intel-svm.c 19 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717
  1. /*
  2. * Copyright © 2015 Intel Corporation.
  3. *
  4. * This program is free software; you can redistribute it and/or modify it
  5. * under the terms and conditions of the GNU General Public License,
  6. * version 2, as published by the Free Software Foundation.
  7. *
  8. * This program is distributed in the hope it will be useful, but WITHOUT
  9. * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  10. * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
  11. * more details.
  12. *
  13. * Authors: David Woodhouse <dwmw2@infradead.org>
  14. */
  15. #include <linux/intel-iommu.h>
  16. #include <linux/mmu_notifier.h>
  17. #include <linux/sched.h>
  18. #include <linux/sched/mm.h>
  19. #include <linux/slab.h>
  20. #include <linux/intel-svm.h>
  21. #include <linux/rculist.h>
  22. #include <linux/pci.h>
  23. #include <linux/pci-ats.h>
  24. #include <linux/dmar.h>
  25. #include <linux/interrupt.h>
  26. #include <asm/page.h>
  27. #define PASID_ENTRY_P BIT_ULL(0)
  28. #define PASID_ENTRY_FLPM_5LP BIT_ULL(9)
  29. #define PASID_ENTRY_SRE BIT_ULL(11)
  30. static irqreturn_t prq_event_thread(int irq, void *d);
  31. struct pasid_entry {
  32. u64 val;
  33. };
  34. struct pasid_state_entry {
  35. u64 val;
  36. };
  37. int intel_svm_alloc_pasid_tables(struct intel_iommu *iommu)
  38. {
  39. struct page *pages;
  40. int order;
  41. if (cpu_feature_enabled(X86_FEATURE_GBPAGES) &&
  42. !cap_fl1gp_support(iommu->cap))
  43. return -EINVAL;
  44. if (cpu_feature_enabled(X86_FEATURE_LA57) &&
  45. !cap_5lp_support(iommu->cap))
  46. return -EINVAL;
  47. /* Start at 2 because it's defined as 2^(1+PSS) */
  48. iommu->pasid_max = 2 << ecap_pss(iommu->ecap);
  49. /* Eventually I'm promised we will get a multi-level PASID table
  50. * and it won't have to be physically contiguous. Until then,
  51. * limit the size because 8MiB contiguous allocations can be hard
  52. * to come by. The limit of 0x20000, which is 1MiB for each of
  53. * the PASID and PASID-state tables, is somewhat arbitrary. */
  54. if (iommu->pasid_max > 0x20000)
  55. iommu->pasid_max = 0x20000;
  56. order = get_order(sizeof(struct pasid_entry) * iommu->pasid_max);
  57. pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, order);
  58. if (!pages) {
  59. pr_warn("IOMMU: %s: Failed to allocate PASID table\n",
  60. iommu->name);
  61. return -ENOMEM;
  62. }
  63. iommu->pasid_table = page_address(pages);
  64. pr_info("%s: Allocated order %d PASID table.\n", iommu->name, order);
  65. if (ecap_dis(iommu->ecap)) {
  66. /* Just making it explicit... */
  67. BUILD_BUG_ON(sizeof(struct pasid_entry) != sizeof(struct pasid_state_entry));
  68. pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, order);
  69. if (pages)
  70. iommu->pasid_state_table = page_address(pages);
  71. else
  72. pr_warn("IOMMU: %s: Failed to allocate PASID state table\n",
  73. iommu->name);
  74. }
  75. idr_init(&iommu->pasid_idr);
  76. return 0;
  77. }
  78. int intel_svm_free_pasid_tables(struct intel_iommu *iommu)
  79. {
  80. int order = get_order(sizeof(struct pasid_entry) * iommu->pasid_max);
  81. if (iommu->pasid_table) {
  82. free_pages((unsigned long)iommu->pasid_table, order);
  83. iommu->pasid_table = NULL;
  84. }
  85. if (iommu->pasid_state_table) {
  86. free_pages((unsigned long)iommu->pasid_state_table, order);
  87. iommu->pasid_state_table = NULL;
  88. }
  89. idr_destroy(&iommu->pasid_idr);
  90. return 0;
  91. }
  92. #define PRQ_ORDER 0
  93. int intel_svm_enable_prq(struct intel_iommu *iommu)
  94. {
  95. struct page *pages;
  96. int irq, ret;
  97. pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, PRQ_ORDER);
  98. if (!pages) {
  99. pr_warn("IOMMU: %s: Failed to allocate page request queue\n",
  100. iommu->name);
  101. return -ENOMEM;
  102. }
  103. iommu->prq = page_address(pages);
  104. irq = dmar_alloc_hwirq(DMAR_UNITS_SUPPORTED + iommu->seq_id, iommu->node, iommu);
  105. if (irq <= 0) {
  106. pr_err("IOMMU: %s: Failed to create IRQ vector for page request queue\n",
  107. iommu->name);
  108. ret = -EINVAL;
  109. err:
  110. free_pages((unsigned long)iommu->prq, PRQ_ORDER);
  111. iommu->prq = NULL;
  112. return ret;
  113. }
  114. iommu->pr_irq = irq;
  115. snprintf(iommu->prq_name, sizeof(iommu->prq_name), "dmar%d-prq", iommu->seq_id);
  116. ret = request_threaded_irq(irq, NULL, prq_event_thread, IRQF_ONESHOT,
  117. iommu->prq_name, iommu);
  118. if (ret) {
  119. pr_err("IOMMU: %s: Failed to request IRQ for page request queue\n",
  120. iommu->name);
  121. dmar_free_hwirq(irq);
  122. iommu->pr_irq = 0;
  123. goto err;
  124. }
  125. dmar_writeq(iommu->reg + DMAR_PQH_REG, 0ULL);
  126. dmar_writeq(iommu->reg + DMAR_PQT_REG, 0ULL);
  127. dmar_writeq(iommu->reg + DMAR_PQA_REG, virt_to_phys(iommu->prq) | PRQ_ORDER);
  128. return 0;
  129. }
  130. int intel_svm_finish_prq(struct intel_iommu *iommu)
  131. {
  132. dmar_writeq(iommu->reg + DMAR_PQH_REG, 0ULL);
  133. dmar_writeq(iommu->reg + DMAR_PQT_REG, 0ULL);
  134. dmar_writeq(iommu->reg + DMAR_PQA_REG, 0ULL);
  135. if (iommu->pr_irq) {
  136. free_irq(iommu->pr_irq, iommu);
  137. dmar_free_hwirq(iommu->pr_irq);
  138. iommu->pr_irq = 0;
  139. }
  140. free_pages((unsigned long)iommu->prq, PRQ_ORDER);
  141. iommu->prq = NULL;
  142. return 0;
  143. }
  144. static void intel_flush_svm_range_dev (struct intel_svm *svm, struct intel_svm_dev *sdev,
  145. unsigned long address, unsigned long pages, int ih, int gl)
  146. {
  147. struct qi_desc desc;
  148. if (pages == -1) {
  149. /* For global kernel pages we have to flush them in *all* PASIDs
  150. * because that's the only option the hardware gives us. Despite
  151. * the fact that they are actually only accessible through one. */
  152. if (gl)
  153. desc.low = QI_EIOTLB_PASID(svm->pasid) | QI_EIOTLB_DID(sdev->did) |
  154. QI_EIOTLB_GRAN(QI_GRAN_ALL_ALL) | QI_EIOTLB_TYPE;
  155. else
  156. desc.low = QI_EIOTLB_PASID(svm->pasid) | QI_EIOTLB_DID(sdev->did) |
  157. QI_EIOTLB_GRAN(QI_GRAN_NONG_PASID) | QI_EIOTLB_TYPE;
  158. desc.high = 0;
  159. } else {
  160. int mask = ilog2(__roundup_pow_of_two(pages));
  161. desc.low = QI_EIOTLB_PASID(svm->pasid) | QI_EIOTLB_DID(sdev->did) |
  162. QI_EIOTLB_GRAN(QI_GRAN_PSI_PASID) | QI_EIOTLB_TYPE;
  163. desc.high = QI_EIOTLB_ADDR(address) | QI_EIOTLB_GL(gl) |
  164. QI_EIOTLB_IH(ih) | QI_EIOTLB_AM(mask);
  165. }
  166. qi_submit_sync(&desc, svm->iommu);
  167. if (sdev->dev_iotlb) {
  168. desc.low = QI_DEV_EIOTLB_PASID(svm->pasid) | QI_DEV_EIOTLB_SID(sdev->sid) |
  169. QI_DEV_EIOTLB_QDEP(sdev->qdep) | QI_DEIOTLB_TYPE;
  170. if (pages == -1) {
  171. desc.high = QI_DEV_EIOTLB_ADDR(-1ULL >> 1) | QI_DEV_EIOTLB_SIZE;
  172. } else if (pages > 1) {
  173. /* The least significant zero bit indicates the size. So,
  174. * for example, an "address" value of 0x12345f000 will
  175. * flush from 0x123440000 to 0x12347ffff (256KiB). */
  176. unsigned long last = address + ((unsigned long)(pages - 1) << VTD_PAGE_SHIFT);
  177. unsigned long mask = __rounddown_pow_of_two(address ^ last);
  178. desc.high = QI_DEV_EIOTLB_ADDR((address & ~mask) | (mask - 1)) | QI_DEV_EIOTLB_SIZE;
  179. } else {
  180. desc.high = QI_DEV_EIOTLB_ADDR(address);
  181. }
  182. qi_submit_sync(&desc, svm->iommu);
  183. }
  184. }
  185. static void intel_flush_svm_range(struct intel_svm *svm, unsigned long address,
  186. unsigned long pages, int ih, int gl)
  187. {
  188. struct intel_svm_dev *sdev;
  189. /* Try deferred invalidate if available */
  190. if (svm->iommu->pasid_state_table &&
  191. !cmpxchg64(&svm->iommu->pasid_state_table[svm->pasid].val, 0, 1ULL << 63))
  192. return;
  193. rcu_read_lock();
  194. list_for_each_entry_rcu(sdev, &svm->devs, list)
  195. intel_flush_svm_range_dev(svm, sdev, address, pages, ih, gl);
  196. rcu_read_unlock();
  197. }
  198. static void intel_change_pte(struct mmu_notifier *mn, struct mm_struct *mm,
  199. unsigned long address, pte_t pte)
  200. {
  201. struct intel_svm *svm = container_of(mn, struct intel_svm, notifier);
  202. intel_flush_svm_range(svm, address, 1, 1, 0);
  203. }
  204. /* Pages have been freed at this point */
  205. static void intel_invalidate_range(struct mmu_notifier *mn,
  206. struct mm_struct *mm,
  207. unsigned long start, unsigned long end)
  208. {
  209. struct intel_svm *svm = container_of(mn, struct intel_svm, notifier);
  210. intel_flush_svm_range(svm, start,
  211. (end - start + PAGE_SIZE - 1) >> VTD_PAGE_SHIFT, 0, 0);
  212. }
  213. static void intel_flush_pasid_dev(struct intel_svm *svm, struct intel_svm_dev *sdev, int pasid)
  214. {
  215. struct qi_desc desc;
  216. desc.high = 0;
  217. desc.low = QI_PC_TYPE | QI_PC_DID(sdev->did) | QI_PC_PASID_SEL | QI_PC_PASID(pasid);
  218. qi_submit_sync(&desc, svm->iommu);
  219. }
  220. static void intel_mm_release(struct mmu_notifier *mn, struct mm_struct *mm)
  221. {
  222. struct intel_svm *svm = container_of(mn, struct intel_svm, notifier);
  223. struct intel_svm_dev *sdev;
  224. /* This might end up being called from exit_mmap(), *before* the page
  225. * tables are cleared. And __mmu_notifier_release() will delete us from
  226. * the list of notifiers so that our invalidate_range() callback doesn't
  227. * get called when the page tables are cleared. So we need to protect
  228. * against hardware accessing those page tables.
  229. *
  230. * We do it by clearing the entry in the PASID table and then flushing
  231. * the IOTLB and the PASID table caches. This might upset hardware;
  232. * perhaps we'll want to point the PASID to a dummy PGD (like the zero
  233. * page) so that we end up taking a fault that the hardware really
  234. * *has* to handle gracefully without affecting other processes.
  235. */
  236. svm->iommu->pasid_table[svm->pasid].val = 0;
  237. wmb();
  238. rcu_read_lock();
  239. list_for_each_entry_rcu(sdev, &svm->devs, list) {
  240. intel_flush_pasid_dev(svm, sdev, svm->pasid);
  241. intel_flush_svm_range_dev(svm, sdev, 0, -1, 0, !svm->mm);
  242. }
  243. rcu_read_unlock();
  244. }
  245. static const struct mmu_notifier_ops intel_mmuops = {
  246. .flags = MMU_INVALIDATE_DOES_NOT_BLOCK,
  247. .release = intel_mm_release,
  248. .change_pte = intel_change_pte,
  249. .invalidate_range = intel_invalidate_range,
  250. };
  251. static DEFINE_MUTEX(pasid_mutex);
  252. static LIST_HEAD(global_svm_list);
  253. int intel_svm_bind_mm(struct device *dev, int *pasid, int flags, struct svm_dev_ops *ops)
  254. {
  255. struct intel_iommu *iommu = intel_svm_device_to_iommu(dev);
  256. struct intel_svm_dev *sdev;
  257. struct intel_svm *svm = NULL;
  258. struct mm_struct *mm = NULL;
  259. u64 pasid_entry_val;
  260. int pasid_max;
  261. int ret;
  262. if (!iommu || !iommu->pasid_table)
  263. return -EINVAL;
  264. if (dev_is_pci(dev)) {
  265. pasid_max = pci_max_pasids(to_pci_dev(dev));
  266. if (pasid_max < 0)
  267. return -EINVAL;
  268. } else
  269. pasid_max = 1 << 20;
  270. if (flags & SVM_FLAG_SUPERVISOR_MODE) {
  271. if (!ecap_srs(iommu->ecap))
  272. return -EINVAL;
  273. } else if (pasid) {
  274. mm = get_task_mm(current);
  275. BUG_ON(!mm);
  276. }
  277. mutex_lock(&pasid_mutex);
  278. if (pasid && !(flags & SVM_FLAG_PRIVATE_PASID)) {
  279. struct intel_svm *t;
  280. list_for_each_entry(t, &global_svm_list, list) {
  281. if (t->mm != mm || (t->flags & SVM_FLAG_PRIVATE_PASID))
  282. continue;
  283. svm = t;
  284. if (svm->pasid >= pasid_max) {
  285. dev_warn(dev,
  286. "Limited PASID width. Cannot use existing PASID %d\n",
  287. svm->pasid);
  288. ret = -ENOSPC;
  289. goto out;
  290. }
  291. list_for_each_entry(sdev, &svm->devs, list) {
  292. if (dev == sdev->dev) {
  293. if (sdev->ops != ops) {
  294. ret = -EBUSY;
  295. goto out;
  296. }
  297. sdev->users++;
  298. goto success;
  299. }
  300. }
  301. break;
  302. }
  303. }
  304. sdev = kzalloc(sizeof(*sdev), GFP_KERNEL);
  305. if (!sdev) {
  306. ret = -ENOMEM;
  307. goto out;
  308. }
  309. sdev->dev = dev;
  310. ret = intel_iommu_enable_pasid(iommu, sdev);
  311. if (ret || !pasid) {
  312. /* If they don't actually want to assign a PASID, this is
  313. * just an enabling check/preparation. */
  314. kfree(sdev);
  315. goto out;
  316. }
  317. /* Finish the setup now we know we're keeping it */
  318. sdev->users = 1;
  319. sdev->ops = ops;
  320. init_rcu_head(&sdev->rcu);
  321. if (!svm) {
  322. svm = kzalloc(sizeof(*svm), GFP_KERNEL);
  323. if (!svm) {
  324. ret = -ENOMEM;
  325. kfree(sdev);
  326. goto out;
  327. }
  328. svm->iommu = iommu;
  329. if (pasid_max > iommu->pasid_max)
  330. pasid_max = iommu->pasid_max;
  331. /* Do not use PASID 0 in caching mode (virtualised IOMMU) */
  332. ret = idr_alloc(&iommu->pasid_idr, svm,
  333. !!cap_caching_mode(iommu->cap),
  334. pasid_max - 1, GFP_KERNEL);
  335. if (ret < 0) {
  336. kfree(svm);
  337. kfree(sdev);
  338. goto out;
  339. }
  340. svm->pasid = ret;
  341. svm->notifier.ops = &intel_mmuops;
  342. svm->mm = mm;
  343. svm->flags = flags;
  344. INIT_LIST_HEAD_RCU(&svm->devs);
  345. INIT_LIST_HEAD(&svm->list);
  346. ret = -ENOMEM;
  347. if (mm) {
  348. ret = mmu_notifier_register(&svm->notifier, mm);
  349. if (ret) {
  350. idr_remove(&svm->iommu->pasid_idr, svm->pasid);
  351. kfree(svm);
  352. kfree(sdev);
  353. goto out;
  354. }
  355. pasid_entry_val = (u64)__pa(mm->pgd) | PASID_ENTRY_P;
  356. } else
  357. pasid_entry_val = (u64)__pa(init_mm.pgd) |
  358. PASID_ENTRY_P | PASID_ENTRY_SRE;
  359. if (cpu_feature_enabled(X86_FEATURE_LA57))
  360. pasid_entry_val |= PASID_ENTRY_FLPM_5LP;
  361. iommu->pasid_table[svm->pasid].val = pasid_entry_val;
  362. wmb();
  363. /*
  364. * Flush PASID cache when a PASID table entry becomes
  365. * present.
  366. */
  367. if (cap_caching_mode(iommu->cap))
  368. intel_flush_pasid_dev(svm, sdev, svm->pasid);
  369. list_add_tail(&svm->list, &global_svm_list);
  370. }
  371. list_add_rcu(&sdev->list, &svm->devs);
  372. success:
  373. *pasid = svm->pasid;
  374. ret = 0;
  375. out:
  376. mutex_unlock(&pasid_mutex);
  377. if (mm)
  378. mmput(mm);
  379. return ret;
  380. }
  381. EXPORT_SYMBOL_GPL(intel_svm_bind_mm);
  382. int intel_svm_unbind_mm(struct device *dev, int pasid)
  383. {
  384. struct intel_svm_dev *sdev;
  385. struct intel_iommu *iommu;
  386. struct intel_svm *svm;
  387. int ret = -EINVAL;
  388. mutex_lock(&pasid_mutex);
  389. iommu = intel_svm_device_to_iommu(dev);
  390. if (!iommu || !iommu->pasid_table)
  391. goto out;
  392. svm = idr_find(&iommu->pasid_idr, pasid);
  393. if (!svm)
  394. goto out;
  395. list_for_each_entry(sdev, &svm->devs, list) {
  396. if (dev == sdev->dev) {
  397. ret = 0;
  398. sdev->users--;
  399. if (!sdev->users) {
  400. list_del_rcu(&sdev->list);
  401. /* Flush the PASID cache and IOTLB for this device.
  402. * Note that we do depend on the hardware *not* using
  403. * the PASID any more. Just as we depend on other
  404. * devices never using PASIDs that they have no right
  405. * to use. We have a *shared* PASID table, because it's
  406. * large and has to be physically contiguous. So it's
  407. * hard to be as defensive as we might like. */
  408. intel_flush_pasid_dev(svm, sdev, svm->pasid);
  409. intel_flush_svm_range_dev(svm, sdev, 0, -1, 0, !svm->mm);
  410. kfree_rcu(sdev, rcu);
  411. if (list_empty(&svm->devs)) {
  412. svm->iommu->pasid_table[svm->pasid].val = 0;
  413. wmb();
  414. idr_remove(&svm->iommu->pasid_idr, svm->pasid);
  415. if (svm->mm)
  416. mmu_notifier_unregister(&svm->notifier, svm->mm);
  417. list_del(&svm->list);
  418. /* We mandate that no page faults may be outstanding
  419. * for the PASID when intel_svm_unbind_mm() is called.
  420. * If that is not obeyed, subtle errors will happen.
  421. * Let's make them less subtle... */
  422. memset(svm, 0x6b, sizeof(*svm));
  423. kfree(svm);
  424. }
  425. }
  426. break;
  427. }
  428. }
  429. out:
  430. mutex_unlock(&pasid_mutex);
  431. return ret;
  432. }
  433. EXPORT_SYMBOL_GPL(intel_svm_unbind_mm);
  434. int intel_svm_is_pasid_valid(struct device *dev, int pasid)
  435. {
  436. struct intel_iommu *iommu;
  437. struct intel_svm *svm;
  438. int ret = -EINVAL;
  439. mutex_lock(&pasid_mutex);
  440. iommu = intel_svm_device_to_iommu(dev);
  441. if (!iommu || !iommu->pasid_table)
  442. goto out;
  443. svm = idr_find(&iommu->pasid_idr, pasid);
  444. if (!svm)
  445. goto out;
  446. /* init_mm is used in this case */
  447. if (!svm->mm)
  448. ret = 1;
  449. else if (atomic_read(&svm->mm->mm_users) > 0)
  450. ret = 1;
  451. else
  452. ret = 0;
  453. out:
  454. mutex_unlock(&pasid_mutex);
  455. return ret;
  456. }
  457. EXPORT_SYMBOL_GPL(intel_svm_is_pasid_valid);
  458. /* Page request queue descriptor */
  459. struct page_req_dsc {
  460. u64 srr:1;
  461. u64 bof:1;
  462. u64 pasid_present:1;
  463. u64 lpig:1;
  464. u64 pasid:20;
  465. u64 bus:8;
  466. u64 private:23;
  467. u64 prg_index:9;
  468. u64 rd_req:1;
  469. u64 wr_req:1;
  470. u64 exe_req:1;
  471. u64 priv_req:1;
  472. u64 devfn:8;
  473. u64 addr:52;
  474. };
  475. #define PRQ_RING_MASK ((0x1000 << PRQ_ORDER) - 0x10)
  476. static bool access_error(struct vm_area_struct *vma, struct page_req_dsc *req)
  477. {
  478. unsigned long requested = 0;
  479. if (req->exe_req)
  480. requested |= VM_EXEC;
  481. if (req->rd_req)
  482. requested |= VM_READ;
  483. if (req->wr_req)
  484. requested |= VM_WRITE;
  485. return (requested & ~vma->vm_flags) != 0;
  486. }
  487. static bool is_canonical_address(u64 addr)
  488. {
  489. int shift = 64 - (__VIRTUAL_MASK_SHIFT + 1);
  490. long saddr = (long) addr;
  491. return (((saddr << shift) >> shift) == saddr);
  492. }
  493. static irqreturn_t prq_event_thread(int irq, void *d)
  494. {
  495. struct intel_iommu *iommu = d;
  496. struct intel_svm *svm = NULL;
  497. int head, tail, handled = 0;
  498. /* Clear PPR bit before reading head/tail registers, to
  499. * ensure that we get a new interrupt if needed. */
  500. writel(DMA_PRS_PPR, iommu->reg + DMAR_PRS_REG);
  501. tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
  502. head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
  503. while (head != tail) {
  504. struct intel_svm_dev *sdev;
  505. struct vm_area_struct *vma;
  506. struct page_req_dsc *req;
  507. struct qi_desc resp;
  508. int ret, result;
  509. u64 address;
  510. handled = 1;
  511. req = &iommu->prq[head / sizeof(*req)];
  512. result = QI_RESP_FAILURE;
  513. address = (u64)req->addr << VTD_PAGE_SHIFT;
  514. if (!req->pasid_present) {
  515. pr_err("%s: Page request without PASID: %08llx %08llx\n",
  516. iommu->name, ((unsigned long long *)req)[0],
  517. ((unsigned long long *)req)[1]);
  518. goto bad_req;
  519. }
  520. if (!svm || svm->pasid != req->pasid) {
  521. rcu_read_lock();
  522. svm = idr_find(&iommu->pasid_idr, req->pasid);
  523. /* It *can't* go away, because the driver is not permitted
  524. * to unbind the mm while any page faults are outstanding.
  525. * So we only need RCU to protect the internal idr code. */
  526. rcu_read_unlock();
  527. if (!svm) {
  528. pr_err("%s: Page request for invalid PASID %d: %08llx %08llx\n",
  529. iommu->name, req->pasid, ((unsigned long long *)req)[0],
  530. ((unsigned long long *)req)[1]);
  531. goto no_pasid;
  532. }
  533. }
  534. result = QI_RESP_INVALID;
  535. /* Since we're using init_mm.pgd directly, we should never take
  536. * any faults on kernel addresses. */
  537. if (!svm->mm)
  538. goto bad_req;
  539. /* If the mm is already defunct, don't handle faults. */
  540. if (!mmget_not_zero(svm->mm))
  541. goto bad_req;
  542. /* If address is not canonical, return invalid response */
  543. if (!is_canonical_address(address))
  544. goto bad_req;
  545. down_read(&svm->mm->mmap_sem);
  546. vma = find_extend_vma(svm->mm, address);
  547. if (!vma || address < vma->vm_start)
  548. goto invalid;
  549. if (access_error(vma, req))
  550. goto invalid;
  551. ret = handle_mm_fault(vma, address,
  552. req->wr_req ? FAULT_FLAG_WRITE : 0);
  553. if (ret & VM_FAULT_ERROR)
  554. goto invalid;
  555. result = QI_RESP_SUCCESS;
  556. invalid:
  557. up_read(&svm->mm->mmap_sem);
  558. mmput(svm->mm);
  559. bad_req:
  560. /* Accounting for major/minor faults? */
  561. rcu_read_lock();
  562. list_for_each_entry_rcu(sdev, &svm->devs, list) {
  563. if (sdev->sid == PCI_DEVID(req->bus, req->devfn))
  564. break;
  565. }
  566. /* Other devices can go away, but the drivers are not permitted
  567. * to unbind while any page faults might be in flight. So it's
  568. * OK to drop the 'lock' here now we have it. */
  569. rcu_read_unlock();
  570. if (WARN_ON(&sdev->list == &svm->devs))
  571. sdev = NULL;
  572. if (sdev && sdev->ops && sdev->ops->fault_cb) {
  573. int rwxp = (req->rd_req << 3) | (req->wr_req << 2) |
  574. (req->exe_req << 1) | (req->priv_req);
  575. sdev->ops->fault_cb(sdev->dev, req->pasid, req->addr, req->private, rwxp, result);
  576. }
  577. /* We get here in the error case where the PASID lookup failed,
  578. and these can be NULL. Do not use them below this point! */
  579. sdev = NULL;
  580. svm = NULL;
  581. no_pasid:
  582. if (req->lpig) {
  583. /* Page Group Response */
  584. resp.low = QI_PGRP_PASID(req->pasid) |
  585. QI_PGRP_DID((req->bus << 8) | req->devfn) |
  586. QI_PGRP_PASID_P(req->pasid_present) |
  587. QI_PGRP_RESP_TYPE;
  588. resp.high = QI_PGRP_IDX(req->prg_index) |
  589. QI_PGRP_PRIV(req->private) | QI_PGRP_RESP_CODE(result);
  590. qi_submit_sync(&resp, iommu);
  591. } else if (req->srr) {
  592. /* Page Stream Response */
  593. resp.low = QI_PSTRM_IDX(req->prg_index) |
  594. QI_PSTRM_PRIV(req->private) | QI_PSTRM_BUS(req->bus) |
  595. QI_PSTRM_PASID(req->pasid) | QI_PSTRM_RESP_TYPE;
  596. resp.high = QI_PSTRM_ADDR(address) | QI_PSTRM_DEVFN(req->devfn) |
  597. QI_PSTRM_RESP_CODE(result);
  598. qi_submit_sync(&resp, iommu);
  599. }
  600. head = (head + sizeof(*req)) & PRQ_RING_MASK;
  601. }
  602. dmar_writeq(iommu->reg + DMAR_PQH_REG, tail);
  603. return IRQ_RETVAL(handled);
  604. }