drivers/iommu/amd_iommu_v2.c

   1 /*
   2  * Copyright (C) 2010-2012 Advanced Micro Devices, Inc.
   3  * Author: Joerg Roedel <joerg.roedel@amd.com>
   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/iommu.h>
  26 #include <linux/wait.h>
  27 #include <linux/pci.h>
  28 #include <linux/gfp.h>
  29
  30 #include "amd_iommu_types.h"
  31 #include "amd_iommu_proto.h"
  32
  33 MODULE_LICENSE("GPL v2");
  34 MODULE_AUTHOR("Joerg Roedel <joerg.roedel@amd.com>");
  35
  36 #define MAX_DEVICES             0x10000
  37 #define PRI_QUEUE_SIZE          512
  38
  39 struct pri_queue {
  40         atomic_t inflight;
  41         bool finish;
  42         int status;
  43 };
  44
  45 struct pasid_state {
  46         struct list_head list;                  /* For global state-list */
  47         atomic_t count;                         /* Reference count */
  48         struct task_struct *task;               /* Task bound to this PASID */
  49         struct mm_struct *mm;                   /* mm_struct for the faults */
  50         struct mmu_notifier mn;                 /* mmu_otifier handle */
  51         struct pri_queue pri[PRI_QUEUE_SIZE];   /* PRI tag states */
  52         struct device_state *device_state;      /* Link to our device_state */
  53         int pasid;                              /* PASID index */
  54         spinlock_t lock;                        /* Protect pri_queues */
  55         wait_queue_head_t wq;                   /* To wait for count == 0 */
  56 };
  57
  58 struct device_state {
  59         atomic_t count;
  60         struct pci_dev *pdev;
  61         struct pasid_state **states;
  62         struct iommu_domain *domain;
  63         int pasid_levels;
  64         int max_pasids;
  65         amd_iommu_invalid_ppr_cb inv_ppr_cb;
  66         amd_iommu_invalidate_ctx inv_ctx_cb;
  67         spinlock_t lock;
  68         wait_queue_head_t wq;
  69 };
  70
  71 struct fault {
  72         struct work_struct work;
  73         struct device_state *dev_state;
  74         struct pasid_state *state;
  75         struct mm_struct *mm;
  76         u64 address;
  77         u16 devid;
  78         u16 pasid;
  79         u16 tag;
  80         u16 finish;
  81         u16 flags;
  82 };
  83
  84 struct device_state **state_table;
  85 static spinlock_t state_lock;
  86
  87 /* List and lock for all pasid_states */
  88 static LIST_HEAD(pasid_state_list);
  89 static DEFINE_SPINLOCK(ps_lock);
  90
  91 static struct workqueue_struct *iommu_wq;
  92
  93 /*
  94  * Empty page table - Used between
  95  * mmu_notifier_invalidate_range_start and
  96  * mmu_notifier_invalidate_range_end
  97  */
  98 static u64 *empty_page_table;
  99
 100 static void free_pasid_states(struct device_state *dev_state);
 101 static void unbind_pasid(struct device_state *dev_state, int pasid);
 102 static int task_exit(struct notifier_block *nb, unsigned long e, void *data);
 103
 104 static u16 device_id(struct pci_dev *pdev)
 105 {
 106         u16 devid;
 107
 108         devid = pdev->bus->number;
 109         devid = (devid << 8) | pdev->devfn;
 110
 111         return devid;
 112 }
 113
 114 static struct device_state *get_device_state(u16 devid)
 115 {
 116         struct device_state *dev_state;
 117         unsigned long flags;
 118
 119         spin_lock_irqsave(&state_lock, flags);
 120         dev_state = state_table[devid];
 121         if (dev_state != NULL)
 122                 atomic_inc(&dev_state->count);
 123         spin_unlock_irqrestore(&state_lock, flags);
 124
 125         return dev_state;
 126 }
 127
 128 static void free_device_state(struct device_state *dev_state)
 129 {
 130         /*
 131          * First detach device from domain - No more PRI requests will arrive
 132          * from that device after it is unbound from the IOMMUv2 domain.
 133          */
 134         iommu_detach_device(dev_state->domain, &dev_state->pdev->dev);
 135
 136         /* Everything is down now, free the IOMMUv2 domain */
 137         iommu_domain_free(dev_state->domain);
 138
 139         /* Finally get rid of the device-state */
 140         kfree(dev_state);
 141 }
 142
 143 static void put_device_state(struct device_state *dev_state)
 144 {
 145         if (atomic_dec_and_test(&dev_state->count))
 146                 wake_up(&dev_state->wq);
 147 }
 148
 149 static void put_device_state_wait(struct device_state *dev_state)
 150 {
 151         DEFINE_WAIT(wait);
 152
 153         prepare_to_wait(&dev_state->wq, &wait, TASK_UNINTERRUPTIBLE);
 154         if (!atomic_dec_and_test(&dev_state->count))
 155                 schedule();
 156         finish_wait(&dev_state->wq, &wait);
 157
 158         free_device_state(dev_state);
 159 }
 160
 161 static struct notifier_block profile_nb = {
 162         .notifier_call = task_exit,
 163 };
 164
 165 static void link_pasid_state(struct pasid_state *pasid_state)
 166 {
 167         spin_lock(&ps_lock);
 168         list_add_tail(&pasid_state->list, &pasid_state_list);
 169         spin_unlock(&ps_lock);
 170 }
 171
 172 static void __unlink_pasid_state(struct pasid_state *pasid_state)
 173 {
 174         list_del(&pasid_state->list);
 175 }
 176
 177 static void unlink_pasid_state(struct pasid_state *pasid_state)
 178 {
 179         spin_lock(&ps_lock);
 180         __unlink_pasid_state(pasid_state);
 181         spin_unlock(&ps_lock);
 182 }
 183
 184 /* Must be called under dev_state->lock */
 185 static struct pasid_state **__get_pasid_state_ptr(struct device_state *dev_state,
 186                                                   int pasid, bool alloc)
 187 {
 188         struct pasid_state **root, **ptr;
 189         int level, index;
 190
 191         level = dev_state->pasid_levels;
 192         root  = dev_state->states;
 193
 194         while (true) {
 195
 196                 index = (pasid >> (9 * level)) & 0x1ff;
 197                 ptr   = &root[index];
 198
 199                 if (level == 0)
 200                         break;
 201
 202                 if (*ptr == NULL) {
 203                         if (!alloc)
 204                                 return NULL;
 205
 206                         *ptr = (void *)get_zeroed_page(GFP_ATOMIC);
 207                         if (*ptr == NULL)
 208                                 return NULL;
 209                 }
 210
 211                 root   = (struct pasid_state **)*ptr;
 212                 level -= 1;
 213         }
 214
 215         return ptr;
 216 }
 217
 218 static int set_pasid_state(struct device_state *dev_state,
 219                            struct pasid_state *pasid_state,
 220                            int pasid)
 221 {
 222         struct pasid_state **ptr;
 223         unsigned long flags;
 224         int ret;
 225
 226         spin_lock_irqsave(&dev_state->lock, flags);
 227         ptr = __get_pasid_state_ptr(dev_state, pasid, true);
 228
 229         ret = -ENOMEM;
 230         if (ptr == NULL)
 231                 goto out_unlock;
 232
 233         ret = -ENOMEM;
 234         if (*ptr != NULL)
 235                 goto out_unlock;
 236
 237         *ptr = pasid_state;
 238
 239         ret = 0;
 240
 241 out_unlock:
 242         spin_unlock_irqrestore(&dev_state->lock, flags);
 243
 244         return ret;
 245 }
 246
 247 static void clear_pasid_state(struct device_state *dev_state, int pasid)
 248 {
 249         struct pasid_state **ptr;
 250         unsigned long flags;
 251
 252         spin_lock_irqsave(&dev_state->lock, flags);
 253         ptr = __get_pasid_state_ptr(dev_state, pasid, true);
 254
 255         if (ptr == NULL)
 256                 goto out_unlock;
 257
 258         *ptr = NULL;
 259
 260 out_unlock:
 261         spin_unlock_irqrestore(&dev_state->lock, flags);
 262 }
 263
 264 static struct pasid_state *get_pasid_state(struct device_state *dev_state,
 265                                            int pasid)
 266 {
 267         struct pasid_state **ptr, *ret = NULL;
 268         unsigned long flags;
 269
 270         spin_lock_irqsave(&dev_state->lock, flags);
 271         ptr = __get_pasid_state_ptr(dev_state, pasid, false);
 272
 273         if (ptr == NULL)
 274                 goto out_unlock;
 275
 276         ret = *ptr;
 277         if (ret)
 278                 atomic_inc(&ret->count);
 279
 280 out_unlock:
 281         spin_unlock_irqrestore(&dev_state->lock, flags);
 282
 283         return ret;
 284 }
 285
 286 static void free_pasid_state(struct pasid_state *pasid_state)
 287 {
 288         kfree(pasid_state);
 289 }
 290
 291 static void put_pasid_state(struct pasid_state *pasid_state)
 292 {
 293         if (atomic_dec_and_test(&pasid_state->count)) {
 294                 put_device_state(pasid_state->device_state);
 295                 wake_up(&pasid_state->wq);
 296         }
 297 }
 298
 299 static void put_pasid_state_wait(struct pasid_state *pasid_state)
 300 {
 301         DEFINE_WAIT(wait);
 302
 303         prepare_to_wait(&pasid_state->wq, &wait, TASK_UNINTERRUPTIBLE);
 304
 305         if (atomic_dec_and_test(&pasid_state->count))
 306                 put_device_state(pasid_state->device_state);
 307         else
 308                 schedule();
 309
 310         finish_wait(&pasid_state->wq, &wait);
 311         mmput(pasid_state->mm);
 312         free_pasid_state(pasid_state);
 313 }
 314
 315 static void __unbind_pasid(struct pasid_state *pasid_state)
 316 {
 317         struct iommu_domain *domain;
 318
 319         domain = pasid_state->device_state->domain;
 320
 321         amd_iommu_domain_clear_gcr3(domain, pasid_state->pasid);
 322         clear_pasid_state(pasid_state->device_state, pasid_state->pasid);
 323
 324         /* Make sure no more pending faults are in the queue */
 325         flush_workqueue(iommu_wq);
 326
 327         mmu_notifier_unregister(&pasid_state->mn, pasid_state->mm);
 328
 329         put_pasid_state(pasid_state); /* Reference taken in bind() function */
 330 }
 331
 332 static void unbind_pasid(struct device_state *dev_state, int pasid)
 333 {
 334         struct pasid_state *pasid_state;
 335
 336         pasid_state = get_pasid_state(dev_state, pasid);
 337         if (pasid_state == NULL)
 338                 return;
 339
 340         unlink_pasid_state(pasid_state);
 341         __unbind_pasid(pasid_state);
 342         put_pasid_state_wait(pasid_state); /* Reference taken in this function */
 343 }
 344
 345 static void free_pasid_states_level1(struct pasid_state **tbl)
 346 {
 347         int i;
 348
 349         for (i = 0; i < 512; ++i) {
 350                 if (tbl[i] == NULL)
 351                         continue;
 352
 353                 free_page((unsigned long)tbl[i]);
 354         }
 355 }
 356
 357 static void free_pasid_states_level2(struct pasid_state **tbl)
 358 {
 359         struct pasid_state **ptr;
 360         int i;
 361
 362         for (i = 0; i < 512; ++i) {
 363                 if (tbl[i] == NULL)
 364                         continue;
 365
 366                 ptr = (struct pasid_state **)tbl[i];
 367                 free_pasid_states_level1(ptr);
 368         }
 369 }
 370
 371 static void free_pasid_states(struct device_state *dev_state)
 372 {
 373         struct pasid_state *pasid_state;
 374         int i;
 375
 376         for (i = 0; i < dev_state->max_pasids; ++i) {
 377                 pasid_state = get_pasid_state(dev_state, i);
 378                 if (pasid_state == NULL)
 379                         continue;
 380
 381                 put_pasid_state(pasid_state);
 382                 unbind_pasid(dev_state, i);
 383         }
 384
 385         if (dev_state->pasid_levels == 2)
 386                 free_pasid_states_level2(dev_state->states);
 387         else if (dev_state->pasid_levels == 1)
 388                 free_pasid_states_level1(dev_state->states);
 389         else if (dev_state->pasid_levels != 0)
 390                 BUG();
 391
 392         free_page((unsigned long)dev_state->states);
 393 }
 394
 395 static struct pasid_state *mn_to_state(struct mmu_notifier *mn)
 396 {
 397         return container_of(mn, struct pasid_state, mn);
 398 }
 399
 400 static void __mn_flush_page(struct mmu_notifier *mn,
 401                             unsigned long address)
 402 {
 403         struct pasid_state *pasid_state;
 404         struct device_state *dev_state;
 405
 406         pasid_state = mn_to_state(mn);
 407         dev_state   = pasid_state->device_state;
 408
 409         amd_iommu_flush_page(dev_state->domain, pasid_state->pasid, address);
 410 }
 411
 412 static int mn_clear_flush_young(struct mmu_notifier *mn,
 413                                 struct mm_struct *mm,
 414                                 unsigned long address)
 415 {
 416         __mn_flush_page(mn, address);
 417
 418         return 0;
 419 }
 420
 421 static void mn_change_pte(struct mmu_notifier *mn,
 422                           struct mm_struct *mm,
 423                           unsigned long address,
 424                           pte_t pte)
 425 {
 426         __mn_flush_page(mn, address);
 427 }
 428
 429 static void mn_invalidate_page(struct mmu_notifier *mn,
 430                                struct mm_struct *mm,
 431                                unsigned long address)
 432 {
 433         __mn_flush_page(mn, address);
 434 }
 435
 436 static void mn_invalidate_range_start(struct mmu_notifier *mn,
 437                                       struct mm_struct *mm,
 438                                       unsigned long start, unsigned long end)
 439 {
 440         struct pasid_state *pasid_state;
 441         struct device_state *dev_state;
 442
 443         pasid_state = mn_to_state(mn);
 444         dev_state   = pasid_state->device_state;
 445
 446         amd_iommu_domain_set_gcr3(dev_state->domain, pasid_state->pasid,
 447                                   __pa(empty_page_table));
 448 }
 449
 450 static void mn_invalidate_range_end(struct mmu_notifier *mn,
 451                                     struct mm_struct *mm,
 452                                     unsigned long start, unsigned long end)
 453 {
 454         struct pasid_state *pasid_state;
 455         struct device_state *dev_state;
 456
 457         pasid_state = mn_to_state(mn);
 458         dev_state   = pasid_state->device_state;
 459
 460         amd_iommu_domain_set_gcr3(dev_state->domain, pasid_state->pasid,
 461                                   __pa(pasid_state->mm->pgd));
 462 }
 463
 464 static struct mmu_notifier_ops iommu_mn = {
 465         .clear_flush_young      = mn_clear_flush_young,
 466         .change_pte             = mn_change_pte,
 467         .invalidate_page        = mn_invalidate_page,
 468         .invalidate_range_start = mn_invalidate_range_start,
 469         .invalidate_range_end   = mn_invalidate_range_end,
 470 };
 471
 472 static void set_pri_tag_status(struct pasid_state *pasid_state,
 473                                u16 tag, int status)
 474 {
 475         unsigned long flags;
 476
 477         spin_lock_irqsave(&pasid_state->lock, flags);
 478         pasid_state->pri[tag].status = status;
 479         spin_unlock_irqrestore(&pasid_state->lock, flags);
 480 }
 481
 482 static void finish_pri_tag(struct device_state *dev_state,
 483                            struct pasid_state *pasid_state,
 484                            u16 tag)
 485 {
 486         unsigned long flags;
 487
 488         spin_lock_irqsave(&pasid_state->lock, flags);
 489         if (atomic_dec_and_test(&pasid_state->pri[tag].inflight) &&
 490             pasid_state->pri[tag].finish) {
 491                 amd_iommu_complete_ppr(dev_state->pdev, pasid_state->pasid,
 492                                        pasid_state->pri[tag].status, tag);
 493                 pasid_state->pri[tag].finish = false;
 494                 pasid_state->pri[tag].status = PPR_SUCCESS;
 495         }
 496         spin_unlock_irqrestore(&pasid_state->lock, flags);
 497 }
 498
 499 static void do_fault(struct work_struct *work)
 500 {
 501         struct fault *fault = container_of(work, struct fault, work);
 502         int npages, write;
 503         struct page *page;
 504
 505         write = !!(fault->flags & PPR_FAULT_WRITE);
 506
 507         npages = get_user_pages(fault->state->task, fault->state->mm,
 508                                 fault->address, 1, write, 0, &page, NULL);
 509
 510         if (npages == 1) {
 511                 put_page(page);
 512         } else if (fault->dev_state->inv_ppr_cb) {
 513                 int status;
 514
 515                 status = fault->dev_state->inv_ppr_cb(fault->dev_state->pdev,
 516                                                       fault->pasid,
 517                                                       fault->address,
 518                                                       fault->flags);
 519                 switch (status) {
 520                 case AMD_IOMMU_INV_PRI_RSP_SUCCESS:
 521                         set_pri_tag_status(fault->state, fault->tag, PPR_SUCCESS);
 522                         break;
 523                 case AMD_IOMMU_INV_PRI_RSP_INVALID:
 524                         set_pri_tag_status(fault->state, fault->tag, PPR_INVALID);
 525                         break;
 526                 case AMD_IOMMU_INV_PRI_RSP_FAIL:
 527                         set_pri_tag_status(fault->state, fault->tag, PPR_FAILURE);
 528                         break;
 529                 default:
 530                         BUG();
 531                 }
 532         } else {
 533                 set_pri_tag_status(fault->state, fault->tag, PPR_INVALID);
 534         }
 535
 536         finish_pri_tag(fault->dev_state, fault->state, fault->tag);
 537
 538         put_pasid_state(fault->state);
 539
 540         kfree(fault);
 541 }
 542
 543 static int ppr_notifier(struct notifier_block *nb, unsigned long e, void *data)
 544 {
 545         struct amd_iommu_fault *iommu_fault;
 546         struct pasid_state *pasid_state;
 547         struct device_state *dev_state;
 548         unsigned long flags;
 549         struct fault *fault;
 550         bool finish;
 551         u16 tag;
 552         int ret;
 553
 554         iommu_fault = data;
 555         tag         = iommu_fault->tag & 0x1ff;
 556         finish      = (iommu_fault->tag >> 9) & 1;
 557
 558         ret = NOTIFY_DONE;
 559         dev_state = get_device_state(iommu_fault->device_id);
 560         if (dev_state == NULL)
 561                 goto out;
 562
 563         pasid_state = get_pasid_state(dev_state, iommu_fault->pasid);
 564         if (pasid_state == NULL) {
 565                 /* We know the device but not the PASID -> send INVALID */
 566                 amd_iommu_complete_ppr(dev_state->pdev, iommu_fault->pasid,
 567                                        PPR_INVALID, tag);
 568                 goto out_drop_state;
 569         }
 570
 571         spin_lock_irqsave(&pasid_state->lock, flags);
 572         atomic_inc(&pasid_state->pri[tag].inflight);
 573         if (finish)
 574                 pasid_state->pri[tag].finish = true;
 575         spin_unlock_irqrestore(&pasid_state->lock, flags);
 576
 577         fault = kzalloc(sizeof(*fault), GFP_ATOMIC);
 578         if (fault == NULL) {
 579                 /* We are OOM - send success and let the device re-fault */
 580                 finish_pri_tag(dev_state, pasid_state, tag);
 581                 goto out_drop_state;
 582         }
 583
 584         fault->dev_state = dev_state;
 585         fault->address   = iommu_fault->address;
 586         fault->state     = pasid_state;
 587         fault->tag       = tag;
 588         fault->finish    = finish;
 589         fault->flags     = iommu_fault->flags;
 590         INIT_WORK(&fault->work, do_fault);
 591
 592         queue_work(iommu_wq, &fault->work);
 593
 594         ret = NOTIFY_OK;
 595
 596 out_drop_state:
 597         put_device_state(dev_state);
 598
 599 out:
 600         return ret;
 601 }
 602
 603 static struct notifier_block ppr_nb = {
 604         .notifier_call = ppr_notifier,
 605 };
 606
 607 static int task_exit(struct notifier_block *nb, unsigned long e, void *data)
 608 {
 609         struct pasid_state *pasid_state;
 610         struct task_struct *task;
 611
 612         task = data;
 613
 614         /*
 615          * Using this notifier is a hack - but there is no other choice
 616          * at the moment. What I really want is a sleeping notifier that
 617          * is called when an MM goes down. But such a notifier doesn't
 618          * exist yet. The notifier needs to sleep because it has to make
 619          * sure that the device does not use the PASID and the address
 620          * space anymore before it is destroyed. This includes waiting
 621          * for pending PRI requests to pass the workqueue. The
 622          * MMU-Notifiers would be a good fit, but they use RCU and so
 623          * they are not allowed to sleep. Lets see how we can solve this
 624          * in a more intelligent way in the future.
 625          */
 626 again:
 627         spin_lock(&ps_lock);
 628         list_for_each_entry(pasid_state, &pasid_state_list, list) {
 629                 struct device_state *dev_state;
 630                 int pasid;
 631
 632                 if (pasid_state->task != task)
 633                         continue;
 634
 635                 /* Drop Lock and unbind */
 636                 spin_unlock(&ps_lock);
 637
 638                 dev_state = pasid_state->device_state;
 639                 pasid     = pasid_state->pasid;
 640
 641                 if (pasid_state->device_state->inv_ctx_cb)
 642                         dev_state->inv_ctx_cb(dev_state->pdev, pasid);
 643
 644                 unbind_pasid(dev_state, pasid);
 645
 646                 /* Task may be in the list multiple times */
 647                 goto again;
 648         }
 649         spin_unlock(&ps_lock);
 650
 651         return NOTIFY_OK;
 652 }
 653
 654 int amd_iommu_bind_pasid(struct pci_dev *pdev, int pasid,
 655                          struct task_struct *task)
 656 {
 657         struct pasid_state *pasid_state;
 658         struct device_state *dev_state;
 659         u16 devid;
 660         int ret;
 661
 662         might_sleep();
 663
 664         if (!amd_iommu_v2_supported())
 665                 return -ENODEV;
 666
 667         devid     = device_id(pdev);
 668         dev_state = get_device_state(devid);
 669
 670         if (dev_state == NULL)
 671                 return -EINVAL;
 672
 673         ret = -EINVAL;
 674         if (pasid < 0 || pasid >= dev_state->max_pasids)
 675                 goto out;
 676
 677         ret = -ENOMEM;
 678         pasid_state = kzalloc(sizeof(*pasid_state), GFP_KERNEL);
 679         if (pasid_state == NULL)
 680                 goto out;
 681
 682         atomic_set(&pasid_state->count, 1);
 683         init_waitqueue_head(&pasid_state->wq);
 684         pasid_state->task         = task;
 685         pasid_state->mm           = get_task_mm(task);
 686         pasid_state->device_state = dev_state;
 687         pasid_state->pasid        = pasid;
 688         pasid_state->mn.ops       = &iommu_mn;
 689
 690         if (pasid_state->mm == NULL)
 691                 goto out_free;
 692
 693         mmu_notifier_register(&pasid_state->mn, pasid_state->mm);
 694
 695         ret = set_pasid_state(dev_state, pasid_state, pasid);
 696         if (ret)
 697                 goto out_unregister;
 698
 699         ret = amd_iommu_domain_set_gcr3(dev_state->domain, pasid,
 700                                         __pa(pasid_state->mm->pgd));
 701         if (ret)
 702                 goto out_clear_state;
 703
 704         link_pasid_state(pasid_state);
 705
 706         return 0;
 707
 708 out_clear_state:
 709         clear_pasid_state(dev_state, pasid);
 710
 711 out_unregister:
 712         mmu_notifier_unregister(&pasid_state->mn, pasid_state->mm);
 713
 714 out_free:
 715         free_pasid_state(pasid_state);
 716
 717 out:
 718         put_device_state(dev_state);
 719
 720         return ret;
 721 }
 722 EXPORT_SYMBOL(amd_iommu_bind_pasid);
 723
 724 void amd_iommu_unbind_pasid(struct pci_dev *pdev, int pasid)
 725 {
 726         struct device_state *dev_state;
 727         u16 devid;
 728
 729         might_sleep();
 730
 731         if (!amd_iommu_v2_supported())
 732                 return;
 733
 734         devid = device_id(pdev);
 735         dev_state = get_device_state(devid);
 736         if (dev_state == NULL)
 737                 return;
 738
 739         if (pasid < 0 || pasid >= dev_state->max_pasids)
 740                 goto out;
 741
 742         unbind_pasid(dev_state, pasid);
 743
 744 out:
 745         put_device_state(dev_state);
 746 }
 747 EXPORT_SYMBOL(amd_iommu_unbind_pasid);
 748
 749 int amd_iommu_init_device(struct pci_dev *pdev, int pasids)
 750 {
 751         struct device_state *dev_state;
 752         unsigned long flags;
 753         int ret, tmp;
 754         u16 devid;
 755
 756         might_sleep();
 757
 758         if (!amd_iommu_v2_supported())
 759                 return -ENODEV;
 760
 761         if (pasids <= 0 || pasids > (PASID_MASK + 1))
 762                 return -EINVAL;
 763
 764         devid = device_id(pdev);
 765
 766         dev_state = kzalloc(sizeof(*dev_state), GFP_KERNEL);
 767         if (dev_state == NULL)
 768                 return -ENOMEM;
 769
 770         spin_lock_init(&dev_state->lock);
 771         init_waitqueue_head(&dev_state->wq);
 772         dev_state->pdev = pdev;
 773
 774         tmp = pasids;
 775         for (dev_state->pasid_levels = 0; (tmp - 1) & ~0x1ff; tmp >>= 9)
 776                 dev_state->pasid_levels += 1;
 777
 778         atomic_set(&dev_state->count, 1);
 779         dev_state->max_pasids = pasids;
 780
 781         ret = -ENOMEM;
 782         dev_state->states = (void *)get_zeroed_page(GFP_KERNEL);
 783         if (dev_state->states == NULL)
 784                 goto out_free_dev_state;
 785
 786         dev_state->domain = iommu_domain_alloc(&pci_bus_type);
 787         if (dev_state->domain == NULL)
 788                 goto out_free_states;
 789
 790         amd_iommu_domain_direct_map(dev_state->domain);
 791
 792         ret = amd_iommu_domain_enable_v2(dev_state->domain, pasids);
 793         if (ret)
 794                 goto out_free_domain;
 795
 796         ret = iommu_attach_device(dev_state->domain, &pdev->dev);
 797         if (ret != 0)
 798                 goto out_free_domain;
 799
 800         spin_lock_irqsave(&state_lock, flags);
 801
 802         if (state_table[devid] != NULL) {
 803                 spin_unlock_irqrestore(&state_lock, flags);
 804                 ret = -EBUSY;
 805                 goto out_free_domain;
 806         }
 807
 808         state_table[devid] = dev_state;
 809
 810         spin_unlock_irqrestore(&state_lock, flags);
 811
 812         return 0;
 813
 814 out_free_domain:
 815         iommu_domain_free(dev_state->domain);
 816
 817 out_free_states:
 818         free_page((unsigned long)dev_state->states);
 819
 820 out_free_dev_state:
 821         kfree(dev_state);
 822
 823         return ret;
 824 }
 825 EXPORT_SYMBOL(amd_iommu_init_device);
 826
 827 void amd_iommu_free_device(struct pci_dev *pdev)
 828 {
 829         struct device_state *dev_state;
 830         unsigned long flags;
 831         u16 devid;
 832
 833         if (!amd_iommu_v2_supported())
 834                 return;
 835
 836         devid = device_id(pdev);
 837
 838         spin_lock_irqsave(&state_lock, flags);
 839
 840         dev_state = state_table[devid];
 841         if (dev_state == NULL) {
 842                 spin_unlock_irqrestore(&state_lock, flags);
 843                 return;
 844         }
 845
 846         state_table[devid] = NULL;
 847
 848         spin_unlock_irqrestore(&state_lock, flags);
 849
 850         /* Get rid of any remaining pasid states */
 851         free_pasid_states(dev_state);
 852
 853         put_device_state_wait(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 = state_table[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 = state_table[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         size_t state_table_size;
 922         int ret;
 923
 924         pr_info("AMD IOMMUv2 driver by Joerg Roedel <joerg.roedel@amd.com>\n");
 925
 926         if (!amd_iommu_v2_supported()) {
 927                 pr_info("AMD IOMMUv2 functionality not available on this sytem\n");
 928                 /*
 929                  * Load anyway to provide the symbols to other modules
 930                  * which may use AMD IOMMUv2 optionally.
 931                  */
 932                 return 0;
 933         }
 934
 935         spin_lock_init(&state_lock);
 936
 937         state_table_size = MAX_DEVICES * sizeof(struct device_state *);
 938         state_table = (void *)__get_free_pages(GFP_KERNEL | __GFP_ZERO,
 939                                                get_order(state_table_size));
 940         if (state_table == NULL)
 941                 return -ENOMEM;
 942
 943         ret = -ENOMEM;
 944         iommu_wq = create_workqueue("amd_iommu_v2");
 945         if (iommu_wq == NULL)
 946                 goto out_free;
 947
 948         ret = -ENOMEM;
 949         empty_page_table = (u64 *)get_zeroed_page(GFP_KERNEL);
 950         if (empty_page_table == NULL)
 951                 goto out_destroy_wq;
 952
 953         amd_iommu_register_ppr_notifier(&ppr_nb);
 954         profile_event_register(PROFILE_TASK_EXIT, &profile_nb);
 955
 956         return 0;
 957
 958 out_destroy_wq:
 959         destroy_workqueue(iommu_wq);
 960
 961 out_free:
 962         free_pages((unsigned long)state_table, get_order(state_table_size));
 963
 964         return ret;
 965 }
 966
 967 static void __exit amd_iommu_v2_exit(void)
 968 {
 969         struct device_state *dev_state;
 970         size_t state_table_size;
 971         int i;
 972
 973         if (!amd_iommu_v2_supported())
 974                 return;
 975
 976         profile_event_unregister(PROFILE_TASK_EXIT, &profile_nb);
 977         amd_iommu_unregister_ppr_notifier(&ppr_nb);
 978
 979         flush_workqueue(iommu_wq);
 980
 981         /*
 982          * The loop below might call flush_workqueue(), so call
 983          * destroy_workqueue() after it
 984          */
 985         for (i = 0; i < MAX_DEVICES; ++i) {
 986                 dev_state = get_device_state(i);
 987
 988                 if (dev_state == NULL)
 989                         continue;
 990
 991                 WARN_ON_ONCE(1);
 992
 993                 put_device_state(dev_state);
 994                 amd_iommu_free_device(dev_state->pdev);
 995         }
 996
 997         destroy_workqueue(iommu_wq);
 998
 999         state_table_size = MAX_DEVICES * sizeof(struct device_state *);
1000         free_pages((unsigned long)state_table, get_order(state_table_size));
1001
1002         free_page((unsigned long)empty_page_table);
1003 }
1004
1005 module_init(amd_iommu_v2_init);
1006 module_exit(amd_iommu_v2_exit);