1 /*
2 * VFIO: IOMMU DMA mapping support for TCE on POWER
3 *
4 * Copyright (C) 2013 IBM Corp. All rights reserved.
5 * Author: Alexey Kardashevskiy <aik@ozlabs.ru>
6 *
7 * This program is free software; you can redistribute it and/or modify
8 * it under the terms of the GNU General Public License version 2 as
9 * published by the Free Software Foundation.
10 *
11 * Derived from original vfio_iommu_type1.c:
12 * Copyright (C) 2012 Red Hat, Inc. All rights reserved.
13 * Author: Alex Williamson <alex.williamson@redhat.com>
14 */
15
16 #include <linux/module.h>
17 #include <linux/pci.h>
18 #include <linux/slab.h>
19 #include <linux/uaccess.h>
20 #include <linux/err.h>
21 #include <linux/vfio.h>
22 #include <linux/vmalloc.h>
23 #include <linux/sched/mm.h>
24 #include <linux/sched/signal.h>
25
26 #include <asm/iommu.h>
27 #include <asm/tce.h>
28 #include <asm/mmu_context.h>
29
30 #define DRIVER_VERSION "0.1"
31 #define DRIVER_AUTHOR "aik@ozlabs.ru"
32 #define DRIVER_DESC "VFIO IOMMU SPAPR TCE"
33
34 static void tce_iommu_detach_group(void *iommu_data,
35 struct iommu_group *iommu_group);
36
try_increment_locked_vm(struct mm_struct * mm,long npages)37 static long try_increment_locked_vm(struct mm_struct *mm, long npages)
38 {
39 long ret = 0, locked, lock_limit;
40
41 if (WARN_ON_ONCE(!mm))
42 return -EPERM;
43
44 if (!npages)
45 return 0;
46
47 down_write(&mm->mmap_sem);
48 locked = mm->locked_vm + npages;
49 lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
50 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
51 ret = -ENOMEM;
52 else
53 mm->locked_vm += npages;
54
55 pr_debug("[%d] RLIMIT_MEMLOCK +%ld %ld/%ld%s\n", current->pid,
56 npages << PAGE_SHIFT,
57 mm->locked_vm << PAGE_SHIFT,
58 rlimit(RLIMIT_MEMLOCK),
59 ret ? " - exceeded" : "");
60
61 up_write(&mm->mmap_sem);
62
63 return ret;
64 }
65
decrement_locked_vm(struct mm_struct * mm,long npages)66 static void decrement_locked_vm(struct mm_struct *mm, long npages)
67 {
68 if (!mm || !npages)
69 return;
70
71 down_write(&mm->mmap_sem);
72 if (WARN_ON_ONCE(npages > mm->locked_vm))
73 npages = mm->locked_vm;
74 mm->locked_vm -= npages;
75 pr_debug("[%d] RLIMIT_MEMLOCK -%ld %ld/%ld\n", current->pid,
76 npages << PAGE_SHIFT,
77 mm->locked_vm << PAGE_SHIFT,
78 rlimit(RLIMIT_MEMLOCK));
79 up_write(&mm->mmap_sem);
80 }
81
82 /*
83 * VFIO IOMMU fd for SPAPR_TCE IOMMU implementation
84 *
85 * This code handles mapping and unmapping of user data buffers
86 * into DMA'ble space using the IOMMU
87 */
88
89 struct tce_iommu_group {
90 struct list_head next;
91 struct iommu_group *grp;
92 };
93
94 /*
95 * A container needs to remember which preregistered region it has
96 * referenced to do proper cleanup at the userspace process exit.
97 */
98 struct tce_iommu_prereg {
99 struct list_head next;
100 struct mm_iommu_table_group_mem_t *mem;
101 };
102
103 /*
104 * The container descriptor supports only a single group per container.
105 * Required by the API as the container is not supplied with the IOMMU group
106 * at the moment of initialization.
107 */
108 struct tce_container {
109 struct mutex lock;
110 bool enabled;
111 bool v2;
112 bool def_window_pending;
113 unsigned long locked_pages;
114 struct mm_struct *mm;
115 struct iommu_table *tables[IOMMU_TABLE_GROUP_MAX_TABLES];
116 struct list_head group_list;
117 struct list_head prereg_list;
118 };
119
tce_iommu_mm_set(struct tce_container * container)120 static long tce_iommu_mm_set(struct tce_container *container)
121 {
122 if (container->mm) {
123 if (container->mm == current->mm)
124 return 0;
125 return -EPERM;
126 }
127 BUG_ON(!current->mm);
128 container->mm = current->mm;
129 atomic_inc(&container->mm->mm_count);
130
131 return 0;
132 }
133
tce_iommu_prereg_free(struct tce_container * container,struct tce_iommu_prereg * tcemem)134 static long tce_iommu_prereg_free(struct tce_container *container,
135 struct tce_iommu_prereg *tcemem)
136 {
137 long ret;
138
139 ret = mm_iommu_put(container->mm, tcemem->mem);
140 if (ret)
141 return ret;
142
143 list_del(&tcemem->next);
144 kfree(tcemem);
145
146 return 0;
147 }
148
tce_iommu_unregister_pages(struct tce_container * container,__u64 vaddr,__u64 size)149 static long tce_iommu_unregister_pages(struct tce_container *container,
150 __u64 vaddr, __u64 size)
151 {
152 struct mm_iommu_table_group_mem_t *mem;
153 struct tce_iommu_prereg *tcemem;
154 bool found = false;
155
156 if ((vaddr & ~PAGE_MASK) || (size & ~PAGE_MASK))
157 return -EINVAL;
158
159 mem = mm_iommu_find(container->mm, vaddr, size >> PAGE_SHIFT);
160 if (!mem)
161 return -ENOENT;
162
163 list_for_each_entry(tcemem, &container->prereg_list, next) {
164 if (tcemem->mem == mem) {
165 found = true;
166 break;
167 }
168 }
169
170 if (!found)
171 return -ENOENT;
172
173 return tce_iommu_prereg_free(container, tcemem);
174 }
175
tce_iommu_register_pages(struct tce_container * container,__u64 vaddr,__u64 size)176 static long tce_iommu_register_pages(struct tce_container *container,
177 __u64 vaddr, __u64 size)
178 {
179 long ret = 0;
180 struct mm_iommu_table_group_mem_t *mem = NULL;
181 struct tce_iommu_prereg *tcemem;
182 unsigned long entries = size >> PAGE_SHIFT;
183
184 if ((vaddr & ~PAGE_MASK) || (size & ~PAGE_MASK) ||
185 ((vaddr + size) < vaddr))
186 return -EINVAL;
187
188 mem = mm_iommu_find(container->mm, vaddr, entries);
189 if (mem) {
190 list_for_each_entry(tcemem, &container->prereg_list, next) {
191 if (tcemem->mem == mem)
192 return -EBUSY;
193 }
194 }
195
196 ret = mm_iommu_get(container->mm, vaddr, entries, &mem);
197 if (ret)
198 return ret;
199
200 tcemem = kzalloc(sizeof(*tcemem), GFP_KERNEL);
201 if (!tcemem) {
202 mm_iommu_put(container->mm, mem);
203 return -ENOMEM;
204 }
205
206 tcemem->mem = mem;
207 list_add(&tcemem->next, &container->prereg_list);
208
209 container->enabled = true;
210
211 return 0;
212 }
213
tce_iommu_userspace_view_alloc(struct iommu_table * tbl,struct mm_struct * mm)214 static long tce_iommu_userspace_view_alloc(struct iommu_table *tbl,
215 struct mm_struct *mm)
216 {
217 unsigned long cb = _ALIGN_UP(sizeof(tbl->it_userspace[0]) *
218 tbl->it_size, PAGE_SIZE);
219 unsigned long *uas;
220 long ret;
221
222 BUG_ON(tbl->it_userspace);
223
224 ret = try_increment_locked_vm(mm, cb >> PAGE_SHIFT);
225 if (ret)
226 return ret;
227
228 uas = vzalloc(cb);
229 if (!uas) {
230 decrement_locked_vm(mm, cb >> PAGE_SHIFT);
231 return -ENOMEM;
232 }
233 tbl->it_userspace = uas;
234
235 return 0;
236 }
237
tce_iommu_userspace_view_free(struct iommu_table * tbl,struct mm_struct * mm)238 static void tce_iommu_userspace_view_free(struct iommu_table *tbl,
239 struct mm_struct *mm)
240 {
241 unsigned long cb = _ALIGN_UP(sizeof(tbl->it_userspace[0]) *
242 tbl->it_size, PAGE_SIZE);
243
244 if (!tbl->it_userspace)
245 return;
246
247 vfree(tbl->it_userspace);
248 tbl->it_userspace = NULL;
249 decrement_locked_vm(mm, cb >> PAGE_SHIFT);
250 }
251
tce_page_is_contained(struct page * page,unsigned page_shift)252 static bool tce_page_is_contained(struct page *page, unsigned page_shift)
253 {
254 /*
255 * Check that the TCE table granularity is not bigger than the size of
256 * a page we just found. Otherwise the hardware can get access to
257 * a bigger memory chunk that it should.
258 */
259 return (PAGE_SHIFT + compound_order(compound_head(page))) >= page_shift;
260 }
261
tce_groups_attached(struct tce_container * container)262 static inline bool tce_groups_attached(struct tce_container *container)
263 {
264 return !list_empty(&container->group_list);
265 }
266
tce_iommu_find_table(struct tce_container * container,phys_addr_t ioba,struct iommu_table ** ptbl)267 static long tce_iommu_find_table(struct tce_container *container,
268 phys_addr_t ioba, struct iommu_table **ptbl)
269 {
270 long i;
271
272 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
273 struct iommu_table *tbl = container->tables[i];
274
275 if (tbl) {
276 unsigned long entry = ioba >> tbl->it_page_shift;
277 unsigned long start = tbl->it_offset;
278 unsigned long end = start + tbl->it_size;
279
280 if ((start <= entry) && (entry < end)) {
281 *ptbl = tbl;
282 return i;
283 }
284 }
285 }
286
287 return -1;
288 }
289
tce_iommu_find_free_table(struct tce_container * container)290 static int tce_iommu_find_free_table(struct tce_container *container)
291 {
292 int i;
293
294 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
295 if (!container->tables[i])
296 return i;
297 }
298
299 return -ENOSPC;
300 }
301
tce_iommu_enable(struct tce_container * container)302 static int tce_iommu_enable(struct tce_container *container)
303 {
304 int ret = 0;
305 unsigned long locked;
306 struct iommu_table_group *table_group;
307 struct tce_iommu_group *tcegrp;
308
309 if (container->enabled)
310 return -EBUSY;
311
312 /*
313 * When userspace pages are mapped into the IOMMU, they are effectively
314 * locked memory, so, theoretically, we need to update the accounting
315 * of locked pages on each map and unmap. For powerpc, the map unmap
316 * paths can be very hot, though, and the accounting would kill
317 * performance, especially since it would be difficult to impossible
318 * to handle the accounting in real mode only.
319 *
320 * To address that, rather than precisely accounting every page, we
321 * instead account for a worst case on locked memory when the iommu is
322 * enabled and disabled. The worst case upper bound on locked memory
323 * is the size of the whole iommu window, which is usually relatively
324 * small (compared to total memory sizes) on POWER hardware.
325 *
326 * Also we don't have a nice way to fail on H_PUT_TCE due to ulimits,
327 * that would effectively kill the guest at random points, much better
328 * enforcing the limit based on the max that the guest can map.
329 *
330 * Unfortunately at the moment it counts whole tables, no matter how
331 * much memory the guest has. I.e. for 4GB guest and 4 IOMMU groups
332 * each with 2GB DMA window, 8GB will be counted here. The reason for
333 * this is that we cannot tell here the amount of RAM used by the guest
334 * as this information is only available from KVM and VFIO is
335 * KVM agnostic.
336 *
337 * So we do not allow enabling a container without a group attached
338 * as there is no way to know how much we should increment
339 * the locked_vm counter.
340 */
341 if (!tce_groups_attached(container))
342 return -ENODEV;
343
344 tcegrp = list_first_entry(&container->group_list,
345 struct tce_iommu_group, next);
346 table_group = iommu_group_get_iommudata(tcegrp->grp);
347 if (!table_group)
348 return -ENODEV;
349
350 if (!table_group->tce32_size)
351 return -EPERM;
352
353 ret = tce_iommu_mm_set(container);
354 if (ret)
355 return ret;
356
357 locked = table_group->tce32_size >> PAGE_SHIFT;
358 ret = try_increment_locked_vm(container->mm, locked);
359 if (ret)
360 return ret;
361
362 container->locked_pages = locked;
363
364 container->enabled = true;
365
366 return ret;
367 }
368
tce_iommu_disable(struct tce_container * container)369 static void tce_iommu_disable(struct tce_container *container)
370 {
371 if (!container->enabled)
372 return;
373
374 container->enabled = false;
375
376 BUG_ON(!container->mm);
377 decrement_locked_vm(container->mm, container->locked_pages);
378 }
379
tce_iommu_open(unsigned long arg)380 static void *tce_iommu_open(unsigned long arg)
381 {
382 struct tce_container *container;
383
384 if ((arg != VFIO_SPAPR_TCE_IOMMU) && (arg != VFIO_SPAPR_TCE_v2_IOMMU)) {
385 pr_err("tce_vfio: Wrong IOMMU type\n");
386 return ERR_PTR(-EINVAL);
387 }
388
389 container = kzalloc(sizeof(*container), GFP_KERNEL);
390 if (!container)
391 return ERR_PTR(-ENOMEM);
392
393 mutex_init(&container->lock);
394 INIT_LIST_HEAD_RCU(&container->group_list);
395 INIT_LIST_HEAD_RCU(&container->prereg_list);
396
397 container->v2 = arg == VFIO_SPAPR_TCE_v2_IOMMU;
398
399 return container;
400 }
401
402 static int tce_iommu_clear(struct tce_container *container,
403 struct iommu_table *tbl,
404 unsigned long entry, unsigned long pages);
405 static void tce_iommu_free_table(struct tce_container *container,
406 struct iommu_table *tbl);
407
tce_iommu_release(void * iommu_data)408 static void tce_iommu_release(void *iommu_data)
409 {
410 struct tce_container *container = iommu_data;
411 struct tce_iommu_group *tcegrp;
412 struct tce_iommu_prereg *tcemem, *tmtmp;
413 long i;
414
415 while (tce_groups_attached(container)) {
416 tcegrp = list_first_entry(&container->group_list,
417 struct tce_iommu_group, next);
418 tce_iommu_detach_group(iommu_data, tcegrp->grp);
419 }
420
421 /*
422 * If VFIO created a table, it was not disposed
423 * by tce_iommu_detach_group() so do it now.
424 */
425 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
426 struct iommu_table *tbl = container->tables[i];
427
428 if (!tbl)
429 continue;
430
431 tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
432 tce_iommu_free_table(container, tbl);
433 }
434
435 list_for_each_entry_safe(tcemem, tmtmp, &container->prereg_list, next)
436 WARN_ON(tce_iommu_prereg_free(container, tcemem));
437
438 tce_iommu_disable(container);
439 if (container->mm)
440 mmdrop(container->mm);
441 mutex_destroy(&container->lock);
442
443 kfree(container);
444 }
445
tce_iommu_unuse_page(struct tce_container * container,unsigned long hpa)446 static void tce_iommu_unuse_page(struct tce_container *container,
447 unsigned long hpa)
448 {
449 struct page *page;
450
451 page = pfn_to_page(hpa >> PAGE_SHIFT);
452 put_page(page);
453 }
454
tce_iommu_prereg_ua_to_hpa(struct tce_container * container,unsigned long tce,unsigned long shift,unsigned long * phpa,struct mm_iommu_table_group_mem_t ** pmem)455 static int tce_iommu_prereg_ua_to_hpa(struct tce_container *container,
456 unsigned long tce, unsigned long shift,
457 unsigned long *phpa, struct mm_iommu_table_group_mem_t **pmem)
458 {
459 long ret = 0;
460 struct mm_iommu_table_group_mem_t *mem;
461
462 mem = mm_iommu_lookup(container->mm, tce, 1ULL << shift);
463 if (!mem)
464 return -EINVAL;
465
466 ret = mm_iommu_ua_to_hpa(mem, tce, shift, phpa);
467 if (ret)
468 return -EINVAL;
469
470 *pmem = mem;
471
472 return 0;
473 }
474
tce_iommu_unuse_page_v2(struct tce_container * container,struct iommu_table * tbl,unsigned long entry)475 static void tce_iommu_unuse_page_v2(struct tce_container *container,
476 struct iommu_table *tbl, unsigned long entry)
477 {
478 struct mm_iommu_table_group_mem_t *mem = NULL;
479 int ret;
480 unsigned long hpa = 0;
481 unsigned long *pua = IOMMU_TABLE_USERSPACE_ENTRY(tbl, entry);
482
483 if (!pua)
484 return;
485
486 ret = tce_iommu_prereg_ua_to_hpa(container, *pua, tbl->it_page_shift,
487 &hpa, &mem);
488 if (ret)
489 pr_debug("%s: tce %lx at #%lx was not cached, ret=%d\n",
490 __func__, *pua, entry, ret);
491 if (mem)
492 mm_iommu_mapped_dec(mem);
493
494 *pua = 0;
495 }
496
tce_iommu_clear(struct tce_container * container,struct iommu_table * tbl,unsigned long entry,unsigned long pages)497 static int tce_iommu_clear(struct tce_container *container,
498 struct iommu_table *tbl,
499 unsigned long entry, unsigned long pages)
500 {
501 unsigned long oldhpa;
502 long ret;
503 enum dma_data_direction direction;
504
505 for ( ; pages; --pages, ++entry) {
506 direction = DMA_NONE;
507 oldhpa = 0;
508 ret = iommu_tce_xchg(tbl, entry, &oldhpa, &direction);
509 if (ret)
510 continue;
511
512 if (direction == DMA_NONE)
513 continue;
514
515 if (container->v2) {
516 tce_iommu_unuse_page_v2(container, tbl, entry);
517 continue;
518 }
519
520 tce_iommu_unuse_page(container, oldhpa);
521 }
522
523 return 0;
524 }
525
tce_iommu_use_page(unsigned long tce,unsigned long * hpa)526 static int tce_iommu_use_page(unsigned long tce, unsigned long *hpa)
527 {
528 struct page *page = NULL;
529 enum dma_data_direction direction = iommu_tce_direction(tce);
530
531 if (get_user_pages_fast(tce & PAGE_MASK, 1,
532 direction != DMA_TO_DEVICE, &page) != 1)
533 return -EFAULT;
534
535 *hpa = __pa((unsigned long) page_address(page));
536
537 return 0;
538 }
539
tce_iommu_build(struct tce_container * container,struct iommu_table * tbl,unsigned long entry,unsigned long tce,unsigned long pages,enum dma_data_direction direction)540 static long tce_iommu_build(struct tce_container *container,
541 struct iommu_table *tbl,
542 unsigned long entry, unsigned long tce, unsigned long pages,
543 enum dma_data_direction direction)
544 {
545 long i, ret = 0;
546 struct page *page;
547 unsigned long hpa;
548 enum dma_data_direction dirtmp;
549
550 for (i = 0; i < pages; ++i) {
551 unsigned long offset = tce & IOMMU_PAGE_MASK(tbl) & ~PAGE_MASK;
552
553 ret = tce_iommu_use_page(tce, &hpa);
554 if (ret)
555 break;
556
557 page = pfn_to_page(hpa >> PAGE_SHIFT);
558 if (!tce_page_is_contained(page, tbl->it_page_shift)) {
559 ret = -EPERM;
560 break;
561 }
562
563 hpa |= offset;
564 dirtmp = direction;
565 ret = iommu_tce_xchg(tbl, entry + i, &hpa, &dirtmp);
566 if (ret) {
567 tce_iommu_unuse_page(container, hpa);
568 pr_err("iommu_tce: %s failed ioba=%lx, tce=%lx, ret=%ld\n",
569 __func__, entry << tbl->it_page_shift,
570 tce, ret);
571 break;
572 }
573
574 if (dirtmp != DMA_NONE)
575 tce_iommu_unuse_page(container, hpa);
576
577 tce += IOMMU_PAGE_SIZE(tbl);
578 }
579
580 if (ret)
581 tce_iommu_clear(container, tbl, entry, i);
582
583 return ret;
584 }
585
tce_iommu_build_v2(struct tce_container * container,struct iommu_table * tbl,unsigned long entry,unsigned long tce,unsigned long pages,enum dma_data_direction direction)586 static long tce_iommu_build_v2(struct tce_container *container,
587 struct iommu_table *tbl,
588 unsigned long entry, unsigned long tce, unsigned long pages,
589 enum dma_data_direction direction)
590 {
591 long i, ret = 0;
592 struct page *page;
593 unsigned long hpa;
594 enum dma_data_direction dirtmp;
595
596 if (!tbl->it_userspace) {
597 ret = tce_iommu_userspace_view_alloc(tbl, container->mm);
598 if (ret)
599 return ret;
600 }
601
602 for (i = 0; i < pages; ++i) {
603 struct mm_iommu_table_group_mem_t *mem = NULL;
604 unsigned long *pua = IOMMU_TABLE_USERSPACE_ENTRY(tbl,
605 entry + i);
606
607 ret = tce_iommu_prereg_ua_to_hpa(container,
608 tce, tbl->it_page_shift, &hpa, &mem);
609 if (ret)
610 break;
611
612 page = pfn_to_page(hpa >> PAGE_SHIFT);
613 if (!tce_page_is_contained(page, tbl->it_page_shift)) {
614 ret = -EPERM;
615 break;
616 }
617
618 /* Preserve offset within IOMMU page */
619 hpa |= tce & IOMMU_PAGE_MASK(tbl) & ~PAGE_MASK;
620 dirtmp = direction;
621
622 /* The registered region is being unregistered */
623 if (mm_iommu_mapped_inc(mem))
624 break;
625
626 ret = iommu_tce_xchg(tbl, entry + i, &hpa, &dirtmp);
627 if (ret) {
628 /* dirtmp cannot be DMA_NONE here */
629 tce_iommu_unuse_page_v2(container, tbl, entry + i);
630 pr_err("iommu_tce: %s failed ioba=%lx, tce=%lx, ret=%ld\n",
631 __func__, entry << tbl->it_page_shift,
632 tce, ret);
633 break;
634 }
635
636 if (dirtmp != DMA_NONE)
637 tce_iommu_unuse_page_v2(container, tbl, entry + i);
638
639 *pua = tce;
640
641 tce += IOMMU_PAGE_SIZE(tbl);
642 }
643
644 if (ret)
645 tce_iommu_clear(container, tbl, entry, i);
646
647 return ret;
648 }
649
tce_iommu_create_table(struct tce_container * container,struct iommu_table_group * table_group,int num,__u32 page_shift,__u64 window_size,__u32 levels,struct iommu_table ** ptbl)650 static long tce_iommu_create_table(struct tce_container *container,
651 struct iommu_table_group *table_group,
652 int num,
653 __u32 page_shift,
654 __u64 window_size,
655 __u32 levels,
656 struct iommu_table **ptbl)
657 {
658 long ret, table_size;
659
660 table_size = table_group->ops->get_table_size(page_shift, window_size,
661 levels);
662 if (!table_size)
663 return -EINVAL;
664
665 ret = try_increment_locked_vm(container->mm, table_size >> PAGE_SHIFT);
666 if (ret)
667 return ret;
668
669 ret = table_group->ops->create_table(table_group, num,
670 page_shift, window_size, levels, ptbl);
671
672 WARN_ON(!ret && !(*ptbl)->it_ops->free);
673 WARN_ON(!ret && ((*ptbl)->it_allocated_size != table_size));
674
675 return ret;
676 }
677
tce_iommu_free_table(struct tce_container * container,struct iommu_table * tbl)678 static void tce_iommu_free_table(struct tce_container *container,
679 struct iommu_table *tbl)
680 {
681 unsigned long pages = tbl->it_allocated_size >> PAGE_SHIFT;
682
683 tce_iommu_userspace_view_free(tbl, container->mm);
684 iommu_tce_table_put(tbl);
685 decrement_locked_vm(container->mm, pages);
686 }
687
tce_iommu_create_window(struct tce_container * container,__u32 page_shift,__u64 window_size,__u32 levels,__u64 * start_addr)688 static long tce_iommu_create_window(struct tce_container *container,
689 __u32 page_shift, __u64 window_size, __u32 levels,
690 __u64 *start_addr)
691 {
692 struct tce_iommu_group *tcegrp;
693 struct iommu_table_group *table_group;
694 struct iommu_table *tbl = NULL;
695 long ret, num;
696
697 num = tce_iommu_find_free_table(container);
698 if (num < 0)
699 return num;
700
701 /* Get the first group for ops::create_table */
702 tcegrp = list_first_entry(&container->group_list,
703 struct tce_iommu_group, next);
704 table_group = iommu_group_get_iommudata(tcegrp->grp);
705 if (!table_group)
706 return -EFAULT;
707
708 if (!(table_group->pgsizes & (1ULL << page_shift)))
709 return -EINVAL;
710
711 if (!table_group->ops->set_window || !table_group->ops->unset_window ||
712 !table_group->ops->get_table_size ||
713 !table_group->ops->create_table)
714 return -EPERM;
715
716 /* Create TCE table */
717 ret = tce_iommu_create_table(container, table_group, num,
718 page_shift, window_size, levels, &tbl);
719 if (ret)
720 return ret;
721
722 BUG_ON(!tbl->it_ops->free);
723
724 /*
725 * Program the table to every group.
726 * Groups have been tested for compatibility at the attach time.
727 */
728 list_for_each_entry(tcegrp, &container->group_list, next) {
729 table_group = iommu_group_get_iommudata(tcegrp->grp);
730
731 ret = table_group->ops->set_window(table_group, num, tbl);
732 if (ret)
733 goto unset_exit;
734 }
735
736 container->tables[num] = tbl;
737
738 /* Return start address assigned by platform in create_table() */
739 *start_addr = tbl->it_offset << tbl->it_page_shift;
740
741 return 0;
742
743 unset_exit:
744 list_for_each_entry(tcegrp, &container->group_list, next) {
745 table_group = iommu_group_get_iommudata(tcegrp->grp);
746 table_group->ops->unset_window(table_group, num);
747 }
748 tce_iommu_free_table(container, tbl);
749
750 return ret;
751 }
752
tce_iommu_remove_window(struct tce_container * container,__u64 start_addr)753 static long tce_iommu_remove_window(struct tce_container *container,
754 __u64 start_addr)
755 {
756 struct iommu_table_group *table_group = NULL;
757 struct iommu_table *tbl;
758 struct tce_iommu_group *tcegrp;
759 int num;
760
761 num = tce_iommu_find_table(container, start_addr, &tbl);
762 if (num < 0)
763 return -EINVAL;
764
765 BUG_ON(!tbl->it_size);
766
767 /* Detach groups from IOMMUs */
768 list_for_each_entry(tcegrp, &container->group_list, next) {
769 table_group = iommu_group_get_iommudata(tcegrp->grp);
770
771 /*
772 * SPAPR TCE IOMMU exposes the default DMA window to
773 * the guest via dma32_window_start/size of
774 * VFIO_IOMMU_SPAPR_TCE_GET_INFO. Some platforms allow
775 * the userspace to remove this window, some do not so
776 * here we check for the platform capability.
777 */
778 if (!table_group->ops || !table_group->ops->unset_window)
779 return -EPERM;
780
781 table_group->ops->unset_window(table_group, num);
782 }
783
784 /* Free table */
785 tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
786 tce_iommu_free_table(container, tbl);
787 container->tables[num] = NULL;
788
789 return 0;
790 }
791
tce_iommu_create_default_window(struct tce_container * container)792 static long tce_iommu_create_default_window(struct tce_container *container)
793 {
794 long ret;
795 __u64 start_addr = 0;
796 struct tce_iommu_group *tcegrp;
797 struct iommu_table_group *table_group;
798
799 if (!container->def_window_pending)
800 return 0;
801
802 if (!tce_groups_attached(container))
803 return -ENODEV;
804
805 tcegrp = list_first_entry(&container->group_list,
806 struct tce_iommu_group, next);
807 table_group = iommu_group_get_iommudata(tcegrp->grp);
808 if (!table_group)
809 return -ENODEV;
810
811 ret = tce_iommu_create_window(container, IOMMU_PAGE_SHIFT_4K,
812 table_group->tce32_size, 1, &start_addr);
813 WARN_ON_ONCE(!ret && start_addr);
814
815 if (!ret)
816 container->def_window_pending = false;
817
818 return ret;
819 }
820
tce_iommu_ioctl(void * iommu_data,unsigned int cmd,unsigned long arg)821 static long tce_iommu_ioctl(void *iommu_data,
822 unsigned int cmd, unsigned long arg)
823 {
824 struct tce_container *container = iommu_data;
825 unsigned long minsz, ddwsz;
826 long ret;
827
828 switch (cmd) {
829 case VFIO_CHECK_EXTENSION:
830 switch (arg) {
831 case VFIO_SPAPR_TCE_IOMMU:
832 case VFIO_SPAPR_TCE_v2_IOMMU:
833 ret = 1;
834 break;
835 default:
836 ret = vfio_spapr_iommu_eeh_ioctl(NULL, cmd, arg);
837 break;
838 }
839
840 return (ret < 0) ? 0 : ret;
841 }
842
843 /*
844 * Sanity check to prevent one userspace from manipulating
845 * another userspace mm.
846 */
847 BUG_ON(!container);
848 if (container->mm && container->mm != current->mm)
849 return -EPERM;
850
851 switch (cmd) {
852 case VFIO_IOMMU_SPAPR_TCE_GET_INFO: {
853 struct vfio_iommu_spapr_tce_info info;
854 struct tce_iommu_group *tcegrp;
855 struct iommu_table_group *table_group;
856
857 if (!tce_groups_attached(container))
858 return -ENXIO;
859
860 tcegrp = list_first_entry(&container->group_list,
861 struct tce_iommu_group, next);
862 table_group = iommu_group_get_iommudata(tcegrp->grp);
863
864 if (!table_group)
865 return -ENXIO;
866
867 minsz = offsetofend(struct vfio_iommu_spapr_tce_info,
868 dma32_window_size);
869
870 if (copy_from_user(&info, (void __user *)arg, minsz))
871 return -EFAULT;
872
873 if (info.argsz < minsz)
874 return -EINVAL;
875
876 info.dma32_window_start = table_group->tce32_start;
877 info.dma32_window_size = table_group->tce32_size;
878 info.flags = 0;
879 memset(&info.ddw, 0, sizeof(info.ddw));
880
881 if (table_group->max_dynamic_windows_supported &&
882 container->v2) {
883 info.flags |= VFIO_IOMMU_SPAPR_INFO_DDW;
884 info.ddw.pgsizes = table_group->pgsizes;
885 info.ddw.max_dynamic_windows_supported =
886 table_group->max_dynamic_windows_supported;
887 info.ddw.levels = table_group->max_levels;
888 }
889
890 ddwsz = offsetofend(struct vfio_iommu_spapr_tce_info, ddw);
891
892 if (info.argsz >= ddwsz)
893 minsz = ddwsz;
894
895 if (copy_to_user((void __user *)arg, &info, minsz))
896 return -EFAULT;
897
898 return 0;
899 }
900 case VFIO_IOMMU_MAP_DMA: {
901 struct vfio_iommu_type1_dma_map param;
902 struct iommu_table *tbl = NULL;
903 long num;
904 enum dma_data_direction direction;
905
906 if (!container->enabled)
907 return -EPERM;
908
909 minsz = offsetofend(struct vfio_iommu_type1_dma_map, size);
910
911 if (copy_from_user(¶m, (void __user *)arg, minsz))
912 return -EFAULT;
913
914 if (param.argsz < minsz)
915 return -EINVAL;
916
917 if (param.flags & ~(VFIO_DMA_MAP_FLAG_READ |
918 VFIO_DMA_MAP_FLAG_WRITE))
919 return -EINVAL;
920
921 ret = tce_iommu_create_default_window(container);
922 if (ret)
923 return ret;
924
925 num = tce_iommu_find_table(container, param.iova, &tbl);
926 if (num < 0)
927 return -ENXIO;
928
929 if ((param.size & ~IOMMU_PAGE_MASK(tbl)) ||
930 (param.vaddr & ~IOMMU_PAGE_MASK(tbl)))
931 return -EINVAL;
932
933 /* iova is checked by the IOMMU API */
934 if (param.flags & VFIO_DMA_MAP_FLAG_READ) {
935 if (param.flags & VFIO_DMA_MAP_FLAG_WRITE)
936 direction = DMA_BIDIRECTIONAL;
937 else
938 direction = DMA_TO_DEVICE;
939 } else {
940 if (param.flags & VFIO_DMA_MAP_FLAG_WRITE)
941 direction = DMA_FROM_DEVICE;
942 else
943 return -EINVAL;
944 }
945
946 ret = iommu_tce_put_param_check(tbl, param.iova, param.vaddr);
947 if (ret)
948 return ret;
949
950 if (container->v2)
951 ret = tce_iommu_build_v2(container, tbl,
952 param.iova >> tbl->it_page_shift,
953 param.vaddr,
954 param.size >> tbl->it_page_shift,
955 direction);
956 else
957 ret = tce_iommu_build(container, tbl,
958 param.iova >> tbl->it_page_shift,
959 param.vaddr,
960 param.size >> tbl->it_page_shift,
961 direction);
962
963 iommu_flush_tce(tbl);
964
965 return ret;
966 }
967 case VFIO_IOMMU_UNMAP_DMA: {
968 struct vfio_iommu_type1_dma_unmap param;
969 struct iommu_table *tbl = NULL;
970 long num;
971
972 if (!container->enabled)
973 return -EPERM;
974
975 minsz = offsetofend(struct vfio_iommu_type1_dma_unmap,
976 size);
977
978 if (copy_from_user(¶m, (void __user *)arg, minsz))
979 return -EFAULT;
980
981 if (param.argsz < minsz)
982 return -EINVAL;
983
984 /* No flag is supported now */
985 if (param.flags)
986 return -EINVAL;
987
988 ret = tce_iommu_create_default_window(container);
989 if (ret)
990 return ret;
991
992 num = tce_iommu_find_table(container, param.iova, &tbl);
993 if (num < 0)
994 return -ENXIO;
995
996 if (param.size & ~IOMMU_PAGE_MASK(tbl))
997 return -EINVAL;
998
999 ret = iommu_tce_clear_param_check(tbl, param.iova, 0,
1000 param.size >> tbl->it_page_shift);
1001 if (ret)
1002 return ret;
1003
1004 ret = tce_iommu_clear(container, tbl,
1005 param.iova >> tbl->it_page_shift,
1006 param.size >> tbl->it_page_shift);
1007 iommu_flush_tce(tbl);
1008
1009 return ret;
1010 }
1011 case VFIO_IOMMU_SPAPR_REGISTER_MEMORY: {
1012 struct vfio_iommu_spapr_register_memory param;
1013
1014 if (!container->v2)
1015 break;
1016
1017 minsz = offsetofend(struct vfio_iommu_spapr_register_memory,
1018 size);
1019
1020 ret = tce_iommu_mm_set(container);
1021 if (ret)
1022 return ret;
1023
1024 if (copy_from_user(¶m, (void __user *)arg, minsz))
1025 return -EFAULT;
1026
1027 if (param.argsz < minsz)
1028 return -EINVAL;
1029
1030 /* No flag is supported now */
1031 if (param.flags)
1032 return -EINVAL;
1033
1034 mutex_lock(&container->lock);
1035 ret = tce_iommu_register_pages(container, param.vaddr,
1036 param.size);
1037 mutex_unlock(&container->lock);
1038
1039 return ret;
1040 }
1041 case VFIO_IOMMU_SPAPR_UNREGISTER_MEMORY: {
1042 struct vfio_iommu_spapr_register_memory param;
1043
1044 if (!container->v2)
1045 break;
1046
1047 if (!container->mm)
1048 return -EPERM;
1049
1050 minsz = offsetofend(struct vfio_iommu_spapr_register_memory,
1051 size);
1052
1053 if (copy_from_user(¶m, (void __user *)arg, minsz))
1054 return -EFAULT;
1055
1056 if (param.argsz < minsz)
1057 return -EINVAL;
1058
1059 /* No flag is supported now */
1060 if (param.flags)
1061 return -EINVAL;
1062
1063 mutex_lock(&container->lock);
1064 ret = tce_iommu_unregister_pages(container, param.vaddr,
1065 param.size);
1066 mutex_unlock(&container->lock);
1067
1068 return ret;
1069 }
1070 case VFIO_IOMMU_ENABLE:
1071 if (container->v2)
1072 break;
1073
1074 mutex_lock(&container->lock);
1075 ret = tce_iommu_enable(container);
1076 mutex_unlock(&container->lock);
1077 return ret;
1078
1079
1080 case VFIO_IOMMU_DISABLE:
1081 if (container->v2)
1082 break;
1083
1084 mutex_lock(&container->lock);
1085 tce_iommu_disable(container);
1086 mutex_unlock(&container->lock);
1087 return 0;
1088
1089 case VFIO_EEH_PE_OP: {
1090 struct tce_iommu_group *tcegrp;
1091
1092 ret = 0;
1093 list_for_each_entry(tcegrp, &container->group_list, next) {
1094 ret = vfio_spapr_iommu_eeh_ioctl(tcegrp->grp,
1095 cmd, arg);
1096 if (ret)
1097 return ret;
1098 }
1099 return ret;
1100 }
1101
1102 case VFIO_IOMMU_SPAPR_TCE_CREATE: {
1103 struct vfio_iommu_spapr_tce_create create;
1104
1105 if (!container->v2)
1106 break;
1107
1108 ret = tce_iommu_mm_set(container);
1109 if (ret)
1110 return ret;
1111
1112 if (!tce_groups_attached(container))
1113 return -ENXIO;
1114
1115 minsz = offsetofend(struct vfio_iommu_spapr_tce_create,
1116 start_addr);
1117
1118 if (copy_from_user(&create, (void __user *)arg, minsz))
1119 return -EFAULT;
1120
1121 if (create.argsz < minsz)
1122 return -EINVAL;
1123
1124 if (create.flags)
1125 return -EINVAL;
1126
1127 mutex_lock(&container->lock);
1128
1129 ret = tce_iommu_create_default_window(container);
1130 if (!ret)
1131 ret = tce_iommu_create_window(container,
1132 create.page_shift,
1133 create.window_size, create.levels,
1134 &create.start_addr);
1135
1136 mutex_unlock(&container->lock);
1137
1138 if (!ret && copy_to_user((void __user *)arg, &create, minsz))
1139 ret = -EFAULT;
1140
1141 return ret;
1142 }
1143 case VFIO_IOMMU_SPAPR_TCE_REMOVE: {
1144 struct vfio_iommu_spapr_tce_remove remove;
1145
1146 if (!container->v2)
1147 break;
1148
1149 ret = tce_iommu_mm_set(container);
1150 if (ret)
1151 return ret;
1152
1153 if (!tce_groups_attached(container))
1154 return -ENXIO;
1155
1156 minsz = offsetofend(struct vfio_iommu_spapr_tce_remove,
1157 start_addr);
1158
1159 if (copy_from_user(&remove, (void __user *)arg, minsz))
1160 return -EFAULT;
1161
1162 if (remove.argsz < minsz)
1163 return -EINVAL;
1164
1165 if (remove.flags)
1166 return -EINVAL;
1167
1168 if (container->def_window_pending && !remove.start_addr) {
1169 container->def_window_pending = false;
1170 return 0;
1171 }
1172
1173 mutex_lock(&container->lock);
1174
1175 ret = tce_iommu_remove_window(container, remove.start_addr);
1176
1177 mutex_unlock(&container->lock);
1178
1179 return ret;
1180 }
1181 }
1182
1183 return -ENOTTY;
1184 }
1185
tce_iommu_release_ownership(struct tce_container * container,struct iommu_table_group * table_group)1186 static void tce_iommu_release_ownership(struct tce_container *container,
1187 struct iommu_table_group *table_group)
1188 {
1189 int i;
1190
1191 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1192 struct iommu_table *tbl = container->tables[i];
1193
1194 if (!tbl)
1195 continue;
1196
1197 tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
1198 tce_iommu_userspace_view_free(tbl, container->mm);
1199 if (tbl->it_map)
1200 iommu_release_ownership(tbl);
1201
1202 container->tables[i] = NULL;
1203 }
1204 }
1205
tce_iommu_take_ownership(struct tce_container * container,struct iommu_table_group * table_group)1206 static int tce_iommu_take_ownership(struct tce_container *container,
1207 struct iommu_table_group *table_group)
1208 {
1209 int i, j, rc = 0;
1210
1211 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1212 struct iommu_table *tbl = table_group->tables[i];
1213
1214 if (!tbl || !tbl->it_map)
1215 continue;
1216
1217 rc = iommu_take_ownership(tbl);
1218 if (rc) {
1219 for (j = 0; j < i; ++j)
1220 iommu_release_ownership(
1221 table_group->tables[j]);
1222
1223 return rc;
1224 }
1225 }
1226
1227 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1228 container->tables[i] = table_group->tables[i];
1229
1230 return 0;
1231 }
1232
tce_iommu_release_ownership_ddw(struct tce_container * container,struct iommu_table_group * table_group)1233 static void tce_iommu_release_ownership_ddw(struct tce_container *container,
1234 struct iommu_table_group *table_group)
1235 {
1236 long i;
1237
1238 if (!table_group->ops->unset_window) {
1239 WARN_ON_ONCE(1);
1240 return;
1241 }
1242
1243 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1244 table_group->ops->unset_window(table_group, i);
1245
1246 table_group->ops->release_ownership(table_group);
1247 }
1248
tce_iommu_take_ownership_ddw(struct tce_container * container,struct iommu_table_group * table_group)1249 static long tce_iommu_take_ownership_ddw(struct tce_container *container,
1250 struct iommu_table_group *table_group)
1251 {
1252 long i, ret = 0;
1253
1254 if (!table_group->ops->create_table || !table_group->ops->set_window ||
1255 !table_group->ops->release_ownership) {
1256 WARN_ON_ONCE(1);
1257 return -EFAULT;
1258 }
1259
1260 table_group->ops->take_ownership(table_group);
1261
1262 /* Set all windows to the new group */
1263 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1264 struct iommu_table *tbl = container->tables[i];
1265
1266 if (!tbl)
1267 continue;
1268
1269 ret = table_group->ops->set_window(table_group, i, tbl);
1270 if (ret)
1271 goto release_exit;
1272 }
1273
1274 return 0;
1275
1276 release_exit:
1277 for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1278 table_group->ops->unset_window(table_group, i);
1279
1280 table_group->ops->release_ownership(table_group);
1281
1282 return ret;
1283 }
1284
tce_iommu_attach_group(void * iommu_data,struct iommu_group * iommu_group)1285 static int tce_iommu_attach_group(void *iommu_data,
1286 struct iommu_group *iommu_group)
1287 {
1288 int ret;
1289 struct tce_container *container = iommu_data;
1290 struct iommu_table_group *table_group;
1291 struct tce_iommu_group *tcegrp = NULL;
1292
1293 mutex_lock(&container->lock);
1294
1295 /* pr_debug("tce_vfio: Attaching group #%u to iommu %p\n",
1296 iommu_group_id(iommu_group), iommu_group); */
1297 table_group = iommu_group_get_iommudata(iommu_group);
1298 if (!table_group) {
1299 ret = -ENODEV;
1300 goto unlock_exit;
1301 }
1302
1303 if (tce_groups_attached(container) && (!table_group->ops ||
1304 !table_group->ops->take_ownership ||
1305 !table_group->ops->release_ownership)) {
1306 ret = -EBUSY;
1307 goto unlock_exit;
1308 }
1309
1310 /* Check if new group has the same iommu_ops (i.e. compatible) */
1311 list_for_each_entry(tcegrp, &container->group_list, next) {
1312 struct iommu_table_group *table_group_tmp;
1313
1314 if (tcegrp->grp == iommu_group) {
1315 pr_warn("tce_vfio: Group %d is already attached\n",
1316 iommu_group_id(iommu_group));
1317 ret = -EBUSY;
1318 goto unlock_exit;
1319 }
1320 table_group_tmp = iommu_group_get_iommudata(tcegrp->grp);
1321 if (table_group_tmp->ops->create_table !=
1322 table_group->ops->create_table) {
1323 pr_warn("tce_vfio: Group %d is incompatible with group %d\n",
1324 iommu_group_id(iommu_group),
1325 iommu_group_id(tcegrp->grp));
1326 ret = -EPERM;
1327 goto unlock_exit;
1328 }
1329 }
1330
1331 tcegrp = kzalloc(sizeof(*tcegrp), GFP_KERNEL);
1332 if (!tcegrp) {
1333 ret = -ENOMEM;
1334 goto unlock_exit;
1335 }
1336
1337 if (!table_group->ops || !table_group->ops->take_ownership ||
1338 !table_group->ops->release_ownership) {
1339 if (container->v2) {
1340 ret = -EPERM;
1341 goto unlock_exit;
1342 }
1343 ret = tce_iommu_take_ownership(container, table_group);
1344 } else {
1345 if (!container->v2) {
1346 ret = -EPERM;
1347 goto unlock_exit;
1348 }
1349 ret = tce_iommu_take_ownership_ddw(container, table_group);
1350 if (!tce_groups_attached(container) && !container->tables[0])
1351 container->def_window_pending = true;
1352 }
1353
1354 if (!ret) {
1355 tcegrp->grp = iommu_group;
1356 list_add(&tcegrp->next, &container->group_list);
1357 }
1358
1359 unlock_exit:
1360 if (ret && tcegrp)
1361 kfree(tcegrp);
1362
1363 mutex_unlock(&container->lock);
1364
1365 return ret;
1366 }
1367
tce_iommu_detach_group(void * iommu_data,struct iommu_group * iommu_group)1368 static void tce_iommu_detach_group(void *iommu_data,
1369 struct iommu_group *iommu_group)
1370 {
1371 struct tce_container *container = iommu_data;
1372 struct iommu_table_group *table_group;
1373 bool found = false;
1374 struct tce_iommu_group *tcegrp;
1375
1376 mutex_lock(&container->lock);
1377
1378 list_for_each_entry(tcegrp, &container->group_list, next) {
1379 if (tcegrp->grp == iommu_group) {
1380 found = true;
1381 break;
1382 }
1383 }
1384
1385 if (!found) {
1386 pr_warn("tce_vfio: detaching unattached group #%u\n",
1387 iommu_group_id(iommu_group));
1388 goto unlock_exit;
1389 }
1390
1391 list_del(&tcegrp->next);
1392 kfree(tcegrp);
1393
1394 table_group = iommu_group_get_iommudata(iommu_group);
1395 BUG_ON(!table_group);
1396
1397 if (!table_group->ops || !table_group->ops->release_ownership)
1398 tce_iommu_release_ownership(container, table_group);
1399 else
1400 tce_iommu_release_ownership_ddw(container, table_group);
1401
1402 unlock_exit:
1403 mutex_unlock(&container->lock);
1404 }
1405
1406 const struct vfio_iommu_driver_ops tce_iommu_driver_ops = {
1407 .name = "iommu-vfio-powerpc",
1408 .owner = THIS_MODULE,
1409 .open = tce_iommu_open,
1410 .release = tce_iommu_release,
1411 .ioctl = tce_iommu_ioctl,
1412 .attach_group = tce_iommu_attach_group,
1413 .detach_group = tce_iommu_detach_group,
1414 };
1415
tce_iommu_init(void)1416 static int __init tce_iommu_init(void)
1417 {
1418 return vfio_register_iommu_driver(&tce_iommu_driver_ops);
1419 }
1420
tce_iommu_cleanup(void)1421 static void __exit tce_iommu_cleanup(void)
1422 {
1423 vfio_unregister_iommu_driver(&tce_iommu_driver_ops);
1424 }
1425
1426 module_init(tce_iommu_init);
1427 module_exit(tce_iommu_cleanup);
1428
1429 MODULE_VERSION(DRIVER_VERSION);
1430 MODULE_LICENSE("GPL v2");
1431 MODULE_AUTHOR(DRIVER_AUTHOR);
1432 MODULE_DESCRIPTION(DRIVER_DESC);
1433
1434