drivers/iommu/intel-svm.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright © 2015 Intel Corporation.
   4  *
   5  * Authors: David Woodhouse <dwmw2@infradead.org>
   6  */
   7
   8 #include <linux/intel-iommu.h>
   9 #include <linux/mmu_notifier.h>
  10 #include <linux/sched.h>
  11 #include <linux/sched/mm.h>
  12 #include <linux/slab.h>
  13 #include <linux/intel-svm.h>
  14 #include <linux/rculist.h>
  15 #include <linux/pci.h>
  16 #include <linux/pci-ats.h>
  17 #include <linux/dmar.h>
  18 #include <linux/interrupt.h>
  19 #include <linux/mm_types.h>
  20 #include <asm/page.h>
  21
  22 #include "intel-pasid.h"
  23
  24 static irqreturn_t prq_event_thread(int irq, void *d);
  25
  26 int intel_svm_init(struct intel_iommu *iommu)
  27 {
  28         if (cpu_feature_enabled(X86_FEATURE_GBPAGES) &&
  29                         !cap_fl1gp_support(iommu->cap))
  30                 return -EINVAL;
  31
  32         if (cpu_feature_enabled(X86_FEATURE_LA57) &&
  33                         !cap_5lp_support(iommu->cap))
  34                 return -EINVAL;
  35
  36         return 0;
  37 }
  38
  39 #define PRQ_ORDER 0
  40
  41 int intel_svm_enable_prq(struct intel_iommu *iommu)
  42 {
  43         struct page *pages;
  44         int irq, ret;
  45
  46         pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, PRQ_ORDER);
  47         if (!pages) {
  48                 pr_warn("IOMMU: %s: Failed to allocate page request queue\n",
  49                         iommu->name);
  50                 return -ENOMEM;
  51         }
  52         iommu->prq = page_address(pages);
  53
  54         irq = dmar_alloc_hwirq(DMAR_UNITS_SUPPORTED + iommu->seq_id, iommu->node, iommu);
  55         if (irq <= 0) {
  56                 pr_err("IOMMU: %s: Failed to create IRQ vector for page request queue\n",
  57                        iommu->name);
  58                 ret = -EINVAL;
  59         err:
  60                 free_pages((unsigned long)iommu->prq, PRQ_ORDER);
  61                 iommu->prq = NULL;
  62                 return ret;
  63         }
  64         iommu->pr_irq = irq;
  65
  66         snprintf(iommu->prq_name, sizeof(iommu->prq_name), "dmar%d-prq", iommu->seq_id);
  67
  68         ret = request_threaded_irq(irq, NULL, prq_event_thread, IRQF_ONESHOT,
  69                                    iommu->prq_name, iommu);
  70         if (ret) {
  71                 pr_err("IOMMU: %s: Failed to request IRQ for page request queue\n",
  72                        iommu->name);
  73                 dmar_free_hwirq(irq);
  74                 iommu->pr_irq = 0;
  75                 goto err;
  76         }
  77         dmar_writeq(iommu->reg + DMAR_PQH_REG, 0ULL);
  78         dmar_writeq(iommu->reg + DMAR_PQT_REG, 0ULL);
  79         dmar_writeq(iommu->reg + DMAR_PQA_REG, virt_to_phys(iommu->prq) | PRQ_ORDER);
  80
  81         return 0;
  82 }
  83
  84 int intel_svm_finish_prq(struct intel_iommu *iommu)
  85 {
  86         dmar_writeq(iommu->reg + DMAR_PQH_REG, 0ULL);
  87         dmar_writeq(iommu->reg + DMAR_PQT_REG, 0ULL);
  88         dmar_writeq(iommu->reg + DMAR_PQA_REG, 0ULL);
  89
  90         if (iommu->pr_irq) {
  91                 free_irq(iommu->pr_irq, iommu);
  92                 dmar_free_hwirq(iommu->pr_irq);
  93                 iommu->pr_irq = 0;
  94         }
  95
  96         free_pages((unsigned long)iommu->prq, PRQ_ORDER);
  97         iommu->prq = NULL;
  98
  99         return 0;
 100 }
 101
 102 static void intel_flush_svm_range_dev (struct intel_svm *svm, struct intel_svm_dev *sdev,
 103                                        unsigned long address, unsigned long pages, int ih, int gl)
 104 {
 105         struct qi_desc desc;
 106
 107         if (pages == -1) {
 108                 /* For global kernel pages we have to flush them in *all* PASIDs
 109                  * because that's the only option the hardware gives us. Despite
 110                  * the fact that they are actually only accessible through one. */
 111                 if (gl)
 112                         desc.qw0 = QI_EIOTLB_PASID(svm->pasid) |
 113                                         QI_EIOTLB_DID(sdev->did) |
 114                                         QI_EIOTLB_GRAN(QI_GRAN_ALL_ALL) |
 115                                         QI_EIOTLB_TYPE;
 116                 else
 117                         desc.qw0 = QI_EIOTLB_PASID(svm->pasid) |
 118                                         QI_EIOTLB_DID(sdev->did) |
 119                                         QI_EIOTLB_GRAN(QI_GRAN_NONG_PASID) |
 120                                         QI_EIOTLB_TYPE;
 121                 desc.qw1 = 0;
 122         } else {
 123                 int mask = ilog2(__roundup_pow_of_two(pages));
 124
 125                 desc.qw0 = QI_EIOTLB_PASID(svm->pasid) |
 126                                 QI_EIOTLB_DID(sdev->did) |
 127                                 QI_EIOTLB_GRAN(QI_GRAN_PSI_PASID) |
 128                                 QI_EIOTLB_TYPE;
 129                 desc.qw1 = QI_EIOTLB_ADDR(address) |
 130                                 QI_EIOTLB_GL(gl) |
 131                                 QI_EIOTLB_IH(ih) |
 132                                 QI_EIOTLB_AM(mask);
 133         }
 134         desc.qw2 = 0;
 135         desc.qw3 = 0;
 136         qi_submit_sync(&desc, svm->iommu);
 137
 138         if (sdev->dev_iotlb) {
 139                 desc.qw0 = QI_DEV_EIOTLB_PASID(svm->pasid) |
 140                                 QI_DEV_EIOTLB_SID(sdev->sid) |
 141                                 QI_DEV_EIOTLB_QDEP(sdev->qdep) |
 142                                 QI_DEIOTLB_TYPE;
 143                 if (pages == -1) {
 144                         desc.qw1 = QI_DEV_EIOTLB_ADDR(-1ULL >> 1) |
 145                                         QI_DEV_EIOTLB_SIZE;
 146                 } else if (pages > 1) {
 147                         /* The least significant zero bit indicates the size. So,
 148                          * for example, an "address" value of 0x12345f000 will
 149                          * flush from 0x123440000 to 0x12347ffff (256KiB). */
 150                         unsigned long last = address + ((unsigned long)(pages - 1) << VTD_PAGE_SHIFT);
 151                         unsigned long mask = __rounddown_pow_of_two(address ^ last);
 152
 153                         desc.qw1 = QI_DEV_EIOTLB_ADDR((address & ~mask) |
 154                                         (mask - 1)) | QI_DEV_EIOTLB_SIZE;
 155                 } else {
 156                         desc.qw1 = QI_DEV_EIOTLB_ADDR(address);
 157                 }
 158                 desc.qw2 = 0;
 159                 desc.qw3 = 0;
 160                 qi_submit_sync(&desc, svm->iommu);
 161         }
 162 }
 163
 164 static void intel_flush_svm_range(struct intel_svm *svm, unsigned long address,
 165                                   unsigned long pages, int ih, int gl)
 166 {
 167         struct intel_svm_dev *sdev;
 168
 169         rcu_read_lock();
 170         list_for_each_entry_rcu(sdev, &svm->devs, list)
 171                 intel_flush_svm_range_dev(svm, sdev, address, pages, ih, gl);
 172         rcu_read_unlock();
 173 }
 174
 175 /* Pages have been freed at this point */
 176 static void intel_invalidate_range(struct mmu_notifier *mn,
 177                                    struct mm_struct *mm,
 178                                    unsigned long start, unsigned long end)
 179 {
 180         struct intel_svm *svm = container_of(mn, struct intel_svm, notifier);
 181
 182         intel_flush_svm_range(svm, start,
 183                               (end - start + PAGE_SIZE - 1) >> VTD_PAGE_SHIFT, 0, 0);
 184 }
 185
 186 static void intel_mm_release(struct mmu_notifier *mn, struct mm_struct *mm)
 187 {
 188         struct intel_svm *svm = container_of(mn, struct intel_svm, notifier);
 189         struct intel_svm_dev *sdev;
 190
 191         /* This might end up being called from exit_mmap(), *before* the page
 192          * tables are cleared. And __mmu_notifier_release() will delete us from
 193          * the list of notifiers so that our invalidate_range() callback doesn't
 194          * get called when the page tables are cleared. So we need to protect
 195          * against hardware accessing those page tables.
 196          *
 197          * We do it by clearing the entry in the PASID table and then flushing
 198          * the IOTLB and the PASID table caches. This might upset hardware;
 199          * perhaps we'll want to point the PASID to a dummy PGD (like the zero
 200          * page) so that we end up taking a fault that the hardware really
 201          * *has* to handle gracefully without affecting other processes.
 202          */
 203         rcu_read_lock();
 204         list_for_each_entry_rcu(sdev, &svm->devs, list) {
 205                 intel_pasid_tear_down_entry(svm->iommu, sdev->dev, svm->pasid);
 206                 intel_flush_svm_range_dev(svm, sdev, 0, -1, 0, !svm->mm);
 207         }
 208         rcu_read_unlock();
 209
 210 }
 211
 212 static const struct mmu_notifier_ops intel_mmuops = {
 213         .release = intel_mm_release,
 214         .invalidate_range = intel_invalidate_range,
 215 };
 216
 217 static DEFINE_MUTEX(pasid_mutex);
 218 static LIST_HEAD(global_svm_list);
 219
 220 int intel_svm_bind_mm(struct device *dev, int *pasid, int flags, struct svm_dev_ops *ops)
 221 {
 222         struct intel_iommu *iommu = intel_svm_device_to_iommu(dev);
 223         struct device_domain_info *info;
 224         struct intel_svm_dev *sdev;
 225         struct intel_svm *svm = NULL;
 226         struct mm_struct *mm = NULL;
 227         int pasid_max;
 228         int ret;
 229
 230         if (!iommu || dmar_disabled)
 231                 return -EINVAL;
 232
 233         if (dev_is_pci(dev)) {
 234                 pasid_max = pci_max_pasids(to_pci_dev(dev));
 235                 if (pasid_max < 0)
 236                         return -EINVAL;
 237         } else
 238                 pasid_max = 1 << 20;
 239
 240         if (flags & SVM_FLAG_SUPERVISOR_MODE) {
 241                 if (!ecap_srs(iommu->ecap))
 242                         return -EINVAL;
 243         } else if (pasid) {
 244                 mm = get_task_mm(current);
 245                 BUG_ON(!mm);
 246         }
 247
 248         mutex_lock(&pasid_mutex);
 249         if (pasid && !(flags & SVM_FLAG_PRIVATE_PASID)) {
 250                 struct intel_svm *t;
 251
 252                 list_for_each_entry(t, &global_svm_list, list) {
 253                         if (t->mm != mm || (t->flags & SVM_FLAG_PRIVATE_PASID))
 254                                 continue;
 255
 256                         svm = t;
 257                         if (svm->pasid >= pasid_max) {
 258                                 dev_warn(dev,
 259                                          "Limited PASID width. Cannot use existing PASID %d\n",
 260                                          svm->pasid);
 261                                 ret = -ENOSPC;
 262                                 goto out;
 263                         }
 264
 265                         list_for_each_entry(sdev, &svm->devs, list) {
 266                                 if (dev == sdev->dev) {
 267                                         if (sdev->ops != ops) {
 268                                                 ret = -EBUSY;
 269                                                 goto out;
 270                                         }
 271                                         sdev->users++;
 272                                         goto success;
 273                                 }
 274                         }
 275
 276                         break;
 277                 }
 278         }
 279
 280         sdev = kzalloc(sizeof(*sdev), GFP_KERNEL);
 281         if (!sdev) {
 282                 ret = -ENOMEM;
 283                 goto out;
 284         }
 285         sdev->dev = dev;
 286
 287         ret = intel_iommu_enable_pasid(iommu, dev);
 288         if (ret || !pasid) {
 289                 /* If they don't actually want to assign a PASID, this is
 290                  * just an enabling check/preparation. */
 291                 kfree(sdev);
 292                 goto out;
 293         }
 294
 295         info = dev->archdata.iommu;
 296         if (!info || !info->pasid_supported) {
 297                 kfree(sdev);
 298                 goto out;
 299         }
 300
 301         sdev->did = FLPT_DEFAULT_DID;
 302         sdev->sid = PCI_DEVID(info->bus, info->devfn);
 303         if (info->ats_enabled) {
 304                 sdev->dev_iotlb = 1;
 305                 sdev->qdep = info->ats_qdep;
 306                 if (sdev->qdep >= QI_DEV_EIOTLB_MAX_INVS)
 307                         sdev->qdep = 0;
 308         }
 309
 310         /* Finish the setup now we know we're keeping it */
 311         sdev->users = 1;
 312         sdev->ops = ops;
 313         init_rcu_head(&sdev->rcu);
 314
 315         if (!svm) {
 316                 svm = kzalloc(sizeof(*svm), GFP_KERNEL);
 317                 if (!svm) {
 318                         ret = -ENOMEM;
 319                         kfree(sdev);
 320                         goto out;
 321                 }
 322                 svm->iommu = iommu;
 323
 324                 if (pasid_max > intel_pasid_max_id)
 325                         pasid_max = intel_pasid_max_id;
 326
 327                 /* Do not use PASID 0 in caching mode (virtualised IOMMU) */
 328                 ret = intel_pasid_alloc_id(svm,
 329                                            !!cap_caching_mode(iommu->cap),
 330                                            pasid_max - 1, GFP_KERNEL);
 331                 if (ret < 0) {
 332                         kfree(svm);
 333                         kfree(sdev);
 334                         goto out;
 335                 }
 336                 svm->pasid = ret;
 337                 svm->notifier.ops = &intel_mmuops;
 338                 svm->mm = mm;
 339                 svm->flags = flags;
 340                 INIT_LIST_HEAD_RCU(&svm->devs);
 341                 INIT_LIST_HEAD(&svm->list);
 342                 ret = -ENOMEM;
 343                 if (mm) {
 344                         ret = mmu_notifier_register(&svm->notifier, mm);
 345                         if (ret) {
 346                                 intel_pasid_free_id(svm->pasid);
 347                                 kfree(svm);
 348                                 kfree(sdev);
 349                                 goto out;
 350                         }
 351                 }
 352
 353                 spin_lock(&iommu->lock);
 354                 ret = intel_pasid_setup_first_level(iommu, dev,
 355                                 mm ? mm->pgd : init_mm.pgd,
 356                                 svm->pasid, FLPT_DEFAULT_DID,
 357                                 mm ? 0 : PASID_FLAG_SUPERVISOR_MODE);
 358                 spin_unlock(&iommu->lock);
 359                 if (ret) {
 360                         if (mm)
 361                                 mmu_notifier_unregister(&svm->notifier, mm);
 362                         intel_pasid_free_id(svm->pasid);
 363                         kfree(svm);
 364                         kfree(sdev);
 365                         goto out;
 366                 }
 367
 368                 list_add_tail(&svm->list, &global_svm_list);
 369         } else {
 370                 /*
 371                  * Binding a new device with existing PASID, need to setup
 372                  * the PASID entry.
 373                  */
 374                 spin_lock(&iommu->lock);
 375                 ret = intel_pasid_setup_first_level(iommu, dev,
 376                                                 mm ? mm->pgd : init_mm.pgd,
 377                                                 svm->pasid, FLPT_DEFAULT_DID,
 378                                                 mm ? 0 : PASID_FLAG_SUPERVISOR_MODE);
 379                 spin_unlock(&iommu->lock);
 380                 if (ret) {
 381                         kfree(sdev);
 382                         goto out;
 383                 }
 384         }
 385         list_add_rcu(&sdev->list, &svm->devs);
 386
 387  success:
 388         *pasid = svm->pasid;
 389         ret = 0;
 390  out:
 391         mutex_unlock(&pasid_mutex);
 392         if (mm)
 393                 mmput(mm);
 394         return ret;
 395 }
 396 EXPORT_SYMBOL_GPL(intel_svm_bind_mm);
 397
 398 int intel_svm_unbind_mm(struct device *dev, int pasid)
 399 {
 400         struct intel_svm_dev *sdev;
 401         struct intel_iommu *iommu;
 402         struct intel_svm *svm;
 403         int ret = -EINVAL;
 404
 405         mutex_lock(&pasid_mutex);
 406         iommu = intel_svm_device_to_iommu(dev);
 407         if (!iommu)
 408                 goto out;
 409
 410         svm = intel_pasid_lookup_id(pasid);
 411         if (!svm)
 412                 goto out;
 413
 414         list_for_each_entry(sdev, &svm->devs, list) {
 415                 if (dev == sdev->dev) {
 416                         ret = 0;
 417                         sdev->users--;
 418                         if (!sdev->users) {
 419                                 list_del_rcu(&sdev->list);
 420                                 /* Flush the PASID cache and IOTLB for this device.
 421                                  * Note that we do depend on the hardware *not* using
 422                                  * the PASID any more. Just as we depend on other
 423                                  * devices never using PASIDs that they have no right
 424                                  * to use. We have a *shared* PASID table, because it's
 425                                  * large and has to be physically contiguous. So it's
 426                                  * hard to be as defensive as we might like. */
 427                                 intel_pasid_tear_down_entry(iommu, dev, svm->pasid);
 428                                 intel_flush_svm_range_dev(svm, sdev, 0, -1, 0, !svm->mm);
 429                                 kfree_rcu(sdev, rcu);
 430
 431                                 if (list_empty(&svm->devs)) {
 432                                         intel_pasid_free_id(svm->pasid);
 433                                         if (svm->mm)
 434                                                 mmu_notifier_unregister(&svm->notifier, svm->mm);
 435
 436                                         list_del(&svm->list);
 437
 438                                         /* We mandate that no page faults may be outstanding
 439                                          * for the PASID when intel_svm_unbind_mm() is called.
 440                                          * If that is not obeyed, subtle errors will happen.
 441                                          * Let's make them less subtle... */
 442                                         memset(svm, 0x6b, sizeof(*svm));
 443                                         kfree(svm);
 444                                 }
 445                         }
 446                         break;
 447                 }
 448         }
 449  out:
 450         mutex_unlock(&pasid_mutex);
 451
 452         return ret;
 453 }
 454 EXPORT_SYMBOL_GPL(intel_svm_unbind_mm);
 455
 456 int intel_svm_is_pasid_valid(struct device *dev, int pasid)
 457 {
 458         struct intel_iommu *iommu;
 459         struct intel_svm *svm;
 460         int ret = -EINVAL;
 461
 462         mutex_lock(&pasid_mutex);
 463         iommu = intel_svm_device_to_iommu(dev);
 464         if (!iommu)
 465                 goto out;
 466
 467         svm = intel_pasid_lookup_id(pasid);
 468         if (!svm)
 469                 goto out;
 470
 471         /* init_mm is used in this case */
 472         if (!svm->mm)
 473                 ret = 1;
 474         else if (atomic_read(&svm->mm->mm_users) > 0)
 475                 ret = 1;
 476         else
 477                 ret = 0;
 478
 479  out:
 480         mutex_unlock(&pasid_mutex);
 481
 482         return ret;
 483 }
 484 EXPORT_SYMBOL_GPL(intel_svm_is_pasid_valid);
 485
 486 /* Page request queue descriptor */
 487 struct page_req_dsc {
 488         union {
 489                 struct {
 490                         u64 type:8;
 491                         u64 pasid_present:1;
 492                         u64 priv_data_present:1;
 493                         u64 rsvd:6;
 494                         u64 rid:16;
 495                         u64 pasid:20;
 496                         u64 exe_req:1;
 497                         u64 pm_req:1;
 498                         u64 rsvd2:10;
 499                 };
 500                 u64 qw_0;
 501         };
 502         union {
 503                 struct {
 504                         u64 rd_req:1;
 505                         u64 wr_req:1;
 506                         u64 lpig:1;
 507                         u64 prg_index:9;
 508                         u64 addr:52;
 509                 };
 510                 u64 qw_1;
 511         };
 512         u64 priv_data[2];
 513 };
 514
 515 #define PRQ_RING_MASK ((0x1000 << PRQ_ORDER) - 0x10)
 516
 517 static bool access_error(struct vm_area_struct *vma, struct page_req_dsc *req)
 518 {
 519         unsigned long requested = 0;
 520
 521         if (req->exe_req)
 522                 requested |= VM_EXEC;
 523
 524         if (req->rd_req)
 525                 requested |= VM_READ;
 526
 527         if (req->wr_req)
 528                 requested |= VM_WRITE;
 529
 530         return (requested & ~vma->vm_flags) != 0;
 531 }
 532
 533 static bool is_canonical_address(u64 addr)
 534 {
 535         int shift = 64 - (__VIRTUAL_MASK_SHIFT + 1);
 536         long saddr = (long) addr;
 537
 538         return (((saddr << shift) >> shift) == saddr);
 539 }
 540
 541 static irqreturn_t prq_event_thread(int irq, void *d)
 542 {
 543         struct intel_iommu *iommu = d;
 544         struct intel_svm *svm = NULL;
 545         int head, tail, handled = 0;
 546
 547         /* Clear PPR bit before reading head/tail registers, to
 548          * ensure that we get a new interrupt if needed. */
 549         writel(DMA_PRS_PPR, iommu->reg + DMAR_PRS_REG);
 550
 551         tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
 552         head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
 553         while (head != tail) {
 554                 struct intel_svm_dev *sdev;
 555                 struct vm_area_struct *vma;
 556                 struct page_req_dsc *req;
 557                 struct qi_desc resp;
 558                 int result;
 559                 vm_fault_t ret;
 560                 u64 address;
 561
 562                 handled = 1;
 563
 564                 req = &iommu->prq[head / sizeof(*req)];
 565
 566                 result = QI_RESP_FAILURE;
 567                 address = (u64)req->addr << VTD_PAGE_SHIFT;
 568                 if (!req->pasid_present) {
 569                         pr_err("%s: Page request without PASID: %08llx %08llx\n",
 570                                iommu->name, ((unsigned long long *)req)[0],
 571                                ((unsigned long long *)req)[1]);
 572                         goto no_pasid;
 573                 }
 574
 575                 if (!svm || svm->pasid != req->pasid) {
 576                         rcu_read_lock();
 577                         svm = intel_pasid_lookup_id(req->pasid);
 578                         /* It *can't* go away, because the driver is not permitted
 579                          * to unbind the mm while any page faults are outstanding.
 580                          * So we only need RCU to protect the internal idr code. */
 581                         rcu_read_unlock();
 582
 583                         if (!svm) {
 584                                 pr_err("%s: Page request for invalid PASID %d: %08llx %08llx\n",
 585                                        iommu->name, req->pasid, ((unsigned long long *)req)[0],
 586                                        ((unsigned long long *)req)[1]);
 587                                 goto no_pasid;
 588                         }
 589                 }
 590
 591                 result = QI_RESP_INVALID;
 592                 /* Since we're using init_mm.pgd directly, we should never take
 593                  * any faults on kernel addresses. */
 594                 if (!svm->mm)
 595                         goto bad_req;
 596                 /* If the mm is already defunct, don't handle faults. */
 597                 if (!mmget_not_zero(svm->mm))
 598                         goto bad_req;
 599
 600                 /* If address is not canonical, return invalid response */
 601                 if (!is_canonical_address(address))
 602                         goto bad_req;
 603
 604                 down_read(&svm->mm->mmap_sem);
 605                 vma = find_extend_vma(svm->mm, address);
 606                 if (!vma || address < vma->vm_start)
 607                         goto invalid;
 608
 609                 if (access_error(vma, req))
 610                         goto invalid;
 611
 612                 ret = handle_mm_fault(vma, address,
 613                                       req->wr_req ? FAULT_FLAG_WRITE : 0);
 614                 if (ret & VM_FAULT_ERROR)
 615                         goto invalid;
 616
 617                 result = QI_RESP_SUCCESS;
 618         invalid:
 619                 up_read(&svm->mm->mmap_sem);
 620                 mmput(svm->mm);
 621         bad_req:
 622                 /* Accounting for major/minor faults? */
 623                 rcu_read_lock();
 624                 list_for_each_entry_rcu(sdev, &svm->devs, list) {
 625                         if (sdev->sid == req->rid)
 626                                 break;
 627                 }
 628                 /* Other devices can go away, but the drivers are not permitted
 629                  * to unbind while any page faults might be in flight. So it's
 630                  * OK to drop the 'lock' here now we have it. */
 631                 rcu_read_unlock();
 632
 633                 if (WARN_ON(&sdev->list == &svm->devs))
 634                         sdev = NULL;
 635
 636                 if (sdev && sdev->ops && sdev->ops->fault_cb) {
 637                         int rwxp = (req->rd_req << 3) | (req->wr_req << 2) |
 638                                 (req->exe_req << 1) | (req->pm_req);
 639                         sdev->ops->fault_cb(sdev->dev, req->pasid, req->addr,
 640                                             req->priv_data, rwxp, result);
 641                 }
 642                 /* We get here in the error case where the PASID lookup failed,
 643                    and these can be NULL. Do not use them below this point! */
 644                 sdev = NULL;
 645                 svm = NULL;
 646         no_pasid:
 647                 if (req->lpig || req->priv_data_present) {
 648                         /*
 649                          * Per VT-d spec. v3.0 ch7.7, system software must
 650                          * respond with page group response if private data
 651                          * is present (PDP) or last page in group (LPIG) bit
 652                          * is set. This is an additional VT-d feature beyond
 653                          * PCI ATS spec.
 654                          */
 655                         resp.qw0 = QI_PGRP_PASID(req->pasid) |
 656                                 QI_PGRP_DID(req->rid) |
 657                                 QI_PGRP_PASID_P(req->pasid_present) |
 658                                 QI_PGRP_PDP(req->pasid_present) |
 659                                 QI_PGRP_RESP_CODE(result) |
 660                                 QI_PGRP_RESP_TYPE;
 661                         resp.qw1 = QI_PGRP_IDX(req->prg_index) |
 662                                 QI_PGRP_LPIG(req->lpig);
 663
 664                         if (req->priv_data_present)
 665                                 memcpy(&resp.qw2, req->priv_data,
 666                                        sizeof(req->priv_data));
 667                 }
 668                 resp.qw2 = 0;
 669                 resp.qw3 = 0;
 670                 qi_submit_sync(&resp, iommu);
 671
 672                 head = (head + sizeof(*req)) & PRQ_RING_MASK;
 673         }
 674
 675         dmar_writeq(iommu->reg + DMAR_PQH_REG, tail);
 676
 677         return IRQ_RETVAL(handled);
 678 }