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/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>
27 static irqreturn_t prq_event_thread(int irq, void *d);
33 struct pasid_state_entry {
37 int intel_svm_alloc_pasid_tables(struct intel_iommu *iommu)
42 order = ecap_pss(iommu->ecap) + 7 - PAGE_SHIFT;
46 pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, order);
48 pr_warn("IOMMU: %s: Failed to allocate PASID table\n",
52 iommu->pasid_table = page_address(pages);
53 pr_info("%s: Allocated order %d PASID table.\n", iommu->name, order);
55 if (ecap_dis(iommu->ecap)) {
56 pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, order);
58 iommu->pasid_state_table = page_address(pages);
60 pr_warn("IOMMU: %s: Failed to allocate PASID state table\n",
64 idr_init(&iommu->pasid_idr);
69 int intel_svm_free_pasid_tables(struct intel_iommu *iommu)
73 order = ecap_pss(iommu->ecap) + 7 - PAGE_SHIFT;
77 if (iommu->pasid_table) {
78 free_pages((unsigned long)iommu->pasid_table, order);
79 iommu->pasid_table = NULL;
81 if (iommu->pasid_state_table) {
82 free_pages((unsigned long)iommu->pasid_state_table, order);
83 iommu->pasid_state_table = NULL;
85 idr_destroy(&iommu->pasid_idr);
91 int intel_svm_enable_prq(struct intel_iommu *iommu)
96 pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, PRQ_ORDER);
98 pr_warn("IOMMU: %s: Failed to allocate page request queue\n",
102 iommu->prq = page_address(pages);
104 irq = dmar_alloc_hwirq(DMAR_UNITS_SUPPORTED + iommu->seq_id, iommu->node, iommu);
106 pr_err("IOMMU: %s: Failed to create IRQ vector for page request queue\n",
110 free_pages((unsigned long)iommu->prq, PRQ_ORDER);
116 snprintf(iommu->prq_name, sizeof(iommu->prq_name), "dmar%d-prq", iommu->seq_id);
118 ret = request_threaded_irq(irq, NULL, prq_event_thread, IRQF_ONESHOT,
119 iommu->prq_name, iommu);
121 pr_err("IOMMU: %s: Failed to request IRQ for page request queue\n",
123 dmar_free_hwirq(irq);
126 dmar_writeq(iommu->reg + DMAR_PQH_REG, 0ULL);
127 dmar_writeq(iommu->reg + DMAR_PQT_REG, 0ULL);
128 dmar_writeq(iommu->reg + DMAR_PQA_REG, virt_to_phys(iommu->prq) | PRQ_ORDER);
133 int intel_svm_finish_prq(struct intel_iommu *iommu)
135 dmar_writeq(iommu->reg + DMAR_PQH_REG, 0ULL);
136 dmar_writeq(iommu->reg + DMAR_PQT_REG, 0ULL);
137 dmar_writeq(iommu->reg + DMAR_PQA_REG, 0ULL);
139 free_irq(iommu->pr_irq, iommu);
140 dmar_free_hwirq(iommu->pr_irq);
143 free_pages((unsigned long)iommu->prq, PRQ_ORDER);
149 static void intel_flush_svm_range_dev (struct intel_svm *svm, struct intel_svm_dev *sdev,
150 unsigned long address, unsigned long pages, int ih, int gl)
155 /* For global kernel pages we have to flush them in *all* PASIDs
156 * because that's the only option the hardware gives us. Despite
157 * the fact that they are actually only accessible through one. */
159 desc.low = QI_EIOTLB_PASID(svm->pasid) | QI_EIOTLB_DID(sdev->did) |
160 QI_EIOTLB_GRAN(QI_GRAN_ALL_ALL) | QI_EIOTLB_TYPE;
162 desc.low = QI_EIOTLB_PASID(svm->pasid) | QI_EIOTLB_DID(sdev->did) |
163 QI_EIOTLB_GRAN(QI_GRAN_NONG_PASID) | QI_EIOTLB_TYPE;
166 int mask = ilog2(__roundup_pow_of_two(pages));
168 desc.low = QI_EIOTLB_PASID(svm->pasid) | QI_EIOTLB_DID(sdev->did) |
169 QI_EIOTLB_GRAN(QI_GRAN_PSI_PASID) | QI_EIOTLB_TYPE;
170 desc.high = QI_EIOTLB_ADDR(address) | QI_EIOTLB_GL(gl) |
171 QI_EIOTLB_IH(ih) | QI_EIOTLB_AM(mask);
173 qi_submit_sync(&desc, svm->iommu);
175 if (sdev->dev_iotlb) {
176 desc.low = QI_DEV_EIOTLB_PASID(svm->pasid) | QI_DEV_EIOTLB_SID(sdev->sid) |
177 QI_DEV_EIOTLB_QDEP(sdev->qdep) | QI_DEIOTLB_TYPE;
179 desc.high = QI_DEV_EIOTLB_ADDR(-1ULL >> 1) | QI_DEV_EIOTLB_SIZE;
180 } else if (pages > 1) {
181 /* The least significant zero bit indicates the size. So,
182 * for example, an "address" value of 0x12345f000 will
183 * flush from 0x123440000 to 0x12347ffff (256KiB). */
184 unsigned long last = address + ((unsigned long)(pages - 1) << VTD_PAGE_SHIFT);
185 unsigned long mask = __rounddown_pow_of_two(address ^ last);;
187 desc.high = QI_DEV_EIOTLB_ADDR((address & ~mask) | (mask - 1)) | QI_DEV_EIOTLB_SIZE;
189 desc.high = QI_DEV_EIOTLB_ADDR(address);
191 qi_submit_sync(&desc, svm->iommu);
195 static void intel_flush_svm_range(struct intel_svm *svm, unsigned long address,
196 unsigned long pages, int ih, int gl)
198 struct intel_svm_dev *sdev;
200 /* Try deferred invalidate if available */
201 if (svm->iommu->pasid_state_table &&
202 !cmpxchg64(&svm->iommu->pasid_state_table[svm->pasid].val, 0, 1ULL << 63))
206 list_for_each_entry_rcu(sdev, &svm->devs, list)
207 intel_flush_svm_range_dev(svm, sdev, address, pages, ih, gl);
211 static void intel_change_pte(struct mmu_notifier *mn, struct mm_struct *mm,
212 unsigned long address, pte_t pte)
214 struct intel_svm *svm = container_of(mn, struct intel_svm, notifier);
216 intel_flush_svm_range(svm, address, 1, 1, 0);
219 static void intel_invalidate_page(struct mmu_notifier *mn, struct mm_struct *mm,
220 unsigned long address)
222 struct intel_svm *svm = container_of(mn, struct intel_svm, notifier);
224 intel_flush_svm_range(svm, address, 1, 1, 0);
227 /* Pages have been freed at this point */
228 static void intel_invalidate_range(struct mmu_notifier *mn,
229 struct mm_struct *mm,
230 unsigned long start, unsigned long end)
232 struct intel_svm *svm = container_of(mn, struct intel_svm, notifier);
234 intel_flush_svm_range(svm, start,
235 (end - start + PAGE_SIZE - 1) >> VTD_PAGE_SHIFT, 0, 0);
239 static void intel_flush_pasid_dev(struct intel_svm *svm, struct intel_svm_dev *sdev, int pasid)
244 desc.low = QI_PC_TYPE | QI_PC_DID(sdev->did) | QI_PC_PASID_SEL | QI_PC_PASID(pasid);
246 qi_submit_sync(&desc, svm->iommu);
249 static void intel_mm_release(struct mmu_notifier *mn, struct mm_struct *mm)
251 struct intel_svm *svm = container_of(mn, struct intel_svm, notifier);
253 svm->iommu->pasid_table[svm->pasid].val = 0;
255 /* There's no need to do any flush because we can't get here if there
256 * are any devices left anyway. */
257 WARN_ON(!list_empty(&svm->devs));
260 static const struct mmu_notifier_ops intel_mmuops = {
261 .release = intel_mm_release,
262 .change_pte = intel_change_pte,
263 .invalidate_page = intel_invalidate_page,
264 .invalidate_range = intel_invalidate_range,
267 static DEFINE_MUTEX(pasid_mutex);
269 int intel_svm_bind_mm(struct device *dev, int *pasid, int flags, struct svm_dev_ops *ops)
271 struct intel_iommu *iommu = intel_svm_device_to_iommu(dev);
272 struct intel_svm_dev *sdev;
273 struct intel_svm *svm = NULL;
274 struct mm_struct *mm = NULL;
281 if (dev_is_pci(dev)) {
282 pasid_max = pci_max_pasids(to_pci_dev(dev));
288 if ((flags & SVM_FLAG_SUPERVISOR_MODE)) {
289 if (!ecap_srs(iommu->ecap))
292 mm = get_task_mm(current);
296 mutex_lock(&pasid_mutex);
297 if (pasid && !(flags & SVM_FLAG_PRIVATE_PASID)) {
300 idr_for_each_entry(&iommu->pasid_idr, svm, i) {
302 (svm->flags & SVM_FLAG_PRIVATE_PASID))
305 if (svm->pasid >= pasid_max) {
307 "Limited PASID width. Cannot use existing PASID %d\n",
313 list_for_each_entry(sdev, &svm->devs, list) {
314 if (dev == sdev->dev) {
315 if (sdev->ops != ops) {
328 sdev = kzalloc(sizeof(*sdev), GFP_KERNEL);
335 ret = intel_iommu_enable_pasid(iommu, sdev);
337 /* If they don't actually want to assign a PASID, this is
338 * just an enabling check/preparation. */
342 /* Finish the setup now we know we're keeping it */
345 init_rcu_head(&sdev->rcu);
348 svm = kzalloc(sizeof(*svm), GFP_KERNEL);
356 if (pasid_max > 2 << ecap_pss(iommu->ecap))
357 pasid_max = 2 << ecap_pss(iommu->ecap);
359 /* Do not use PASID 0 in caching mode (virtualised IOMMU) */
360 ret = idr_alloc(&iommu->pasid_idr, svm,
361 !!cap_caching_mode(iommu->cap),
362 pasid_max - 1, GFP_KERNEL);
368 svm->notifier.ops = &intel_mmuops;
371 INIT_LIST_HEAD_RCU(&svm->devs);
374 ret = mmu_notifier_register(&svm->notifier, mm);
376 idr_remove(&svm->iommu->pasid_idr, svm->pasid);
381 iommu->pasid_table[svm->pasid].val = (u64)__pa(mm->pgd) | 1;
384 iommu->pasid_table[svm->pasid].val = (u64)__pa(init_mm.pgd) | 1 | (1ULL << 11);
386 /* In caching mode, we still have to flush with PASID 0 when
387 * a PASID table entry becomes present. Not entirely clear
388 * *why* that would be the case — surely we could just issue
389 * a flush with the PASID value that we've changed? The PASID
390 * is the index into the table, after all. It's not like domain
391 * IDs in the case of the equivalent context-entry change in
392 * caching mode. And for that matter it's not entirely clear why
393 * a VMM would be in the business of caching the PASID table
394 * anyway. Surely that can be left entirely to the guest? */
395 if (cap_caching_mode(iommu->cap))
396 intel_flush_pasid_dev(svm, sdev, 0);
398 list_add_rcu(&sdev->list, &svm->devs);
404 mutex_unlock(&pasid_mutex);
409 EXPORT_SYMBOL_GPL(intel_svm_bind_mm);
411 int intel_svm_unbind_mm(struct device *dev, int pasid)
413 struct intel_svm_dev *sdev;
414 struct intel_iommu *iommu;
415 struct intel_svm *svm;
418 mutex_lock(&pasid_mutex);
419 iommu = intel_svm_device_to_iommu(dev);
420 if (!iommu || !iommu->pasid_table)
423 svm = idr_find(&iommu->pasid_idr, pasid);
427 list_for_each_entry(sdev, &svm->devs, list) {
428 if (dev == sdev->dev) {
432 list_del_rcu(&sdev->list);
433 /* Flush the PASID cache and IOTLB for this device.
434 * Note that we do depend on the hardware *not* using
435 * the PASID any more. Just as we depend on other
436 * devices never using PASIDs that they have no right
437 * to use. We have a *shared* PASID table, because it's
438 * large and has to be physically contiguous. So it's
439 * hard to be as defensive as we might like. */
440 intel_flush_pasid_dev(svm, sdev, svm->pasid);
441 intel_flush_svm_range_dev(svm, sdev, 0, -1, 0, !svm->mm);
442 kfree_rcu(sdev, rcu);
444 if (list_empty(&svm->devs)) {
445 mmu_notifier_unregister(&svm->notifier, svm->mm);
447 idr_remove(&svm->iommu->pasid_idr, svm->pasid);
450 /* We mandate that no page faults may be outstanding
451 * for the PASID when intel_svm_unbind_mm() is called.
452 * If that is not obeyed, subtle errors will happen.
453 * Let's make them less subtle... */
454 memset(svm, 0x6b, sizeof(*svm));
462 mutex_unlock(&pasid_mutex);
466 EXPORT_SYMBOL_GPL(intel_svm_unbind_mm);
468 /* Page request queue descriptor */
469 struct page_req_dsc {
486 #define PRQ_RING_MASK ((0x1000 << PRQ_ORDER) - 0x10)
488 static bool access_error(struct vm_area_struct *vma, struct page_req_dsc *req)
490 unsigned long requested = 0;
493 requested |= VM_EXEC;
496 requested |= VM_READ;
499 requested |= VM_WRITE;
501 return (requested & ~vma->vm_flags) != 0;
504 static irqreturn_t prq_event_thread(int irq, void *d)
506 struct intel_iommu *iommu = d;
507 struct intel_svm *svm = NULL;
508 int head, tail, handled = 0;
510 tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
511 head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
512 while (head != tail) {
513 struct intel_svm_dev *sdev;
514 struct vm_area_struct *vma;
515 struct page_req_dsc *req;
522 req = &iommu->prq[head / sizeof(*req)];
524 result = QI_RESP_FAILURE;
525 address = (u64)req->addr << VTD_PAGE_SHIFT;
526 if (!req->pasid_present) {
527 pr_err("%s: Page request without PASID: %08llx %08llx\n",
528 iommu->name, ((unsigned long long *)req)[0],
529 ((unsigned long long *)req)[1]);
533 if (!svm || svm->pasid != req->pasid) {
535 svm = idr_find(&iommu->pasid_idr, req->pasid);
536 /* It *can't* go away, because the driver is not permitted
537 * to unbind the mm while any page faults are outstanding.
538 * So we only need RCU to protect the internal idr code. */
542 pr_err("%s: Page request for invalid PASID %d: %08llx %08llx\n",
543 iommu->name, req->pasid, ((unsigned long long *)req)[0],
544 ((unsigned long long *)req)[1]);
549 result = QI_RESP_INVALID;
550 /* Since we're using init_mm.pgd directly, we should never take
551 * any faults on kernel addresses. */
554 down_read(&svm->mm->mmap_sem);
555 vma = find_extend_vma(svm->mm, address);
556 if (!vma || address < vma->vm_start)
559 if (access_error(vma, req))
562 ret = handle_mm_fault(svm->mm, vma, address,
563 req->wr_req ? FAULT_FLAG_WRITE : 0);
564 if (ret & VM_FAULT_ERROR)
567 result = QI_RESP_SUCCESS;
569 up_read(&svm->mm->mmap_sem);
571 /* Accounting for major/minor faults? */
573 list_for_each_entry_rcu(sdev, &svm->devs, list) {
574 if (sdev->sid == PCI_DEVID(req->bus, req->devfn))
577 /* Other devices can go away, but the drivers are not permitted
578 * to unbind while any page faults might be in flight. So it's
579 * OK to drop the 'lock' here now we have it. */
582 if (WARN_ON(&sdev->list == &svm->devs))
585 if (sdev && sdev->ops && sdev->ops->fault_cb) {
586 int rwxp = (req->rd_req << 3) | (req->wr_req << 2) |
587 (req->exe_req << 1) | (req->priv_req);
588 sdev->ops->fault_cb(sdev->dev, req->pasid, req->addr, req->private, rwxp, result);
590 /* We get here in the error case where the PASID lookup failed,
591 and these can be NULL. Do not use them below this point! */
596 /* Page Group Response */
597 resp.low = QI_PGRP_PASID(req->pasid) |
598 QI_PGRP_DID((req->bus << 8) | req->devfn) |
599 QI_PGRP_PASID_P(req->pasid_present) |
601 resp.high = QI_PGRP_IDX(req->prg_index) |
602 QI_PGRP_PRIV(req->private) | QI_PGRP_RESP_CODE(result);
604 qi_submit_sync(&resp, iommu);
605 } else if (req->srr) {
606 /* Page Stream Response */
607 resp.low = QI_PSTRM_IDX(req->prg_index) |
608 QI_PSTRM_PRIV(req->private) | QI_PSTRM_BUS(req->bus) |
609 QI_PSTRM_PASID(req->pasid) | QI_PSTRM_RESP_TYPE;
610 resp.high = QI_PSTRM_ADDR(address) | QI_PSTRM_DEVFN(req->devfn) |
611 QI_PSTRM_RESP_CODE(result);
613 qi_submit_sync(&resp, iommu);
616 head = (head + sizeof(*req)) & PRQ_RING_MASK;
619 dmar_writeq(iommu->reg + DMAR_PQH_REG, tail);
621 return IRQ_RETVAL(handled);