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 
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 
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 
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 
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 
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 	long ret;
156 
157 	if ((vaddr & ~PAGE_MASK) || (size & ~PAGE_MASK))
158 		return -EINVAL;
159 
160 	mem = mm_iommu_get(container->mm, vaddr, size >> PAGE_SHIFT);
161 	if (!mem)
162 		return -ENOENT;
163 
164 	list_for_each_entry(tcemem, &container->prereg_list, next) {
165 		if (tcemem->mem == mem) {
166 			found = true;
167 			break;
168 		}
169 	}
170 
171 	if (!found)
172 		ret = -ENOENT;
173 	else
174 		ret = tce_iommu_prereg_free(container, tcemem);
175 
176 	mm_iommu_put(container->mm, mem);
177 
178 	return ret;
179 }
180 
181 static long tce_iommu_register_pages(struct tce_container *container,
182 		__u64 vaddr, __u64 size)
183 {
184 	long ret = 0;
185 	struct mm_iommu_table_group_mem_t *mem = NULL;
186 	struct tce_iommu_prereg *tcemem;
187 	unsigned long entries = size >> PAGE_SHIFT;
188 
189 	if ((vaddr & ~PAGE_MASK) || (size & ~PAGE_MASK) ||
190 			((vaddr + size) < vaddr))
191 		return -EINVAL;
192 
193 	mem = mm_iommu_get(container->mm, vaddr, entries);
194 	if (mem) {
195 		list_for_each_entry(tcemem, &container->prereg_list, next) {
196 			if (tcemem->mem == mem) {
197 				ret = -EBUSY;
198 				goto put_exit;
199 			}
200 		}
201 	} else {
202 		ret = mm_iommu_new(container->mm, vaddr, entries, &mem);
203 		if (ret)
204 			return ret;
205 	}
206 
207 	tcemem = kzalloc(sizeof(*tcemem), GFP_KERNEL);
208 	if (!tcemem) {
209 		ret = -ENOMEM;
210 		goto put_exit;
211 	}
212 
213 	tcemem->mem = mem;
214 	list_add(&tcemem->next, &container->prereg_list);
215 
216 	container->enabled = true;
217 
218 	return 0;
219 
220 put_exit:
221 	mm_iommu_put(container->mm, mem);
222 	return ret;
223 }
224 
225 static bool tce_page_is_contained(struct mm_struct *mm, unsigned long hpa,
226 		unsigned int page_shift)
227 {
228 	struct page *page;
229 	unsigned long size = 0;
230 
231 	if (mm_iommu_is_devmem(mm, hpa, page_shift, &size))
232 		return size == (1UL << page_shift);
233 
234 	page = pfn_to_page(hpa >> PAGE_SHIFT);
235 	/*
236 	 * Check that the TCE table granularity is not bigger than the size of
237 	 * a page we just found. Otherwise the hardware can get access to
238 	 * a bigger memory chunk that it should.
239 	 */
240 	return (PAGE_SHIFT + compound_order(compound_head(page))) >= page_shift;
241 }
242 
243 static inline bool tce_groups_attached(struct tce_container *container)
244 {
245 	return !list_empty(&container->group_list);
246 }
247 
248 static long tce_iommu_find_table(struct tce_container *container,
249 		phys_addr_t ioba, struct iommu_table **ptbl)
250 {
251 	long i;
252 
253 	for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
254 		struct iommu_table *tbl = container->tables[i];
255 
256 		if (tbl) {
257 			unsigned long entry = ioba >> tbl->it_page_shift;
258 			unsigned long start = tbl->it_offset;
259 			unsigned long end = start + tbl->it_size;
260 
261 			if ((start <= entry) && (entry < end)) {
262 				*ptbl = tbl;
263 				return i;
264 			}
265 		}
266 	}
267 
268 	return -1;
269 }
270 
271 static int tce_iommu_find_free_table(struct tce_container *container)
272 {
273 	int i;
274 
275 	for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
276 		if (!container->tables[i])
277 			return i;
278 	}
279 
280 	return -ENOSPC;
281 }
282 
283 static int tce_iommu_enable(struct tce_container *container)
284 {
285 	int ret = 0;
286 	unsigned long locked;
287 	struct iommu_table_group *table_group;
288 	struct tce_iommu_group *tcegrp;
289 
290 	if (container->enabled)
291 		return -EBUSY;
292 
293 	/*
294 	 * When userspace pages are mapped into the IOMMU, they are effectively
295 	 * locked memory, so, theoretically, we need to update the accounting
296 	 * of locked pages on each map and unmap.  For powerpc, the map unmap
297 	 * paths can be very hot, though, and the accounting would kill
298 	 * performance, especially since it would be difficult to impossible
299 	 * to handle the accounting in real mode only.
300 	 *
301 	 * To address that, rather than precisely accounting every page, we
302 	 * instead account for a worst case on locked memory when the iommu is
303 	 * enabled and disabled.  The worst case upper bound on locked memory
304 	 * is the size of the whole iommu window, which is usually relatively
305 	 * small (compared to total memory sizes) on POWER hardware.
306 	 *
307 	 * Also we don't have a nice way to fail on H_PUT_TCE due to ulimits,
308 	 * that would effectively kill the guest at random points, much better
309 	 * enforcing the limit based on the max that the guest can map.
310 	 *
311 	 * Unfortunately at the moment it counts whole tables, no matter how
312 	 * much memory the guest has. I.e. for 4GB guest and 4 IOMMU groups
313 	 * each with 2GB DMA window, 8GB will be counted here. The reason for
314 	 * this is that we cannot tell here the amount of RAM used by the guest
315 	 * as this information is only available from KVM and VFIO is
316 	 * KVM agnostic.
317 	 *
318 	 * So we do not allow enabling a container without a group attached
319 	 * as there is no way to know how much we should increment
320 	 * the locked_vm counter.
321 	 */
322 	if (!tce_groups_attached(container))
323 		return -ENODEV;
324 
325 	tcegrp = list_first_entry(&container->group_list,
326 			struct tce_iommu_group, next);
327 	table_group = iommu_group_get_iommudata(tcegrp->grp);
328 	if (!table_group)
329 		return -ENODEV;
330 
331 	if (!table_group->tce32_size)
332 		return -EPERM;
333 
334 	ret = tce_iommu_mm_set(container);
335 	if (ret)
336 		return ret;
337 
338 	locked = table_group->tce32_size >> PAGE_SHIFT;
339 	ret = try_increment_locked_vm(container->mm, locked);
340 	if (ret)
341 		return ret;
342 
343 	container->locked_pages = locked;
344 
345 	container->enabled = true;
346 
347 	return ret;
348 }
349 
350 static void tce_iommu_disable(struct tce_container *container)
351 {
352 	if (!container->enabled)
353 		return;
354 
355 	container->enabled = false;
356 
357 	BUG_ON(!container->mm);
358 	decrement_locked_vm(container->mm, container->locked_pages);
359 }
360 
361 static void *tce_iommu_open(unsigned long arg)
362 {
363 	struct tce_container *container;
364 
365 	if ((arg != VFIO_SPAPR_TCE_IOMMU) && (arg != VFIO_SPAPR_TCE_v2_IOMMU)) {
366 		pr_err("tce_vfio: Wrong IOMMU type\n");
367 		return ERR_PTR(-EINVAL);
368 	}
369 
370 	container = kzalloc(sizeof(*container), GFP_KERNEL);
371 	if (!container)
372 		return ERR_PTR(-ENOMEM);
373 
374 	mutex_init(&container->lock);
375 	INIT_LIST_HEAD_RCU(&container->group_list);
376 	INIT_LIST_HEAD_RCU(&container->prereg_list);
377 
378 	container->v2 = arg == VFIO_SPAPR_TCE_v2_IOMMU;
379 
380 	return container;
381 }
382 
383 static int tce_iommu_clear(struct tce_container *container,
384 		struct iommu_table *tbl,
385 		unsigned long entry, unsigned long pages);
386 static void tce_iommu_free_table(struct tce_container *container,
387 		struct iommu_table *tbl);
388 
389 static void tce_iommu_release(void *iommu_data)
390 {
391 	struct tce_container *container = iommu_data;
392 	struct tce_iommu_group *tcegrp;
393 	struct tce_iommu_prereg *tcemem, *tmtmp;
394 	long i;
395 
396 	while (tce_groups_attached(container)) {
397 		tcegrp = list_first_entry(&container->group_list,
398 				struct tce_iommu_group, next);
399 		tce_iommu_detach_group(iommu_data, tcegrp->grp);
400 	}
401 
402 	/*
403 	 * If VFIO created a table, it was not disposed
404 	 * by tce_iommu_detach_group() so do it now.
405 	 */
406 	for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
407 		struct iommu_table *tbl = container->tables[i];
408 
409 		if (!tbl)
410 			continue;
411 
412 		tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
413 		tce_iommu_free_table(container, tbl);
414 	}
415 
416 	list_for_each_entry_safe(tcemem, tmtmp, &container->prereg_list, next)
417 		WARN_ON(tce_iommu_prereg_free(container, tcemem));
418 
419 	tce_iommu_disable(container);
420 	if (container->mm)
421 		mmdrop(container->mm);
422 	mutex_destroy(&container->lock);
423 
424 	kfree(container);
425 }
426 
427 static void tce_iommu_unuse_page(struct tce_container *container,
428 		unsigned long hpa)
429 {
430 	struct page *page;
431 
432 	page = pfn_to_page(hpa >> PAGE_SHIFT);
433 	put_page(page);
434 }
435 
436 static int tce_iommu_prereg_ua_to_hpa(struct tce_container *container,
437 		unsigned long tce, unsigned long shift,
438 		unsigned long *phpa, struct mm_iommu_table_group_mem_t **pmem)
439 {
440 	long ret = 0;
441 	struct mm_iommu_table_group_mem_t *mem;
442 
443 	mem = mm_iommu_lookup(container->mm, tce, 1ULL << shift);
444 	if (!mem)
445 		return -EINVAL;
446 
447 	ret = mm_iommu_ua_to_hpa(mem, tce, shift, phpa);
448 	if (ret)
449 		return -EINVAL;
450 
451 	*pmem = mem;
452 
453 	return 0;
454 }
455 
456 static void tce_iommu_unuse_page_v2(struct tce_container *container,
457 		struct iommu_table *tbl, unsigned long entry)
458 {
459 	struct mm_iommu_table_group_mem_t *mem = NULL;
460 	int ret;
461 	unsigned long hpa = 0;
462 	__be64 *pua = IOMMU_TABLE_USERSPACE_ENTRY_RO(tbl, entry);
463 
464 	if (!pua)
465 		return;
466 
467 	ret = tce_iommu_prereg_ua_to_hpa(container, be64_to_cpu(*pua),
468 			tbl->it_page_shift, &hpa, &mem);
469 	if (ret)
470 		pr_debug("%s: tce %llx at #%lx was not cached, ret=%d\n",
471 				__func__, be64_to_cpu(*pua), entry, ret);
472 	if (mem)
473 		mm_iommu_mapped_dec(mem);
474 
475 	*pua = cpu_to_be64(0);
476 }
477 
478 static int tce_iommu_clear(struct tce_container *container,
479 		struct iommu_table *tbl,
480 		unsigned long entry, unsigned long pages)
481 {
482 	unsigned long oldhpa;
483 	long ret;
484 	enum dma_data_direction direction;
485 	unsigned long lastentry = entry + pages;
486 
487 	for ( ; entry < lastentry; ++entry) {
488 		if (tbl->it_indirect_levels && tbl->it_userspace) {
489 			/*
490 			 * For multilevel tables, we can take a shortcut here
491 			 * and skip some TCEs as we know that the userspace
492 			 * addresses cache is a mirror of the real TCE table
493 			 * and if it is missing some indirect levels, then
494 			 * the hardware table does not have them allocated
495 			 * either and therefore does not require updating.
496 			 */
497 			__be64 *pua = IOMMU_TABLE_USERSPACE_ENTRY_RO(tbl,
498 					entry);
499 			if (!pua) {
500 				/* align to level_size which is power of two */
501 				entry |= tbl->it_level_size - 1;
502 				continue;
503 			}
504 		}
505 
506 		cond_resched();
507 
508 		direction = DMA_NONE;
509 		oldhpa = 0;
510 		ret = iommu_tce_xchg(container->mm, tbl, entry, &oldhpa,
511 				&direction);
512 		if (ret)
513 			continue;
514 
515 		if (direction == DMA_NONE)
516 			continue;
517 
518 		if (container->v2) {
519 			tce_iommu_unuse_page_v2(container, tbl, entry);
520 			continue;
521 		}
522 
523 		tce_iommu_unuse_page(container, oldhpa);
524 	}
525 
526 	return 0;
527 }
528 
529 static int tce_iommu_use_page(unsigned long tce, unsigned long *hpa)
530 {
531 	struct page *page = NULL;
532 	enum dma_data_direction direction = iommu_tce_direction(tce);
533 
534 	if (get_user_pages_fast(tce & PAGE_MASK, 1,
535 			direction != DMA_TO_DEVICE ? FOLL_WRITE : 0,
536 			&page) != 1)
537 		return -EFAULT;
538 
539 	*hpa = __pa((unsigned long) page_address(page));
540 
541 	return 0;
542 }
543 
544 static long tce_iommu_build(struct tce_container *container,
545 		struct iommu_table *tbl,
546 		unsigned long entry, unsigned long tce, unsigned long pages,
547 		enum dma_data_direction direction)
548 {
549 	long i, ret = 0;
550 	unsigned long hpa;
551 	enum dma_data_direction dirtmp;
552 
553 	for (i = 0; i < pages; ++i) {
554 		unsigned long offset = tce & IOMMU_PAGE_MASK(tbl) & ~PAGE_MASK;
555 
556 		ret = tce_iommu_use_page(tce, &hpa);
557 		if (ret)
558 			break;
559 
560 		if (!tce_page_is_contained(container->mm, hpa,
561 				tbl->it_page_shift)) {
562 			ret = -EPERM;
563 			break;
564 		}
565 
566 		hpa |= offset;
567 		dirtmp = direction;
568 		ret = iommu_tce_xchg(container->mm, tbl, entry + i, &hpa,
569 				&dirtmp);
570 		if (ret) {
571 			tce_iommu_unuse_page(container, hpa);
572 			pr_err("iommu_tce: %s failed ioba=%lx, tce=%lx, ret=%ld\n",
573 					__func__, entry << tbl->it_page_shift,
574 					tce, ret);
575 			break;
576 		}
577 
578 		if (dirtmp != DMA_NONE)
579 			tce_iommu_unuse_page(container, hpa);
580 
581 		tce += IOMMU_PAGE_SIZE(tbl);
582 	}
583 
584 	if (ret)
585 		tce_iommu_clear(container, tbl, entry, i);
586 
587 	return ret;
588 }
589 
590 static long tce_iommu_build_v2(struct tce_container *container,
591 		struct iommu_table *tbl,
592 		unsigned long entry, unsigned long tce, unsigned long pages,
593 		enum dma_data_direction direction)
594 {
595 	long i, ret = 0;
596 	unsigned long hpa;
597 	enum dma_data_direction dirtmp;
598 
599 	for (i = 0; i < pages; ++i) {
600 		struct mm_iommu_table_group_mem_t *mem = NULL;
601 		__be64 *pua = IOMMU_TABLE_USERSPACE_ENTRY(tbl, entry + i);
602 
603 		ret = tce_iommu_prereg_ua_to_hpa(container,
604 				tce, tbl->it_page_shift, &hpa, &mem);
605 		if (ret)
606 			break;
607 
608 		if (!tce_page_is_contained(container->mm, hpa,
609 				tbl->it_page_shift)) {
610 			ret = -EPERM;
611 			break;
612 		}
613 
614 		/* Preserve offset within IOMMU page */
615 		hpa |= tce & IOMMU_PAGE_MASK(tbl) & ~PAGE_MASK;
616 		dirtmp = direction;
617 
618 		/* The registered region is being unregistered */
619 		if (mm_iommu_mapped_inc(mem))
620 			break;
621 
622 		ret = iommu_tce_xchg(container->mm, tbl, entry + i, &hpa,
623 				&dirtmp);
624 		if (ret) {
625 			/* dirtmp cannot be DMA_NONE here */
626 			tce_iommu_unuse_page_v2(container, tbl, entry + i);
627 			pr_err("iommu_tce: %s failed ioba=%lx, tce=%lx, ret=%ld\n",
628 					__func__, entry << tbl->it_page_shift,
629 					tce, ret);
630 			break;
631 		}
632 
633 		if (dirtmp != DMA_NONE)
634 			tce_iommu_unuse_page_v2(container, tbl, entry + i);
635 
636 		*pua = cpu_to_be64(tce);
637 
638 		tce += IOMMU_PAGE_SIZE(tbl);
639 	}
640 
641 	if (ret)
642 		tce_iommu_clear(container, tbl, entry, i);
643 
644 	return ret;
645 }
646 
647 static long tce_iommu_create_table(struct tce_container *container,
648 			struct iommu_table_group *table_group,
649 			int num,
650 			__u32 page_shift,
651 			__u64 window_size,
652 			__u32 levels,
653 			struct iommu_table **ptbl)
654 {
655 	long ret, table_size;
656 
657 	table_size = table_group->ops->get_table_size(page_shift, window_size,
658 			levels);
659 	if (!table_size)
660 		return -EINVAL;
661 
662 	ret = try_increment_locked_vm(container->mm, table_size >> PAGE_SHIFT);
663 	if (ret)
664 		return ret;
665 
666 	ret = table_group->ops->create_table(table_group, num,
667 			page_shift, window_size, levels, ptbl);
668 
669 	WARN_ON(!ret && !(*ptbl)->it_ops->free);
670 	WARN_ON(!ret && ((*ptbl)->it_allocated_size > table_size));
671 
672 	return ret;
673 }
674 
675 static void tce_iommu_free_table(struct tce_container *container,
676 		struct iommu_table *tbl)
677 {
678 	unsigned long pages = tbl->it_allocated_size >> PAGE_SHIFT;
679 
680 	iommu_tce_table_put(tbl);
681 	decrement_locked_vm(container->mm, pages);
682 }
683 
684 static long tce_iommu_create_window(struct tce_container *container,
685 		__u32 page_shift, __u64 window_size, __u32 levels,
686 		__u64 *start_addr)
687 {
688 	struct tce_iommu_group *tcegrp;
689 	struct iommu_table_group *table_group;
690 	struct iommu_table *tbl = NULL;
691 	long ret, num;
692 
693 	num = tce_iommu_find_free_table(container);
694 	if (num < 0)
695 		return num;
696 
697 	/* Get the first group for ops::create_table */
698 	tcegrp = list_first_entry(&container->group_list,
699 			struct tce_iommu_group, next);
700 	table_group = iommu_group_get_iommudata(tcegrp->grp);
701 	if (!table_group)
702 		return -EFAULT;
703 
704 	if (!(table_group->pgsizes & (1ULL << page_shift)))
705 		return -EINVAL;
706 
707 	if (!table_group->ops->set_window || !table_group->ops->unset_window ||
708 			!table_group->ops->get_table_size ||
709 			!table_group->ops->create_table)
710 		return -EPERM;
711 
712 	/* Create TCE table */
713 	ret = tce_iommu_create_table(container, table_group, num,
714 			page_shift, window_size, levels, &tbl);
715 	if (ret)
716 		return ret;
717 
718 	BUG_ON(!tbl->it_ops->free);
719 
720 	/*
721 	 * Program the table to every group.
722 	 * Groups have been tested for compatibility at the attach time.
723 	 */
724 	list_for_each_entry(tcegrp, &container->group_list, next) {
725 		table_group = iommu_group_get_iommudata(tcegrp->grp);
726 
727 		ret = table_group->ops->set_window(table_group, num, tbl);
728 		if (ret)
729 			goto unset_exit;
730 	}
731 
732 	container->tables[num] = tbl;
733 
734 	/* Return start address assigned by platform in create_table() */
735 	*start_addr = tbl->it_offset << tbl->it_page_shift;
736 
737 	return 0;
738 
739 unset_exit:
740 	list_for_each_entry(tcegrp, &container->group_list, next) {
741 		table_group = iommu_group_get_iommudata(tcegrp->grp);
742 		table_group->ops->unset_window(table_group, num);
743 	}
744 	tce_iommu_free_table(container, tbl);
745 
746 	return ret;
747 }
748 
749 static long tce_iommu_remove_window(struct tce_container *container,
750 		__u64 start_addr)
751 {
752 	struct iommu_table_group *table_group = NULL;
753 	struct iommu_table *tbl;
754 	struct tce_iommu_group *tcegrp;
755 	int num;
756 
757 	num = tce_iommu_find_table(container, start_addr, &tbl);
758 	if (num < 0)
759 		return -EINVAL;
760 
761 	BUG_ON(!tbl->it_size);
762 
763 	/* Detach groups from IOMMUs */
764 	list_for_each_entry(tcegrp, &container->group_list, next) {
765 		table_group = iommu_group_get_iommudata(tcegrp->grp);
766 
767 		/*
768 		 * SPAPR TCE IOMMU exposes the default DMA window to
769 		 * the guest via dma32_window_start/size of
770 		 * VFIO_IOMMU_SPAPR_TCE_GET_INFO. Some platforms allow
771 		 * the userspace to remove this window, some do not so
772 		 * here we check for the platform capability.
773 		 */
774 		if (!table_group->ops || !table_group->ops->unset_window)
775 			return -EPERM;
776 
777 		table_group->ops->unset_window(table_group, num);
778 	}
779 
780 	/* Free table */
781 	tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
782 	tce_iommu_free_table(container, tbl);
783 	container->tables[num] = NULL;
784 
785 	return 0;
786 }
787 
788 static long tce_iommu_create_default_window(struct tce_container *container)
789 {
790 	long ret;
791 	__u64 start_addr = 0;
792 	struct tce_iommu_group *tcegrp;
793 	struct iommu_table_group *table_group;
794 
795 	if (!container->def_window_pending)
796 		return 0;
797 
798 	if (!tce_groups_attached(container))
799 		return -ENODEV;
800 
801 	tcegrp = list_first_entry(&container->group_list,
802 			struct tce_iommu_group, next);
803 	table_group = iommu_group_get_iommudata(tcegrp->grp);
804 	if (!table_group)
805 		return -ENODEV;
806 
807 	ret = tce_iommu_create_window(container, IOMMU_PAGE_SHIFT_4K,
808 			table_group->tce32_size, 1, &start_addr);
809 	WARN_ON_ONCE(!ret && start_addr);
810 
811 	if (!ret)
812 		container->def_window_pending = false;
813 
814 	return ret;
815 }
816 
817 static long tce_iommu_ioctl(void *iommu_data,
818 				 unsigned int cmd, unsigned long arg)
819 {
820 	struct tce_container *container = iommu_data;
821 	unsigned long minsz, ddwsz;
822 	long ret;
823 
824 	switch (cmd) {
825 	case VFIO_CHECK_EXTENSION:
826 		switch (arg) {
827 		case VFIO_SPAPR_TCE_IOMMU:
828 		case VFIO_SPAPR_TCE_v2_IOMMU:
829 			ret = 1;
830 			break;
831 		default:
832 			ret = vfio_spapr_iommu_eeh_ioctl(NULL, cmd, arg);
833 			break;
834 		}
835 
836 		return (ret < 0) ? 0 : ret;
837 	}
838 
839 	/*
840 	 * Sanity check to prevent one userspace from manipulating
841 	 * another userspace mm.
842 	 */
843 	BUG_ON(!container);
844 	if (container->mm && container->mm != current->mm)
845 		return -EPERM;
846 
847 	switch (cmd) {
848 	case VFIO_IOMMU_SPAPR_TCE_GET_INFO: {
849 		struct vfio_iommu_spapr_tce_info info;
850 		struct tce_iommu_group *tcegrp;
851 		struct iommu_table_group *table_group;
852 
853 		if (!tce_groups_attached(container))
854 			return -ENXIO;
855 
856 		tcegrp = list_first_entry(&container->group_list,
857 				struct tce_iommu_group, next);
858 		table_group = iommu_group_get_iommudata(tcegrp->grp);
859 
860 		if (!table_group)
861 			return -ENXIO;
862 
863 		minsz = offsetofend(struct vfio_iommu_spapr_tce_info,
864 				dma32_window_size);
865 
866 		if (copy_from_user(&info, (void __user *)arg, minsz))
867 			return -EFAULT;
868 
869 		if (info.argsz < minsz)
870 			return -EINVAL;
871 
872 		info.dma32_window_start = table_group->tce32_start;
873 		info.dma32_window_size = table_group->tce32_size;
874 		info.flags = 0;
875 		memset(&info.ddw, 0, sizeof(info.ddw));
876 
877 		if (table_group->max_dynamic_windows_supported &&
878 				container->v2) {
879 			info.flags |= VFIO_IOMMU_SPAPR_INFO_DDW;
880 			info.ddw.pgsizes = table_group->pgsizes;
881 			info.ddw.max_dynamic_windows_supported =
882 				table_group->max_dynamic_windows_supported;
883 			info.ddw.levels = table_group->max_levels;
884 		}
885 
886 		ddwsz = offsetofend(struct vfio_iommu_spapr_tce_info, ddw);
887 
888 		if (info.argsz >= ddwsz)
889 			minsz = ddwsz;
890 
891 		if (copy_to_user((void __user *)arg, &info, minsz))
892 			return -EFAULT;
893 
894 		return 0;
895 	}
896 	case VFIO_IOMMU_MAP_DMA: {
897 		struct vfio_iommu_type1_dma_map param;
898 		struct iommu_table *tbl = NULL;
899 		long num;
900 		enum dma_data_direction direction;
901 
902 		if (!container->enabled)
903 			return -EPERM;
904 
905 		minsz = offsetofend(struct vfio_iommu_type1_dma_map, size);
906 
907 		if (copy_from_user(&param, (void __user *)arg, minsz))
908 			return -EFAULT;
909 
910 		if (param.argsz < minsz)
911 			return -EINVAL;
912 
913 		if (param.flags & ~(VFIO_DMA_MAP_FLAG_READ |
914 				VFIO_DMA_MAP_FLAG_WRITE))
915 			return -EINVAL;
916 
917 		ret = tce_iommu_create_default_window(container);
918 		if (ret)
919 			return ret;
920 
921 		num = tce_iommu_find_table(container, param.iova, &tbl);
922 		if (num < 0)
923 			return -ENXIO;
924 
925 		if ((param.size & ~IOMMU_PAGE_MASK(tbl)) ||
926 				(param.vaddr & ~IOMMU_PAGE_MASK(tbl)))
927 			return -EINVAL;
928 
929 		/* iova is checked by the IOMMU API */
930 		if (param.flags & VFIO_DMA_MAP_FLAG_READ) {
931 			if (param.flags & VFIO_DMA_MAP_FLAG_WRITE)
932 				direction = DMA_BIDIRECTIONAL;
933 			else
934 				direction = DMA_TO_DEVICE;
935 		} else {
936 			if (param.flags & VFIO_DMA_MAP_FLAG_WRITE)
937 				direction = DMA_FROM_DEVICE;
938 			else
939 				return -EINVAL;
940 		}
941 
942 		ret = iommu_tce_put_param_check(tbl, param.iova, param.vaddr);
943 		if (ret)
944 			return ret;
945 
946 		if (container->v2)
947 			ret = tce_iommu_build_v2(container, tbl,
948 					param.iova >> tbl->it_page_shift,
949 					param.vaddr,
950 					param.size >> tbl->it_page_shift,
951 					direction);
952 		else
953 			ret = tce_iommu_build(container, tbl,
954 					param.iova >> tbl->it_page_shift,
955 					param.vaddr,
956 					param.size >> tbl->it_page_shift,
957 					direction);
958 
959 		iommu_flush_tce(tbl);
960 
961 		return ret;
962 	}
963 	case VFIO_IOMMU_UNMAP_DMA: {
964 		struct vfio_iommu_type1_dma_unmap param;
965 		struct iommu_table *tbl = NULL;
966 		long num;
967 
968 		if (!container->enabled)
969 			return -EPERM;
970 
971 		minsz = offsetofend(struct vfio_iommu_type1_dma_unmap,
972 				size);
973 
974 		if (copy_from_user(&param, (void __user *)arg, minsz))
975 			return -EFAULT;
976 
977 		if (param.argsz < minsz)
978 			return -EINVAL;
979 
980 		/* No flag is supported now */
981 		if (param.flags)
982 			return -EINVAL;
983 
984 		ret = tce_iommu_create_default_window(container);
985 		if (ret)
986 			return ret;
987 
988 		num = tce_iommu_find_table(container, param.iova, &tbl);
989 		if (num < 0)
990 			return -ENXIO;
991 
992 		if (param.size & ~IOMMU_PAGE_MASK(tbl))
993 			return -EINVAL;
994 
995 		ret = iommu_tce_clear_param_check(tbl, param.iova, 0,
996 				param.size >> tbl->it_page_shift);
997 		if (ret)
998 			return ret;
999 
1000 		ret = tce_iommu_clear(container, tbl,
1001 				param.iova >> tbl->it_page_shift,
1002 				param.size >> tbl->it_page_shift);
1003 		iommu_flush_tce(tbl);
1004 
1005 		return ret;
1006 	}
1007 	case VFIO_IOMMU_SPAPR_REGISTER_MEMORY: {
1008 		struct vfio_iommu_spapr_register_memory param;
1009 
1010 		if (!container->v2)
1011 			break;
1012 
1013 		minsz = offsetofend(struct vfio_iommu_spapr_register_memory,
1014 				size);
1015 
1016 		ret = tce_iommu_mm_set(container);
1017 		if (ret)
1018 			return ret;
1019 
1020 		if (copy_from_user(&param, (void __user *)arg, minsz))
1021 			return -EFAULT;
1022 
1023 		if (param.argsz < minsz)
1024 			return -EINVAL;
1025 
1026 		/* No flag is supported now */
1027 		if (param.flags)
1028 			return -EINVAL;
1029 
1030 		mutex_lock(&container->lock);
1031 		ret = tce_iommu_register_pages(container, param.vaddr,
1032 				param.size);
1033 		mutex_unlock(&container->lock);
1034 
1035 		return ret;
1036 	}
1037 	case VFIO_IOMMU_SPAPR_UNREGISTER_MEMORY: {
1038 		struct vfio_iommu_spapr_register_memory param;
1039 
1040 		if (!container->v2)
1041 			break;
1042 
1043 		if (!container->mm)
1044 			return -EPERM;
1045 
1046 		minsz = offsetofend(struct vfio_iommu_spapr_register_memory,
1047 				size);
1048 
1049 		if (copy_from_user(&param, (void __user *)arg, minsz))
1050 			return -EFAULT;
1051 
1052 		if (param.argsz < minsz)
1053 			return -EINVAL;
1054 
1055 		/* No flag is supported now */
1056 		if (param.flags)
1057 			return -EINVAL;
1058 
1059 		mutex_lock(&container->lock);
1060 		ret = tce_iommu_unregister_pages(container, param.vaddr,
1061 				param.size);
1062 		mutex_unlock(&container->lock);
1063 
1064 		return ret;
1065 	}
1066 	case VFIO_IOMMU_ENABLE:
1067 		if (container->v2)
1068 			break;
1069 
1070 		mutex_lock(&container->lock);
1071 		ret = tce_iommu_enable(container);
1072 		mutex_unlock(&container->lock);
1073 		return ret;
1074 
1075 
1076 	case VFIO_IOMMU_DISABLE:
1077 		if (container->v2)
1078 			break;
1079 
1080 		mutex_lock(&container->lock);
1081 		tce_iommu_disable(container);
1082 		mutex_unlock(&container->lock);
1083 		return 0;
1084 
1085 	case VFIO_EEH_PE_OP: {
1086 		struct tce_iommu_group *tcegrp;
1087 
1088 		ret = 0;
1089 		list_for_each_entry(tcegrp, &container->group_list, next) {
1090 			ret = vfio_spapr_iommu_eeh_ioctl(tcegrp->grp,
1091 					cmd, arg);
1092 			if (ret)
1093 				return ret;
1094 		}
1095 		return ret;
1096 	}
1097 
1098 	case VFIO_IOMMU_SPAPR_TCE_CREATE: {
1099 		struct vfio_iommu_spapr_tce_create create;
1100 
1101 		if (!container->v2)
1102 			break;
1103 
1104 		ret = tce_iommu_mm_set(container);
1105 		if (ret)
1106 			return ret;
1107 
1108 		if (!tce_groups_attached(container))
1109 			return -ENXIO;
1110 
1111 		minsz = offsetofend(struct vfio_iommu_spapr_tce_create,
1112 				start_addr);
1113 
1114 		if (copy_from_user(&create, (void __user *)arg, minsz))
1115 			return -EFAULT;
1116 
1117 		if (create.argsz < minsz)
1118 			return -EINVAL;
1119 
1120 		if (create.flags)
1121 			return -EINVAL;
1122 
1123 		mutex_lock(&container->lock);
1124 
1125 		ret = tce_iommu_create_default_window(container);
1126 		if (!ret)
1127 			ret = tce_iommu_create_window(container,
1128 					create.page_shift,
1129 					create.window_size, create.levels,
1130 					&create.start_addr);
1131 
1132 		mutex_unlock(&container->lock);
1133 
1134 		if (!ret && copy_to_user((void __user *)arg, &create, minsz))
1135 			ret = -EFAULT;
1136 
1137 		return ret;
1138 	}
1139 	case VFIO_IOMMU_SPAPR_TCE_REMOVE: {
1140 		struct vfio_iommu_spapr_tce_remove remove;
1141 
1142 		if (!container->v2)
1143 			break;
1144 
1145 		ret = tce_iommu_mm_set(container);
1146 		if (ret)
1147 			return ret;
1148 
1149 		if (!tce_groups_attached(container))
1150 			return -ENXIO;
1151 
1152 		minsz = offsetofend(struct vfio_iommu_spapr_tce_remove,
1153 				start_addr);
1154 
1155 		if (copy_from_user(&remove, (void __user *)arg, minsz))
1156 			return -EFAULT;
1157 
1158 		if (remove.argsz < minsz)
1159 			return -EINVAL;
1160 
1161 		if (remove.flags)
1162 			return -EINVAL;
1163 
1164 		if (container->def_window_pending && !remove.start_addr) {
1165 			container->def_window_pending = false;
1166 			return 0;
1167 		}
1168 
1169 		mutex_lock(&container->lock);
1170 
1171 		ret = tce_iommu_remove_window(container, remove.start_addr);
1172 
1173 		mutex_unlock(&container->lock);
1174 
1175 		return ret;
1176 	}
1177 	}
1178 
1179 	return -ENOTTY;
1180 }
1181 
1182 static void tce_iommu_release_ownership(struct tce_container *container,
1183 		struct iommu_table_group *table_group)
1184 {
1185 	int i;
1186 
1187 	for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1188 		struct iommu_table *tbl = container->tables[i];
1189 
1190 		if (!tbl)
1191 			continue;
1192 
1193 		tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
1194 		if (tbl->it_map)
1195 			iommu_release_ownership(tbl);
1196 
1197 		container->tables[i] = NULL;
1198 	}
1199 }
1200 
1201 static int tce_iommu_take_ownership(struct tce_container *container,
1202 		struct iommu_table_group *table_group)
1203 {
1204 	int i, j, rc = 0;
1205 
1206 	for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1207 		struct iommu_table *tbl = table_group->tables[i];
1208 
1209 		if (!tbl || !tbl->it_map)
1210 			continue;
1211 
1212 		rc = iommu_take_ownership(tbl);
1213 		if (rc) {
1214 			for (j = 0; j < i; ++j)
1215 				iommu_release_ownership(
1216 						table_group->tables[j]);
1217 
1218 			return rc;
1219 		}
1220 	}
1221 
1222 	for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1223 		container->tables[i] = table_group->tables[i];
1224 
1225 	return 0;
1226 }
1227 
1228 static void tce_iommu_release_ownership_ddw(struct tce_container *container,
1229 		struct iommu_table_group *table_group)
1230 {
1231 	long i;
1232 
1233 	if (!table_group->ops->unset_window) {
1234 		WARN_ON_ONCE(1);
1235 		return;
1236 	}
1237 
1238 	for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1239 		if (container->tables[i])
1240 			table_group->ops->unset_window(table_group, i);
1241 
1242 	table_group->ops->release_ownership(table_group);
1243 }
1244 
1245 static long tce_iommu_take_ownership_ddw(struct tce_container *container,
1246 		struct iommu_table_group *table_group)
1247 {
1248 	long i, ret = 0;
1249 
1250 	if (!table_group->ops->create_table || !table_group->ops->set_window ||
1251 			!table_group->ops->release_ownership) {
1252 		WARN_ON_ONCE(1);
1253 		return -EFAULT;
1254 	}
1255 
1256 	table_group->ops->take_ownership(table_group);
1257 
1258 	/* Set all windows to the new group */
1259 	for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1260 		struct iommu_table *tbl = container->tables[i];
1261 
1262 		if (!tbl)
1263 			continue;
1264 
1265 		ret = table_group->ops->set_window(table_group, i, tbl);
1266 		if (ret)
1267 			goto release_exit;
1268 	}
1269 
1270 	return 0;
1271 
1272 release_exit:
1273 	for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1274 		table_group->ops->unset_window(table_group, i);
1275 
1276 	table_group->ops->release_ownership(table_group);
1277 
1278 	return ret;
1279 }
1280 
1281 static int tce_iommu_attach_group(void *iommu_data,
1282 		struct iommu_group *iommu_group)
1283 {
1284 	int ret;
1285 	struct tce_container *container = iommu_data;
1286 	struct iommu_table_group *table_group;
1287 	struct tce_iommu_group *tcegrp = NULL;
1288 
1289 	mutex_lock(&container->lock);
1290 
1291 	/* pr_debug("tce_vfio: Attaching group #%u to iommu %p\n",
1292 			iommu_group_id(iommu_group), iommu_group); */
1293 	table_group = iommu_group_get_iommudata(iommu_group);
1294 	if (!table_group) {
1295 		ret = -ENODEV;
1296 		goto unlock_exit;
1297 	}
1298 
1299 	if (tce_groups_attached(container) && (!table_group->ops ||
1300 			!table_group->ops->take_ownership ||
1301 			!table_group->ops->release_ownership)) {
1302 		ret = -EBUSY;
1303 		goto unlock_exit;
1304 	}
1305 
1306 	/* Check if new group has the same iommu_ops (i.e. compatible) */
1307 	list_for_each_entry(tcegrp, &container->group_list, next) {
1308 		struct iommu_table_group *table_group_tmp;
1309 
1310 		if (tcegrp->grp == iommu_group) {
1311 			pr_warn("tce_vfio: Group %d is already attached\n",
1312 					iommu_group_id(iommu_group));
1313 			ret = -EBUSY;
1314 			goto unlock_exit;
1315 		}
1316 		table_group_tmp = iommu_group_get_iommudata(tcegrp->grp);
1317 		if (table_group_tmp->ops->create_table !=
1318 				table_group->ops->create_table) {
1319 			pr_warn("tce_vfio: Group %d is incompatible with group %d\n",
1320 					iommu_group_id(iommu_group),
1321 					iommu_group_id(tcegrp->grp));
1322 			ret = -EPERM;
1323 			goto unlock_exit;
1324 		}
1325 	}
1326 
1327 	tcegrp = kzalloc(sizeof(*tcegrp), GFP_KERNEL);
1328 	if (!tcegrp) {
1329 		ret = -ENOMEM;
1330 		goto unlock_exit;
1331 	}
1332 
1333 	if (!table_group->ops || !table_group->ops->take_ownership ||
1334 			!table_group->ops->release_ownership) {
1335 		if (container->v2) {
1336 			ret = -EPERM;
1337 			goto unlock_exit;
1338 		}
1339 		ret = tce_iommu_take_ownership(container, table_group);
1340 	} else {
1341 		if (!container->v2) {
1342 			ret = -EPERM;
1343 			goto unlock_exit;
1344 		}
1345 		ret = tce_iommu_take_ownership_ddw(container, table_group);
1346 		if (!tce_groups_attached(container) && !container->tables[0])
1347 			container->def_window_pending = true;
1348 	}
1349 
1350 	if (!ret) {
1351 		tcegrp->grp = iommu_group;
1352 		list_add(&tcegrp->next, &container->group_list);
1353 	}
1354 
1355 unlock_exit:
1356 	if (ret && tcegrp)
1357 		kfree(tcegrp);
1358 
1359 	mutex_unlock(&container->lock);
1360 
1361 	return ret;
1362 }
1363 
1364 static void tce_iommu_detach_group(void *iommu_data,
1365 		struct iommu_group *iommu_group)
1366 {
1367 	struct tce_container *container = iommu_data;
1368 	struct iommu_table_group *table_group;
1369 	bool found = false;
1370 	struct tce_iommu_group *tcegrp;
1371 
1372 	mutex_lock(&container->lock);
1373 
1374 	list_for_each_entry(tcegrp, &container->group_list, next) {
1375 		if (tcegrp->grp == iommu_group) {
1376 			found = true;
1377 			break;
1378 		}
1379 	}
1380 
1381 	if (!found) {
1382 		pr_warn("tce_vfio: detaching unattached group #%u\n",
1383 				iommu_group_id(iommu_group));
1384 		goto unlock_exit;
1385 	}
1386 
1387 	list_del(&tcegrp->next);
1388 	kfree(tcegrp);
1389 
1390 	table_group = iommu_group_get_iommudata(iommu_group);
1391 	BUG_ON(!table_group);
1392 
1393 	if (!table_group->ops || !table_group->ops->release_ownership)
1394 		tce_iommu_release_ownership(container, table_group);
1395 	else
1396 		tce_iommu_release_ownership_ddw(container, table_group);
1397 
1398 unlock_exit:
1399 	mutex_unlock(&container->lock);
1400 }
1401 
1402 static const struct vfio_iommu_driver_ops tce_iommu_driver_ops = {
1403 	.name		= "iommu-vfio-powerpc",
1404 	.owner		= THIS_MODULE,
1405 	.open		= tce_iommu_open,
1406 	.release	= tce_iommu_release,
1407 	.ioctl		= tce_iommu_ioctl,
1408 	.attach_group	= tce_iommu_attach_group,
1409 	.detach_group	= tce_iommu_detach_group,
1410 };
1411 
1412 static int __init tce_iommu_init(void)
1413 {
1414 	return vfio_register_iommu_driver(&tce_iommu_driver_ops);
1415 }
1416 
1417 static void __exit tce_iommu_cleanup(void)
1418 {
1419 	vfio_unregister_iommu_driver(&tce_iommu_driver_ops);
1420 }
1421 
1422 module_init(tce_iommu_init);
1423 module_exit(tce_iommu_cleanup);
1424 
1425 MODULE_VERSION(DRIVER_VERSION);
1426 MODULE_LICENSE("GPL v2");
1427 MODULE_AUTHOR(DRIVER_AUTHOR);
1428 MODULE_DESCRIPTION(DRIVER_DESC);
1429 
1430