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>
27 #include <linux/mm_types.h>
30 #include "intel-pasid.h"
32 static irqreturn_t prq_event_thread(int irq, void *d);
34 int intel_svm_init(struct intel_iommu *iommu)
36 if (cpu_feature_enabled(X86_FEATURE_GBPAGES) &&
37 !cap_fl1gp_support(iommu->cap))
40 if (cpu_feature_enabled(X86_FEATURE_LA57) &&
41 !cap_5lp_support(iommu->cap))
49 int intel_svm_enable_prq(struct intel_iommu *iommu)
54 pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, PRQ_ORDER);
56 pr_warn("IOMMU: %s: Failed to allocate page request queue\n",
60 iommu->prq = page_address(pages);
62 irq = dmar_alloc_hwirq(DMAR_UNITS_SUPPORTED + iommu->seq_id, iommu->node, iommu);
64 pr_err("IOMMU: %s: Failed to create IRQ vector for page request queue\n",
68 free_pages((unsigned long)iommu->prq, PRQ_ORDER);
74 snprintf(iommu->prq_name, sizeof(iommu->prq_name), "dmar%d-prq", iommu->seq_id);
76 ret = request_threaded_irq(irq, NULL, prq_event_thread, IRQF_ONESHOT,
77 iommu->prq_name, iommu);
79 pr_err("IOMMU: %s: Failed to request IRQ for page request queue\n",
85 dmar_writeq(iommu->reg + DMAR_PQH_REG, 0ULL);
86 dmar_writeq(iommu->reg + DMAR_PQT_REG, 0ULL);
87 dmar_writeq(iommu->reg + DMAR_PQA_REG, virt_to_phys(iommu->prq) | PRQ_ORDER);
92 int intel_svm_finish_prq(struct intel_iommu *iommu)
94 dmar_writeq(iommu->reg + DMAR_PQH_REG, 0ULL);
95 dmar_writeq(iommu->reg + DMAR_PQT_REG, 0ULL);
96 dmar_writeq(iommu->reg + DMAR_PQA_REG, 0ULL);
99 free_irq(iommu->pr_irq, iommu);
100 dmar_free_hwirq(iommu->pr_irq);
104 free_pages((unsigned long)iommu->prq, PRQ_ORDER);
110 static void intel_flush_svm_range_dev (struct intel_svm *svm, struct intel_svm_dev *sdev,
111 unsigned long address, unsigned long pages, int ih, int gl)
116 /* For global kernel pages we have to flush them in *all* PASIDs
117 * because that's the only option the hardware gives us. Despite
118 * the fact that they are actually only accessible through one. */
120 desc.qw0 = QI_EIOTLB_PASID(svm->pasid) |
121 QI_EIOTLB_DID(sdev->did) |
122 QI_EIOTLB_GRAN(QI_GRAN_ALL_ALL) |
125 desc.qw0 = QI_EIOTLB_PASID(svm->pasid) |
126 QI_EIOTLB_DID(sdev->did) |
127 QI_EIOTLB_GRAN(QI_GRAN_NONG_PASID) |
131 int mask = ilog2(__roundup_pow_of_two(pages));
133 desc.qw0 = QI_EIOTLB_PASID(svm->pasid) |
134 QI_EIOTLB_DID(sdev->did) |
135 QI_EIOTLB_GRAN(QI_GRAN_PSI_PASID) |
137 desc.qw1 = QI_EIOTLB_ADDR(address) |
144 qi_submit_sync(&desc, svm->iommu);
146 if (sdev->dev_iotlb) {
147 desc.qw0 = QI_DEV_EIOTLB_PASID(svm->pasid) |
148 QI_DEV_EIOTLB_SID(sdev->sid) |
149 QI_DEV_EIOTLB_QDEP(sdev->qdep) |
152 desc.qw1 = QI_DEV_EIOTLB_ADDR(-1ULL >> 1) |
154 } else if (pages > 1) {
155 /* The least significant zero bit indicates the size. So,
156 * for example, an "address" value of 0x12345f000 will
157 * flush from 0x123440000 to 0x12347ffff (256KiB). */
158 unsigned long last = address + ((unsigned long)(pages - 1) << VTD_PAGE_SHIFT);
159 unsigned long mask = __rounddown_pow_of_two(address ^ last);
161 desc.qw1 = QI_DEV_EIOTLB_ADDR((address & ~mask) |
162 (mask - 1)) | QI_DEV_EIOTLB_SIZE;
164 desc.qw1 = QI_DEV_EIOTLB_ADDR(address);
168 qi_submit_sync(&desc, svm->iommu);
172 static void intel_flush_svm_range(struct intel_svm *svm, unsigned long address,
173 unsigned long pages, int ih, int gl)
175 struct intel_svm_dev *sdev;
178 list_for_each_entry_rcu(sdev, &svm->devs, list)
179 intel_flush_svm_range_dev(svm, sdev, address, pages, ih, gl);
183 /* Pages have been freed at this point */
184 static void intel_invalidate_range(struct mmu_notifier *mn,
185 struct mm_struct *mm,
186 unsigned long start, unsigned long end)
188 struct intel_svm *svm = container_of(mn, struct intel_svm, notifier);
190 intel_flush_svm_range(svm, start,
191 (end - start + PAGE_SIZE - 1) >> VTD_PAGE_SHIFT, 0, 0);
194 static void intel_mm_release(struct mmu_notifier *mn, struct mm_struct *mm)
196 struct intel_svm *svm = container_of(mn, struct intel_svm, notifier);
197 struct intel_svm_dev *sdev;
199 /* This might end up being called from exit_mmap(), *before* the page
200 * tables are cleared. And __mmu_notifier_release() will delete us from
201 * the list of notifiers so that our invalidate_range() callback doesn't
202 * get called when the page tables are cleared. So we need to protect
203 * against hardware accessing those page tables.
205 * We do it by clearing the entry in the PASID table and then flushing
206 * the IOTLB and the PASID table caches. This might upset hardware;
207 * perhaps we'll want to point the PASID to a dummy PGD (like the zero
208 * page) so that we end up taking a fault that the hardware really
209 * *has* to handle gracefully without affecting other processes.
212 list_for_each_entry_rcu(sdev, &svm->devs, list) {
213 intel_pasid_tear_down_entry(svm->iommu, sdev->dev, svm->pasid);
214 intel_flush_svm_range_dev(svm, sdev, 0, -1, 0, !svm->mm);
220 static const struct mmu_notifier_ops intel_mmuops = {
221 .release = intel_mm_release,
222 .invalidate_range = intel_invalidate_range,
225 static DEFINE_MUTEX(pasid_mutex);
226 static LIST_HEAD(global_svm_list);
228 int intel_svm_bind_mm(struct device *dev, int *pasid, int flags, struct svm_dev_ops *ops)
230 struct intel_iommu *iommu = intel_svm_device_to_iommu(dev);
231 struct device_domain_info *info;
232 struct intel_svm_dev *sdev;
233 struct intel_svm *svm = NULL;
234 struct mm_struct *mm = NULL;
238 if (!iommu || dmar_disabled)
241 if (dev_is_pci(dev)) {
242 pasid_max = pci_max_pasids(to_pci_dev(dev));
248 if (flags & SVM_FLAG_SUPERVISOR_MODE) {
249 if (!ecap_srs(iommu->ecap))
252 mm = get_task_mm(current);
256 mutex_lock(&pasid_mutex);
257 if (pasid && !(flags & SVM_FLAG_PRIVATE_PASID)) {
260 list_for_each_entry(t, &global_svm_list, list) {
261 if (t->mm != mm || (t->flags & SVM_FLAG_PRIVATE_PASID))
265 if (svm->pasid >= pasid_max) {
267 "Limited PASID width. Cannot use existing PASID %d\n",
273 list_for_each_entry(sdev, &svm->devs, list) {
274 if (dev == sdev->dev) {
275 if (sdev->ops != ops) {
288 sdev = kzalloc(sizeof(*sdev), GFP_KERNEL);
295 ret = intel_iommu_enable_pasid(iommu, dev);
297 /* If they don't actually want to assign a PASID, this is
298 * just an enabling check/preparation. */
303 info = dev->archdata.iommu;
304 if (!info || !info->pasid_supported) {
309 sdev->did = FLPT_DEFAULT_DID;
310 sdev->sid = PCI_DEVID(info->bus, info->devfn);
311 if (info->ats_enabled) {
313 sdev->qdep = info->ats_qdep;
314 if (sdev->qdep >= QI_DEV_EIOTLB_MAX_INVS)
318 /* Finish the setup now we know we're keeping it */
321 init_rcu_head(&sdev->rcu);
324 svm = kzalloc(sizeof(*svm), GFP_KERNEL);
332 if (pasid_max > intel_pasid_max_id)
333 pasid_max = intel_pasid_max_id;
335 /* Do not use PASID 0 in caching mode (virtualised IOMMU) */
336 ret = intel_pasid_alloc_id(svm,
337 !!cap_caching_mode(iommu->cap),
338 pasid_max - 1, GFP_KERNEL);
345 svm->notifier.ops = &intel_mmuops;
348 INIT_LIST_HEAD_RCU(&svm->devs);
349 INIT_LIST_HEAD(&svm->list);
352 ret = mmu_notifier_register(&svm->notifier, mm);
354 intel_pasid_free_id(svm->pasid);
361 spin_lock(&iommu->lock);
362 ret = intel_pasid_setup_first_level(iommu, dev,
363 mm ? mm->pgd : init_mm.pgd,
364 svm->pasid, FLPT_DEFAULT_DID,
365 mm ? 0 : PASID_FLAG_SUPERVISOR_MODE);
366 spin_unlock(&iommu->lock);
369 mmu_notifier_unregister(&svm->notifier, mm);
370 intel_pasid_free_id(svm->pasid);
376 list_add_tail(&svm->list, &global_svm_list);
378 list_add_rcu(&sdev->list, &svm->devs);
384 mutex_unlock(&pasid_mutex);
389 EXPORT_SYMBOL_GPL(intel_svm_bind_mm);
391 int intel_svm_unbind_mm(struct device *dev, int pasid)
393 struct intel_svm_dev *sdev;
394 struct intel_iommu *iommu;
395 struct intel_svm *svm;
398 mutex_lock(&pasid_mutex);
399 iommu = intel_svm_device_to_iommu(dev);
403 svm = intel_pasid_lookup_id(pasid);
407 list_for_each_entry(sdev, &svm->devs, list) {
408 if (dev == sdev->dev) {
412 list_del_rcu(&sdev->list);
413 /* Flush the PASID cache and IOTLB for this device.
414 * Note that we do depend on the hardware *not* using
415 * the PASID any more. Just as we depend on other
416 * devices never using PASIDs that they have no right
417 * to use. We have a *shared* PASID table, because it's
418 * large and has to be physically contiguous. So it's
419 * hard to be as defensive as we might like. */
420 intel_pasid_tear_down_entry(iommu, dev, svm->pasid);
421 intel_flush_svm_range_dev(svm, sdev, 0, -1, 0, !svm->mm);
422 kfree_rcu(sdev, rcu);
424 if (list_empty(&svm->devs)) {
425 intel_pasid_free_id(svm->pasid);
427 mmu_notifier_unregister(&svm->notifier, svm->mm);
429 list_del(&svm->list);
431 /* We mandate that no page faults may be outstanding
432 * for the PASID when intel_svm_unbind_mm() is called.
433 * If that is not obeyed, subtle errors will happen.
434 * Let's make them less subtle... */
435 memset(svm, 0x6b, sizeof(*svm));
443 mutex_unlock(&pasid_mutex);
447 EXPORT_SYMBOL_GPL(intel_svm_unbind_mm);
449 int intel_svm_is_pasid_valid(struct device *dev, int pasid)
451 struct intel_iommu *iommu;
452 struct intel_svm *svm;
455 mutex_lock(&pasid_mutex);
456 iommu = intel_svm_device_to_iommu(dev);
460 svm = intel_pasid_lookup_id(pasid);
464 /* init_mm is used in this case */
467 else if (atomic_read(&svm->mm->mm_users) > 0)
473 mutex_unlock(&pasid_mutex);
477 EXPORT_SYMBOL_GPL(intel_svm_is_pasid_valid);
479 /* Page request queue descriptor */
480 struct page_req_dsc {
485 u64 priv_data_present:1;
508 #define PRQ_RING_MASK ((0x1000 << PRQ_ORDER) - 0x10)
510 static bool access_error(struct vm_area_struct *vma, struct page_req_dsc *req)
512 unsigned long requested = 0;
515 requested |= VM_EXEC;
518 requested |= VM_READ;
521 requested |= VM_WRITE;
523 return (requested & ~vma->vm_flags) != 0;
526 static bool is_canonical_address(u64 addr)
528 int shift = 64 - (__VIRTUAL_MASK_SHIFT + 1);
529 long saddr = (long) addr;
531 return (((saddr << shift) >> shift) == saddr);
534 static irqreturn_t prq_event_thread(int irq, void *d)
536 struct intel_iommu *iommu = d;
537 struct intel_svm *svm = NULL;
538 int head, tail, handled = 0;
540 /* Clear PPR bit before reading head/tail registers, to
541 * ensure that we get a new interrupt if needed. */
542 writel(DMA_PRS_PPR, iommu->reg + DMAR_PRS_REG);
544 tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
545 head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
546 while (head != tail) {
547 struct intel_svm_dev *sdev;
548 struct vm_area_struct *vma;
549 struct page_req_dsc *req;
557 req = &iommu->prq[head / sizeof(*req)];
559 result = QI_RESP_FAILURE;
560 address = (u64)req->addr << VTD_PAGE_SHIFT;
561 if (!req->pasid_present) {
562 pr_err("%s: Page request without PASID: %08llx %08llx\n",
563 iommu->name, ((unsigned long long *)req)[0],
564 ((unsigned long long *)req)[1]);
568 if (!svm || svm->pasid != req->pasid) {
570 svm = intel_pasid_lookup_id(req->pasid);
571 /* It *can't* go away, because the driver is not permitted
572 * to unbind the mm while any page faults are outstanding.
573 * So we only need RCU to protect the internal idr code. */
577 pr_err("%s: Page request for invalid PASID %d: %08llx %08llx\n",
578 iommu->name, req->pasid, ((unsigned long long *)req)[0],
579 ((unsigned long long *)req)[1]);
584 result = QI_RESP_INVALID;
585 /* Since we're using init_mm.pgd directly, we should never take
586 * any faults on kernel addresses. */
589 /* If the mm is already defunct, don't handle faults. */
590 if (!mmget_not_zero(svm->mm))
593 /* If address is not canonical, return invalid response */
594 if (!is_canonical_address(address))
597 down_read(&svm->mm->mmap_sem);
598 vma = find_extend_vma(svm->mm, address);
599 if (!vma || address < vma->vm_start)
602 if (access_error(vma, req))
605 ret = handle_mm_fault(vma, address,
606 req->wr_req ? FAULT_FLAG_WRITE : 0);
607 if (ret & VM_FAULT_ERROR)
610 result = QI_RESP_SUCCESS;
612 up_read(&svm->mm->mmap_sem);
615 /* Accounting for major/minor faults? */
617 list_for_each_entry_rcu(sdev, &svm->devs, list) {
618 if (sdev->sid == req->rid)
621 /* Other devices can go away, but the drivers are not permitted
622 * to unbind while any page faults might be in flight. So it's
623 * OK to drop the 'lock' here now we have it. */
626 if (WARN_ON(&sdev->list == &svm->devs))
629 if (sdev && sdev->ops && sdev->ops->fault_cb) {
630 int rwxp = (req->rd_req << 3) | (req->wr_req << 2) |
631 (req->exe_req << 1) | (req->pm_req);
632 sdev->ops->fault_cb(sdev->dev, req->pasid, req->addr,
633 req->priv_data, rwxp, result);
635 /* We get here in the error case where the PASID lookup failed,
636 and these can be NULL. Do not use them below this point! */
640 if (req->lpig || req->priv_data_present) {
642 * Per VT-d spec. v3.0 ch7.7, system software must
643 * respond with page group response if private data
644 * is present (PDP) or last page in group (LPIG) bit
645 * is set. This is an additional VT-d feature beyond
648 resp.qw0 = QI_PGRP_PASID(req->pasid) |
649 QI_PGRP_DID(req->rid) |
650 QI_PGRP_PASID_P(req->pasid_present) |
651 QI_PGRP_PDP(req->pasid_present) |
652 QI_PGRP_RESP_CODE(result) |
654 resp.qw1 = QI_PGRP_IDX(req->prg_index) |
655 QI_PGRP_LPIG(req->lpig);
657 if (req->priv_data_present)
658 memcpy(&resp.qw2, req->priv_data,
659 sizeof(req->priv_data));
663 qi_submit_sync(&resp, iommu);
665 head = (head + sizeof(*req)) & PRQ_RING_MASK;
668 dmar_writeq(iommu->reg + DMAR_PQH_REG, tail);
670 return IRQ_RETVAL(handled);