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 #include "vfio.h"
24 
25 #include <asm/iommu.h>
26 #include <asm/tce.h>
27 #include <asm/mmu_context.h>
28 
29 #define DRIVER_VERSION  "0.1"
30 #define DRIVER_AUTHOR   "aik@ozlabs.ru"
31 #define DRIVER_DESC     "VFIO IOMMU SPAPR TCE"
32 
33 static void tce_iommu_detach_group(void *iommu_data,
34 		struct iommu_group *iommu_group);
35 
36 /*
37  * VFIO IOMMU fd for SPAPR_TCE IOMMU implementation
38  *
39  * This code handles mapping and unmapping of user data buffers
40  * into DMA'ble space using the IOMMU
41  */
42 
43 struct tce_iommu_group {
44 	struct list_head next;
45 	struct iommu_group *grp;
46 };
47 
48 /*
49  * A container needs to remember which preregistered region  it has
50  * referenced to do proper cleanup at the userspace process exit.
51  */
52 struct tce_iommu_prereg {
53 	struct list_head next;
54 	struct mm_iommu_table_group_mem_t *mem;
55 };
56 
57 /*
58  * The container descriptor supports only a single group per container.
59  * Required by the API as the container is not supplied with the IOMMU group
60  * at the moment of initialization.
61  */
62 struct tce_container {
63 	struct mutex lock;
64 	bool enabled;
65 	bool v2;
66 	bool def_window_pending;
67 	unsigned long locked_pages;
68 	struct mm_struct *mm;
69 	struct iommu_table *tables[IOMMU_TABLE_GROUP_MAX_TABLES];
70 	struct list_head group_list;
71 	struct list_head prereg_list;
72 };
73 
74 static long tce_iommu_mm_set(struct tce_container *container)
75 {
76 	if (container->mm) {
77 		if (container->mm == current->mm)
78 			return 0;
79 		return -EPERM;
80 	}
81 	BUG_ON(!current->mm);
82 	container->mm = current->mm;
83 	mmgrab(container->mm);
84 
85 	return 0;
86 }
87 
88 static long tce_iommu_prereg_free(struct tce_container *container,
89 		struct tce_iommu_prereg *tcemem)
90 {
91 	long ret;
92 
93 	ret = mm_iommu_put(container->mm, tcemem->mem);
94 	if (ret)
95 		return ret;
96 
97 	list_del(&tcemem->next);
98 	kfree(tcemem);
99 
100 	return 0;
101 }
102 
103 static long tce_iommu_unregister_pages(struct tce_container *container,
104 		__u64 vaddr, __u64 size)
105 {
106 	struct mm_iommu_table_group_mem_t *mem;
107 	struct tce_iommu_prereg *tcemem;
108 	bool found = false;
109 	long ret;
110 
111 	if ((vaddr & ~PAGE_MASK) || (size & ~PAGE_MASK))
112 		return -EINVAL;
113 
114 	mem = mm_iommu_get(container->mm, vaddr, size >> PAGE_SHIFT);
115 	if (!mem)
116 		return -ENOENT;
117 
118 	list_for_each_entry(tcemem, &container->prereg_list, next) {
119 		if (tcemem->mem == mem) {
120 			found = true;
121 			break;
122 		}
123 	}
124 
125 	if (!found)
126 		ret = -ENOENT;
127 	else
128 		ret = tce_iommu_prereg_free(container, tcemem);
129 
130 	mm_iommu_put(container->mm, mem);
131 
132 	return ret;
133 }
134 
135 static long tce_iommu_register_pages(struct tce_container *container,
136 		__u64 vaddr, __u64 size)
137 {
138 	long ret = 0;
139 	struct mm_iommu_table_group_mem_t *mem = NULL;
140 	struct tce_iommu_prereg *tcemem;
141 	unsigned long entries = size >> PAGE_SHIFT;
142 
143 	if ((vaddr & ~PAGE_MASK) || (size & ~PAGE_MASK) ||
144 			((vaddr + size) < vaddr))
145 		return -EINVAL;
146 
147 	mem = mm_iommu_get(container->mm, vaddr, entries);
148 	if (mem) {
149 		list_for_each_entry(tcemem, &container->prereg_list, next) {
150 			if (tcemem->mem == mem) {
151 				ret = -EBUSY;
152 				goto put_exit;
153 			}
154 		}
155 	} else {
156 		ret = mm_iommu_new(container->mm, vaddr, entries, &mem);
157 		if (ret)
158 			return ret;
159 	}
160 
161 	tcemem = kzalloc(sizeof(*tcemem), GFP_KERNEL);
162 	if (!tcemem) {
163 		ret = -ENOMEM;
164 		goto put_exit;
165 	}
166 
167 	tcemem->mem = mem;
168 	list_add(&tcemem->next, &container->prereg_list);
169 
170 	container->enabled = true;
171 
172 	return 0;
173 
174 put_exit:
175 	mm_iommu_put(container->mm, mem);
176 	return ret;
177 }
178 
179 static bool tce_page_is_contained(struct mm_struct *mm, unsigned long hpa,
180 		unsigned int it_page_shift)
181 {
182 	struct page *page;
183 	unsigned long size = 0;
184 
185 	if (mm_iommu_is_devmem(mm, hpa, it_page_shift, &size))
186 		return size == (1UL << it_page_shift);
187 
188 	page = pfn_to_page(hpa >> PAGE_SHIFT);
189 	/*
190 	 * Check that the TCE table granularity is not bigger than the size of
191 	 * a page we just found. Otherwise the hardware can get access to
192 	 * a bigger memory chunk that it should.
193 	 */
194 	return page_shift(compound_head(page)) >= it_page_shift;
195 }
196 
197 static inline bool tce_groups_attached(struct tce_container *container)
198 {
199 	return !list_empty(&container->group_list);
200 }
201 
202 static long tce_iommu_find_table(struct tce_container *container,
203 		phys_addr_t ioba, struct iommu_table **ptbl)
204 {
205 	long i;
206 
207 	for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
208 		struct iommu_table *tbl = container->tables[i];
209 
210 		if (tbl) {
211 			unsigned long entry = ioba >> tbl->it_page_shift;
212 			unsigned long start = tbl->it_offset;
213 			unsigned long end = start + tbl->it_size;
214 
215 			if ((start <= entry) && (entry < end)) {
216 				*ptbl = tbl;
217 				return i;
218 			}
219 		}
220 	}
221 
222 	return -1;
223 }
224 
225 static int tce_iommu_find_free_table(struct tce_container *container)
226 {
227 	int i;
228 
229 	for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
230 		if (!container->tables[i])
231 			return i;
232 	}
233 
234 	return -ENOSPC;
235 }
236 
237 static int tce_iommu_enable(struct tce_container *container)
238 {
239 	int ret = 0;
240 	unsigned long locked;
241 	struct iommu_table_group *table_group;
242 	struct tce_iommu_group *tcegrp;
243 
244 	if (container->enabled)
245 		return -EBUSY;
246 
247 	/*
248 	 * When userspace pages are mapped into the IOMMU, they are effectively
249 	 * locked memory, so, theoretically, we need to update the accounting
250 	 * of locked pages on each map and unmap.  For powerpc, the map unmap
251 	 * paths can be very hot, though, and the accounting would kill
252 	 * performance, especially since it would be difficult to impossible
253 	 * to handle the accounting in real mode only.
254 	 *
255 	 * To address that, rather than precisely accounting every page, we
256 	 * instead account for a worst case on locked memory when the iommu is
257 	 * enabled and disabled.  The worst case upper bound on locked memory
258 	 * is the size of the whole iommu window, which is usually relatively
259 	 * small (compared to total memory sizes) on POWER hardware.
260 	 *
261 	 * Also we don't have a nice way to fail on H_PUT_TCE due to ulimits,
262 	 * that would effectively kill the guest at random points, much better
263 	 * enforcing the limit based on the max that the guest can map.
264 	 *
265 	 * Unfortunately at the moment it counts whole tables, no matter how
266 	 * much memory the guest has. I.e. for 4GB guest and 4 IOMMU groups
267 	 * each with 2GB DMA window, 8GB will be counted here. The reason for
268 	 * this is that we cannot tell here the amount of RAM used by the guest
269 	 * as this information is only available from KVM and VFIO is
270 	 * KVM agnostic.
271 	 *
272 	 * So we do not allow enabling a container without a group attached
273 	 * as there is no way to know how much we should increment
274 	 * the locked_vm counter.
275 	 */
276 	if (!tce_groups_attached(container))
277 		return -ENODEV;
278 
279 	tcegrp = list_first_entry(&container->group_list,
280 			struct tce_iommu_group, next);
281 	table_group = iommu_group_get_iommudata(tcegrp->grp);
282 	if (!table_group)
283 		return -ENODEV;
284 
285 	if (!table_group->tce32_size)
286 		return -EPERM;
287 
288 	ret = tce_iommu_mm_set(container);
289 	if (ret)
290 		return ret;
291 
292 	locked = table_group->tce32_size >> PAGE_SHIFT;
293 	ret = account_locked_vm(container->mm, locked, true);
294 	if (ret)
295 		return ret;
296 
297 	container->locked_pages = locked;
298 
299 	container->enabled = true;
300 
301 	return ret;
302 }
303 
304 static void tce_iommu_disable(struct tce_container *container)
305 {
306 	if (!container->enabled)
307 		return;
308 
309 	container->enabled = false;
310 
311 	BUG_ON(!container->mm);
312 	account_locked_vm(container->mm, container->locked_pages, false);
313 }
314 
315 static void *tce_iommu_open(unsigned long arg)
316 {
317 	struct tce_container *container;
318 
319 	if ((arg != VFIO_SPAPR_TCE_IOMMU) && (arg != VFIO_SPAPR_TCE_v2_IOMMU)) {
320 		pr_err("tce_vfio: Wrong IOMMU type\n");
321 		return ERR_PTR(-EINVAL);
322 	}
323 
324 	container = kzalloc(sizeof(*container), GFP_KERNEL);
325 	if (!container)
326 		return ERR_PTR(-ENOMEM);
327 
328 	mutex_init(&container->lock);
329 	INIT_LIST_HEAD_RCU(&container->group_list);
330 	INIT_LIST_HEAD_RCU(&container->prereg_list);
331 
332 	container->v2 = arg == VFIO_SPAPR_TCE_v2_IOMMU;
333 
334 	return container;
335 }
336 
337 static int tce_iommu_clear(struct tce_container *container,
338 		struct iommu_table *tbl,
339 		unsigned long entry, unsigned long pages);
340 static void tce_iommu_free_table(struct tce_container *container,
341 		struct iommu_table *tbl);
342 
343 static void tce_iommu_release(void *iommu_data)
344 {
345 	struct tce_container *container = iommu_data;
346 	struct tce_iommu_group *tcegrp;
347 	struct tce_iommu_prereg *tcemem, *tmtmp;
348 	long i;
349 
350 	while (tce_groups_attached(container)) {
351 		tcegrp = list_first_entry(&container->group_list,
352 				struct tce_iommu_group, next);
353 		tce_iommu_detach_group(iommu_data, tcegrp->grp);
354 	}
355 
356 	/*
357 	 * If VFIO created a table, it was not disposed
358 	 * by tce_iommu_detach_group() so do it now.
359 	 */
360 	for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
361 		struct iommu_table *tbl = container->tables[i];
362 
363 		if (!tbl)
364 			continue;
365 
366 		tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
367 		tce_iommu_free_table(container, tbl);
368 	}
369 
370 	list_for_each_entry_safe(tcemem, tmtmp, &container->prereg_list, next)
371 		WARN_ON(tce_iommu_prereg_free(container, tcemem));
372 
373 	tce_iommu_disable(container);
374 	if (container->mm)
375 		mmdrop(container->mm);
376 	mutex_destroy(&container->lock);
377 
378 	kfree(container);
379 }
380 
381 static void tce_iommu_unuse_page(struct tce_container *container,
382 		unsigned long hpa)
383 {
384 	struct page *page;
385 
386 	page = pfn_to_page(hpa >> PAGE_SHIFT);
387 	unpin_user_page(page);
388 }
389 
390 static int tce_iommu_prereg_ua_to_hpa(struct tce_container *container,
391 		unsigned long tce, unsigned long shift,
392 		unsigned long *phpa, struct mm_iommu_table_group_mem_t **pmem)
393 {
394 	long ret = 0;
395 	struct mm_iommu_table_group_mem_t *mem;
396 
397 	mem = mm_iommu_lookup(container->mm, tce, 1ULL << shift);
398 	if (!mem)
399 		return -EINVAL;
400 
401 	ret = mm_iommu_ua_to_hpa(mem, tce, shift, phpa);
402 	if (ret)
403 		return -EINVAL;
404 
405 	*pmem = mem;
406 
407 	return 0;
408 }
409 
410 static void tce_iommu_unuse_page_v2(struct tce_container *container,
411 		struct iommu_table *tbl, unsigned long entry)
412 {
413 	struct mm_iommu_table_group_mem_t *mem = NULL;
414 	int ret;
415 	unsigned long hpa = 0;
416 	__be64 *pua = IOMMU_TABLE_USERSPACE_ENTRY_RO(tbl, entry);
417 
418 	if (!pua)
419 		return;
420 
421 	ret = tce_iommu_prereg_ua_to_hpa(container, be64_to_cpu(*pua),
422 			tbl->it_page_shift, &hpa, &mem);
423 	if (ret)
424 		pr_debug("%s: tce %llx at #%lx was not cached, ret=%d\n",
425 				__func__, be64_to_cpu(*pua), entry, ret);
426 	if (mem)
427 		mm_iommu_mapped_dec(mem);
428 
429 	*pua = cpu_to_be64(0);
430 }
431 
432 static int tce_iommu_clear(struct tce_container *container,
433 		struct iommu_table *tbl,
434 		unsigned long entry, unsigned long pages)
435 {
436 	unsigned long oldhpa;
437 	long ret;
438 	enum dma_data_direction direction;
439 	unsigned long lastentry = entry + pages, firstentry = entry;
440 
441 	for ( ; entry < lastentry; ++entry) {
442 		if (tbl->it_indirect_levels && tbl->it_userspace) {
443 			/*
444 			 * For multilevel tables, we can take a shortcut here
445 			 * and skip some TCEs as we know that the userspace
446 			 * addresses cache is a mirror of the real TCE table
447 			 * and if it is missing some indirect levels, then
448 			 * the hardware table does not have them allocated
449 			 * either and therefore does not require updating.
450 			 */
451 			__be64 *pua = IOMMU_TABLE_USERSPACE_ENTRY_RO(tbl,
452 					entry);
453 			if (!pua) {
454 				/* align to level_size which is power of two */
455 				entry |= tbl->it_level_size - 1;
456 				continue;
457 			}
458 		}
459 
460 		cond_resched();
461 
462 		direction = DMA_NONE;
463 		oldhpa = 0;
464 		ret = iommu_tce_xchg_no_kill(container->mm, tbl, entry, &oldhpa,
465 				&direction);
466 		if (ret)
467 			continue;
468 
469 		if (direction == DMA_NONE)
470 			continue;
471 
472 		if (container->v2) {
473 			tce_iommu_unuse_page_v2(container, tbl, entry);
474 			continue;
475 		}
476 
477 		tce_iommu_unuse_page(container, oldhpa);
478 	}
479 
480 	iommu_tce_kill(tbl, firstentry, pages);
481 
482 	return 0;
483 }
484 
485 static int tce_iommu_use_page(unsigned long tce, unsigned long *hpa)
486 {
487 	struct page *page = NULL;
488 	enum dma_data_direction direction = iommu_tce_direction(tce);
489 
490 	if (pin_user_pages_fast(tce & PAGE_MASK, 1,
491 			direction != DMA_TO_DEVICE ? FOLL_WRITE : 0,
492 			&page) != 1)
493 		return -EFAULT;
494 
495 	*hpa = __pa((unsigned long) page_address(page));
496 
497 	return 0;
498 }
499 
500 static long tce_iommu_build(struct tce_container *container,
501 		struct iommu_table *tbl,
502 		unsigned long entry, unsigned long tce, unsigned long pages,
503 		enum dma_data_direction direction)
504 {
505 	long i, ret = 0;
506 	unsigned long hpa;
507 	enum dma_data_direction dirtmp;
508 
509 	for (i = 0; i < pages; ++i) {
510 		unsigned long offset = tce & IOMMU_PAGE_MASK(tbl) & ~PAGE_MASK;
511 
512 		ret = tce_iommu_use_page(tce, &hpa);
513 		if (ret)
514 			break;
515 
516 		if (!tce_page_is_contained(container->mm, hpa,
517 				tbl->it_page_shift)) {
518 			ret = -EPERM;
519 			break;
520 		}
521 
522 		hpa |= offset;
523 		dirtmp = direction;
524 		ret = iommu_tce_xchg_no_kill(container->mm, tbl, entry + i,
525 				&hpa, &dirtmp);
526 		if (ret) {
527 			tce_iommu_unuse_page(container, hpa);
528 			pr_err("iommu_tce: %s failed ioba=%lx, tce=%lx, ret=%ld\n",
529 					__func__, entry << tbl->it_page_shift,
530 					tce, ret);
531 			break;
532 		}
533 
534 		if (dirtmp != DMA_NONE)
535 			tce_iommu_unuse_page(container, hpa);
536 
537 		tce += IOMMU_PAGE_SIZE(tbl);
538 	}
539 
540 	if (ret)
541 		tce_iommu_clear(container, tbl, entry, i);
542 	else
543 		iommu_tce_kill(tbl, entry, pages);
544 
545 	return ret;
546 }
547 
548 static long tce_iommu_build_v2(struct tce_container *container,
549 		struct iommu_table *tbl,
550 		unsigned long entry, unsigned long tce, unsigned long pages,
551 		enum dma_data_direction direction)
552 {
553 	long i, ret = 0;
554 	unsigned long hpa;
555 	enum dma_data_direction dirtmp;
556 
557 	for (i = 0; i < pages; ++i) {
558 		struct mm_iommu_table_group_mem_t *mem = NULL;
559 		__be64 *pua = IOMMU_TABLE_USERSPACE_ENTRY(tbl, entry + i);
560 
561 		ret = tce_iommu_prereg_ua_to_hpa(container,
562 				tce, tbl->it_page_shift, &hpa, &mem);
563 		if (ret)
564 			break;
565 
566 		if (!tce_page_is_contained(container->mm, hpa,
567 				tbl->it_page_shift)) {
568 			ret = -EPERM;
569 			break;
570 		}
571 
572 		/* Preserve offset within IOMMU page */
573 		hpa |= tce & IOMMU_PAGE_MASK(tbl) & ~PAGE_MASK;
574 		dirtmp = direction;
575 
576 		/* The registered region is being unregistered */
577 		if (mm_iommu_mapped_inc(mem))
578 			break;
579 
580 		ret = iommu_tce_xchg_no_kill(container->mm, tbl, entry + i,
581 				&hpa, &dirtmp);
582 		if (ret) {
583 			/* dirtmp cannot be DMA_NONE here */
584 			tce_iommu_unuse_page_v2(container, tbl, entry + i);
585 			pr_err("iommu_tce: %s failed ioba=%lx, tce=%lx, ret=%ld\n",
586 					__func__, entry << tbl->it_page_shift,
587 					tce, ret);
588 			break;
589 		}
590 
591 		if (dirtmp != DMA_NONE)
592 			tce_iommu_unuse_page_v2(container, tbl, entry + i);
593 
594 		*pua = cpu_to_be64(tce);
595 
596 		tce += IOMMU_PAGE_SIZE(tbl);
597 	}
598 
599 	if (ret)
600 		tce_iommu_clear(container, tbl, entry, i);
601 	else
602 		iommu_tce_kill(tbl, entry, pages);
603 
604 	return ret;
605 }
606 
607 static long tce_iommu_create_table(struct tce_container *container,
608 			struct iommu_table_group *table_group,
609 			int num,
610 			__u32 page_shift,
611 			__u64 window_size,
612 			__u32 levels,
613 			struct iommu_table **ptbl)
614 {
615 	long ret, table_size;
616 
617 	table_size = table_group->ops->get_table_size(page_shift, window_size,
618 			levels);
619 	if (!table_size)
620 		return -EINVAL;
621 
622 	ret = account_locked_vm(container->mm, table_size >> PAGE_SHIFT, true);
623 	if (ret)
624 		return ret;
625 
626 	ret = table_group->ops->create_table(table_group, num,
627 			page_shift, window_size, levels, ptbl);
628 
629 	WARN_ON(!ret && !(*ptbl)->it_ops->free);
630 	WARN_ON(!ret && ((*ptbl)->it_allocated_size > table_size));
631 
632 	return ret;
633 }
634 
635 static void tce_iommu_free_table(struct tce_container *container,
636 		struct iommu_table *tbl)
637 {
638 	unsigned long pages = tbl->it_allocated_size >> PAGE_SHIFT;
639 
640 	iommu_tce_table_put(tbl);
641 	account_locked_vm(container->mm, pages, false);
642 }
643 
644 static long tce_iommu_create_window(struct tce_container *container,
645 		__u32 page_shift, __u64 window_size, __u32 levels,
646 		__u64 *start_addr)
647 {
648 	struct tce_iommu_group *tcegrp;
649 	struct iommu_table_group *table_group;
650 	struct iommu_table *tbl = NULL;
651 	long ret, num;
652 
653 	num = tce_iommu_find_free_table(container);
654 	if (num < 0)
655 		return num;
656 
657 	/* Get the first group for ops::create_table */
658 	tcegrp = list_first_entry(&container->group_list,
659 			struct tce_iommu_group, next);
660 	table_group = iommu_group_get_iommudata(tcegrp->grp);
661 	if (!table_group)
662 		return -EFAULT;
663 
664 	if (!(table_group->pgsizes & (1ULL << page_shift)))
665 		return -EINVAL;
666 
667 	if (!table_group->ops->set_window || !table_group->ops->unset_window ||
668 			!table_group->ops->get_table_size ||
669 			!table_group->ops->create_table)
670 		return -EPERM;
671 
672 	/* Create TCE table */
673 	ret = tce_iommu_create_table(container, table_group, num,
674 			page_shift, window_size, levels, &tbl);
675 	if (ret)
676 		return ret;
677 
678 	BUG_ON(!tbl->it_ops->free);
679 
680 	/*
681 	 * Program the table to every group.
682 	 * Groups have been tested for compatibility at the attach time.
683 	 */
684 	list_for_each_entry(tcegrp, &container->group_list, next) {
685 		table_group = iommu_group_get_iommudata(tcegrp->grp);
686 
687 		ret = table_group->ops->set_window(table_group, num, tbl);
688 		if (ret)
689 			goto unset_exit;
690 	}
691 
692 	container->tables[num] = tbl;
693 
694 	/* Return start address assigned by platform in create_table() */
695 	*start_addr = tbl->it_offset << tbl->it_page_shift;
696 
697 	return 0;
698 
699 unset_exit:
700 	list_for_each_entry(tcegrp, &container->group_list, next) {
701 		table_group = iommu_group_get_iommudata(tcegrp->grp);
702 		table_group->ops->unset_window(table_group, num);
703 	}
704 	tce_iommu_free_table(container, tbl);
705 
706 	return ret;
707 }
708 
709 static long tce_iommu_remove_window(struct tce_container *container,
710 		__u64 start_addr)
711 {
712 	struct iommu_table_group *table_group = NULL;
713 	struct iommu_table *tbl;
714 	struct tce_iommu_group *tcegrp;
715 	int num;
716 
717 	num = tce_iommu_find_table(container, start_addr, &tbl);
718 	if (num < 0)
719 		return -EINVAL;
720 
721 	BUG_ON(!tbl->it_size);
722 
723 	/* Detach groups from IOMMUs */
724 	list_for_each_entry(tcegrp, &container->group_list, next) {
725 		table_group = iommu_group_get_iommudata(tcegrp->grp);
726 
727 		/*
728 		 * SPAPR TCE IOMMU exposes the default DMA window to
729 		 * the guest via dma32_window_start/size of
730 		 * VFIO_IOMMU_SPAPR_TCE_GET_INFO. Some platforms allow
731 		 * the userspace to remove this window, some do not so
732 		 * here we check for the platform capability.
733 		 */
734 		if (!table_group->ops || !table_group->ops->unset_window)
735 			return -EPERM;
736 
737 		table_group->ops->unset_window(table_group, num);
738 	}
739 
740 	/* Free table */
741 	tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
742 	tce_iommu_free_table(container, tbl);
743 	container->tables[num] = NULL;
744 
745 	return 0;
746 }
747 
748 static long tce_iommu_create_default_window(struct tce_container *container)
749 {
750 	long ret;
751 	__u64 start_addr = 0;
752 	struct tce_iommu_group *tcegrp;
753 	struct iommu_table_group *table_group;
754 
755 	if (!container->def_window_pending)
756 		return 0;
757 
758 	if (!tce_groups_attached(container))
759 		return -ENODEV;
760 
761 	tcegrp = list_first_entry(&container->group_list,
762 			struct tce_iommu_group, next);
763 	table_group = iommu_group_get_iommudata(tcegrp->grp);
764 	if (!table_group)
765 		return -ENODEV;
766 
767 	ret = tce_iommu_create_window(container, IOMMU_PAGE_SHIFT_4K,
768 			table_group->tce32_size, 1, &start_addr);
769 	WARN_ON_ONCE(!ret && start_addr);
770 
771 	if (!ret)
772 		container->def_window_pending = false;
773 
774 	return ret;
775 }
776 
777 static long tce_iommu_ioctl(void *iommu_data,
778 				 unsigned int cmd, unsigned long arg)
779 {
780 	struct tce_container *container = iommu_data;
781 	unsigned long minsz, ddwsz;
782 	long ret;
783 
784 	switch (cmd) {
785 	case VFIO_CHECK_EXTENSION:
786 		switch (arg) {
787 		case VFIO_SPAPR_TCE_IOMMU:
788 		case VFIO_SPAPR_TCE_v2_IOMMU:
789 			ret = 1;
790 			break;
791 		default:
792 			ret = vfio_spapr_iommu_eeh_ioctl(NULL, cmd, arg);
793 			break;
794 		}
795 
796 		return (ret < 0) ? 0 : ret;
797 	}
798 
799 	/*
800 	 * Sanity check to prevent one userspace from manipulating
801 	 * another userspace mm.
802 	 */
803 	BUG_ON(!container);
804 	if (container->mm && container->mm != current->mm)
805 		return -EPERM;
806 
807 	switch (cmd) {
808 	case VFIO_IOMMU_SPAPR_TCE_GET_INFO: {
809 		struct vfio_iommu_spapr_tce_info info;
810 		struct tce_iommu_group *tcegrp;
811 		struct iommu_table_group *table_group;
812 
813 		if (!tce_groups_attached(container))
814 			return -ENXIO;
815 
816 		tcegrp = list_first_entry(&container->group_list,
817 				struct tce_iommu_group, next);
818 		table_group = iommu_group_get_iommudata(tcegrp->grp);
819 
820 		if (!table_group)
821 			return -ENXIO;
822 
823 		minsz = offsetofend(struct vfio_iommu_spapr_tce_info,
824 				dma32_window_size);
825 
826 		if (copy_from_user(&info, (void __user *)arg, minsz))
827 			return -EFAULT;
828 
829 		if (info.argsz < minsz)
830 			return -EINVAL;
831 
832 		info.dma32_window_start = table_group->tce32_start;
833 		info.dma32_window_size = table_group->tce32_size;
834 		info.flags = 0;
835 		memset(&info.ddw, 0, sizeof(info.ddw));
836 
837 		if (table_group->max_dynamic_windows_supported &&
838 				container->v2) {
839 			info.flags |= VFIO_IOMMU_SPAPR_INFO_DDW;
840 			info.ddw.pgsizes = table_group->pgsizes;
841 			info.ddw.max_dynamic_windows_supported =
842 				table_group->max_dynamic_windows_supported;
843 			info.ddw.levels = table_group->max_levels;
844 		}
845 
846 		ddwsz = offsetofend(struct vfio_iommu_spapr_tce_info, ddw);
847 
848 		if (info.argsz >= ddwsz)
849 			minsz = ddwsz;
850 
851 		if (copy_to_user((void __user *)arg, &info, minsz))
852 			return -EFAULT;
853 
854 		return 0;
855 	}
856 	case VFIO_IOMMU_MAP_DMA: {
857 		struct vfio_iommu_type1_dma_map param;
858 		struct iommu_table *tbl = NULL;
859 		long num;
860 		enum dma_data_direction direction;
861 
862 		if (!container->enabled)
863 			return -EPERM;
864 
865 		minsz = offsetofend(struct vfio_iommu_type1_dma_map, size);
866 
867 		if (copy_from_user(&param, (void __user *)arg, minsz))
868 			return -EFAULT;
869 
870 		if (param.argsz < minsz)
871 			return -EINVAL;
872 
873 		if (param.flags & ~(VFIO_DMA_MAP_FLAG_READ |
874 				VFIO_DMA_MAP_FLAG_WRITE))
875 			return -EINVAL;
876 
877 		ret = tce_iommu_create_default_window(container);
878 		if (ret)
879 			return ret;
880 
881 		num = tce_iommu_find_table(container, param.iova, &tbl);
882 		if (num < 0)
883 			return -ENXIO;
884 
885 		if ((param.size & ~IOMMU_PAGE_MASK(tbl)) ||
886 				(param.vaddr & ~IOMMU_PAGE_MASK(tbl)))
887 			return -EINVAL;
888 
889 		/* iova is checked by the IOMMU API */
890 		if (param.flags & VFIO_DMA_MAP_FLAG_READ) {
891 			if (param.flags & VFIO_DMA_MAP_FLAG_WRITE)
892 				direction = DMA_BIDIRECTIONAL;
893 			else
894 				direction = DMA_TO_DEVICE;
895 		} else {
896 			if (param.flags & VFIO_DMA_MAP_FLAG_WRITE)
897 				direction = DMA_FROM_DEVICE;
898 			else
899 				return -EINVAL;
900 		}
901 
902 		ret = iommu_tce_put_param_check(tbl, param.iova, param.vaddr);
903 		if (ret)
904 			return ret;
905 
906 		if (container->v2)
907 			ret = tce_iommu_build_v2(container, tbl,
908 					param.iova >> tbl->it_page_shift,
909 					param.vaddr,
910 					param.size >> tbl->it_page_shift,
911 					direction);
912 		else
913 			ret = tce_iommu_build(container, tbl,
914 					param.iova >> tbl->it_page_shift,
915 					param.vaddr,
916 					param.size >> tbl->it_page_shift,
917 					direction);
918 
919 		iommu_flush_tce(tbl);
920 
921 		return ret;
922 	}
923 	case VFIO_IOMMU_UNMAP_DMA: {
924 		struct vfio_iommu_type1_dma_unmap param;
925 		struct iommu_table *tbl = NULL;
926 		long num;
927 
928 		if (!container->enabled)
929 			return -EPERM;
930 
931 		minsz = offsetofend(struct vfio_iommu_type1_dma_unmap,
932 				size);
933 
934 		if (copy_from_user(&param, (void __user *)arg, minsz))
935 			return -EFAULT;
936 
937 		if (param.argsz < minsz)
938 			return -EINVAL;
939 
940 		/* No flag is supported now */
941 		if (param.flags)
942 			return -EINVAL;
943 
944 		ret = tce_iommu_create_default_window(container);
945 		if (ret)
946 			return ret;
947 
948 		num = tce_iommu_find_table(container, param.iova, &tbl);
949 		if (num < 0)
950 			return -ENXIO;
951 
952 		if (param.size & ~IOMMU_PAGE_MASK(tbl))
953 			return -EINVAL;
954 
955 		ret = iommu_tce_clear_param_check(tbl, param.iova, 0,
956 				param.size >> tbl->it_page_shift);
957 		if (ret)
958 			return ret;
959 
960 		ret = tce_iommu_clear(container, tbl,
961 				param.iova >> tbl->it_page_shift,
962 				param.size >> tbl->it_page_shift);
963 		iommu_flush_tce(tbl);
964 
965 		return ret;
966 	}
967 	case VFIO_IOMMU_SPAPR_REGISTER_MEMORY: {
968 		struct vfio_iommu_spapr_register_memory param;
969 
970 		if (!container->v2)
971 			break;
972 
973 		minsz = offsetofend(struct vfio_iommu_spapr_register_memory,
974 				size);
975 
976 		ret = tce_iommu_mm_set(container);
977 		if (ret)
978 			return ret;
979 
980 		if (copy_from_user(&param, (void __user *)arg, minsz))
981 			return -EFAULT;
982 
983 		if (param.argsz < minsz)
984 			return -EINVAL;
985 
986 		/* No flag is supported now */
987 		if (param.flags)
988 			return -EINVAL;
989 
990 		mutex_lock(&container->lock);
991 		ret = tce_iommu_register_pages(container, param.vaddr,
992 				param.size);
993 		mutex_unlock(&container->lock);
994 
995 		return ret;
996 	}
997 	case VFIO_IOMMU_SPAPR_UNREGISTER_MEMORY: {
998 		struct vfio_iommu_spapr_register_memory param;
999 
1000 		if (!container->v2)
1001 			break;
1002 
1003 		if (!container->mm)
1004 			return -EPERM;
1005 
1006 		minsz = offsetofend(struct vfio_iommu_spapr_register_memory,
1007 				size);
1008 
1009 		if (copy_from_user(&param, (void __user *)arg, minsz))
1010 			return -EFAULT;
1011 
1012 		if (param.argsz < minsz)
1013 			return -EINVAL;
1014 
1015 		/* No flag is supported now */
1016 		if (param.flags)
1017 			return -EINVAL;
1018 
1019 		mutex_lock(&container->lock);
1020 		ret = tce_iommu_unregister_pages(container, param.vaddr,
1021 				param.size);
1022 		mutex_unlock(&container->lock);
1023 
1024 		return ret;
1025 	}
1026 	case VFIO_IOMMU_ENABLE:
1027 		if (container->v2)
1028 			break;
1029 
1030 		mutex_lock(&container->lock);
1031 		ret = tce_iommu_enable(container);
1032 		mutex_unlock(&container->lock);
1033 		return ret;
1034 
1035 
1036 	case VFIO_IOMMU_DISABLE:
1037 		if (container->v2)
1038 			break;
1039 
1040 		mutex_lock(&container->lock);
1041 		tce_iommu_disable(container);
1042 		mutex_unlock(&container->lock);
1043 		return 0;
1044 
1045 	case VFIO_EEH_PE_OP: {
1046 		struct tce_iommu_group *tcegrp;
1047 
1048 		ret = 0;
1049 		list_for_each_entry(tcegrp, &container->group_list, next) {
1050 			ret = vfio_spapr_iommu_eeh_ioctl(tcegrp->grp,
1051 					cmd, arg);
1052 			if (ret)
1053 				return ret;
1054 		}
1055 		return ret;
1056 	}
1057 
1058 	case VFIO_IOMMU_SPAPR_TCE_CREATE: {
1059 		struct vfio_iommu_spapr_tce_create create;
1060 
1061 		if (!container->v2)
1062 			break;
1063 
1064 		ret = tce_iommu_mm_set(container);
1065 		if (ret)
1066 			return ret;
1067 
1068 		if (!tce_groups_attached(container))
1069 			return -ENXIO;
1070 
1071 		minsz = offsetofend(struct vfio_iommu_spapr_tce_create,
1072 				start_addr);
1073 
1074 		if (copy_from_user(&create, (void __user *)arg, minsz))
1075 			return -EFAULT;
1076 
1077 		if (create.argsz < minsz)
1078 			return -EINVAL;
1079 
1080 		if (create.flags)
1081 			return -EINVAL;
1082 
1083 		mutex_lock(&container->lock);
1084 
1085 		ret = tce_iommu_create_default_window(container);
1086 		if (!ret)
1087 			ret = tce_iommu_create_window(container,
1088 					create.page_shift,
1089 					create.window_size, create.levels,
1090 					&create.start_addr);
1091 
1092 		mutex_unlock(&container->lock);
1093 
1094 		if (!ret && copy_to_user((void __user *)arg, &create, minsz))
1095 			ret = -EFAULT;
1096 
1097 		return ret;
1098 	}
1099 	case VFIO_IOMMU_SPAPR_TCE_REMOVE: {
1100 		struct vfio_iommu_spapr_tce_remove remove;
1101 
1102 		if (!container->v2)
1103 			break;
1104 
1105 		ret = tce_iommu_mm_set(container);
1106 		if (ret)
1107 			return ret;
1108 
1109 		if (!tce_groups_attached(container))
1110 			return -ENXIO;
1111 
1112 		minsz = offsetofend(struct vfio_iommu_spapr_tce_remove,
1113 				start_addr);
1114 
1115 		if (copy_from_user(&remove, (void __user *)arg, minsz))
1116 			return -EFAULT;
1117 
1118 		if (remove.argsz < minsz)
1119 			return -EINVAL;
1120 
1121 		if (remove.flags)
1122 			return -EINVAL;
1123 
1124 		if (container->def_window_pending && !remove.start_addr) {
1125 			container->def_window_pending = false;
1126 			return 0;
1127 		}
1128 
1129 		mutex_lock(&container->lock);
1130 
1131 		ret = tce_iommu_remove_window(container, remove.start_addr);
1132 
1133 		mutex_unlock(&container->lock);
1134 
1135 		return ret;
1136 	}
1137 	}
1138 
1139 	return -ENOTTY;
1140 }
1141 
1142 static void tce_iommu_release_ownership(struct tce_container *container,
1143 		struct iommu_table_group *table_group)
1144 {
1145 	int i;
1146 
1147 	for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1148 		struct iommu_table *tbl = container->tables[i];
1149 
1150 		if (!tbl)
1151 			continue;
1152 
1153 		tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
1154 		if (tbl->it_map)
1155 			iommu_release_ownership(tbl);
1156 
1157 		container->tables[i] = NULL;
1158 	}
1159 }
1160 
1161 static int tce_iommu_take_ownership(struct tce_container *container,
1162 		struct iommu_table_group *table_group)
1163 {
1164 	int i, j, rc = 0;
1165 
1166 	for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1167 		struct iommu_table *tbl = table_group->tables[i];
1168 
1169 		if (!tbl || !tbl->it_map)
1170 			continue;
1171 
1172 		rc = iommu_take_ownership(tbl);
1173 		if (rc) {
1174 			for (j = 0; j < i; ++j)
1175 				iommu_release_ownership(
1176 						table_group->tables[j]);
1177 
1178 			return rc;
1179 		}
1180 	}
1181 
1182 	for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1183 		container->tables[i] = table_group->tables[i];
1184 
1185 	return 0;
1186 }
1187 
1188 static void tce_iommu_release_ownership_ddw(struct tce_container *container,
1189 		struct iommu_table_group *table_group)
1190 {
1191 	long i;
1192 
1193 	if (!table_group->ops->unset_window) {
1194 		WARN_ON_ONCE(1);
1195 		return;
1196 	}
1197 
1198 	for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1199 		if (container->tables[i])
1200 			table_group->ops->unset_window(table_group, i);
1201 
1202 	table_group->ops->release_ownership(table_group);
1203 }
1204 
1205 static long tce_iommu_take_ownership_ddw(struct tce_container *container,
1206 		struct iommu_table_group *table_group)
1207 {
1208 	long i, ret = 0;
1209 
1210 	if (!table_group->ops->create_table || !table_group->ops->set_window ||
1211 			!table_group->ops->release_ownership) {
1212 		WARN_ON_ONCE(1);
1213 		return -EFAULT;
1214 	}
1215 
1216 	table_group->ops->take_ownership(table_group);
1217 
1218 	/* Set all windows to the new group */
1219 	for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1220 		struct iommu_table *tbl = container->tables[i];
1221 
1222 		if (!tbl)
1223 			continue;
1224 
1225 		ret = table_group->ops->set_window(table_group, i, tbl);
1226 		if (ret)
1227 			goto release_exit;
1228 	}
1229 
1230 	return 0;
1231 
1232 release_exit:
1233 	for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1234 		table_group->ops->unset_window(table_group, i);
1235 
1236 	table_group->ops->release_ownership(table_group);
1237 
1238 	return ret;
1239 }
1240 
1241 static int tce_iommu_attach_group(void *iommu_data,
1242 		struct iommu_group *iommu_group, enum vfio_group_type type)
1243 {
1244 	int ret = 0;
1245 	struct tce_container *container = iommu_data;
1246 	struct iommu_table_group *table_group;
1247 	struct tce_iommu_group *tcegrp = NULL;
1248 
1249 	if (type == VFIO_EMULATED_IOMMU)
1250 		return -EINVAL;
1251 
1252 	mutex_lock(&container->lock);
1253 
1254 	/* pr_debug("tce_vfio: Attaching group #%u to iommu %p\n",
1255 			iommu_group_id(iommu_group), iommu_group); */
1256 	table_group = iommu_group_get_iommudata(iommu_group);
1257 	if (!table_group) {
1258 		ret = -ENODEV;
1259 		goto unlock_exit;
1260 	}
1261 
1262 	if (tce_groups_attached(container) && (!table_group->ops ||
1263 			!table_group->ops->take_ownership ||
1264 			!table_group->ops->release_ownership)) {
1265 		ret = -EBUSY;
1266 		goto unlock_exit;
1267 	}
1268 
1269 	/* Check if new group has the same iommu_ops (i.e. compatible) */
1270 	list_for_each_entry(tcegrp, &container->group_list, next) {
1271 		struct iommu_table_group *table_group_tmp;
1272 
1273 		if (tcegrp->grp == iommu_group) {
1274 			pr_warn("tce_vfio: Group %d is already attached\n",
1275 					iommu_group_id(iommu_group));
1276 			ret = -EBUSY;
1277 			goto unlock_exit;
1278 		}
1279 		table_group_tmp = iommu_group_get_iommudata(tcegrp->grp);
1280 		if (table_group_tmp->ops->create_table !=
1281 				table_group->ops->create_table) {
1282 			pr_warn("tce_vfio: Group %d is incompatible with group %d\n",
1283 					iommu_group_id(iommu_group),
1284 					iommu_group_id(tcegrp->grp));
1285 			ret = -EPERM;
1286 			goto unlock_exit;
1287 		}
1288 	}
1289 
1290 	tcegrp = kzalloc(sizeof(*tcegrp), GFP_KERNEL);
1291 	if (!tcegrp) {
1292 		ret = -ENOMEM;
1293 		goto unlock_exit;
1294 	}
1295 
1296 	if (!table_group->ops || !table_group->ops->take_ownership ||
1297 			!table_group->ops->release_ownership) {
1298 		if (container->v2) {
1299 			ret = -EPERM;
1300 			goto free_exit;
1301 		}
1302 		ret = tce_iommu_take_ownership(container, table_group);
1303 	} else {
1304 		if (!container->v2) {
1305 			ret = -EPERM;
1306 			goto free_exit;
1307 		}
1308 		ret = tce_iommu_take_ownership_ddw(container, table_group);
1309 		if (!tce_groups_attached(container) && !container->tables[0])
1310 			container->def_window_pending = true;
1311 	}
1312 
1313 	if (!ret) {
1314 		tcegrp->grp = iommu_group;
1315 		list_add(&tcegrp->next, &container->group_list);
1316 	}
1317 
1318 free_exit:
1319 	if (ret && tcegrp)
1320 		kfree(tcegrp);
1321 
1322 unlock_exit:
1323 	mutex_unlock(&container->lock);
1324 
1325 	return ret;
1326 }
1327 
1328 static void tce_iommu_detach_group(void *iommu_data,
1329 		struct iommu_group *iommu_group)
1330 {
1331 	struct tce_container *container = iommu_data;
1332 	struct iommu_table_group *table_group;
1333 	bool found = false;
1334 	struct tce_iommu_group *tcegrp;
1335 
1336 	mutex_lock(&container->lock);
1337 
1338 	list_for_each_entry(tcegrp, &container->group_list, next) {
1339 		if (tcegrp->grp == iommu_group) {
1340 			found = true;
1341 			break;
1342 		}
1343 	}
1344 
1345 	if (!found) {
1346 		pr_warn("tce_vfio: detaching unattached group #%u\n",
1347 				iommu_group_id(iommu_group));
1348 		goto unlock_exit;
1349 	}
1350 
1351 	list_del(&tcegrp->next);
1352 	kfree(tcegrp);
1353 
1354 	table_group = iommu_group_get_iommudata(iommu_group);
1355 	BUG_ON(!table_group);
1356 
1357 	if (!table_group->ops || !table_group->ops->release_ownership)
1358 		tce_iommu_release_ownership(container, table_group);
1359 	else
1360 		tce_iommu_release_ownership_ddw(container, table_group);
1361 
1362 unlock_exit:
1363 	mutex_unlock(&container->lock);
1364 }
1365 
1366 static const struct vfio_iommu_driver_ops tce_iommu_driver_ops = {
1367 	.name		= "iommu-vfio-powerpc",
1368 	.owner		= THIS_MODULE,
1369 	.open		= tce_iommu_open,
1370 	.release	= tce_iommu_release,
1371 	.ioctl		= tce_iommu_ioctl,
1372 	.attach_group	= tce_iommu_attach_group,
1373 	.detach_group	= tce_iommu_detach_group,
1374 };
1375 
1376 static int __init tce_iommu_init(void)
1377 {
1378 	return vfio_register_iommu_driver(&tce_iommu_driver_ops);
1379 }
1380 
1381 static void __exit tce_iommu_cleanup(void)
1382 {
1383 	vfio_unregister_iommu_driver(&tce_iommu_driver_ops);
1384 }
1385 
1386 module_init(tce_iommu_init);
1387 module_exit(tce_iommu_cleanup);
1388 
1389 MODULE_VERSION(DRIVER_VERSION);
1390 MODULE_LICENSE("GPL v2");
1391 MODULE_AUTHOR(DRIVER_AUTHOR);
1392 MODULE_DESCRIPTION(DRIVER_DESC);
1393 
1394