1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * VFIO: IOMMU DMA mapping support for TCE on POWER
4  *
5  * Copyright (C) 2013 IBM Corp.  All rights reserved.
6  *     Author: Alexey Kardashevskiy <aik@ozlabs.ru>
7  *
8  * Derived from original vfio_iommu_type1.c:
9  * Copyright (C) 2012 Red Hat, Inc.  All rights reserved.
10  *     Author: Alex Williamson <alex.williamson@redhat.com>
11  */
12 
13 #include <linux/module.h>
14 #include <linux/pci.h>
15 #include <linux/slab.h>
16 #include <linux/uaccess.h>
17 #include <linux/err.h>
18 #include <linux/vfio.h>
19 #include <linux/vmalloc.h>
20 #include <linux/sched/mm.h>
21 #include <linux/sched/signal.h>
22 #include <linux/mm.h>
23 
24 #include <asm/iommu.h>
25 #include <asm/tce.h>
26 #include <asm/mmu_context.h>
27 
28 #define DRIVER_VERSION  "0.1"
29 #define DRIVER_AUTHOR   "aik@ozlabs.ru"
30 #define DRIVER_DESC     "VFIO IOMMU SPAPR TCE"
31 
32 static void tce_iommu_detach_group(void *iommu_data,
33 		struct iommu_group *iommu_group);
34 
35 /*
36  * VFIO IOMMU fd for SPAPR_TCE IOMMU implementation
37  *
38  * This code handles mapping and unmapping of user data buffers
39  * into DMA'ble space using the IOMMU
40  */
41 
42 struct tce_iommu_group {
43 	struct list_head next;
44 	struct iommu_group *grp;
45 };
46 
47 /*
48  * A container needs to remember which preregistered region  it has
49  * referenced to do proper cleanup at the userspace process exit.
50  */
51 struct tce_iommu_prereg {
52 	struct list_head next;
53 	struct mm_iommu_table_group_mem_t *mem;
54 };
55 
56 /*
57  * The container descriptor supports only a single group per container.
58  * Required by the API as the container is not supplied with the IOMMU group
59  * at the moment of initialization.
60  */
61 struct tce_container {
62 	struct mutex lock;
63 	bool enabled;
64 	bool v2;
65 	bool def_window_pending;
66 	unsigned long locked_pages;
67 	struct mm_struct *mm;
68 	struct iommu_table *tables[IOMMU_TABLE_GROUP_MAX_TABLES];
69 	struct list_head group_list;
70 	struct list_head prereg_list;
71 };
72 
73 static long tce_iommu_mm_set(struct tce_container *container)
74 {
75 	if (container->mm) {
76 		if (container->mm == current->mm)
77 			return 0;
78 		return -EPERM;
79 	}
80 	BUG_ON(!current->mm);
81 	container->mm = current->mm;
82 	atomic_inc(&container->mm->mm_count);
83 
84 	return 0;
85 }
86 
87 static long tce_iommu_prereg_free(struct tce_container *container,
88 		struct tce_iommu_prereg *tcemem)
89 {
90 	long ret;
91 
92 	ret = mm_iommu_put(container->mm, tcemem->mem);
93 	if (ret)
94 		return ret;
95 
96 	list_del(&tcemem->next);
97 	kfree(tcemem);
98 
99 	return 0;
100 }
101 
102 static long tce_iommu_unregister_pages(struct tce_container *container,
103 		__u64 vaddr, __u64 size)
104 {
105 	struct mm_iommu_table_group_mem_t *mem;
106 	struct tce_iommu_prereg *tcemem;
107 	bool found = false;
108 	long ret;
109 
110 	if ((vaddr & ~PAGE_MASK) || (size & ~PAGE_MASK))
111 		return -EINVAL;
112 
113 	mem = mm_iommu_get(container->mm, vaddr, size >> PAGE_SHIFT);
114 	if (!mem)
115 		return -ENOENT;
116 
117 	list_for_each_entry(tcemem, &container->prereg_list, next) {
118 		if (tcemem->mem == mem) {
119 			found = true;
120 			break;
121 		}
122 	}
123 
124 	if (!found)
125 		ret = -ENOENT;
126 	else
127 		ret = tce_iommu_prereg_free(container, tcemem);
128 
129 	mm_iommu_put(container->mm, mem);
130 
131 	return ret;
132 }
133 
134 static long tce_iommu_register_pages(struct tce_container *container,
135 		__u64 vaddr, __u64 size)
136 {
137 	long ret = 0;
138 	struct mm_iommu_table_group_mem_t *mem = NULL;
139 	struct tce_iommu_prereg *tcemem;
140 	unsigned long entries = size >> PAGE_SHIFT;
141 
142 	if ((vaddr & ~PAGE_MASK) || (size & ~PAGE_MASK) ||
143 			((vaddr + size) < vaddr))
144 		return -EINVAL;
145 
146 	mem = mm_iommu_get(container->mm, vaddr, entries);
147 	if (mem) {
148 		list_for_each_entry(tcemem, &container->prereg_list, next) {
149 			if (tcemem->mem == mem) {
150 				ret = -EBUSY;
151 				goto put_exit;
152 			}
153 		}
154 	} else {
155 		ret = mm_iommu_new(container->mm, vaddr, entries, &mem);
156 		if (ret)
157 			return ret;
158 	}
159 
160 	tcemem = kzalloc(sizeof(*tcemem), GFP_KERNEL);
161 	if (!tcemem) {
162 		ret = -ENOMEM;
163 		goto put_exit;
164 	}
165 
166 	tcemem->mem = mem;
167 	list_add(&tcemem->next, &container->prereg_list);
168 
169 	container->enabled = true;
170 
171 	return 0;
172 
173 put_exit:
174 	mm_iommu_put(container->mm, mem);
175 	return ret;
176 }
177 
178 static bool tce_page_is_contained(struct mm_struct *mm, unsigned long hpa,
179 		unsigned int page_shift)
180 {
181 	struct page *page;
182 	unsigned long size = 0;
183 
184 	if (mm_iommu_is_devmem(mm, hpa, page_shift, &size))
185 		return size == (1UL << page_shift);
186 
187 	page = pfn_to_page(hpa >> PAGE_SHIFT);
188 	/*
189 	 * Check that the TCE table granularity is not bigger than the size of
190 	 * a page we just found. Otherwise the hardware can get access to
191 	 * a bigger memory chunk that it should.
192 	 */
193 	return (PAGE_SHIFT + compound_order(compound_head(page))) >= page_shift;
194 }
195 
196 static inline bool tce_groups_attached(struct tce_container *container)
197 {
198 	return !list_empty(&container->group_list);
199 }
200 
201 static long tce_iommu_find_table(struct tce_container *container,
202 		phys_addr_t ioba, struct iommu_table **ptbl)
203 {
204 	long i;
205 
206 	for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
207 		struct iommu_table *tbl = container->tables[i];
208 
209 		if (tbl) {
210 			unsigned long entry = ioba >> tbl->it_page_shift;
211 			unsigned long start = tbl->it_offset;
212 			unsigned long end = start + tbl->it_size;
213 
214 			if ((start <= entry) && (entry < end)) {
215 				*ptbl = tbl;
216 				return i;
217 			}
218 		}
219 	}
220 
221 	return -1;
222 }
223 
224 static int tce_iommu_find_free_table(struct tce_container *container)
225 {
226 	int i;
227 
228 	for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
229 		if (!container->tables[i])
230 			return i;
231 	}
232 
233 	return -ENOSPC;
234 }
235 
236 static int tce_iommu_enable(struct tce_container *container)
237 {
238 	int ret = 0;
239 	unsigned long locked;
240 	struct iommu_table_group *table_group;
241 	struct tce_iommu_group *tcegrp;
242 
243 	if (container->enabled)
244 		return -EBUSY;
245 
246 	/*
247 	 * When userspace pages are mapped into the IOMMU, they are effectively
248 	 * locked memory, so, theoretically, we need to update the accounting
249 	 * of locked pages on each map and unmap.  For powerpc, the map unmap
250 	 * paths can be very hot, though, and the accounting would kill
251 	 * performance, especially since it would be difficult to impossible
252 	 * to handle the accounting in real mode only.
253 	 *
254 	 * To address that, rather than precisely accounting every page, we
255 	 * instead account for a worst case on locked memory when the iommu is
256 	 * enabled and disabled.  The worst case upper bound on locked memory
257 	 * is the size of the whole iommu window, which is usually relatively
258 	 * small (compared to total memory sizes) on POWER hardware.
259 	 *
260 	 * Also we don't have a nice way to fail on H_PUT_TCE due to ulimits,
261 	 * that would effectively kill the guest at random points, much better
262 	 * enforcing the limit based on the max that the guest can map.
263 	 *
264 	 * Unfortunately at the moment it counts whole tables, no matter how
265 	 * much memory the guest has. I.e. for 4GB guest and 4 IOMMU groups
266 	 * each with 2GB DMA window, 8GB will be counted here. The reason for
267 	 * this is that we cannot tell here the amount of RAM used by the guest
268 	 * as this information is only available from KVM and VFIO is
269 	 * KVM agnostic.
270 	 *
271 	 * So we do not allow enabling a container without a group attached
272 	 * as there is no way to know how much we should increment
273 	 * the locked_vm counter.
274 	 */
275 	if (!tce_groups_attached(container))
276 		return -ENODEV;
277 
278 	tcegrp = list_first_entry(&container->group_list,
279 			struct tce_iommu_group, next);
280 	table_group = iommu_group_get_iommudata(tcegrp->grp);
281 	if (!table_group)
282 		return -ENODEV;
283 
284 	if (!table_group->tce32_size)
285 		return -EPERM;
286 
287 	ret = tce_iommu_mm_set(container);
288 	if (ret)
289 		return ret;
290 
291 	locked = table_group->tce32_size >> PAGE_SHIFT;
292 	ret = account_locked_vm(container->mm, locked, true);
293 	if (ret)
294 		return ret;
295 
296 	container->locked_pages = locked;
297 
298 	container->enabled = true;
299 
300 	return ret;
301 }
302 
303 static void tce_iommu_disable(struct tce_container *container)
304 {
305 	if (!container->enabled)
306 		return;
307 
308 	container->enabled = false;
309 
310 	BUG_ON(!container->mm);
311 	account_locked_vm(container->mm, container->locked_pages, false);
312 }
313 
314 static void *tce_iommu_open(unsigned long arg)
315 {
316 	struct tce_container *container;
317 
318 	if ((arg != VFIO_SPAPR_TCE_IOMMU) && (arg != VFIO_SPAPR_TCE_v2_IOMMU)) {
319 		pr_err("tce_vfio: Wrong IOMMU type\n");
320 		return ERR_PTR(-EINVAL);
321 	}
322 
323 	container = kzalloc(sizeof(*container), GFP_KERNEL);
324 	if (!container)
325 		return ERR_PTR(-ENOMEM);
326 
327 	mutex_init(&container->lock);
328 	INIT_LIST_HEAD_RCU(&container->group_list);
329 	INIT_LIST_HEAD_RCU(&container->prereg_list);
330 
331 	container->v2 = arg == VFIO_SPAPR_TCE_v2_IOMMU;
332 
333 	return container;
334 }
335 
336 static int tce_iommu_clear(struct tce_container *container,
337 		struct iommu_table *tbl,
338 		unsigned long entry, unsigned long pages);
339 static void tce_iommu_free_table(struct tce_container *container,
340 		struct iommu_table *tbl);
341 
342 static void tce_iommu_release(void *iommu_data)
343 {
344 	struct tce_container *container = iommu_data;
345 	struct tce_iommu_group *tcegrp;
346 	struct tce_iommu_prereg *tcemem, *tmtmp;
347 	long i;
348 
349 	while (tce_groups_attached(container)) {
350 		tcegrp = list_first_entry(&container->group_list,
351 				struct tce_iommu_group, next);
352 		tce_iommu_detach_group(iommu_data, tcegrp->grp);
353 	}
354 
355 	/*
356 	 * If VFIO created a table, it was not disposed
357 	 * by tce_iommu_detach_group() so do it now.
358 	 */
359 	for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
360 		struct iommu_table *tbl = container->tables[i];
361 
362 		if (!tbl)
363 			continue;
364 
365 		tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
366 		tce_iommu_free_table(container, tbl);
367 	}
368 
369 	list_for_each_entry_safe(tcemem, tmtmp, &container->prereg_list, next)
370 		WARN_ON(tce_iommu_prereg_free(container, tcemem));
371 
372 	tce_iommu_disable(container);
373 	if (container->mm)
374 		mmdrop(container->mm);
375 	mutex_destroy(&container->lock);
376 
377 	kfree(container);
378 }
379 
380 static void tce_iommu_unuse_page(struct tce_container *container,
381 		unsigned long hpa)
382 {
383 	struct page *page;
384 
385 	page = pfn_to_page(hpa >> PAGE_SHIFT);
386 	put_page(page);
387 }
388 
389 static int tce_iommu_prereg_ua_to_hpa(struct tce_container *container,
390 		unsigned long tce, unsigned long shift,
391 		unsigned long *phpa, struct mm_iommu_table_group_mem_t **pmem)
392 {
393 	long ret = 0;
394 	struct mm_iommu_table_group_mem_t *mem;
395 
396 	mem = mm_iommu_lookup(container->mm, tce, 1ULL << shift);
397 	if (!mem)
398 		return -EINVAL;
399 
400 	ret = mm_iommu_ua_to_hpa(mem, tce, shift, phpa);
401 	if (ret)
402 		return -EINVAL;
403 
404 	*pmem = mem;
405 
406 	return 0;
407 }
408 
409 static void tce_iommu_unuse_page_v2(struct tce_container *container,
410 		struct iommu_table *tbl, unsigned long entry)
411 {
412 	struct mm_iommu_table_group_mem_t *mem = NULL;
413 	int ret;
414 	unsigned long hpa = 0;
415 	__be64 *pua = IOMMU_TABLE_USERSPACE_ENTRY_RO(tbl, entry);
416 
417 	if (!pua)
418 		return;
419 
420 	ret = tce_iommu_prereg_ua_to_hpa(container, be64_to_cpu(*pua),
421 			tbl->it_page_shift, &hpa, &mem);
422 	if (ret)
423 		pr_debug("%s: tce %llx at #%lx was not cached, ret=%d\n",
424 				__func__, be64_to_cpu(*pua), entry, ret);
425 	if (mem)
426 		mm_iommu_mapped_dec(mem);
427 
428 	*pua = cpu_to_be64(0);
429 }
430 
431 static int tce_iommu_clear(struct tce_container *container,
432 		struct iommu_table *tbl,
433 		unsigned long entry, unsigned long pages)
434 {
435 	unsigned long oldhpa;
436 	long ret;
437 	enum dma_data_direction direction;
438 	unsigned long lastentry = entry + pages;
439 
440 	for ( ; entry < lastentry; ++entry) {
441 		if (tbl->it_indirect_levels && tbl->it_userspace) {
442 			/*
443 			 * For multilevel tables, we can take a shortcut here
444 			 * and skip some TCEs as we know that the userspace
445 			 * addresses cache is a mirror of the real TCE table
446 			 * and if it is missing some indirect levels, then
447 			 * the hardware table does not have them allocated
448 			 * either and therefore does not require updating.
449 			 */
450 			__be64 *pua = IOMMU_TABLE_USERSPACE_ENTRY_RO(tbl,
451 					entry);
452 			if (!pua) {
453 				/* align to level_size which is power of two */
454 				entry |= tbl->it_level_size - 1;
455 				continue;
456 			}
457 		}
458 
459 		cond_resched();
460 
461 		direction = DMA_NONE;
462 		oldhpa = 0;
463 		ret = iommu_tce_xchg(container->mm, tbl, entry, &oldhpa,
464 				&direction);
465 		if (ret)
466 			continue;
467 
468 		if (direction == DMA_NONE)
469 			continue;
470 
471 		if (container->v2) {
472 			tce_iommu_unuse_page_v2(container, tbl, entry);
473 			continue;
474 		}
475 
476 		tce_iommu_unuse_page(container, oldhpa);
477 	}
478 
479 	return 0;
480 }
481 
482 static int tce_iommu_use_page(unsigned long tce, unsigned long *hpa)
483 {
484 	struct page *page = NULL;
485 	enum dma_data_direction direction = iommu_tce_direction(tce);
486 
487 	if (get_user_pages_fast(tce & PAGE_MASK, 1,
488 			direction != DMA_TO_DEVICE ? FOLL_WRITE : 0,
489 			&page) != 1)
490 		return -EFAULT;
491 
492 	*hpa = __pa((unsigned long) page_address(page));
493 
494 	return 0;
495 }
496 
497 static long tce_iommu_build(struct tce_container *container,
498 		struct iommu_table *tbl,
499 		unsigned long entry, unsigned long tce, unsigned long pages,
500 		enum dma_data_direction direction)
501 {
502 	long i, ret = 0;
503 	unsigned long hpa;
504 	enum dma_data_direction dirtmp;
505 
506 	for (i = 0; i < pages; ++i) {
507 		unsigned long offset = tce & IOMMU_PAGE_MASK(tbl) & ~PAGE_MASK;
508 
509 		ret = tce_iommu_use_page(tce, &hpa);
510 		if (ret)
511 			break;
512 
513 		if (!tce_page_is_contained(container->mm, hpa,
514 				tbl->it_page_shift)) {
515 			ret = -EPERM;
516 			break;
517 		}
518 
519 		hpa |= offset;
520 		dirtmp = direction;
521 		ret = iommu_tce_xchg(container->mm, tbl, entry + i, &hpa,
522 				&dirtmp);
523 		if (ret) {
524 			tce_iommu_unuse_page(container, hpa);
525 			pr_err("iommu_tce: %s failed ioba=%lx, tce=%lx, ret=%ld\n",
526 					__func__, entry << tbl->it_page_shift,
527 					tce, ret);
528 			break;
529 		}
530 
531 		if (dirtmp != DMA_NONE)
532 			tce_iommu_unuse_page(container, hpa);
533 
534 		tce += IOMMU_PAGE_SIZE(tbl);
535 	}
536 
537 	if (ret)
538 		tce_iommu_clear(container, tbl, entry, i);
539 
540 	return ret;
541 }
542 
543 static long tce_iommu_build_v2(struct tce_container *container,
544 		struct iommu_table *tbl,
545 		unsigned long entry, unsigned long tce, unsigned long pages,
546 		enum dma_data_direction direction)
547 {
548 	long i, ret = 0;
549 	unsigned long hpa;
550 	enum dma_data_direction dirtmp;
551 
552 	for (i = 0; i < pages; ++i) {
553 		struct mm_iommu_table_group_mem_t *mem = NULL;
554 		__be64 *pua = IOMMU_TABLE_USERSPACE_ENTRY(tbl, entry + i);
555 
556 		ret = tce_iommu_prereg_ua_to_hpa(container,
557 				tce, tbl->it_page_shift, &hpa, &mem);
558 		if (ret)
559 			break;
560 
561 		if (!tce_page_is_contained(container->mm, hpa,
562 				tbl->it_page_shift)) {
563 			ret = -EPERM;
564 			break;
565 		}
566 
567 		/* Preserve offset within IOMMU page */
568 		hpa |= tce & IOMMU_PAGE_MASK(tbl) & ~PAGE_MASK;
569 		dirtmp = direction;
570 
571 		/* The registered region is being unregistered */
572 		if (mm_iommu_mapped_inc(mem))
573 			break;
574 
575 		ret = iommu_tce_xchg(container->mm, tbl, entry + i, &hpa,
576 				&dirtmp);
577 		if (ret) {
578 			/* dirtmp cannot be DMA_NONE here */
579 			tce_iommu_unuse_page_v2(container, tbl, entry + i);
580 			pr_err("iommu_tce: %s failed ioba=%lx, tce=%lx, ret=%ld\n",
581 					__func__, entry << tbl->it_page_shift,
582 					tce, ret);
583 			break;
584 		}
585 
586 		if (dirtmp != DMA_NONE)
587 			tce_iommu_unuse_page_v2(container, tbl, entry + i);
588 
589 		*pua = cpu_to_be64(tce);
590 
591 		tce += IOMMU_PAGE_SIZE(tbl);
592 	}
593 
594 	if (ret)
595 		tce_iommu_clear(container, tbl, entry, i);
596 
597 	return ret;
598 }
599 
600 static long tce_iommu_create_table(struct tce_container *container,
601 			struct iommu_table_group *table_group,
602 			int num,
603 			__u32 page_shift,
604 			__u64 window_size,
605 			__u32 levels,
606 			struct iommu_table **ptbl)
607 {
608 	long ret, table_size;
609 
610 	table_size = table_group->ops->get_table_size(page_shift, window_size,
611 			levels);
612 	if (!table_size)
613 		return -EINVAL;
614 
615 	ret = account_locked_vm(container->mm, table_size >> PAGE_SHIFT, true);
616 	if (ret)
617 		return ret;
618 
619 	ret = table_group->ops->create_table(table_group, num,
620 			page_shift, window_size, levels, ptbl);
621 
622 	WARN_ON(!ret && !(*ptbl)->it_ops->free);
623 	WARN_ON(!ret && ((*ptbl)->it_allocated_size > table_size));
624 
625 	return ret;
626 }
627 
628 static void tce_iommu_free_table(struct tce_container *container,
629 		struct iommu_table *tbl)
630 {
631 	unsigned long pages = tbl->it_allocated_size >> PAGE_SHIFT;
632 
633 	iommu_tce_table_put(tbl);
634 	account_locked_vm(container->mm, pages, false);
635 }
636 
637 static long tce_iommu_create_window(struct tce_container *container,
638 		__u32 page_shift, __u64 window_size, __u32 levels,
639 		__u64 *start_addr)
640 {
641 	struct tce_iommu_group *tcegrp;
642 	struct iommu_table_group *table_group;
643 	struct iommu_table *tbl = NULL;
644 	long ret, num;
645 
646 	num = tce_iommu_find_free_table(container);
647 	if (num < 0)
648 		return num;
649 
650 	/* Get the first group for ops::create_table */
651 	tcegrp = list_first_entry(&container->group_list,
652 			struct tce_iommu_group, next);
653 	table_group = iommu_group_get_iommudata(tcegrp->grp);
654 	if (!table_group)
655 		return -EFAULT;
656 
657 	if (!(table_group->pgsizes & (1ULL << page_shift)))
658 		return -EINVAL;
659 
660 	if (!table_group->ops->set_window || !table_group->ops->unset_window ||
661 			!table_group->ops->get_table_size ||
662 			!table_group->ops->create_table)
663 		return -EPERM;
664 
665 	/* Create TCE table */
666 	ret = tce_iommu_create_table(container, table_group, num,
667 			page_shift, window_size, levels, &tbl);
668 	if (ret)
669 		return ret;
670 
671 	BUG_ON(!tbl->it_ops->free);
672 
673 	/*
674 	 * Program the table to every group.
675 	 * Groups have been tested for compatibility at the attach time.
676 	 */
677 	list_for_each_entry(tcegrp, &container->group_list, next) {
678 		table_group = iommu_group_get_iommudata(tcegrp->grp);
679 
680 		ret = table_group->ops->set_window(table_group, num, tbl);
681 		if (ret)
682 			goto unset_exit;
683 	}
684 
685 	container->tables[num] = tbl;
686 
687 	/* Return start address assigned by platform in create_table() */
688 	*start_addr = tbl->it_offset << tbl->it_page_shift;
689 
690 	return 0;
691 
692 unset_exit:
693 	list_for_each_entry(tcegrp, &container->group_list, next) {
694 		table_group = iommu_group_get_iommudata(tcegrp->grp);
695 		table_group->ops->unset_window(table_group, num);
696 	}
697 	tce_iommu_free_table(container, tbl);
698 
699 	return ret;
700 }
701 
702 static long tce_iommu_remove_window(struct tce_container *container,
703 		__u64 start_addr)
704 {
705 	struct iommu_table_group *table_group = NULL;
706 	struct iommu_table *tbl;
707 	struct tce_iommu_group *tcegrp;
708 	int num;
709 
710 	num = tce_iommu_find_table(container, start_addr, &tbl);
711 	if (num < 0)
712 		return -EINVAL;
713 
714 	BUG_ON(!tbl->it_size);
715 
716 	/* Detach groups from IOMMUs */
717 	list_for_each_entry(tcegrp, &container->group_list, next) {
718 		table_group = iommu_group_get_iommudata(tcegrp->grp);
719 
720 		/*
721 		 * SPAPR TCE IOMMU exposes the default DMA window to
722 		 * the guest via dma32_window_start/size of
723 		 * VFIO_IOMMU_SPAPR_TCE_GET_INFO. Some platforms allow
724 		 * the userspace to remove this window, some do not so
725 		 * here we check for the platform capability.
726 		 */
727 		if (!table_group->ops || !table_group->ops->unset_window)
728 			return -EPERM;
729 
730 		table_group->ops->unset_window(table_group, num);
731 	}
732 
733 	/* Free table */
734 	tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
735 	tce_iommu_free_table(container, tbl);
736 	container->tables[num] = NULL;
737 
738 	return 0;
739 }
740 
741 static long tce_iommu_create_default_window(struct tce_container *container)
742 {
743 	long ret;
744 	__u64 start_addr = 0;
745 	struct tce_iommu_group *tcegrp;
746 	struct iommu_table_group *table_group;
747 
748 	if (!container->def_window_pending)
749 		return 0;
750 
751 	if (!tce_groups_attached(container))
752 		return -ENODEV;
753 
754 	tcegrp = list_first_entry(&container->group_list,
755 			struct tce_iommu_group, next);
756 	table_group = iommu_group_get_iommudata(tcegrp->grp);
757 	if (!table_group)
758 		return -ENODEV;
759 
760 	ret = tce_iommu_create_window(container, IOMMU_PAGE_SHIFT_4K,
761 			table_group->tce32_size, 1, &start_addr);
762 	WARN_ON_ONCE(!ret && start_addr);
763 
764 	if (!ret)
765 		container->def_window_pending = false;
766 
767 	return ret;
768 }
769 
770 static long tce_iommu_ioctl(void *iommu_data,
771 				 unsigned int cmd, unsigned long arg)
772 {
773 	struct tce_container *container = iommu_data;
774 	unsigned long minsz, ddwsz;
775 	long ret;
776 
777 	switch (cmd) {
778 	case VFIO_CHECK_EXTENSION:
779 		switch (arg) {
780 		case VFIO_SPAPR_TCE_IOMMU:
781 		case VFIO_SPAPR_TCE_v2_IOMMU:
782 			ret = 1;
783 			break;
784 		default:
785 			ret = vfio_spapr_iommu_eeh_ioctl(NULL, cmd, arg);
786 			break;
787 		}
788 
789 		return (ret < 0) ? 0 : ret;
790 	}
791 
792 	/*
793 	 * Sanity check to prevent one userspace from manipulating
794 	 * another userspace mm.
795 	 */
796 	BUG_ON(!container);
797 	if (container->mm && container->mm != current->mm)
798 		return -EPERM;
799 
800 	switch (cmd) {
801 	case VFIO_IOMMU_SPAPR_TCE_GET_INFO: {
802 		struct vfio_iommu_spapr_tce_info info;
803 		struct tce_iommu_group *tcegrp;
804 		struct iommu_table_group *table_group;
805 
806 		if (!tce_groups_attached(container))
807 			return -ENXIO;
808 
809 		tcegrp = list_first_entry(&container->group_list,
810 				struct tce_iommu_group, next);
811 		table_group = iommu_group_get_iommudata(tcegrp->grp);
812 
813 		if (!table_group)
814 			return -ENXIO;
815 
816 		minsz = offsetofend(struct vfio_iommu_spapr_tce_info,
817 				dma32_window_size);
818 
819 		if (copy_from_user(&info, (void __user *)arg, minsz))
820 			return -EFAULT;
821 
822 		if (info.argsz < minsz)
823 			return -EINVAL;
824 
825 		info.dma32_window_start = table_group->tce32_start;
826 		info.dma32_window_size = table_group->tce32_size;
827 		info.flags = 0;
828 		memset(&info.ddw, 0, sizeof(info.ddw));
829 
830 		if (table_group->max_dynamic_windows_supported &&
831 				container->v2) {
832 			info.flags |= VFIO_IOMMU_SPAPR_INFO_DDW;
833 			info.ddw.pgsizes = table_group->pgsizes;
834 			info.ddw.max_dynamic_windows_supported =
835 				table_group->max_dynamic_windows_supported;
836 			info.ddw.levels = table_group->max_levels;
837 		}
838 
839 		ddwsz = offsetofend(struct vfio_iommu_spapr_tce_info, ddw);
840 
841 		if (info.argsz >= ddwsz)
842 			minsz = ddwsz;
843 
844 		if (copy_to_user((void __user *)arg, &info, minsz))
845 			return -EFAULT;
846 
847 		return 0;
848 	}
849 	case VFIO_IOMMU_MAP_DMA: {
850 		struct vfio_iommu_type1_dma_map param;
851 		struct iommu_table *tbl = NULL;
852 		long num;
853 		enum dma_data_direction direction;
854 
855 		if (!container->enabled)
856 			return -EPERM;
857 
858 		minsz = offsetofend(struct vfio_iommu_type1_dma_map, size);
859 
860 		if (copy_from_user(&param, (void __user *)arg, minsz))
861 			return -EFAULT;
862 
863 		if (param.argsz < minsz)
864 			return -EINVAL;
865 
866 		if (param.flags & ~(VFIO_DMA_MAP_FLAG_READ |
867 				VFIO_DMA_MAP_FLAG_WRITE))
868 			return -EINVAL;
869 
870 		ret = tce_iommu_create_default_window(container);
871 		if (ret)
872 			return ret;
873 
874 		num = tce_iommu_find_table(container, param.iova, &tbl);
875 		if (num < 0)
876 			return -ENXIO;
877 
878 		if ((param.size & ~IOMMU_PAGE_MASK(tbl)) ||
879 				(param.vaddr & ~IOMMU_PAGE_MASK(tbl)))
880 			return -EINVAL;
881 
882 		/* iova is checked by the IOMMU API */
883 		if (param.flags & VFIO_DMA_MAP_FLAG_READ) {
884 			if (param.flags & VFIO_DMA_MAP_FLAG_WRITE)
885 				direction = DMA_BIDIRECTIONAL;
886 			else
887 				direction = DMA_TO_DEVICE;
888 		} else {
889 			if (param.flags & VFIO_DMA_MAP_FLAG_WRITE)
890 				direction = DMA_FROM_DEVICE;
891 			else
892 				return -EINVAL;
893 		}
894 
895 		ret = iommu_tce_put_param_check(tbl, param.iova, param.vaddr);
896 		if (ret)
897 			return ret;
898 
899 		if (container->v2)
900 			ret = tce_iommu_build_v2(container, tbl,
901 					param.iova >> tbl->it_page_shift,
902 					param.vaddr,
903 					param.size >> tbl->it_page_shift,
904 					direction);
905 		else
906 			ret = tce_iommu_build(container, tbl,
907 					param.iova >> tbl->it_page_shift,
908 					param.vaddr,
909 					param.size >> tbl->it_page_shift,
910 					direction);
911 
912 		iommu_flush_tce(tbl);
913 
914 		return ret;
915 	}
916 	case VFIO_IOMMU_UNMAP_DMA: {
917 		struct vfio_iommu_type1_dma_unmap param;
918 		struct iommu_table *tbl = NULL;
919 		long num;
920 
921 		if (!container->enabled)
922 			return -EPERM;
923 
924 		minsz = offsetofend(struct vfio_iommu_type1_dma_unmap,
925 				size);
926 
927 		if (copy_from_user(&param, (void __user *)arg, minsz))
928 			return -EFAULT;
929 
930 		if (param.argsz < minsz)
931 			return -EINVAL;
932 
933 		/* No flag is supported now */
934 		if (param.flags)
935 			return -EINVAL;
936 
937 		ret = tce_iommu_create_default_window(container);
938 		if (ret)
939 			return ret;
940 
941 		num = tce_iommu_find_table(container, param.iova, &tbl);
942 		if (num < 0)
943 			return -ENXIO;
944 
945 		if (param.size & ~IOMMU_PAGE_MASK(tbl))
946 			return -EINVAL;
947 
948 		ret = iommu_tce_clear_param_check(tbl, param.iova, 0,
949 				param.size >> tbl->it_page_shift);
950 		if (ret)
951 			return ret;
952 
953 		ret = tce_iommu_clear(container, tbl,
954 				param.iova >> tbl->it_page_shift,
955 				param.size >> tbl->it_page_shift);
956 		iommu_flush_tce(tbl);
957 
958 		return ret;
959 	}
960 	case VFIO_IOMMU_SPAPR_REGISTER_MEMORY: {
961 		struct vfio_iommu_spapr_register_memory param;
962 
963 		if (!container->v2)
964 			break;
965 
966 		minsz = offsetofend(struct vfio_iommu_spapr_register_memory,
967 				size);
968 
969 		ret = tce_iommu_mm_set(container);
970 		if (ret)
971 			return ret;
972 
973 		if (copy_from_user(&param, (void __user *)arg, minsz))
974 			return -EFAULT;
975 
976 		if (param.argsz < minsz)
977 			return -EINVAL;
978 
979 		/* No flag is supported now */
980 		if (param.flags)
981 			return -EINVAL;
982 
983 		mutex_lock(&container->lock);
984 		ret = tce_iommu_register_pages(container, param.vaddr,
985 				param.size);
986 		mutex_unlock(&container->lock);
987 
988 		return ret;
989 	}
990 	case VFIO_IOMMU_SPAPR_UNREGISTER_MEMORY: {
991 		struct vfio_iommu_spapr_register_memory param;
992 
993 		if (!container->v2)
994 			break;
995 
996 		if (!container->mm)
997 			return -EPERM;
998 
999 		minsz = offsetofend(struct vfio_iommu_spapr_register_memory,
1000 				size);
1001 
1002 		if (copy_from_user(&param, (void __user *)arg, minsz))
1003 			return -EFAULT;
1004 
1005 		if (param.argsz < minsz)
1006 			return -EINVAL;
1007 
1008 		/* No flag is supported now */
1009 		if (param.flags)
1010 			return -EINVAL;
1011 
1012 		mutex_lock(&container->lock);
1013 		ret = tce_iommu_unregister_pages(container, param.vaddr,
1014 				param.size);
1015 		mutex_unlock(&container->lock);
1016 
1017 		return ret;
1018 	}
1019 	case VFIO_IOMMU_ENABLE:
1020 		if (container->v2)
1021 			break;
1022 
1023 		mutex_lock(&container->lock);
1024 		ret = tce_iommu_enable(container);
1025 		mutex_unlock(&container->lock);
1026 		return ret;
1027 
1028 
1029 	case VFIO_IOMMU_DISABLE:
1030 		if (container->v2)
1031 			break;
1032 
1033 		mutex_lock(&container->lock);
1034 		tce_iommu_disable(container);
1035 		mutex_unlock(&container->lock);
1036 		return 0;
1037 
1038 	case VFIO_EEH_PE_OP: {
1039 		struct tce_iommu_group *tcegrp;
1040 
1041 		ret = 0;
1042 		list_for_each_entry(tcegrp, &container->group_list, next) {
1043 			ret = vfio_spapr_iommu_eeh_ioctl(tcegrp->grp,
1044 					cmd, arg);
1045 			if (ret)
1046 				return ret;
1047 		}
1048 		return ret;
1049 	}
1050 
1051 	case VFIO_IOMMU_SPAPR_TCE_CREATE: {
1052 		struct vfio_iommu_spapr_tce_create create;
1053 
1054 		if (!container->v2)
1055 			break;
1056 
1057 		ret = tce_iommu_mm_set(container);
1058 		if (ret)
1059 			return ret;
1060 
1061 		if (!tce_groups_attached(container))
1062 			return -ENXIO;
1063 
1064 		minsz = offsetofend(struct vfio_iommu_spapr_tce_create,
1065 				start_addr);
1066 
1067 		if (copy_from_user(&create, (void __user *)arg, minsz))
1068 			return -EFAULT;
1069 
1070 		if (create.argsz < minsz)
1071 			return -EINVAL;
1072 
1073 		if (create.flags)
1074 			return -EINVAL;
1075 
1076 		mutex_lock(&container->lock);
1077 
1078 		ret = tce_iommu_create_default_window(container);
1079 		if (!ret)
1080 			ret = tce_iommu_create_window(container,
1081 					create.page_shift,
1082 					create.window_size, create.levels,
1083 					&create.start_addr);
1084 
1085 		mutex_unlock(&container->lock);
1086 
1087 		if (!ret && copy_to_user((void __user *)arg, &create, minsz))
1088 			ret = -EFAULT;
1089 
1090 		return ret;
1091 	}
1092 	case VFIO_IOMMU_SPAPR_TCE_REMOVE: {
1093 		struct vfio_iommu_spapr_tce_remove remove;
1094 
1095 		if (!container->v2)
1096 			break;
1097 
1098 		ret = tce_iommu_mm_set(container);
1099 		if (ret)
1100 			return ret;
1101 
1102 		if (!tce_groups_attached(container))
1103 			return -ENXIO;
1104 
1105 		minsz = offsetofend(struct vfio_iommu_spapr_tce_remove,
1106 				start_addr);
1107 
1108 		if (copy_from_user(&remove, (void __user *)arg, minsz))
1109 			return -EFAULT;
1110 
1111 		if (remove.argsz < minsz)
1112 			return -EINVAL;
1113 
1114 		if (remove.flags)
1115 			return -EINVAL;
1116 
1117 		if (container->def_window_pending && !remove.start_addr) {
1118 			container->def_window_pending = false;
1119 			return 0;
1120 		}
1121 
1122 		mutex_lock(&container->lock);
1123 
1124 		ret = tce_iommu_remove_window(container, remove.start_addr);
1125 
1126 		mutex_unlock(&container->lock);
1127 
1128 		return ret;
1129 	}
1130 	}
1131 
1132 	return -ENOTTY;
1133 }
1134 
1135 static void tce_iommu_release_ownership(struct tce_container *container,
1136 		struct iommu_table_group *table_group)
1137 {
1138 	int i;
1139 
1140 	for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1141 		struct iommu_table *tbl = container->tables[i];
1142 
1143 		if (!tbl)
1144 			continue;
1145 
1146 		tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
1147 		if (tbl->it_map)
1148 			iommu_release_ownership(tbl);
1149 
1150 		container->tables[i] = NULL;
1151 	}
1152 }
1153 
1154 static int tce_iommu_take_ownership(struct tce_container *container,
1155 		struct iommu_table_group *table_group)
1156 {
1157 	int i, j, rc = 0;
1158 
1159 	for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1160 		struct iommu_table *tbl = table_group->tables[i];
1161 
1162 		if (!tbl || !tbl->it_map)
1163 			continue;
1164 
1165 		rc = iommu_take_ownership(tbl);
1166 		if (rc) {
1167 			for (j = 0; j < i; ++j)
1168 				iommu_release_ownership(
1169 						table_group->tables[j]);
1170 
1171 			return rc;
1172 		}
1173 	}
1174 
1175 	for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1176 		container->tables[i] = table_group->tables[i];
1177 
1178 	return 0;
1179 }
1180 
1181 static void tce_iommu_release_ownership_ddw(struct tce_container *container,
1182 		struct iommu_table_group *table_group)
1183 {
1184 	long i;
1185 
1186 	if (!table_group->ops->unset_window) {
1187 		WARN_ON_ONCE(1);
1188 		return;
1189 	}
1190 
1191 	for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1192 		if (container->tables[i])
1193 			table_group->ops->unset_window(table_group, i);
1194 
1195 	table_group->ops->release_ownership(table_group);
1196 }
1197 
1198 static long tce_iommu_take_ownership_ddw(struct tce_container *container,
1199 		struct iommu_table_group *table_group)
1200 {
1201 	long i, ret = 0;
1202 
1203 	if (!table_group->ops->create_table || !table_group->ops->set_window ||
1204 			!table_group->ops->release_ownership) {
1205 		WARN_ON_ONCE(1);
1206 		return -EFAULT;
1207 	}
1208 
1209 	table_group->ops->take_ownership(table_group);
1210 
1211 	/* Set all windows to the new group */
1212 	for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1213 		struct iommu_table *tbl = container->tables[i];
1214 
1215 		if (!tbl)
1216 			continue;
1217 
1218 		ret = table_group->ops->set_window(table_group, i, tbl);
1219 		if (ret)
1220 			goto release_exit;
1221 	}
1222 
1223 	return 0;
1224 
1225 release_exit:
1226 	for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1227 		table_group->ops->unset_window(table_group, i);
1228 
1229 	table_group->ops->release_ownership(table_group);
1230 
1231 	return ret;
1232 }
1233 
1234 static int tce_iommu_attach_group(void *iommu_data,
1235 		struct iommu_group *iommu_group)
1236 {
1237 	int ret;
1238 	struct tce_container *container = iommu_data;
1239 	struct iommu_table_group *table_group;
1240 	struct tce_iommu_group *tcegrp = NULL;
1241 
1242 	mutex_lock(&container->lock);
1243 
1244 	/* pr_debug("tce_vfio: Attaching group #%u to iommu %p\n",
1245 			iommu_group_id(iommu_group), iommu_group); */
1246 	table_group = iommu_group_get_iommudata(iommu_group);
1247 	if (!table_group) {
1248 		ret = -ENODEV;
1249 		goto unlock_exit;
1250 	}
1251 
1252 	if (tce_groups_attached(container) && (!table_group->ops ||
1253 			!table_group->ops->take_ownership ||
1254 			!table_group->ops->release_ownership)) {
1255 		ret = -EBUSY;
1256 		goto unlock_exit;
1257 	}
1258 
1259 	/* Check if new group has the same iommu_ops (i.e. compatible) */
1260 	list_for_each_entry(tcegrp, &container->group_list, next) {
1261 		struct iommu_table_group *table_group_tmp;
1262 
1263 		if (tcegrp->grp == iommu_group) {
1264 			pr_warn("tce_vfio: Group %d is already attached\n",
1265 					iommu_group_id(iommu_group));
1266 			ret = -EBUSY;
1267 			goto unlock_exit;
1268 		}
1269 		table_group_tmp = iommu_group_get_iommudata(tcegrp->grp);
1270 		if (table_group_tmp->ops->create_table !=
1271 				table_group->ops->create_table) {
1272 			pr_warn("tce_vfio: Group %d is incompatible with group %d\n",
1273 					iommu_group_id(iommu_group),
1274 					iommu_group_id(tcegrp->grp));
1275 			ret = -EPERM;
1276 			goto unlock_exit;
1277 		}
1278 	}
1279 
1280 	tcegrp = kzalloc(sizeof(*tcegrp), GFP_KERNEL);
1281 	if (!tcegrp) {
1282 		ret = -ENOMEM;
1283 		goto unlock_exit;
1284 	}
1285 
1286 	if (!table_group->ops || !table_group->ops->take_ownership ||
1287 			!table_group->ops->release_ownership) {
1288 		if (container->v2) {
1289 			ret = -EPERM;
1290 			goto unlock_exit;
1291 		}
1292 		ret = tce_iommu_take_ownership(container, table_group);
1293 	} else {
1294 		if (!container->v2) {
1295 			ret = -EPERM;
1296 			goto unlock_exit;
1297 		}
1298 		ret = tce_iommu_take_ownership_ddw(container, table_group);
1299 		if (!tce_groups_attached(container) && !container->tables[0])
1300 			container->def_window_pending = true;
1301 	}
1302 
1303 	if (!ret) {
1304 		tcegrp->grp = iommu_group;
1305 		list_add(&tcegrp->next, &container->group_list);
1306 	}
1307 
1308 unlock_exit:
1309 	if (ret && tcegrp)
1310 		kfree(tcegrp);
1311 
1312 	mutex_unlock(&container->lock);
1313 
1314 	return ret;
1315 }
1316 
1317 static void tce_iommu_detach_group(void *iommu_data,
1318 		struct iommu_group *iommu_group)
1319 {
1320 	struct tce_container *container = iommu_data;
1321 	struct iommu_table_group *table_group;
1322 	bool found = false;
1323 	struct tce_iommu_group *tcegrp;
1324 
1325 	mutex_lock(&container->lock);
1326 
1327 	list_for_each_entry(tcegrp, &container->group_list, next) {
1328 		if (tcegrp->grp == iommu_group) {
1329 			found = true;
1330 			break;
1331 		}
1332 	}
1333 
1334 	if (!found) {
1335 		pr_warn("tce_vfio: detaching unattached group #%u\n",
1336 				iommu_group_id(iommu_group));
1337 		goto unlock_exit;
1338 	}
1339 
1340 	list_del(&tcegrp->next);
1341 	kfree(tcegrp);
1342 
1343 	table_group = iommu_group_get_iommudata(iommu_group);
1344 	BUG_ON(!table_group);
1345 
1346 	if (!table_group->ops || !table_group->ops->release_ownership)
1347 		tce_iommu_release_ownership(container, table_group);
1348 	else
1349 		tce_iommu_release_ownership_ddw(container, table_group);
1350 
1351 unlock_exit:
1352 	mutex_unlock(&container->lock);
1353 }
1354 
1355 static const struct vfio_iommu_driver_ops tce_iommu_driver_ops = {
1356 	.name		= "iommu-vfio-powerpc",
1357 	.owner		= THIS_MODULE,
1358 	.open		= tce_iommu_open,
1359 	.release	= tce_iommu_release,
1360 	.ioctl		= tce_iommu_ioctl,
1361 	.attach_group	= tce_iommu_attach_group,
1362 	.detach_group	= tce_iommu_detach_group,
1363 };
1364 
1365 static int __init tce_iommu_init(void)
1366 {
1367 	return vfio_register_iommu_driver(&tce_iommu_driver_ops);
1368 }
1369 
1370 static void __exit tce_iommu_cleanup(void)
1371 {
1372 	vfio_unregister_iommu_driver(&tce_iommu_driver_ops);
1373 }
1374 
1375 module_init(tce_iommu_init);
1376 module_exit(tce_iommu_cleanup);
1377 
1378 MODULE_VERSION(DRIVER_VERSION);
1379 MODULE_LICENSE("GPL v2");
1380 MODULE_AUTHOR(DRIVER_AUTHOR);
1381 MODULE_DESCRIPTION(DRIVER_DESC);
1382 
1383