm_struct *mm,
 385                                 unsigned long start,
 386                                 unsigned long end)
 387 {
 388         for (; start < end; start += PAGE_SIZE)
 389                 __mn_flush_page(mn, start);
 390
 391         return 0;
 392 }
 393
 394 static void mn_invalidate_range(struct mmu_notifier *mn,
 395                                 struct mm_struct *mm,
 396                                 unsigned long start, unsigned long end)
 397 {
 398         struct pasid_state *pasid_state;
 399         struct device_state *dev_state;
 400
 401         pasid_state = mn_to_state(mn);
 402         dev_state   = pasid_state->device_state;
 403
 404         if ((start ^ (end - 1)) < PAGE_SIZE)
 405                 amd_iommu_flush_page(dev_state->domain, pasid_state->pasid,
 406                                      start);
 407         else
 408                 amd_iommu_flush_tlb(dev_state->domain, pasid_state->pasid);
 409 }
 410
 411 static void mn_release(struct mmu_notifier *mn, struct mm_struct *mm)
 412 {
 413         struct pasid_state *pasid_state;
 414         struct device_state *dev_state;
 415         bool run_inv_ctx_cb;
 416
 417         might_sleep();
 418
 419         pasid_state    = mn_to_state(mn);
 420         dev_state      = pasid_state->device_state;
 421         run_inv_ctx_cb = !pasid_state->invalid;
 422
 423         if (run_inv_ctx_cb && dev_state->inv_ctx_cb)
 424                 dev_state->inv_ctx_cb(dev_state->pdev, pasid_state->pasid);
 425
 426         unbind_pasid(pasid_state);
 427 }
 428
 429 static const struct mmu_notifier_ops iommu_mn = {
 430         .release                = mn_release,
 431         .clear_flush_young      = mn_clear_flush_young,
 432         .invalidate_range       = mn_invalidate_range,
 433 };
 434
 435 static void set_pri_tag_status(struct pasid_state *pasid_state,
 436                                u16 tag, int status)
 437 {
 438         unsigned long flags;
 439
 440         spin_lock_irqsave(&pasid_state->lock, flags);
 441         pasid_state->pri[tag].status = status;
 442         spin_unlock_irqrestore(&pasid_state->lock, flags);
 443 }
 444
 445 static void finish_pri_tag(struct device_state *dev_state,
 446                            struct pasid_state *pasid_state,
 447                            u16 tag)
 448 {
 449         unsigned long flags;
 450
 451         spin_lock_irqsave(&pasid_state->lock, flags);
 452         if (atomic_dec_and_test(&pasid_state->pri[tag].inflight) &&
 453             pasid_state->pri[tag].finish) {
 454                 amd_iommu_complete_ppr(dev_state->pdev, pasid_state->pasid,
 455                                        pasid_state->pri[tag].status, tag);
 456                 pasid_state->pri[tag].finish = false;
 457                 pasid_state->pri[tag].status = PPR_SUCCESS;
 458         }
 459         spin_unlock_irqrestore(&pasid_state->lock, flags);
 460 }
 461
 462 static void handle_fault_error(struct fault *fault)
 463 {
 464         int status;
 465
 466         if (!fault->dev_state->inv_ppr_cb) {
 467                 set_pri_tag_status(fault->state, fault->tag, PPR_INVALID);
 468                 return;
 469         }
 470
 471         status = fault->dev_state->inv_ppr_cb(fault->dev_state->pdev,
 472                                               fault->pasid,
 473                                               fault->address,
 474                                               fault->flags);
 475         switch (status) {
 476         case AMD_IOMMU_INV_PRI_RSP_SUCCESS:
 477                 set_pri_tag_status(fault->state, fault->tag, PPR_SUCCESS);
 478                 break;
 479         case AMD_IOMMU_INV_PRI_RSP_INVALID:
 480                 set_pri_tag_status(fault->state, fault->tag, PPR_INVALID);
 481                 break;
 482         case AMD_IOMMU_INV_PRI_RSP_FAIL:
 483                 set_pri_tag_status(fault->state, fault->tag, PPR_FAILURE);
 484                 break;
 485         default:
 486                 BUG();
 487         }
 488 }
 489
 490 static bool access_error(struct vm_area_struct *vma, struct fault *fault)
 491 {
 492         unsigned long requested = 0;
 493
 494         if (fault->flags & PPR_FAULT_EXEC)
 495                 requested |= VM_EXEC;
 496
 497         if (fault->flags & PPR_FAULT_READ)
 498                 requested |= VM_READ;
 499
 500         if (fault->flags & PPR_FAULT_WRITE)
 501                 requested |= VM_WRITE;
 502
 503         return (requested & ~vma->vm_flags) != 0;
 504 }
 505
 506 static void do_fault(struct work_struct *work)
 507 {
 508         struct fault *fault = container_of(work, struct fault, work);
 509         struct vm_area_struct *vma;
 510         int ret = VM_FAULT_ERROR;
 511         unsigned int flags = 0;
 512         struct mm_struct *mm;
 513         u64 address;
 514
 515         mm = fault->state->mm;
 516         address = fault->address;
 517
 518         if (fault->flags & PPR_FAULT_USER)
 519                 flags |= FAULT_FLAG_USER;
 520         if (fault->flags & PPR_FAULT_WRITE)
 521                 flags |= FAULT_FLAG_WRITE;
 522         flags |= FAULT_FLAG_REMOTE;
 523
 524         down_read(&mm->mmap_sem);
 525         vma = find_extend_vma(mm, address);
 526         if (!vma || address < vma->vm_start)
 527                 /* failed to get a vma in the right range */
 528                 goto out;
 529
 530         /* Check if we have the right permissions on the vma */
 531         if (access_error(vma, fault))
 532                 goto out;
 533
 534         ret = handle_mm_fault(vma, address, flags);
 535 out:
 536         up_read(&mm->mmap_sem);
 537
 538         if (ret & VM_FAULT_ERROR)
 539                 /* failed to service fault */
 540                 handle_fault_error(fault);
 541
 542         finish_pri_tag(fault->dev_state, fault->state, fault->tag);
 543
 544         put_pasid_state(fault->state);
 545
 546         kfree(fault);
 547 }
 548
 549 static int ppr_notifier(struct notifier_block *nb, unsigned long e, void *data)
 550 {
 551         struct amd_iommu_fault *iommu_fault;
 552         struct pasid_state *pasid_state;
 553         struct device_state *dev_state;
 554         unsigned long flags;
 555         struct fault *fault;
 556         bool finish;
 557         u16 tag, devid;
 558         int ret;
 559         struct iommu_dev_data *dev_data;
 560         struct pci_dev *pdev = NULL;
 561
 562         iommu_fault = data;
 563         tag         = iommu_fault->tag & 0x1ff;
 564         finish      = (iommu_fault->tag >> 9) & 1;
 565
 566         devid = iommu_fault->device_id;
 567         pdev = pci_get_bus_and_slot(PCI_BUS_NUM(devid), devid & 0xff);
 568         if (!pdev)
 569                 return -ENODEV;
 570         dev_data = get_dev_data(&pdev->dev);
 571
 572         /* In kdump kernel pci dev is not initialized yet -> send INVALID */
 573         ret = NOTIFY_DONE;
 574         if (translation_pre_enabled(amd_iommu_rlookup_table[devid])
 575                 && dev_data->defer_attach) {
 576                 amd_iommu_complete_ppr(pdev, iommu_fault->pasid,
 577                                        PPR_INVALID, tag);
 578                 goto out;
 579         }
 580
 581         dev_state = get_device_state(iommu_fault->device_id);
 582         if (dev_state == NULL)
 583                 goto out;
 584
 585         pasid_state = get_pasid_state(dev_state, iommu_fault->pasid);
 586         if (pasid_state == NULL || pasid_state->invalid) {
 587                 /* We know the device but not the PASID -> send INVALID */
 588                 amd_iommu_complete_ppr(dev_state->pdev, iommu_fault->pasid,
 589                                        PPR_INVALID, tag);
 590                 goto out_drop_state;
 591         }
 592
 593         spin_lock_irqsave(&pasid_state->lock, flags);
 594         atomic_inc(&pasid_state->pri[tag].inflight);
 595         if (finish)
 596                 pasid_state->pri[tag].finish = true;
 597         spin_unlock_irqrestore(&pasid_state->lock, flags);
 598
 599         fault = kzalloc(sizeof(*fault), GFP_ATOMIC);
 600         if (fault == NULL) {
 601                 /* We are OOM - send success and let the device re-fault */
 602                 finish_pri_tag(dev_state, pasid_state, tag);
 603                 goto out_drop_state;
 604         }
 605
 606         fault->dev_state = dev_state;
 607         fault->address   = iommu_fault->address;
 608         fault->state     = pasid_state;
 609         fault->tag       = tag;
 610         fault->finish    = finish;
 611         fault->pasid     = iommu_fault->pasid;
 612         fault->flags     = iommu_fault->flags;
 613         INIT_WORK(&fault->work, do_fault);
 614
 615         queue_work(iommu_wq, &fault->work);
 616
 617         ret = NOTIFY_OK;
 618
 619 out_drop_state:
 620
 621         if (ret != NOTIFY_OK && pasid_state)
 622                 put_pasid_state(pasid_state);
 623
 624         put_device_state(dev_state);
 625
 626 out:
 627         return ret;
 628 }
 629
 630 static struct notifier_block ppr_nb = {
 631         .notifier_call = ppr_notifier,
 632 };
 633
 634 int amd_iommu_bind_pasid(struct pci_dev *pdev, int pasid,
 635                          struct task_struct *task)
 636 {
 637         struct pasid_state *pasid_state;
 638         struct device_state *dev_state;
 639         struct mm_struct *mm;
 640         u16 devid;
 641         int ret;
 642
 643         might_sleep();
 644
 645         if (!amd_iommu_v2_supported())
 646                 return -ENODEV;
 647
 648         devid     = device_id(pdev);
 649         dev_state = get_device_state(devid);
 650
 651         if (dev_state == NULL)
 652                 return -EINVAL;
 653
 654         ret = -EINVAL;
 655         if (pasid < 0 || pasid >= dev_state->max_pasids)
 656                 goto out;
 657
 658         ret = -ENOMEM;
 659         pasid_state = kzalloc(sizeof(*pasid_state), GFP_KERNEL);
 660         if (pasid_state == NULL)
 661                 goto out;
 662
 663
 664         atomic_set(&pasid_state->count, 1);
 665         init_waitqueue_head(&pasid_state->wq);
 666         spin_lock_init(&pasid_state->lock);
 667
 668         mm                        = get_task_mm(task);
 669         pasid_state->mm           = mm;
 670         pasid_state->device_state = dev_state;
 671         pasid_state->pasid        = pasid;
 672         pasid_state->invalid      = true; /* Mark as valid only if we are
 673                                              done with setting up the pasid */
 674         pasid_state->mn.ops       = &iommu_mn;
 675
 676         if (pasid_state->mm == NULL)
 677                 goto out_free;
 678
 679         mmu_notifier_register(&pasid_state->mn, mm);
 680
 681         ret = set_pasid_state(dev_state, pasid_state, pasid);
 682         if (ret)
 683                 goto out_unregister;
 684
 685         ret = amd_iommu_domain_set_gcr3(dev_state->domain, pasid,
 686                                         __pa(pasid_state->mm->pgd));
 687         if (ret)
 688                 goto out_clear_state;
 689
 690         /* Now we are ready to handle faults */
 691         pasid_state->invalid = false;
 692
 693         /*
 694          * Drop the reference to the mm_struct here. We rely on the
 695          * mmu_notifier release call-back to inform us when the mm
 696          * is going away.
 697          */
 698         mmput(mm);
 699
 700         return 0;
 701
 702 out_clear_state:
 703         clear_pasid_state(dev_state, pasid);
 704
 705 out_unregister:
 706         mmu_notifier_unregister(&pasid_state->mn, mm);
 707         mmput(mm);
 708
 709 out_free:
 710         free_pasid_state(pasid_state);
 711
 712 out:
 713         put_device_state(dev_state);
 714
 715         return ret;
 716 }
 717 EXPORT_SYMBOL(amd_iommu_bind_pasid);
 718
 719 void amd_iommu_unbind_pasid(struct pci_dev *pdev, int pasid)
 720 {
 721         struct pasid_state *pasid_state;
 722         struct device_state *dev_state;
 723         u16 devid;
 724
 725         might_sleep();
 726
 727         if (!amd_iommu_v2_supported())
 728                 return;
 729
 730         devid = device_id(pdev);
 731         dev_state = get_device_state(devid);
 732         if (dev_state == NULL)
 733                 return;
 734
 735         if (pasid < 0 || pasid >= dev_state->max_pasids)
 736                 goto out;
 737
 738         pasid_state = get_pasid_state(dev_state, pasid);
 739         if (pasid_state == NULL)
 740                 goto out;
 741         /*
 742          * Drop reference taken here. We are safe because we still hold
 743          * the reference taken in the amd_iommu_bind_pasid function.
 744          */
 745         put_pasid_state(pasid_state);
 746
 747         /* Clear the pasid state so that the pasid can be re-used */
 748         clear_pasid_state(dev_state, pasid_state->pasid);
 749
 750         /*
 751          * Call mmu_notifier_unregister to drop our reference
 752          * to pasid_state->mm
 753          */
 754         mmu_notifier_unregister(&pasid_state->mn, pasid_state->mm);
 755
 756         put_pasid_state_wait(pasid_state); /* Reference taken in
 757                                               amd_iommu_bind_pasid */
 758 out:
 759         /* Drop reference taken in this function */
 760         put_device_state(dev_state);
 761
 762         /* Drop reference taken in amd_iommu_bind_pasid */
 763         put_device_state(dev_state);
 764 }
 765 EXPORT_SYMBOL(amd_iommu_unbind_pasid);
 766
 767 int amd_iommu_init_device(struct pci_dev *pdev, int pasids)
 768 {
 769         struct device_state *dev_state;
 770         struct iommu_group *group;
 771         unsigned long flags;
 772         int ret, tmp;
 773         u16 devid;
 774
 775         might_sleep();
 776
 777         if (!amd_iommu_v2_supported())
 778                 return -ENODEV;
 779
 780         if (pasids <= 0 || pasids > (PASID_MASK + 1))
 781                 return -EINVAL;
 782
 783         devid = device_id(pdev);
 784
 785         dev_state = kzalloc(sizeof(*dev_state), GFP_KERNEL);
 786         if (dev_state == NULL)
 787                 return -ENOMEM;
 788
 789         spin_lock_init(&dev_state->lock);
 790         init_waitqueue_head(&dev_state->wq);
 791         dev_state->pdev  = pdev;
 792         dev_state->devid = devid;
 793
 794         tmp = pasids;
 795         for (dev_state->pasid_levels = 0; (tmp - 1) & ~0x1ff; tmp >>= 9)
 796                 dev_state->pasid_levels += 1;
 797
 798         atomic_set(&dev_state->count, 1);
 799         dev_state->max_pasids = pasids;
 800
 801         ret = -ENOMEM;
 802         dev_state->states = (void *)get_zeroed_page(GFP_KERNEL);
 803         if (dev_state->states == NULL)
 804                 goto out_free_dev_state;
 805
 806         dev_state->domain = iommu_domain_alloc(&pci_bus_type);
 807         if (dev_state->domain == NULL)
 808                 goto out_free_states;
 809
 810         amd_iommu_domain_direct_map(dev_state->domain);
 811
 812         ret = amd_iommu_domain_enable_v2(dev_state->domain, pasids);
 813         if (ret)
 814                 goto out_free_domain;
 815
 816         group = iommu_group_get(&pdev->dev);
 817         if (!group) {
 818                 ret = -EINVAL;
 819                 goto out_free_domain;
 820         }
 821
 822         ret = iommu_attach_group(dev_state->domain, group);
 823         if (ret != 0)
 824                 goto out_drop_group;
 825
 826         iommu_group_put(group);
 827
 828         spin_lock_irqsave(&state_lock, flags);
 829
 830         if (__get_device_state(devid) != NULL) {
 831                 spin_unlock_irqrestore(&state_lock, flags);
 832                 ret = -EBUSY;
 833                 goto out_free_domain;
 834         }
 835
 836         list_add_tail(&dev_state->list, &state_list);
 837
 838         spin_unlock_irqrestore(&state_lock, flags);
 839
 840         return 0;
 841
 842 out_drop_group:
 843         iommu_group_put(group);
 844
 845 out_free_domain:
 846         iommu_domain_free(dev_state->domain);
 847
 848 out_free_states:
 849         free_page((unsigned long)dev_state->states);
 850
 851 out_free_dev_state:
 852         kfree(dev_state);
 853
 854         return ret;
 855 }
 856 EXPORT_SYMBOL(amd_iommu_init_device);
 857
 858 void amd_iommu_free_device(struct pci_dev *pdev)
 859 {
 860         struct device_state *dev_state;
 861         unsigned long flags;
 862         u16 devid;
 863
 864         if (!amd_iommu_v2_supported())
 865                 return;
 866
 867         devid = device_id(pdev);
 868
 869         spin_lock_irqsave(&state_lock, flags);
 870
 871         dev_state = __get_device_state(devid);
 872         if (dev_state == NULL) {
 873                 spin_unlock_irqrestore(&state_lock, flags);
 874                 return;
 875         }
 876
 877         list_del(&dev_state->list);
 878
 879         spin_unlock_irqrestore(&state_lock, flags);
 880
 881         /* Get rid of any remaining pasid states */
 882         free_pasid_states(dev_state);
 883
 884         put_device_state(dev_state);
 885         /*
 886          * Wait until the last reference is dropped before freeing
 887          * the device state.
 888          */
 889         wait_event(dev_state->wq, !atomic_read(&dev_state->count));
 890         free_device_state(dev_state);
 891 }
 892 EXPORT_SYMBOL(amd_iommu_free_device);
 893
 894 int amd_iommu_set_invalid_ppr_cb(struct pci_dev *pdev,
 895                                  amd_iommu_invalid_ppr_cb cb)
 896 {
 897         struct device_state *dev_state;
 898         unsigned long flags;
 899         u16 devid;
 900         int ret;
 901
 902         if (!amd_iommu_v2_supported())
 903                 return -ENODEV;
 904
 905         devid = device_id(pdev);
 906
 907         spin_lock_irqsave(&state_lock, flags);
 908
 909         ret = -EINVAL;
 910         dev_state = __get_device_state(devid);
 911         if (dev_state == NULL)
 912                 goto out_unlock;
 913
 914         dev_state->inv_ppr_cb = cb;
 915
 916         ret = 0;
 917
 918 out_unlock:
 919         spin_unlock_irqrestore(&state_lock, flags);
 920
 921         return ret;
 922 }
 923 EXPORT_SYMBOL(amd_iommu_set_invalid_ppr_cb);
 924
 925 int amd_iommu_set_invalidate_ctx_cb(struct pci_dev *pdev,
 926                                     amd_iommu_invalidate_ctx cb)
 927 {
 928         struct device_state *dev_state;
 929         unsigned long flags;
 930         u16 devid;
 931         int ret;
 932
 933         if (!amd_iommu_v2_supported())
 934                 return -ENODEV;
 935
 936         devid = device_id(pdev);
 937
 938         spin_lock_irqsave(&state_lock, flags);
 939
 940         ret = -EINVAL;
 941         dev_state = __get_device_state(devid);
 942         if (dev_state == NULL)
 943                 goto out_unlock;
 944
 945         dev_state->inv_ctx_cb = cb;
 946
 947         ret = 0;
 948
 949 out_unlock:
 950         spin_unlock_irqrestore(&state_lock, flags);
 951
 952         return ret;
 953 }
 954 EXPORT_SYMBOL(amd_iommu_set_invalidate_ctx_cb);
 955
 956 static int __init amd_iommu_v2_init(void)
 957 {
 958         int ret;
 959
 960         pr_info("AMD IOMMUv2 driver by Joerg Roedel <jroedel@suse.de>\n");
 961
 962         if (!amd_iommu_v2_supported()) {
 963                 pr_info("AMD IOMMUv2 functionality not available on this system\n");
 964                 /*
 965                  * Load anyway to provide the symbols to other modules
 966                  * which may use AMD IOMMUv2 optionally.
 967                  */
 968                 return 0;
 969         }
 970
 971         spin_lock_init(&state_lock);
 972
 973         ret = -ENOMEM;
 974         iommu_wq = alloc_workqueue("amd_iommu_v2", WQ_MEM_RECLAIM, 0);
 975         if (iommu_wq == NULL)
 976                 goto out;
 977
 978         amd_iommu_register_ppr_notifier(&ppr_nb);
 979
 980         return 0;
 981
 982 out:
 983         return ret;
 984 }
 985
 986 static void __exit amd_iommu_v2_exit(void)
 987 {
 988         struct device_state *dev_state;
 989         int i;
 990
 991         if (!amd_iommu_v2_supported())
 992                 return;
 993
 994         amd_iommu_unregister_ppr_notifier(&ppr_nb);
 995
 996         flush_workqueue(iommu_wq);
 997
 998         /*
 999          * The loop below might call flush_workqueue(), so call
1000          * destroy_workqueue() after it
1001          */
1002         for (i = 0; i < MAX_DEVICES; ++i) {
1003                 dev_state = get_device_state(i);
1004
1005                 if (dev_state == NULL)
1006                         continue;
1007
1008                 WARN_ON_ONCE(1);
1009
1010                 put_device_state(dev_state);
1011                 amd_iommu_free_device(dev_state->pdev);
1012         }
1013
1014         destroy_workqueue(iommu_wq);
1015 }
1016
1017 module_init(amd_iommu_v2_init);
1018 module_exit(amd_iommu_v2_exit);