Merge tag 'iio-fixes-for-4.4a' of ssh://ra.kernel.org/pub/scm/linux/kernel/git/jic23...
[firefly-linux-kernel-4.4.55.git] / drivers / iommu / intel-svm.c
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
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>
26
27 static irqreturn_t prq_event_thread(int irq, void *d);
28
29 struct pasid_entry {
30         u64 val;
31 };
32
33 struct pasid_state_entry {
34         u64 val;
35 };
36
37 int intel_svm_alloc_pasid_tables(struct intel_iommu *iommu)
38 {
39         struct page *pages;
40         int order;
41
42         order = ecap_pss(iommu->ecap) + 7 - PAGE_SHIFT;
43         if (order < 0)
44                 order = 0;
45
46         pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, order);
47         if (!pages) {
48                 pr_warn("IOMMU: %s: Failed to allocate PASID table\n",
49                         iommu->name);
50                 return -ENOMEM;
51         }
52         iommu->pasid_table = page_address(pages);
53         pr_info("%s: Allocated order %d PASID table.\n", iommu->name, order);
54
55         if (ecap_dis(iommu->ecap)) {
56                 pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, order);
57                 if (pages)
58                         iommu->pasid_state_table = page_address(pages);
59                 else
60                         pr_warn("IOMMU: %s: Failed to allocate PASID state table\n",
61                                 iommu->name);
62         }
63
64         idr_init(&iommu->pasid_idr);
65
66         return 0;
67 }
68
69 int intel_svm_free_pasid_tables(struct intel_iommu *iommu)
70 {
71         int order;
72
73         order = ecap_pss(iommu->ecap) + 7 - PAGE_SHIFT;
74         if (order < 0)
75                 order = 0;
76
77         if (iommu->pasid_table) {
78                 free_pages((unsigned long)iommu->pasid_table, order);
79                 iommu->pasid_table = NULL;
80         }
81         if (iommu->pasid_state_table) {
82                 free_pages((unsigned long)iommu->pasid_state_table, order);
83                 iommu->pasid_state_table = NULL;
84         }
85         idr_destroy(&iommu->pasid_idr);
86         return 0;
87 }
88
89 #define PRQ_ORDER 0
90
91 int intel_svm_enable_prq(struct intel_iommu *iommu)
92 {
93         struct page *pages;
94         int irq, ret;
95
96         pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, PRQ_ORDER);
97         if (!pages) {
98                 pr_warn("IOMMU: %s: Failed to allocate page request queue\n",
99                         iommu->name);
100                 return -ENOMEM;
101         }
102         iommu->prq = page_address(pages);
103
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
116         snprintf(iommu->prq_name, sizeof(iommu->prq_name), "dmar%d-prq", iommu->seq_id);
117
118         ret = request_threaded_irq(irq, NULL, prq_event_thread, IRQF_ONESHOT,
119                                    iommu->prq_name, iommu);
120         if (ret) {
121                 pr_err("IOMMU: %s: Failed to request IRQ for page request queue\n",
122                        iommu->name);
123                 dmar_free_hwirq(irq);
124                 goto err;
125         }
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);
129
130         return 0;
131 }
132
133 int intel_svm_finish_prq(struct intel_iommu *iommu)
134 {
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);
138
139         free_irq(iommu->pr_irq, iommu);
140         dmar_free_hwirq(iommu->pr_irq);
141         iommu->pr_irq = 0;
142
143         free_pages((unsigned long)iommu->prq, PRQ_ORDER);
144         iommu->prq = NULL;
145
146         return 0;
147 }
148
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)
151 {
152         struct qi_desc desc;
153
154         if (pages == -1) {
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. */
158                 if (gl)
159                         desc.low = QI_EIOTLB_PASID(svm->pasid) | QI_EIOTLB_DID(sdev->did) |
160                                 QI_EIOTLB_GRAN(QI_GRAN_ALL_ALL) | QI_EIOTLB_TYPE;
161                 else
162                         desc.low = QI_EIOTLB_PASID(svm->pasid) | QI_EIOTLB_DID(sdev->did) |
163                                 QI_EIOTLB_GRAN(QI_GRAN_NONG_PASID) | QI_EIOTLB_TYPE;
164                 desc.high = 0;
165         } else {
166                 int mask = ilog2(__roundup_pow_of_two(pages));
167
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);
172         }
173         qi_submit_sync(&desc, svm->iommu);
174
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;
178                 if (pages == -1) {
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);;
186
187                         desc.high = QI_DEV_EIOTLB_ADDR((address & ~mask) | (mask - 1)) | QI_DEV_EIOTLB_SIZE;
188                 } else {
189                         desc.high = QI_DEV_EIOTLB_ADDR(address);
190                 }
191                 qi_submit_sync(&desc, svm->iommu);
192         }
193 }
194
195 static void intel_flush_svm_range(struct intel_svm *svm, unsigned long address,
196                                   unsigned long pages, int ih, int gl)
197 {
198         struct intel_svm_dev *sdev;
199
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))
203                 return;
204
205         rcu_read_lock();
206         list_for_each_entry_rcu(sdev, &svm->devs, list)
207                 intel_flush_svm_range_dev(svm, sdev, address, pages, ih, gl);
208         rcu_read_unlock();
209 }
210
211 static void intel_change_pte(struct mmu_notifier *mn, struct mm_struct *mm,
212                              unsigned long address, pte_t pte)
213 {
214         struct intel_svm *svm = container_of(mn, struct intel_svm, notifier);
215
216         intel_flush_svm_range(svm, address, 1, 1, 0);
217 }
218
219 static void intel_invalidate_page(struct mmu_notifier *mn, struct mm_struct *mm,
220                                   unsigned long address)
221 {
222         struct intel_svm *svm = container_of(mn, struct intel_svm, notifier);
223
224         intel_flush_svm_range(svm, address, 1, 1, 0);
225 }
226
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)
231 {
232         struct intel_svm *svm = container_of(mn, struct intel_svm, notifier);
233
234         intel_flush_svm_range(svm, start,
235                               (end - start + PAGE_SIZE - 1) >> VTD_PAGE_SHIFT, 0, 0);
236 }
237
238
239 static void intel_flush_pasid_dev(struct intel_svm *svm, struct intel_svm_dev *sdev, int pasid)
240 {
241         struct qi_desc desc;
242
243         desc.high = 0;
244         desc.low = QI_PC_TYPE | QI_PC_DID(sdev->did) | QI_PC_PASID_SEL | QI_PC_PASID(pasid);
245
246         qi_submit_sync(&desc, svm->iommu);
247 }
248
249 static void intel_mm_release(struct mmu_notifier *mn, struct mm_struct *mm)
250 {
251         struct intel_svm *svm = container_of(mn, struct intel_svm, notifier);
252
253         svm->iommu->pasid_table[svm->pasid].val = 0;
254
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));
258 }
259
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,
265 };
266
267 static DEFINE_MUTEX(pasid_mutex);
268
269 int intel_svm_bind_mm(struct device *dev, int *pasid, int flags, struct svm_dev_ops *ops)
270 {
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;
275         int pasid_max;
276         int ret;
277
278         if (WARN_ON(!iommu))
279                 return -EINVAL;
280
281         if (dev_is_pci(dev)) {
282                 pasid_max = pci_max_pasids(to_pci_dev(dev));
283                 if (pasid_max < 0)
284                         return -EINVAL;
285         } else
286                 pasid_max = 1 << 20;
287
288         if ((flags & SVM_FLAG_SUPERVISOR_MODE)) {
289                 if (!ecap_srs(iommu->ecap))
290                         return -EINVAL;
291         } else if (pasid) {
292                 mm = get_task_mm(current);
293                 BUG_ON(!mm);
294         }
295
296         mutex_lock(&pasid_mutex);
297         if (pasid && !(flags & SVM_FLAG_PRIVATE_PASID)) {
298                 int i;
299
300                 idr_for_each_entry(&iommu->pasid_idr, svm, i) {
301                         if (svm->mm != mm ||
302                             (svm->flags & SVM_FLAG_PRIVATE_PASID))
303                                 continue;
304
305                         if (svm->pasid >= pasid_max) {
306                                 dev_warn(dev,
307                                          "Limited PASID width. Cannot use existing PASID %d\n",
308                                          svm->pasid);
309                                 ret = -ENOSPC;
310                                 goto out;
311                         }
312
313                         list_for_each_entry(sdev, &svm->devs, list) {
314                                 if (dev == sdev->dev) {
315                                         if (sdev->ops != ops) {
316                                                 ret = -EBUSY;
317                                                 goto out;
318                                         }
319                                         sdev->users++;
320                                         goto success;
321                                 }
322                         }
323
324                         break;
325                 }
326         }
327
328         sdev = kzalloc(sizeof(*sdev), GFP_KERNEL);
329         if (!sdev) {
330                 ret = -ENOMEM;
331                 goto out;
332         }
333         sdev->dev = dev;
334
335         ret = intel_iommu_enable_pasid(iommu, sdev);
336         if (ret || !pasid) {
337                 /* If they don't actually want to assign a PASID, this is
338                  * just an enabling check/preparation. */
339                 kfree(sdev);
340                 goto out;
341         }
342         /* Finish the setup now we know we're keeping it */
343         sdev->users = 1;
344         sdev->ops = ops;
345         init_rcu_head(&sdev->rcu);
346
347         if (!svm) {
348                 svm = kzalloc(sizeof(*svm), GFP_KERNEL);
349                 if (!svm) {
350                         ret = -ENOMEM;
351                         kfree(sdev);
352                         goto out;
353                 }
354                 svm->iommu = iommu;
355
356                 if (pasid_max > 2 << ecap_pss(iommu->ecap))
357                         pasid_max = 2 << ecap_pss(iommu->ecap);
358
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);
363                 if (ret < 0) {
364                         kfree(svm);
365                         goto out;
366                 }
367                 svm->pasid = ret;
368                 svm->notifier.ops = &intel_mmuops;
369                 svm->mm = mm;
370                 svm->flags = flags;
371                 INIT_LIST_HEAD_RCU(&svm->devs);
372                 ret = -ENOMEM;
373                 if (mm) {
374                         ret = mmu_notifier_register(&svm->notifier, mm);
375                         if (ret) {
376                                 idr_remove(&svm->iommu->pasid_idr, svm->pasid);
377                                 kfree(svm);
378                                 kfree(sdev);
379                                 goto out;
380                         }
381                         iommu->pasid_table[svm->pasid].val = (u64)__pa(mm->pgd) | 1;
382                         mm = NULL;
383                 } else
384                         iommu->pasid_table[svm->pasid].val = (u64)__pa(init_mm.pgd) | 1 | (1ULL << 11);
385                 wmb();
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);
397         }
398         list_add_rcu(&sdev->list, &svm->devs);
399
400  success:
401         *pasid = svm->pasid;
402         ret = 0;
403  out:
404         mutex_unlock(&pasid_mutex);
405         if (mm)
406                 mmput(mm);
407         return ret;
408 }
409 EXPORT_SYMBOL_GPL(intel_svm_bind_mm);
410
411 int intel_svm_unbind_mm(struct device *dev, int pasid)
412 {
413         struct intel_svm_dev *sdev;
414         struct intel_iommu *iommu;
415         struct intel_svm *svm;
416         int ret = -EINVAL;
417
418         mutex_lock(&pasid_mutex);
419         iommu = intel_svm_device_to_iommu(dev);
420         if (!iommu || !iommu->pasid_table)
421                 goto out;
422
423         svm = idr_find(&iommu->pasid_idr, pasid);
424         if (!svm)
425                 goto out;
426
427         list_for_each_entry(sdev, &svm->devs, list) {
428                 if (dev == sdev->dev) {
429                         ret = 0;
430                         sdev->users--;
431                         if (!sdev->users) {
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);
443
444                                 if (list_empty(&svm->devs)) {
445                                         mmu_notifier_unregister(&svm->notifier, svm->mm);
446
447                                         idr_remove(&svm->iommu->pasid_idr, svm->pasid);
448                                         if (svm->mm)
449                                                 mmput(svm->mm);
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));
455                                         kfree(svm);
456                                 }
457                         }
458                         break;
459                 }
460         }
461  out:
462         mutex_unlock(&pasid_mutex);
463
464         return ret;
465 }
466 EXPORT_SYMBOL_GPL(intel_svm_unbind_mm);
467
468 /* Page request queue descriptor */
469 struct page_req_dsc {
470         u64 srr:1;
471         u64 bof:1;
472         u64 pasid_present:1;
473         u64 lpig:1;
474         u64 pasid:20;
475         u64 bus:8;
476         u64 private:23;
477         u64 prg_index:9;
478         u64 rd_req:1;
479         u64 wr_req:1;
480         u64 exe_req:1;
481         u64 priv_req:1;
482         u64 devfn:8;
483         u64 addr:52;
484 };
485
486 #define PRQ_RING_MASK ((0x1000 << PRQ_ORDER) - 0x10)
487 static irqreturn_t prq_event_thread(int irq, void *d)
488 {
489         struct intel_iommu *iommu = d;
490         struct intel_svm *svm = NULL;
491         int head, tail, handled = 0;
492
493         tail = dmar_readq(iommu->reg + DMAR_PQT_REG) & PRQ_RING_MASK;
494         head = dmar_readq(iommu->reg + DMAR_PQH_REG) & PRQ_RING_MASK;
495         while (head != tail) {
496                 struct intel_svm_dev *sdev;
497                 struct vm_area_struct *vma;
498                 struct page_req_dsc *req;
499                 struct qi_desc resp;
500                 int ret, result;
501                 u64 address;
502
503                 handled = 1;
504
505                 req = &iommu->prq[head / sizeof(*req)];
506
507                 result = QI_RESP_FAILURE;
508                 address = (u64)req->addr << VTD_PAGE_SHIFT;
509                 if (!req->pasid_present) {
510                         pr_err("%s: Page request without PASID: %08llx %08llx\n",
511                                iommu->name, ((unsigned long long *)req)[0],
512                                ((unsigned long long *)req)[1]);
513                         goto bad_req;
514                 }
515
516                 if (!svm || svm->pasid != req->pasid) {
517                         rcu_read_lock();
518                         svm = idr_find(&iommu->pasid_idr, req->pasid);
519                         /* It *can't* go away, because the driver is not permitted
520                          * to unbind the mm while any page faults are outstanding.
521                          * So we only need RCU to protect the internal idr code. */
522                         rcu_read_unlock();
523
524                         if (!svm) {
525                                 pr_err("%s: Page request for invalid PASID %d: %08llx %08llx\n",
526                                        iommu->name, req->pasid, ((unsigned long long *)req)[0],
527                                        ((unsigned long long *)req)[1]);
528                                 goto no_pasid;
529                         }
530                 }
531
532                 result = QI_RESP_INVALID;
533                 /* Since we're using init_mm.pgd directly, we should never take
534                  * any faults on kernel addresses. */
535                 if (!svm->mm)
536                         goto bad_req;
537                 down_read(&svm->mm->mmap_sem);
538                 vma = find_extend_vma(svm->mm, address);
539                 if (!vma || address < vma->vm_start)
540                         goto invalid;
541
542                 ret = handle_mm_fault(svm->mm, vma, address,
543                                       req->wr_req ? FAULT_FLAG_WRITE : 0);
544                 if (ret & VM_FAULT_ERROR)
545                         goto invalid;
546
547                 result = QI_RESP_SUCCESS;
548         invalid:
549                 up_read(&svm->mm->mmap_sem);
550         bad_req:
551                 /* Accounting for major/minor faults? */
552                 rcu_read_lock();
553                 list_for_each_entry_rcu(sdev, &svm->devs, list) {
554                         if (sdev->sid == PCI_DEVID(req->bus, req->devfn))
555                                 break;
556                 }
557                 /* Other devices can go away, but the drivers are not permitted
558                  * to unbind while any page faults might be in flight. So it's
559                  * OK to drop the 'lock' here now we have it. */
560                 rcu_read_unlock();
561
562                 if (WARN_ON(&sdev->list == &svm->devs))
563                         sdev = NULL;
564
565                 if (sdev && sdev->ops && sdev->ops->fault_cb) {
566                         int rwxp = (req->rd_req << 3) | (req->wr_req << 2) |
567                                 (req->exe_req << 1) | (req->priv_req);
568                         sdev->ops->fault_cb(sdev->dev, req->pasid, req->addr, req->private, rwxp, result);
569                 }
570                 /* We get here in the error case where the PASID lookup failed,
571                    and these can be NULL. Do not use them below this point! */
572                 sdev = NULL;
573                 svm = NULL;
574         no_pasid:
575                 if (req->lpig) {
576                         /* Page Group Response */
577                         resp.low = QI_PGRP_PASID(req->pasid) |
578                                 QI_PGRP_DID((req->bus << 8) | req->devfn) |
579                                 QI_PGRP_PASID_P(req->pasid_present) |
580                                 QI_PGRP_RESP_TYPE;
581                         resp.high = QI_PGRP_IDX(req->prg_index) |
582                                 QI_PGRP_PRIV(req->private) | QI_PGRP_RESP_CODE(result);
583
584                         qi_submit_sync(&resp, iommu);
585                 } else if (req->srr) {
586                         /* Page Stream Response */
587                         resp.low = QI_PSTRM_IDX(req->prg_index) |
588                                 QI_PSTRM_PRIV(req->private) | QI_PSTRM_BUS(req->bus) |
589                                 QI_PSTRM_PASID(req->pasid) | QI_PSTRM_RESP_TYPE;
590                         resp.high = QI_PSTRM_ADDR(address) | QI_PSTRM_DEVFN(req->devfn) |
591                                 QI_PSTRM_RESP_CODE(result);
592
593                         qi_submit_sync(&resp, iommu);
594                 }
595
596                 head = (head + sizeof(*req)) & PRQ_RING_MASK;
597         }
598
599         dmar_writeq(iommu->reg + DMAR_PQH_REG, tail);
600
601         return IRQ_RETVAL(handled);
602 }