2 * Copyright © 2015 Intel Corporation.
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.
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
13 * Authors: David Woodhouse <dwmw2@infradead.org>
16 #include <linux/intel-iommu.h>
17 #include <linux/mmu_notifier.h>
18 #include <linux/sched.h>
19 #include <linux/sched/mm.h>
20 #include <linux/slab.h>
21 #include <linux/intel-svm.h>
22 #include <linux/rculist.h>
23 #include <linux/pci.h>
24 #include <linux/pci-ats.h>
25 #include <linux/dmar.h>
26 #include <linux/interrupt.h>
29 #include "intel-pasid.h"
31 #define PASID_ENTRY_P BIT_ULL(0)
32 #define PASID_ENTRY_FLPM_5LP BIT_ULL(9)
33 #define PASID_ENTRY_SRE BIT_ULL(11)
35 static irqreturn_t prq_event_thread(int irq, void *d);
37 struct pasid_state_entry {
41 int intel_svm_init(struct intel_iommu *iommu)
46 if (cpu_feature_enabled(X86_FEATURE_GBPAGES) &&
47 !cap_fl1gp_support(iommu->cap))
50 if (cpu_feature_enabled(X86_FEATURE_LA57) &&
51 !cap_5lp_support(iommu->cap))
54 /* Start at 2 because it's defined as 2^(1+PSS) */
55 iommu->pasid_max = 2 << ecap_pss(iommu->ecap);
57 /* Eventually I'm promised we will get a multi-level PASID table
58 * and it won't have to be physically contiguous. Until then,
59 * limit the size because 8MiB contiguous allocations can be hard
60 * to come by. The limit of 0x20000, which is 1MiB for each of
61 * the PASID and PASID-state tables, is somewhat arbitrary. */
62 if (iommu->pasid_max > 0x20000)
63 iommu->pasid_max = 0x20000;
65 order = get_order(sizeof(struct pasid_entry) * iommu->pasid_max);
66 if (ecap_dis(iommu->ecap)) {
67 /* Just making it explicit... */
68 BUILD_BUG_ON(sizeof(struct pasid_entry) != sizeof(struct pasid_state_entry));
69 pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, order);
71 iommu->pasid_state_table = page_address(pages);
73 pr_warn("IOMMU: %s: Failed to allocate PASID state table\n",
80 int intel_svm_exit(struct intel_iommu *iommu)
82 int order = get_order(sizeof(struct pasid_entry) * iommu->pasid_max);
84 if (iommu->pasid_state_table) {
85 free_pages((unsigned long)iommu->pasid_state_table, order);
86 iommu->pasid_state_table = NULL;
94 int intel_svm_enable_prq(struct intel_iommu *iommu)
99 pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, PRQ_ORDER);
101 pr_warn("IOMMU: %s: Failed to allocate page request queue\n",
105 iommu->prq = page_address(pages);
107 irq = dmar_alloc_hwirq(DMAR_UNITS_SUPPORTED + iommu->seq_id, iommu->node, iommu);
109 pr_err("IOMMU: %s: Failed to create IRQ vector for page request queue\n",
113 free_pages((unsigned long)iommu->prq, PRQ_ORDER);
119 snprintf(iommu->prq_name, sizeof(iommu->prq_name), "dmar%d-prq", iommu->seq_id);
121 ret = request_threaded_irq(irq, NULL, prq_event_thread, IRQF_ONESHOT,
122 iommu->prq_name, iommu);
124 pr_err("IOMMU: %s: Failed to request IRQ for page request queue\n",
126 dmar_free_hwirq(irq);
130 dmar_writeq(iommu->reg + DMAR_PQH_REG, 0ULL);
131 dmar_writeq(iommu->reg + DMAR_PQT_REG, 0ULL);
132 dmar_writeq(iommu->reg + DMAR_PQA_REG, virt_to_phys(iommu->prq) | PRQ_ORDER);
137 int intel_svm_finish_prq(struct intel_iommu *iommu)
139 dmar_writeq(iommu->reg + DMAR_PQH_REG, 0ULL);
140 dmar_writeq(iommu->reg + DMAR_PQT_REG, 0ULL);
141 dmar_writeq(iommu->reg + DMAR_PQA_REG, 0ULL);
144 free_irq(iommu->pr_irq, iommu);
145 dmar_free_hwirq(iommu->pr_irq);
149 free_pages((unsigned long)iommu->prq, PRQ_ORDER);
155 static void intel_flush_svm_range_dev (struct intel_svm *svm, struct intel_svm_dev *sdev,
156 unsigned long address, unsigned long pages, int ih, int gl)
161 /* For global kernel pages we have to flush them in *all* PASIDs
162 * because that's the only option the hardware gives us. Despite
163 * the fact that they are actually only accessible through one. */
165 desc.low = QI_EIOTLB_PASID(svm->pasid) | QI_EIOTLB_DID(sdev->did) |
166 QI_EIOTLB_GRAN(QI_GRAN_ALL_ALL) | QI_EIOTLB_TYPE;
168 desc.low = QI_EIOTLB_PASID(svm->pasid) | QI_EIOTLB_DID(sdev->did) |
169 QI_EIOTLB_GRAN(QI_GRAN_NONG_PASID) | QI_EIOTLB_TYPE;
172 int mask = ilog2(__roundup_pow_of_two(pages));
174 desc.low = QI_EIOTLB_PASID(svm->pasid) | QI_EIOTLB_DID(sdev->did) |
175 QI_EIOTLB_GRAN(QI_GRAN_PSI_PASID) | QI_EIOTLB_TYPE;
176 desc.high = QI_EIOTLB_ADDR(address) | QI_EIOTLB_GL(gl) |
177 QI_EIOTLB_IH(ih) | QI_EIOTLB_AM(mask);
179 qi_submit_sync(&desc, svm->iommu);
181 if (sdev->dev_iotlb) {
182 desc.low = QI_DEV_EIOTLB_PASID(svm->pasid) | QI_DEV_EIOTLB_SID(sdev->sid) |
183 QI_DEV_EIOTLB_QDEP(sdev->qdep) | QI_DEIOTLB_TYPE;
185 desc.high = QI_DEV_EIOTLB_ADDR(-1ULL >> 1) | QI_DEV_EIOTLB_SIZE;
186 } else if (pages > 1) {
187 /* The least significant zero bit indicates the size. So,
188 * for example, an "address" value of 0x12345f000 will
189 * flush from 0x123440000 to 0x12347ffff (256KiB). */
190 unsigned long last = address + ((unsigned long)(pages - 1) << VTD_PAGE_SHIFT);
191 unsigned long mask = __rounddown_pow_of_two(address ^ last);
193 desc.high = QI_DEV_EIOTLB_ADDR((address & ~mask) | (mask - 1)) | QI_DEV_EIOTLB_SIZE;
195 desc.high = QI_DEV_EIOTLB_ADDR(address);
197 qi_submit_sync(&desc, svm->iommu);
201 static void intel_flush_svm_range(struct intel_svm *svm, unsigned long address,
202 unsigned long pages, int ih, int gl)
204 struct intel_svm_dev *sdev;
206 /* Try deferred invalidate if available */
207 if (svm->iommu->pasid_state_table &&
208 !cmpxchg64(&svm->iommu->pasid_state_table[svm->pasid].val, 0, 1ULL << 63))
212 list_for_each_entry_rcu(sdev, &svm->devs, list)
213 intel_flush_svm_range_dev(svm, sdev, address, pages, ih, gl);
217 static void intel_change_pte(struct mmu_notifier *mn, struct mm_struct *mm,
218 unsigned long address, pte_t pte)
220 struct intel_svm *svm = container_of(mn, struct intel_svm, notifier);
222 intel_flush_svm_range(svm, address, 1, 1, 0);
225 /* Pages have been freed at this point */
226 static void intel_invalidate_range(struct mmu_notifier *mn,
227 struct mm_struct *mm,
228 unsigned long start, unsigned long end)
230 struct intel_svm *svm = container_of(mn, struct intel_svm, notifier);
232 intel_flush_svm_range(svm, start,
233 (end - start + PAGE_SIZE - 1) >> VTD_PAGE_SHIFT, 0, 0);
237 static void intel_flush_pasid_dev(struct intel_svm *svm, struct intel_svm_dev *sdev, int pasid)
242 desc.low = QI_PC_TYPE | QI_PC_DID(sdev->did) | QI_PC_PASID_SEL | QI_PC_PASID(pasid);
244 qi_submit_sync(&desc, svm->iommu);
247 static void intel_mm_release(struct mmu_notifier *mn, struct mm_struct *mm)
249 struct intel_svm *svm = container_of(mn, struct intel_svm, notifier);
250 struct intel_svm_dev *sdev;
252 /* This might end up being called from exit_mmap(), *before* the page
253 * tables are cleared. And __mmu_notifier_release() will delete us from
254 * the list of notifiers so that our invalidate_range() callback doesn't
255 * get called when the page tables are cleared. So we need to protect
256 * against hardware accessing those page tables.
258 * We do it by clearing the entry in the PASID table and then flushing
259 * the IOTLB and the PASID table caches. This might upset hardware;
260 * perhaps we'll want to point the PASID to a dummy PGD (like the zero
261 * page) so that we end up taking a fault that the hardware really
262 * *has* to handle gracefully without affecting other processes.
265 list_for_each_entry_rcu(sdev, &svm->devs, list) {
266 intel_pasid_clear_entry(sdev->dev, svm->pasid);
267 intel_flush_pasid_dev(svm, sdev, svm->pasid);
268 intel_flush_svm_range_dev(svm, sdev, 0, -1, 0, !svm->mm);
274 static const struct mmu_notifier_ops intel_mmuops = {
275 .flags = MMU_INVALIDATE_DOES_NOT_BLOCK,
276 .release = intel_mm_release,
277 .change_pte = intel_change_pte,
278 .invalidate_range = intel_invalidate_range,
281 static DEFINE_MUTEX(pasid_mutex);
282 static LIST_HEAD(global_svm_list);
284 int intel_svm_bind_mm(struct device *dev, int *pasid, int flags, struct svm_dev_ops *ops)
286 struct intel_iommu *iommu = intel_svm_device_to_iommu(dev);
287 struct pasid_entry *entry;
288 struct intel_svm_dev *sdev;
289 struct intel_svm *svm = NULL;
290 struct mm_struct *mm = NULL;
298 if (dev_is_pci(dev)) {
299 pasid_max = pci_max_pasids(to_pci_dev(dev));
305 if (flags & SVM_FLAG_SUPERVISOR_MODE) {
306 if (!ecap_srs(iommu->ecap))
309 mm = get_task_mm(current);
313 mutex_lock(&pasid_mutex);
314 if (pasid && !(flags & SVM_FLAG_PRIVATE_PASID)) {
317 list_for_each_entry(t, &global_svm_list, list) {
318 if (t->mm != mm || (t->flags & SVM_FLAG_PRIVATE_PASID))
322 if (svm->pasid >= pasid_max) {
324 "Limited PASID width. Cannot use existing PASID %d\n",
330 list_for_each_entry(sdev, &svm->devs, list) {
331 if (dev == sdev->dev) {
332 if (sdev->ops != ops) {
345 sdev = kzalloc(sizeof(*sdev), GFP_KERNEL);
352 ret = intel_iommu_enable_pasid(iommu, sdev);
354 /* If they don't actually want to assign a PASID, this is
355 * just an enabling check/preparation. */
359 /* Finish the setup now we know we're keeping it */
362 init_rcu_head(&sdev->rcu);
365 svm = kzalloc(sizeof(*svm), GFP_KERNEL);
373 if (pasid_max > intel_pasid_max_id)
374 pasid_max = intel_pasid_max_id;
376 /* Do not use PASID 0 in caching mode (virtualised IOMMU) */
377 ret = intel_pasid_alloc_id(svm,
378 !!cap_caching_mode(iommu->cap),
379 pasid_max - 1, GFP_KERNEL);
386 svm->notifier.ops = &intel_mmuops;
389 INIT_LIST_HEAD_RCU(&svm->devs);
390 INIT_LIST_HEAD(&svm->list);
393 ret = mmu_notifier_register(&svm->notifier, mm);
395 intel_pasid_free_id(svm->pasid);
400 pasid_entry_val = (u64)__pa(mm->pgd) | PASID_ENTRY_P;
402 pasid_entry_val = (u64)__pa(init_mm.pgd) |
403 PASID_ENTRY_P | PASID_ENTRY_SRE;
404 if (cpu_feature_enabled(X86_FEATURE_LA57))
405 pasid_entry_val |= PASID_ENTRY_FLPM_5LP;
407 entry = intel_pasid_get_entry(dev, svm->pasid);
408 entry->val = pasid_entry_val;
413 * Flush PASID cache when a PASID table entry becomes
416 if (cap_caching_mode(iommu->cap))
417 intel_flush_pasid_dev(svm, sdev, svm->pasid);
419 list_add_tail(&svm->list, &global_svm_list);
421 list_add_rcu(&sdev->list, &svm->devs);
427 mutex_unlock(&pasid_mutex);
432 EXPORT_SYMBOL_GPL(intel_svm_bind_mm);
434 int intel_svm_unbind_mm(struct device *dev, int pasid)
436 struct intel_svm_dev *sdev;
437 struct intel_iommu *iommu;
438 struct intel_svm *svm;
441 mutex_lock(&pasid_mutex);
442 iommu = intel_svm_device_to_iommu(dev);
446 svm = intel_pasid_lookup_id(pasid);
450 list_for_each_entry(sdev, &svm->devs, list) {
451 if (dev == sdev->dev) {
455 list_del_rcu(&sdev->list);
456 /* Flush the PASID cache and IOTLB for this device.
457 * Note that we do depend on the hardware *not* using
458 * the PASID any more. Just as we depend on other
459 * devices never using PASIDs that they have no right
460 * to use. We have a *shared* PASID table, because it's
461 * large and has to be physically contiguous. So it's
462 * hard to be as defensive as we might like. */
463 intel_flush_pasid_dev(svm, sdev, svm->pasid);
464 intel_flush_svm_range_dev(svm, sdev, 0, -1, 0, !svm->mm);
465 kfree_rcu(sdev, rcu);
466 intel_pasid_clear_entry(dev, svm->pasid);
468 if (list_empty(&svm->devs)) {
469 intel_pasid_free_id(svm->pasid);
471 mmu_notifier_unregister(&svm->notifier, svm->mm);
473 list_del(&svm->list);
475 /* We mandate that no page faults may be outstanding
476 * for the PASID when intel_svm_unbind_mm() is called.
477 * If that is not obeyed, subtle errors will happen.
478 * Let's make them less subtle... */
479 memset(svm, 0x6b, sizeof(*svm));
487 mutex_unlock(&pasid_mutex);
491 EXPORT_SYMBOL_GPL(intel_svm_unbind_mm);
493 int intel_svm_is_pasid_valid(struct device *dev, int pasid)
495 struct intel_iommu *iommu;
496 struct intel_svm *svm;
499 mutex_lock(&pasid_mutex);
500 iommu = intel_svm_device_to_iommu(dev);
504 svm = intel_pasid_lookup_id(pasid);
508 /* init_mm is used in this case */
511 else if (atomic_read(&svm->mm->mm_users) > 0)
517 mutex_unlock(&pasid_mutex);
521 EXPORT_SYMBOL_GPL(intel_svm_is_pasid_valid);
523 /* Page request queue descriptor */
524 struct page_req_dsc {
541 #define PRQ_RING_MASK ((0x1000 << PRQ_ORDER) - 0x10)
543 static bool access_error(struct vm_area_struct *vma, struct page_req_dsc *req)
545 unsigned long requested = 0;
548 requested |= VM_EXEC;
551 requested |= VM_READ;
554 requested |= VM_WRITE;
556 return (requested & ~vma->vm_flags) != 0;
559 static bool is_canonical_address(u64 addr)
561 int shift = 64 - (__VIRTUAL_MASK_SHIFT + 1);
562 long saddr = (long) addr;
564 return (((saddr << shift) >> shift) == saddr);
567 static irqreturn_t prq_event_thread(int irq, void *d)
569 struct intel_iommu *iommu = d;
570 struct intel_svm *svm = NULL;
571 int head, tail, handled = 0;
573 /* Clear PPR bit before reading head/tail registers, to
574 * ensure that we get a new interrupt if needed. */
575 writel(DMA_PRS_PPR, iommu->reg + DMAR_PRS_REG);
577 tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
578 head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
579 while (head != tail) {
580 struct intel_svm_dev *sdev;
581 struct vm_area_struct *vma;
582 struct page_req_dsc *req;
589 req = &iommu->prq[head / sizeof(*req)];
591 result = QI_RESP_FAILURE;
592 address = (u64)req->addr << VTD_PAGE_SHIFT;
593 if (!req->pasid_present) {
594 pr_err("%s: Page request without PASID: %08llx %08llx\n",
595 iommu->name, ((unsigned long long *)req)[0],
596 ((unsigned long long *)req)[1]);
600 if (!svm || svm->pasid != req->pasid) {
602 svm = intel_pasid_lookup_id(req->pasid);
603 /* It *can't* go away, because the driver is not permitted
604 * to unbind the mm while any page faults are outstanding.
605 * So we only need RCU to protect the internal idr code. */
609 pr_err("%s: Page request for invalid PASID %d: %08llx %08llx\n",
610 iommu->name, req->pasid, ((unsigned long long *)req)[0],
611 ((unsigned long long *)req)[1]);
616 result = QI_RESP_INVALID;
617 /* Since we're using init_mm.pgd directly, we should never take
618 * any faults on kernel addresses. */
621 /* If the mm is already defunct, don't handle faults. */
622 if (!mmget_not_zero(svm->mm))
625 /* If address is not canonical, return invalid response */
626 if (!is_canonical_address(address))
629 down_read(&svm->mm->mmap_sem);
630 vma = find_extend_vma(svm->mm, address);
631 if (!vma || address < vma->vm_start)
634 if (access_error(vma, req))
637 ret = handle_mm_fault(vma, address,
638 req->wr_req ? FAULT_FLAG_WRITE : 0);
639 if (ret & VM_FAULT_ERROR)
642 result = QI_RESP_SUCCESS;
644 up_read(&svm->mm->mmap_sem);
647 /* Accounting for major/minor faults? */
649 list_for_each_entry_rcu(sdev, &svm->devs, list) {
650 if (sdev->sid == PCI_DEVID(req->bus, req->devfn))
653 /* Other devices can go away, but the drivers are not permitted
654 * to unbind while any page faults might be in flight. So it's
655 * OK to drop the 'lock' here now we have it. */
658 if (WARN_ON(&sdev->list == &svm->devs))
661 if (sdev && sdev->ops && sdev->ops->fault_cb) {
662 int rwxp = (req->rd_req << 3) | (req->wr_req << 2) |
663 (req->exe_req << 1) | (req->priv_req);
664 sdev->ops->fault_cb(sdev->dev, req->pasid, req->addr, req->private, rwxp, result);
666 /* We get here in the error case where the PASID lookup failed,
667 and these can be NULL. Do not use them below this point! */
672 /* Page Group Response */
673 resp.low = QI_PGRP_PASID(req->pasid) |
674 QI_PGRP_DID((req->bus << 8) | req->devfn) |
675 QI_PGRP_PASID_P(req->pasid_present) |
677 resp.high = QI_PGRP_IDX(req->prg_index) |
678 QI_PGRP_PRIV(req->private) | QI_PGRP_RESP_CODE(result);
680 qi_submit_sync(&resp, iommu);
681 } else if (req->srr) {
682 /* Page Stream Response */
683 resp.low = QI_PSTRM_IDX(req->prg_index) |
684 QI_PSTRM_PRIV(req->private) | QI_PSTRM_BUS(req->bus) |
685 QI_PSTRM_PASID(req->pasid) | QI_PSTRM_RESP_TYPE;
686 resp.high = QI_PSTRM_ADDR(address) | QI_PSTRM_DEVFN(req->devfn) |
687 QI_PSTRM_RESP_CODE(result);
689 qi_submit_sync(&resp, iommu);
692 head = (head + sizeof(*req)) & PRQ_RING_MASK;
695 dmar_writeq(iommu->reg + DMAR_PQH_REG, tail);
697 return IRQ_RETVAL(handled);