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, &page) != 1)
536 		return -EFAULT;
537 
538 	*hpa = __pa((unsigned long) page_address(page));
539 
540 	return 0;
541 }
542 
543 static long tce_iommu_build(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 		unsigned long offset = tce & IOMMU_PAGE_MASK(tbl) & ~PAGE_MASK;
554 
555 		ret = tce_iommu_use_page(tce, &hpa);
556 		if (ret)
557 			break;
558 
559 		if (!tce_page_is_contained(container->mm, hpa,
560 				tbl->it_page_shift)) {
561 			ret = -EPERM;
562 			break;
563 		}
564 
565 		hpa |= offset;
566 		dirtmp = direction;
567 		ret = iommu_tce_xchg(container->mm, tbl, entry + i, &hpa,
568 				&dirtmp);
569 		if (ret) {
570 			tce_iommu_unuse_page(container, hpa);
571 			pr_err("iommu_tce: %s failed ioba=%lx, tce=%lx, ret=%ld\n",
572 					__func__, entry << tbl->it_page_shift,
573 					tce, ret);
574 			break;
575 		}
576 
577 		if (dirtmp != DMA_NONE)
578 			tce_iommu_unuse_page(container, hpa);
579 
580 		tce += IOMMU_PAGE_SIZE(tbl);
581 	}
582 
583 	if (ret)
584 		tce_iommu_clear(container, tbl, entry, i);
585 
586 	return ret;
587 }
588 
589 static long tce_iommu_build_v2(struct tce_container *container,
590 		struct iommu_table *tbl,
591 		unsigned long entry, unsigned long tce, unsigned long pages,
592 		enum dma_data_direction direction)
593 {
594 	long i, ret = 0;
595 	unsigned long hpa;
596 	enum dma_data_direction dirtmp;
597 
598 	for (i = 0; i < pages; ++i) {
599 		struct mm_iommu_table_group_mem_t *mem = NULL;
600 		__be64 *pua = IOMMU_TABLE_USERSPACE_ENTRY(tbl, entry + i);
601 
602 		ret = tce_iommu_prereg_ua_to_hpa(container,
603 				tce, tbl->it_page_shift, &hpa, &mem);
604 		if (ret)
605 			break;
606 
607 		if (!tce_page_is_contained(container->mm, hpa,
608 				tbl->it_page_shift)) {
609 			ret = -EPERM;
610 			break;
611 		}
612 
613 		/* Preserve offset within IOMMU page */
614 		hpa |= tce & IOMMU_PAGE_MASK(tbl) & ~PAGE_MASK;
615 		dirtmp = direction;
616 
617 		/* The registered region is being unregistered */
618 		if (mm_iommu_mapped_inc(mem))
619 			break;
620 
621 		ret = iommu_tce_xchg(container->mm, tbl, entry + i, &hpa,
622 				&dirtmp);
623 		if (ret) {
624 			/* dirtmp cannot be DMA_NONE here */
625 			tce_iommu_unuse_page_v2(container, tbl, entry + i);
626 			pr_err("iommu_tce: %s failed ioba=%lx, tce=%lx, ret=%ld\n",
627 					__func__, entry << tbl->it_page_shift,
628 					tce, ret);
629 			break;
630 		}
631 
632 		if (dirtmp != DMA_NONE)
633 			tce_iommu_unuse_page_v2(container, tbl, entry + i);
634 
635 		*pua = cpu_to_be64(tce);
636 
637 		tce += IOMMU_PAGE_SIZE(tbl);
638 	}
639 
640 	if (ret)
641 		tce_iommu_clear(container, tbl, entry, i);
642 
643 	return ret;
644 }
645 
646 static long tce_iommu_create_table(struct tce_container *container,
647 			struct iommu_table_group *table_group,
648 			int num,
649 			__u32 page_shift,
650 			__u64 window_size,
651 			__u32 levels,
652 			struct iommu_table **ptbl)
653 {
654 	long ret, table_size;
655 
656 	table_size = table_group->ops->get_table_size(page_shift, window_size,
657 			levels);
658 	if (!table_size)
659 		return -EINVAL;
660 
661 	ret = try_increment_locked_vm(container->mm, table_size >> PAGE_SHIFT);
662 	if (ret)
663 		return ret;
664 
665 	ret = table_group->ops->create_table(table_group, num,
666 			page_shift, window_size, levels, ptbl);
667 
668 	WARN_ON(!ret && !(*ptbl)->it_ops->free);
669 	WARN_ON(!ret && ((*ptbl)->it_allocated_size > table_size));
670 
671 	return ret;
672 }
673 
674 static void tce_iommu_free_table(struct tce_container *container,
675 		struct iommu_table *tbl)
676 {
677 	unsigned long pages = tbl->it_allocated_size >> PAGE_SHIFT;
678 
679 	iommu_tce_table_put(tbl);
680 	decrement_locked_vm(container->mm, pages);
681 }
682 
683 static long tce_iommu_create_window(struct tce_container *container,
684 		__u32 page_shift, __u64 window_size, __u32 levels,
685 		__u64 *start_addr)
686 {
687 	struct tce_iommu_group *tcegrp;
688 	struct iommu_table_group *table_group;
689 	struct iommu_table *tbl = NULL;
690 	long ret, num;
691 
692 	num = tce_iommu_find_free_table(container);
693 	if (num < 0)
694 		return num;
695 
696 	/* Get the first group for ops::create_table */
697 	tcegrp = list_first_entry(&container->group_list,
698 			struct tce_iommu_group, next);
699 	table_group = iommu_group_get_iommudata(tcegrp->grp);
700 	if (!table_group)
701 		return -EFAULT;
702 
703 	if (!(table_group->pgsizes & (1ULL << page_shift)))
704 		return -EINVAL;
705 
706 	if (!table_group->ops->set_window || !table_group->ops->unset_window ||
707 			!table_group->ops->get_table_size ||
708 			!table_group->ops->create_table)
709 		return -EPERM;
710 
711 	/* Create TCE table */
712 	ret = tce_iommu_create_table(container, table_group, num,
713 			page_shift, window_size, levels, &tbl);
714 	if (ret)
715 		return ret;
716 
717 	BUG_ON(!tbl->it_ops->free);
718 
719 	/*
720 	 * Program the table to every group.
721 	 * Groups have been tested for compatibility at the attach time.
722 	 */
723 	list_for_each_entry(tcegrp, &container->group_list, next) {
724 		table_group = iommu_group_get_iommudata(tcegrp->grp);
725 
726 		ret = table_group->ops->set_window(table_group, num, tbl);
727 		if (ret)
728 			goto unset_exit;
729 	}
730 
731 	container->tables[num] = tbl;
732 
733 	/* Return start address assigned by platform in create_table() */
734 	*start_addr = tbl->it_offset << tbl->it_page_shift;
735 
736 	return 0;
737 
738 unset_exit:
739 	list_for_each_entry(tcegrp, &container->group_list, next) {
740 		table_group = iommu_group_get_iommudata(tcegrp->grp);
741 		table_group->ops->unset_window(table_group, num);
742 	}
743 	tce_iommu_free_table(container, tbl);
744 
745 	return ret;
746 }
747 
748 static long tce_iommu_remove_window(struct tce_container *container,
749 		__u64 start_addr)
750 {
751 	struct iommu_table_group *table_group = NULL;
752 	struct iommu_table *tbl;
753 	struct tce_iommu_group *tcegrp;
754 	int num;
755 
756 	num = tce_iommu_find_table(container, start_addr, &tbl);
757 	if (num < 0)
758 		return -EINVAL;
759 
760 	BUG_ON(!tbl->it_size);
761 
762 	/* Detach groups from IOMMUs */
763 	list_for_each_entry(tcegrp, &container->group_list, next) {
764 		table_group = iommu_group_get_iommudata(tcegrp->grp);
765 
766 		/*
767 		 * SPAPR TCE IOMMU exposes the default DMA window to
768 		 * the guest via dma32_window_start/size of
769 		 * VFIO_IOMMU_SPAPR_TCE_GET_INFO. Some platforms allow
770 		 * the userspace to remove this window, some do not so
771 		 * here we check for the platform capability.
772 		 */
773 		if (!table_group->ops || !table_group->ops->unset_window)
774 			return -EPERM;
775 
776 		table_group->ops->unset_window(table_group, num);
777 	}
778 
779 	/* Free table */
780 	tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
781 	tce_iommu_free_table(container, tbl);
782 	container->tables[num] = NULL;
783 
784 	return 0;
785 }
786 
787 static long tce_iommu_create_default_window(struct tce_container *container)
788 {
789 	long ret;
790 	__u64 start_addr = 0;
791 	struct tce_iommu_group *tcegrp;
792 	struct iommu_table_group *table_group;
793 
794 	if (!container->def_window_pending)
795 		return 0;
796 
797 	if (!tce_groups_attached(container))
798 		return -ENODEV;
799 
800 	tcegrp = list_first_entry(&container->group_list,
801 			struct tce_iommu_group, next);
802 	table_group = iommu_group_get_iommudata(tcegrp->grp);
803 	if (!table_group)
804 		return -ENODEV;
805 
806 	ret = tce_iommu_create_window(container, IOMMU_PAGE_SHIFT_4K,
807 			table_group->tce32_size, 1, &start_addr);
808 	WARN_ON_ONCE(!ret && start_addr);
809 
810 	if (!ret)
811 		container->def_window_pending = false;
812 
813 	return ret;
814 }
815 
816 static long tce_iommu_ioctl(void *iommu_data,
817 				 unsigned int cmd, unsigned long arg)
818 {
819 	struct tce_container *container = iommu_data;
820 	unsigned long minsz, ddwsz;
821 	long ret;
822 
823 	switch (cmd) {
824 	case VFIO_CHECK_EXTENSION:
825 		switch (arg) {
826 		case VFIO_SPAPR_TCE_IOMMU:
827 		case VFIO_SPAPR_TCE_v2_IOMMU:
828 			ret = 1;
829 			break;
830 		default:
831 			ret = vfio_spapr_iommu_eeh_ioctl(NULL, cmd, arg);
832 			break;
833 		}
834 
835 		return (ret < 0) ? 0 : ret;
836 	}
837 
838 	/*
839 	 * Sanity check to prevent one userspace from manipulating
840 	 * another userspace mm.
841 	 */
842 	BUG_ON(!container);
843 	if (container->mm && container->mm != current->mm)
844 		return -EPERM;
845 
846 	switch (cmd) {
847 	case VFIO_IOMMU_SPAPR_TCE_GET_INFO: {
848 		struct vfio_iommu_spapr_tce_info info;
849 		struct tce_iommu_group *tcegrp;
850 		struct iommu_table_group *table_group;
851 
852 		if (!tce_groups_attached(container))
853 			return -ENXIO;
854 
855 		tcegrp = list_first_entry(&container->group_list,
856 				struct tce_iommu_group, next);
857 		table_group = iommu_group_get_iommudata(tcegrp->grp);
858 
859 		if (!table_group)
860 			return -ENXIO;
861 
862 		minsz = offsetofend(struct vfio_iommu_spapr_tce_info,
863 				dma32_window_size);
864 
865 		if (copy_from_user(&info, (void __user *)arg, minsz))
866 			return -EFAULT;
867 
868 		if (info.argsz < minsz)
869 			return -EINVAL;
870 
871 		info.dma32_window_start = table_group->tce32_start;
872 		info.dma32_window_size = table_group->tce32_size;
873 		info.flags = 0;
874 		memset(&info.ddw, 0, sizeof(info.ddw));
875 
876 		if (table_group->max_dynamic_windows_supported &&
877 				container->v2) {
878 			info.flags |= VFIO_IOMMU_SPAPR_INFO_DDW;
879 			info.ddw.pgsizes = table_group->pgsizes;
880 			info.ddw.max_dynamic_windows_supported =
881 				table_group->max_dynamic_windows_supported;
882 			info.ddw.levels = table_group->max_levels;
883 		}
884 
885 		ddwsz = offsetofend(struct vfio_iommu_spapr_tce_info, ddw);
886 
887 		if (info.argsz >= ddwsz)
888 			minsz = ddwsz;
889 
890 		if (copy_to_user((void __user *)arg, &info, minsz))
891 			return -EFAULT;
892 
893 		return 0;
894 	}
895 	case VFIO_IOMMU_MAP_DMA: {
896 		struct vfio_iommu_type1_dma_map param;
897 		struct iommu_table *tbl = NULL;
898 		long num;
899 		enum dma_data_direction direction;
900 
901 		if (!container->enabled)
902 			return -EPERM;
903 
904 		minsz = offsetofend(struct vfio_iommu_type1_dma_map, size);
905 
906 		if (copy_from_user(&param, (void __user *)arg, minsz))
907 			return -EFAULT;
908 
909 		if (param.argsz < minsz)
910 			return -EINVAL;
911 
912 		if (param.flags & ~(VFIO_DMA_MAP_FLAG_READ |
913 				VFIO_DMA_MAP_FLAG_WRITE))
914 			return -EINVAL;
915 
916 		ret = tce_iommu_create_default_window(container);
917 		if (ret)
918 			return ret;
919 
920 		num = tce_iommu_find_table(container, param.iova, &tbl);
921 		if (num < 0)
922 			return -ENXIO;
923 
924 		if ((param.size & ~IOMMU_PAGE_MASK(tbl)) ||
925 				(param.vaddr & ~IOMMU_PAGE_MASK(tbl)))
926 			return -EINVAL;
927 
928 		/* iova is checked by the IOMMU API */
929 		if (param.flags & VFIO_DMA_MAP_FLAG_READ) {
930 			if (param.flags & VFIO_DMA_MAP_FLAG_WRITE)
931 				direction = DMA_BIDIRECTIONAL;
932 			else
933 				direction = DMA_TO_DEVICE;
934 		} else {
935 			if (param.flags & VFIO_DMA_MAP_FLAG_WRITE)
936 				direction = DMA_FROM_DEVICE;
937 			else
938 				return -EINVAL;
939 		}
940 
941 		ret = iommu_tce_put_param_check(tbl, param.iova, param.vaddr);
942 		if (ret)
943 			return ret;
944 
945 		if (container->v2)
946 			ret = tce_iommu_build_v2(container, tbl,
947 					param.iova >> tbl->it_page_shift,
948 					param.vaddr,
949 					param.size >> tbl->it_page_shift,
950 					direction);
951 		else
952 			ret = tce_iommu_build(container, tbl,
953 					param.iova >> tbl->it_page_shift,
954 					param.vaddr,
955 					param.size >> tbl->it_page_shift,
956 					direction);
957 
958 		iommu_flush_tce(tbl);
959 
960 		return ret;
961 	}
962 	case VFIO_IOMMU_UNMAP_DMA: {
963 		struct vfio_iommu_type1_dma_unmap param;
964 		struct iommu_table *tbl = NULL;
965 		long num;
966 
967 		if (!container->enabled)
968 			return -EPERM;
969 
970 		minsz = offsetofend(struct vfio_iommu_type1_dma_unmap,
971 				size);
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 		ret = tce_iommu_create_default_window(container);
984 		if (ret)
985 			return ret;
986 
987 		num = tce_iommu_find_table(container, param.iova, &tbl);
988 		if (num < 0)
989 			return -ENXIO;
990 
991 		if (param.size & ~IOMMU_PAGE_MASK(tbl))
992 			return -EINVAL;
993 
994 		ret = iommu_tce_clear_param_check(tbl, param.iova, 0,
995 				param.size >> tbl->it_page_shift);
996 		if (ret)
997 			return ret;
998 
999 		ret = tce_iommu_clear(container, tbl,
1000 				param.iova >> tbl->it_page_shift,
1001 				param.size >> tbl->it_page_shift);
1002 		iommu_flush_tce(tbl);
1003 
1004 		return ret;
1005 	}
1006 	case VFIO_IOMMU_SPAPR_REGISTER_MEMORY: {
1007 		struct vfio_iommu_spapr_register_memory param;
1008 
1009 		if (!container->v2)
1010 			break;
1011 
1012 		minsz = offsetofend(struct vfio_iommu_spapr_register_memory,
1013 				size);
1014 
1015 		ret = tce_iommu_mm_set(container);
1016 		if (ret)
1017 			return ret;
1018 
1019 		if (copy_from_user(&param, (void __user *)arg, minsz))
1020 			return -EFAULT;
1021 
1022 		if (param.argsz < minsz)
1023 			return -EINVAL;
1024 
1025 		/* No flag is supported now */
1026 		if (param.flags)
1027 			return -EINVAL;
1028 
1029 		mutex_lock(&container->lock);
1030 		ret = tce_iommu_register_pages(container, param.vaddr,
1031 				param.size);
1032 		mutex_unlock(&container->lock);
1033 
1034 		return ret;
1035 	}
1036 	case VFIO_IOMMU_SPAPR_UNREGISTER_MEMORY: {
1037 		struct vfio_iommu_spapr_register_memory param;
1038 
1039 		if (!container->v2)
1040 			break;
1041 
1042 		if (!container->mm)
1043 			return -EPERM;
1044 
1045 		minsz = offsetofend(struct vfio_iommu_spapr_register_memory,
1046 				size);
1047 
1048 		if (copy_from_user(&param, (void __user *)arg, minsz))
1049 			return -EFAULT;
1050 
1051 		if (param.argsz < minsz)
1052 			return -EINVAL;
1053 
1054 		/* No flag is supported now */
1055 		if (param.flags)
1056 			return -EINVAL;
1057 
1058 		mutex_lock(&container->lock);
1059 		ret = tce_iommu_unregister_pages(container, param.vaddr,
1060 				param.size);
1061 		mutex_unlock(&container->lock);
1062 
1063 		return ret;
1064 	}
1065 	case VFIO_IOMMU_ENABLE:
1066 		if (container->v2)
1067 			break;
1068 
1069 		mutex_lock(&container->lock);
1070 		ret = tce_iommu_enable(container);
1071 		mutex_unlock(&container->lock);
1072 		return ret;
1073 
1074 
1075 	case VFIO_IOMMU_DISABLE:
1076 		if (container->v2)
1077 			break;
1078 
1079 		mutex_lock(&container->lock);
1080 		tce_iommu_disable(container);
1081 		mutex_unlock(&container->lock);
1082 		return 0;
1083 
1084 	case VFIO_EEH_PE_OP: {
1085 		struct tce_iommu_group *tcegrp;
1086 
1087 		ret = 0;
1088 		list_for_each_entry(tcegrp, &container->group_list, next) {
1089 			ret = vfio_spapr_iommu_eeh_ioctl(tcegrp->grp,
1090 					cmd, arg);
1091 			if (ret)
1092 				return ret;
1093 		}
1094 		return ret;
1095 	}
1096 
1097 	case VFIO_IOMMU_SPAPR_TCE_CREATE: {
1098 		struct vfio_iommu_spapr_tce_create create;
1099 
1100 		if (!container->v2)
1101 			break;
1102 
1103 		ret = tce_iommu_mm_set(container);
1104 		if (ret)
1105 			return ret;
1106 
1107 		if (!tce_groups_attached(container))
1108 			return -ENXIO;
1109 
1110 		minsz = offsetofend(struct vfio_iommu_spapr_tce_create,
1111 				start_addr);
1112 
1113 		if (copy_from_user(&create, (void __user *)arg, minsz))
1114 			return -EFAULT;
1115 
1116 		if (create.argsz < minsz)
1117 			return -EINVAL;
1118 
1119 		if (create.flags)
1120 			return -EINVAL;
1121 
1122 		mutex_lock(&container->lock);
1123 
1124 		ret = tce_iommu_create_default_window(container);
1125 		if (!ret)
1126 			ret = tce_iommu_create_window(container,
1127 					create.page_shift,
1128 					create.window_size, create.levels,
1129 					&create.start_addr);
1130 
1131 		mutex_unlock(&container->lock);
1132 
1133 		if (!ret && copy_to_user((void __user *)arg, &create, minsz))
1134 			ret = -EFAULT;
1135 
1136 		return ret;
1137 	}
1138 	case VFIO_IOMMU_SPAPR_TCE_REMOVE: {
1139 		struct vfio_iommu_spapr_tce_remove remove;
1140 
1141 		if (!container->v2)
1142 			break;
1143 
1144 		ret = tce_iommu_mm_set(container);
1145 		if (ret)
1146 			return ret;
1147 
1148 		if (!tce_groups_attached(container))
1149 			return -ENXIO;
1150 
1151 		minsz = offsetofend(struct vfio_iommu_spapr_tce_remove,
1152 				start_addr);
1153 
1154 		if (copy_from_user(&remove, (void __user *)arg, minsz))
1155 			return -EFAULT;
1156 
1157 		if (remove.argsz < minsz)
1158 			return -EINVAL;
1159 
1160 		if (remove.flags)
1161 			return -EINVAL;
1162 
1163 		if (container->def_window_pending && !remove.start_addr) {
1164 			container->def_window_pending = false;
1165 			return 0;
1166 		}
1167 
1168 		mutex_lock(&container->lock);
1169 
1170 		ret = tce_iommu_remove_window(container, remove.start_addr);
1171 
1172 		mutex_unlock(&container->lock);
1173 
1174 		return ret;
1175 	}
1176 	}
1177 
1178 	return -ENOTTY;
1179 }
1180 
1181 static void tce_iommu_release_ownership(struct tce_container *container,
1182 		struct iommu_table_group *table_group)
1183 {
1184 	int i;
1185 
1186 	for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1187 		struct iommu_table *tbl = container->tables[i];
1188 
1189 		if (!tbl)
1190 			continue;
1191 
1192 		tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
1193 		if (tbl->it_map)
1194 			iommu_release_ownership(tbl);
1195 
1196 		container->tables[i] = NULL;
1197 	}
1198 }
1199 
1200 static int tce_iommu_take_ownership(struct tce_container *container,
1201 		struct iommu_table_group *table_group)
1202 {
1203 	int i, j, rc = 0;
1204 
1205 	for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1206 		struct iommu_table *tbl = table_group->tables[i];
1207 
1208 		if (!tbl || !tbl->it_map)
1209 			continue;
1210 
1211 		rc = iommu_take_ownership(tbl);
1212 		if (rc) {
1213 			for (j = 0; j < i; ++j)
1214 				iommu_release_ownership(
1215 						table_group->tables[j]);
1216 
1217 			return rc;
1218 		}
1219 	}
1220 
1221 	for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1222 		container->tables[i] = table_group->tables[i];
1223 
1224 	return 0;
1225 }
1226 
1227 static void tce_iommu_release_ownership_ddw(struct tce_container *container,
1228 		struct iommu_table_group *table_group)
1229 {
1230 	long i;
1231 
1232 	if (!table_group->ops->unset_window) {
1233 		WARN_ON_ONCE(1);
1234 		return;
1235 	}
1236 
1237 	for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1238 		if (container->tables[i])
1239 			table_group->ops->unset_window(table_group, i);
1240 
1241 	table_group->ops->release_ownership(table_group);
1242 }
1243 
1244 static long tce_iommu_take_ownership_ddw(struct tce_container *container,
1245 		struct iommu_table_group *table_group)
1246 {
1247 	long i, ret = 0;
1248 
1249 	if (!table_group->ops->create_table || !table_group->ops->set_window ||
1250 			!table_group->ops->release_ownership) {
1251 		WARN_ON_ONCE(1);
1252 		return -EFAULT;
1253 	}
1254 
1255 	table_group->ops->take_ownership(table_group);
1256 
1257 	/* Set all windows to the new group */
1258 	for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1259 		struct iommu_table *tbl = container->tables[i];
1260 
1261 		if (!tbl)
1262 			continue;
1263 
1264 		ret = table_group->ops->set_window(table_group, i, tbl);
1265 		if (ret)
1266 			goto release_exit;
1267 	}
1268 
1269 	return 0;
1270 
1271 release_exit:
1272 	for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1273 		table_group->ops->unset_window(table_group, i);
1274 
1275 	table_group->ops->release_ownership(table_group);
1276 
1277 	return ret;
1278 }
1279 
1280 static int tce_iommu_attach_group(void *iommu_data,
1281 		struct iommu_group *iommu_group)
1282 {
1283 	int ret;
1284 	struct tce_container *container = iommu_data;
1285 	struct iommu_table_group *table_group;
1286 	struct tce_iommu_group *tcegrp = NULL;
1287 
1288 	mutex_lock(&container->lock);
1289 
1290 	/* pr_debug("tce_vfio: Attaching group #%u to iommu %p\n",
1291 			iommu_group_id(iommu_group), iommu_group); */
1292 	table_group = iommu_group_get_iommudata(iommu_group);
1293 	if (!table_group) {
1294 		ret = -ENODEV;
1295 		goto unlock_exit;
1296 	}
1297 
1298 	if (tce_groups_attached(container) && (!table_group->ops ||
1299 			!table_group->ops->take_ownership ||
1300 			!table_group->ops->release_ownership)) {
1301 		ret = -EBUSY;
1302 		goto unlock_exit;
1303 	}
1304 
1305 	/* Check if new group has the same iommu_ops (i.e. compatible) */
1306 	list_for_each_entry(tcegrp, &container->group_list, next) {
1307 		struct iommu_table_group *table_group_tmp;
1308 
1309 		if (tcegrp->grp == iommu_group) {
1310 			pr_warn("tce_vfio: Group %d is already attached\n",
1311 					iommu_group_id(iommu_group));
1312 			ret = -EBUSY;
1313 			goto unlock_exit;
1314 		}
1315 		table_group_tmp = iommu_group_get_iommudata(tcegrp->grp);
1316 		if (table_group_tmp->ops->create_table !=
1317 				table_group->ops->create_table) {
1318 			pr_warn("tce_vfio: Group %d is incompatible with group %d\n",
1319 					iommu_group_id(iommu_group),
1320 					iommu_group_id(tcegrp->grp));
1321 			ret = -EPERM;
1322 			goto unlock_exit;
1323 		}
1324 	}
1325 
1326 	tcegrp = kzalloc(sizeof(*tcegrp), GFP_KERNEL);
1327 	if (!tcegrp) {
1328 		ret = -ENOMEM;
1329 		goto unlock_exit;
1330 	}
1331 
1332 	if (!table_group->ops || !table_group->ops->take_ownership ||
1333 			!table_group->ops->release_ownership) {
1334 		if (container->v2) {
1335 			ret = -EPERM;
1336 			goto unlock_exit;
1337 		}
1338 		ret = tce_iommu_take_ownership(container, table_group);
1339 	} else {
1340 		if (!container->v2) {
1341 			ret = -EPERM;
1342 			goto unlock_exit;
1343 		}
1344 		ret = tce_iommu_take_ownership_ddw(container, table_group);
1345 		if (!tce_groups_attached(container) && !container->tables[0])
1346 			container->def_window_pending = true;
1347 	}
1348 
1349 	if (!ret) {
1350 		tcegrp->grp = iommu_group;
1351 		list_add(&tcegrp->next, &container->group_list);
1352 	}
1353 
1354 unlock_exit:
1355 	if (ret && tcegrp)
1356 		kfree(tcegrp);
1357 
1358 	mutex_unlock(&container->lock);
1359 
1360 	return ret;
1361 }
1362 
1363 static void tce_iommu_detach_group(void *iommu_data,
1364 		struct iommu_group *iommu_group)
1365 {
1366 	struct tce_container *container = iommu_data;
1367 	struct iommu_table_group *table_group;
1368 	bool found = false;
1369 	struct tce_iommu_group *tcegrp;
1370 
1371 	mutex_lock(&container->lock);
1372 
1373 	list_for_each_entry(tcegrp, &container->group_list, next) {
1374 		if (tcegrp->grp == iommu_group) {
1375 			found = true;
1376 			break;
1377 		}
1378 	}
1379 
1380 	if (!found) {
1381 		pr_warn("tce_vfio: detaching unattached group #%u\n",
1382 				iommu_group_id(iommu_group));
1383 		goto unlock_exit;
1384 	}
1385 
1386 	list_del(&tcegrp->next);
1387 	kfree(tcegrp);
1388 
1389 	table_group = iommu_group_get_iommudata(iommu_group);
1390 	BUG_ON(!table_group);
1391 
1392 	if (!table_group->ops || !table_group->ops->release_ownership)
1393 		tce_iommu_release_ownership(container, table_group);
1394 	else
1395 		tce_iommu_release_ownership_ddw(container, table_group);
1396 
1397 unlock_exit:
1398 	mutex_unlock(&container->lock);
1399 }
1400 
1401 static const struct vfio_iommu_driver_ops tce_iommu_driver_ops = {
1402 	.name		= "iommu-vfio-powerpc",
1403 	.owner		= THIS_MODULE,
1404 	.open		= tce_iommu_open,
1405 	.release	= tce_iommu_release,
1406 	.ioctl		= tce_iommu_ioctl,
1407 	.attach_group	= tce_iommu_attach_group,
1408 	.detach_group	= tce_iommu_detach_group,
1409 };
1410 
1411 static int __init tce_iommu_init(void)
1412 {
1413 	return vfio_register_iommu_driver(&tce_iommu_driver_ops);
1414 }
1415 
1416 static void __exit tce_iommu_cleanup(void)
1417 {
1418 	vfio_unregister_iommu_driver(&tce_iommu_driver_ops);
1419 }
1420 
1421 module_init(tce_iommu_init);
1422 module_exit(tce_iommu_cleanup);
1423 
1424 MODULE_VERSION(DRIVER_VERSION);
1425 MODULE_LICENSE("GPL v2");
1426 MODULE_AUTHOR(DRIVER_AUTHOR);
1427 MODULE_DESCRIPTION(DRIVER_DESC);
1428 
1429