xref: /linux/drivers/vfio/vfio_iommu_spapr_tce.c (revision ca55b2fef3a9373fcfc30f82fd26bc7fccbda732)
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 <asm/iommu.h>
24 #include <asm/tce.h>
25 #include <asm/mmu_context.h>
26 
27 #define DRIVER_VERSION  "0.1"
28 #define DRIVER_AUTHOR   "aik@ozlabs.ru"
29 #define DRIVER_DESC     "VFIO IOMMU SPAPR TCE"
30 
31 static void tce_iommu_detach_group(void *iommu_data,
32 		struct iommu_group *iommu_group);
33 
34 static long try_increment_locked_vm(long npages)
35 {
36 	long ret = 0, locked, lock_limit;
37 
38 	if (!current || !current->mm)
39 		return -ESRCH; /* process exited */
40 
41 	if (!npages)
42 		return 0;
43 
44 	down_write(&current->mm->mmap_sem);
45 	locked = current->mm->locked_vm + npages;
46 	lock_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
47 	if (locked > lock_limit && !capable(CAP_IPC_LOCK))
48 		ret = -ENOMEM;
49 	else
50 		current->mm->locked_vm += npages;
51 
52 	pr_debug("[%d] RLIMIT_MEMLOCK +%ld %ld/%ld%s\n", current->pid,
53 			npages << PAGE_SHIFT,
54 			current->mm->locked_vm << PAGE_SHIFT,
55 			rlimit(RLIMIT_MEMLOCK),
56 			ret ? " - exceeded" : "");
57 
58 	up_write(&current->mm->mmap_sem);
59 
60 	return ret;
61 }
62 
63 static void decrement_locked_vm(long npages)
64 {
65 	if (!current || !current->mm || !npages)
66 		return; /* process exited */
67 
68 	down_write(&current->mm->mmap_sem);
69 	if (WARN_ON_ONCE(npages > current->mm->locked_vm))
70 		npages = current->mm->locked_vm;
71 	current->mm->locked_vm -= npages;
72 	pr_debug("[%d] RLIMIT_MEMLOCK -%ld %ld/%ld\n", current->pid,
73 			npages << PAGE_SHIFT,
74 			current->mm->locked_vm << PAGE_SHIFT,
75 			rlimit(RLIMIT_MEMLOCK));
76 	up_write(&current->mm->mmap_sem);
77 }
78 
79 /*
80  * VFIO IOMMU fd for SPAPR_TCE IOMMU implementation
81  *
82  * This code handles mapping and unmapping of user data buffers
83  * into DMA'ble space using the IOMMU
84  */
85 
86 struct tce_iommu_group {
87 	struct list_head next;
88 	struct iommu_group *grp;
89 };
90 
91 /*
92  * The container descriptor supports only a single group per container.
93  * Required by the API as the container is not supplied with the IOMMU group
94  * at the moment of initialization.
95  */
96 struct tce_container {
97 	struct mutex lock;
98 	bool enabled;
99 	bool v2;
100 	unsigned long locked_pages;
101 	struct iommu_table *tables[IOMMU_TABLE_GROUP_MAX_TABLES];
102 	struct list_head group_list;
103 };
104 
105 static long tce_iommu_unregister_pages(struct tce_container *container,
106 		__u64 vaddr, __u64 size)
107 {
108 	struct mm_iommu_table_group_mem_t *mem;
109 
110 	if ((vaddr & ~PAGE_MASK) || (size & ~PAGE_MASK))
111 		return -EINVAL;
112 
113 	mem = mm_iommu_find(vaddr, size >> PAGE_SHIFT);
114 	if (!mem)
115 		return -ENOENT;
116 
117 	return mm_iommu_put(mem);
118 }
119 
120 static long tce_iommu_register_pages(struct tce_container *container,
121 		__u64 vaddr, __u64 size)
122 {
123 	long ret = 0;
124 	struct mm_iommu_table_group_mem_t *mem = NULL;
125 	unsigned long entries = size >> PAGE_SHIFT;
126 
127 	if ((vaddr & ~PAGE_MASK) || (size & ~PAGE_MASK) ||
128 			((vaddr + size) < vaddr))
129 		return -EINVAL;
130 
131 	ret = mm_iommu_get(vaddr, entries, &mem);
132 	if (ret)
133 		return ret;
134 
135 	container->enabled = true;
136 
137 	return 0;
138 }
139 
140 static long tce_iommu_userspace_view_alloc(struct iommu_table *tbl)
141 {
142 	unsigned long cb = _ALIGN_UP(sizeof(tbl->it_userspace[0]) *
143 			tbl->it_size, PAGE_SIZE);
144 	unsigned long *uas;
145 	long ret;
146 
147 	BUG_ON(tbl->it_userspace);
148 
149 	ret = try_increment_locked_vm(cb >> PAGE_SHIFT);
150 	if (ret)
151 		return ret;
152 
153 	uas = vzalloc(cb);
154 	if (!uas) {
155 		decrement_locked_vm(cb >> PAGE_SHIFT);
156 		return -ENOMEM;
157 	}
158 	tbl->it_userspace = uas;
159 
160 	return 0;
161 }
162 
163 static void tce_iommu_userspace_view_free(struct iommu_table *tbl)
164 {
165 	unsigned long cb = _ALIGN_UP(sizeof(tbl->it_userspace[0]) *
166 			tbl->it_size, PAGE_SIZE);
167 
168 	if (!tbl->it_userspace)
169 		return;
170 
171 	vfree(tbl->it_userspace);
172 	tbl->it_userspace = NULL;
173 	decrement_locked_vm(cb >> PAGE_SHIFT);
174 }
175 
176 static bool tce_page_is_contained(struct page *page, unsigned page_shift)
177 {
178 	/*
179 	 * Check that the TCE table granularity is not bigger than the size of
180 	 * a page we just found. Otherwise the hardware can get access to
181 	 * a bigger memory chunk that it should.
182 	 */
183 	return (PAGE_SHIFT + compound_order(compound_head(page))) >= page_shift;
184 }
185 
186 static inline bool tce_groups_attached(struct tce_container *container)
187 {
188 	return !list_empty(&container->group_list);
189 }
190 
191 static long tce_iommu_find_table(struct tce_container *container,
192 		phys_addr_t ioba, struct iommu_table **ptbl)
193 {
194 	long i;
195 
196 	for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
197 		struct iommu_table *tbl = container->tables[i];
198 
199 		if (tbl) {
200 			unsigned long entry = ioba >> tbl->it_page_shift;
201 			unsigned long start = tbl->it_offset;
202 			unsigned long end = start + tbl->it_size;
203 
204 			if ((start <= entry) && (entry < end)) {
205 				*ptbl = tbl;
206 				return i;
207 			}
208 		}
209 	}
210 
211 	return -1;
212 }
213 
214 static int tce_iommu_find_free_table(struct tce_container *container)
215 {
216 	int i;
217 
218 	for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
219 		if (!container->tables[i])
220 			return i;
221 	}
222 
223 	return -ENOSPC;
224 }
225 
226 static int tce_iommu_enable(struct tce_container *container)
227 {
228 	int ret = 0;
229 	unsigned long locked;
230 	struct iommu_table_group *table_group;
231 	struct tce_iommu_group *tcegrp;
232 
233 	if (!current->mm)
234 		return -ESRCH; /* process exited */
235 
236 	if (container->enabled)
237 		return -EBUSY;
238 
239 	/*
240 	 * When userspace pages are mapped into the IOMMU, they are effectively
241 	 * locked memory, so, theoretically, we need to update the accounting
242 	 * of locked pages on each map and unmap.  For powerpc, the map unmap
243 	 * paths can be very hot, though, and the accounting would kill
244 	 * performance, especially since it would be difficult to impossible
245 	 * to handle the accounting in real mode only.
246 	 *
247 	 * To address that, rather than precisely accounting every page, we
248 	 * instead account for a worst case on locked memory when the iommu is
249 	 * enabled and disabled.  The worst case upper bound on locked memory
250 	 * is the size of the whole iommu window, which is usually relatively
251 	 * small (compared to total memory sizes) on POWER hardware.
252 	 *
253 	 * Also we don't have a nice way to fail on H_PUT_TCE due to ulimits,
254 	 * that would effectively kill the guest at random points, much better
255 	 * enforcing the limit based on the max that the guest can map.
256 	 *
257 	 * Unfortunately at the moment it counts whole tables, no matter how
258 	 * much memory the guest has. I.e. for 4GB guest and 4 IOMMU groups
259 	 * each with 2GB DMA window, 8GB will be counted here. The reason for
260 	 * this is that we cannot tell here the amount of RAM used by the guest
261 	 * as this information is only available from KVM and VFIO is
262 	 * KVM agnostic.
263 	 *
264 	 * So we do not allow enabling a container without a group attached
265 	 * as there is no way to know how much we should increment
266 	 * the locked_vm counter.
267 	 */
268 	if (!tce_groups_attached(container))
269 		return -ENODEV;
270 
271 	tcegrp = list_first_entry(&container->group_list,
272 			struct tce_iommu_group, next);
273 	table_group = iommu_group_get_iommudata(tcegrp->grp);
274 	if (!table_group)
275 		return -ENODEV;
276 
277 	if (!table_group->tce32_size)
278 		return -EPERM;
279 
280 	locked = table_group->tce32_size >> PAGE_SHIFT;
281 	ret = try_increment_locked_vm(locked);
282 	if (ret)
283 		return ret;
284 
285 	container->locked_pages = locked;
286 
287 	container->enabled = true;
288 
289 	return ret;
290 }
291 
292 static void tce_iommu_disable(struct tce_container *container)
293 {
294 	if (!container->enabled)
295 		return;
296 
297 	container->enabled = false;
298 
299 	if (!current->mm)
300 		return;
301 
302 	decrement_locked_vm(container->locked_pages);
303 }
304 
305 static void *tce_iommu_open(unsigned long arg)
306 {
307 	struct tce_container *container;
308 
309 	if ((arg != VFIO_SPAPR_TCE_IOMMU) && (arg != VFIO_SPAPR_TCE_v2_IOMMU)) {
310 		pr_err("tce_vfio: Wrong IOMMU type\n");
311 		return ERR_PTR(-EINVAL);
312 	}
313 
314 	container = kzalloc(sizeof(*container), GFP_KERNEL);
315 	if (!container)
316 		return ERR_PTR(-ENOMEM);
317 
318 	mutex_init(&container->lock);
319 	INIT_LIST_HEAD_RCU(&container->group_list);
320 
321 	container->v2 = arg == VFIO_SPAPR_TCE_v2_IOMMU;
322 
323 	return container;
324 }
325 
326 static int tce_iommu_clear(struct tce_container *container,
327 		struct iommu_table *tbl,
328 		unsigned long entry, unsigned long pages);
329 static void tce_iommu_free_table(struct iommu_table *tbl);
330 
331 static void tce_iommu_release(void *iommu_data)
332 {
333 	struct tce_container *container = iommu_data;
334 	struct iommu_table_group *table_group;
335 	struct tce_iommu_group *tcegrp;
336 	long i;
337 
338 	while (tce_groups_attached(container)) {
339 		tcegrp = list_first_entry(&container->group_list,
340 				struct tce_iommu_group, next);
341 		table_group = iommu_group_get_iommudata(tcegrp->grp);
342 		tce_iommu_detach_group(iommu_data, tcegrp->grp);
343 	}
344 
345 	/*
346 	 * If VFIO created a table, it was not disposed
347 	 * by tce_iommu_detach_group() so do it now.
348 	 */
349 	for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
350 		struct iommu_table *tbl = container->tables[i];
351 
352 		if (!tbl)
353 			continue;
354 
355 		tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
356 		tce_iommu_free_table(tbl);
357 	}
358 
359 	tce_iommu_disable(container);
360 	mutex_destroy(&container->lock);
361 
362 	kfree(container);
363 }
364 
365 static void tce_iommu_unuse_page(struct tce_container *container,
366 		unsigned long hpa)
367 {
368 	struct page *page;
369 
370 	page = pfn_to_page(hpa >> PAGE_SHIFT);
371 	put_page(page);
372 }
373 
374 static int tce_iommu_prereg_ua_to_hpa(unsigned long tce, unsigned long size,
375 		unsigned long *phpa, struct mm_iommu_table_group_mem_t **pmem)
376 {
377 	long ret = 0;
378 	struct mm_iommu_table_group_mem_t *mem;
379 
380 	mem = mm_iommu_lookup(tce, size);
381 	if (!mem)
382 		return -EINVAL;
383 
384 	ret = mm_iommu_ua_to_hpa(mem, tce, phpa);
385 	if (ret)
386 		return -EINVAL;
387 
388 	*pmem = mem;
389 
390 	return 0;
391 }
392 
393 static void tce_iommu_unuse_page_v2(struct iommu_table *tbl,
394 		unsigned long entry)
395 {
396 	struct mm_iommu_table_group_mem_t *mem = NULL;
397 	int ret;
398 	unsigned long hpa = 0;
399 	unsigned long *pua = IOMMU_TABLE_USERSPACE_ENTRY(tbl, entry);
400 
401 	if (!pua || !current || !current->mm)
402 		return;
403 
404 	ret = tce_iommu_prereg_ua_to_hpa(*pua, IOMMU_PAGE_SIZE(tbl),
405 			&hpa, &mem);
406 	if (ret)
407 		pr_debug("%s: tce %lx at #%lx was not cached, ret=%d\n",
408 				__func__, *pua, entry, ret);
409 	if (mem)
410 		mm_iommu_mapped_dec(mem);
411 
412 	*pua = 0;
413 }
414 
415 static int tce_iommu_clear(struct tce_container *container,
416 		struct iommu_table *tbl,
417 		unsigned long entry, unsigned long pages)
418 {
419 	unsigned long oldhpa;
420 	long ret;
421 	enum dma_data_direction direction;
422 
423 	for ( ; pages; --pages, ++entry) {
424 		direction = DMA_NONE;
425 		oldhpa = 0;
426 		ret = iommu_tce_xchg(tbl, entry, &oldhpa, &direction);
427 		if (ret)
428 			continue;
429 
430 		if (direction == DMA_NONE)
431 			continue;
432 
433 		if (container->v2) {
434 			tce_iommu_unuse_page_v2(tbl, entry);
435 			continue;
436 		}
437 
438 		tce_iommu_unuse_page(container, oldhpa);
439 	}
440 
441 	return 0;
442 }
443 
444 static int tce_iommu_use_page(unsigned long tce, unsigned long *hpa)
445 {
446 	struct page *page = NULL;
447 	enum dma_data_direction direction = iommu_tce_direction(tce);
448 
449 	if (get_user_pages_fast(tce & PAGE_MASK, 1,
450 			direction != DMA_TO_DEVICE, &page) != 1)
451 		return -EFAULT;
452 
453 	*hpa = __pa((unsigned long) page_address(page));
454 
455 	return 0;
456 }
457 
458 static long tce_iommu_build(struct tce_container *container,
459 		struct iommu_table *tbl,
460 		unsigned long entry, unsigned long tce, unsigned long pages,
461 		enum dma_data_direction direction)
462 {
463 	long i, ret = 0;
464 	struct page *page;
465 	unsigned long hpa;
466 	enum dma_data_direction dirtmp;
467 
468 	for (i = 0; i < pages; ++i) {
469 		unsigned long offset = tce & IOMMU_PAGE_MASK(tbl) & ~PAGE_MASK;
470 
471 		ret = tce_iommu_use_page(tce, &hpa);
472 		if (ret)
473 			break;
474 
475 		page = pfn_to_page(hpa >> PAGE_SHIFT);
476 		if (!tce_page_is_contained(page, tbl->it_page_shift)) {
477 			ret = -EPERM;
478 			break;
479 		}
480 
481 		hpa |= offset;
482 		dirtmp = direction;
483 		ret = iommu_tce_xchg(tbl, entry + i, &hpa, &dirtmp);
484 		if (ret) {
485 			tce_iommu_unuse_page(container, hpa);
486 			pr_err("iommu_tce: %s failed ioba=%lx, tce=%lx, ret=%ld\n",
487 					__func__, entry << tbl->it_page_shift,
488 					tce, ret);
489 			break;
490 		}
491 
492 		if (dirtmp != DMA_NONE)
493 			tce_iommu_unuse_page(container, hpa);
494 
495 		tce += IOMMU_PAGE_SIZE(tbl);
496 	}
497 
498 	if (ret)
499 		tce_iommu_clear(container, tbl, entry, i);
500 
501 	return ret;
502 }
503 
504 static long tce_iommu_build_v2(struct tce_container *container,
505 		struct iommu_table *tbl,
506 		unsigned long entry, unsigned long tce, unsigned long pages,
507 		enum dma_data_direction direction)
508 {
509 	long i, ret = 0;
510 	struct page *page;
511 	unsigned long hpa;
512 	enum dma_data_direction dirtmp;
513 
514 	for (i = 0; i < pages; ++i) {
515 		struct mm_iommu_table_group_mem_t *mem = NULL;
516 		unsigned long *pua = IOMMU_TABLE_USERSPACE_ENTRY(tbl,
517 				entry + i);
518 
519 		ret = tce_iommu_prereg_ua_to_hpa(tce, IOMMU_PAGE_SIZE(tbl),
520 				&hpa, &mem);
521 		if (ret)
522 			break;
523 
524 		page = pfn_to_page(hpa >> PAGE_SHIFT);
525 		if (!tce_page_is_contained(page, tbl->it_page_shift)) {
526 			ret = -EPERM;
527 			break;
528 		}
529 
530 		/* Preserve offset within IOMMU page */
531 		hpa |= tce & IOMMU_PAGE_MASK(tbl) & ~PAGE_MASK;
532 		dirtmp = direction;
533 
534 		/* The registered region is being unregistered */
535 		if (mm_iommu_mapped_inc(mem))
536 			break;
537 
538 		ret = iommu_tce_xchg(tbl, entry + i, &hpa, &dirtmp);
539 		if (ret) {
540 			/* dirtmp cannot be DMA_NONE here */
541 			tce_iommu_unuse_page_v2(tbl, entry + i);
542 			pr_err("iommu_tce: %s failed ioba=%lx, tce=%lx, ret=%ld\n",
543 					__func__, entry << tbl->it_page_shift,
544 					tce, ret);
545 			break;
546 		}
547 
548 		if (dirtmp != DMA_NONE)
549 			tce_iommu_unuse_page_v2(tbl, entry + i);
550 
551 		*pua = tce;
552 
553 		tce += IOMMU_PAGE_SIZE(tbl);
554 	}
555 
556 	if (ret)
557 		tce_iommu_clear(container, tbl, entry, i);
558 
559 	return ret;
560 }
561 
562 static long tce_iommu_create_table(struct tce_container *container,
563 			struct iommu_table_group *table_group,
564 			int num,
565 			__u32 page_shift,
566 			__u64 window_size,
567 			__u32 levels,
568 			struct iommu_table **ptbl)
569 {
570 	long ret, table_size;
571 
572 	table_size = table_group->ops->get_table_size(page_shift, window_size,
573 			levels);
574 	if (!table_size)
575 		return -EINVAL;
576 
577 	ret = try_increment_locked_vm(table_size >> PAGE_SHIFT);
578 	if (ret)
579 		return ret;
580 
581 	ret = table_group->ops->create_table(table_group, num,
582 			page_shift, window_size, levels, ptbl);
583 
584 	WARN_ON(!ret && !(*ptbl)->it_ops->free);
585 	WARN_ON(!ret && ((*ptbl)->it_allocated_size != table_size));
586 
587 	if (!ret && container->v2) {
588 		ret = tce_iommu_userspace_view_alloc(*ptbl);
589 		if (ret)
590 			(*ptbl)->it_ops->free(*ptbl);
591 	}
592 
593 	if (ret)
594 		decrement_locked_vm(table_size >> PAGE_SHIFT);
595 
596 	return ret;
597 }
598 
599 static void tce_iommu_free_table(struct iommu_table *tbl)
600 {
601 	unsigned long pages = tbl->it_allocated_size >> PAGE_SHIFT;
602 
603 	tce_iommu_userspace_view_free(tbl);
604 	tbl->it_ops->free(tbl);
605 	decrement_locked_vm(pages);
606 }
607 
608 static long tce_iommu_create_window(struct tce_container *container,
609 		__u32 page_shift, __u64 window_size, __u32 levels,
610 		__u64 *start_addr)
611 {
612 	struct tce_iommu_group *tcegrp;
613 	struct iommu_table_group *table_group;
614 	struct iommu_table *tbl = NULL;
615 	long ret, num;
616 
617 	num = tce_iommu_find_free_table(container);
618 	if (num < 0)
619 		return num;
620 
621 	/* Get the first group for ops::create_table */
622 	tcegrp = list_first_entry(&container->group_list,
623 			struct tce_iommu_group, next);
624 	table_group = iommu_group_get_iommudata(tcegrp->grp);
625 	if (!table_group)
626 		return -EFAULT;
627 
628 	if (!(table_group->pgsizes & (1ULL << page_shift)))
629 		return -EINVAL;
630 
631 	if (!table_group->ops->set_window || !table_group->ops->unset_window ||
632 			!table_group->ops->get_table_size ||
633 			!table_group->ops->create_table)
634 		return -EPERM;
635 
636 	/* Create TCE table */
637 	ret = tce_iommu_create_table(container, table_group, num,
638 			page_shift, window_size, levels, &tbl);
639 	if (ret)
640 		return ret;
641 
642 	BUG_ON(!tbl->it_ops->free);
643 
644 	/*
645 	 * Program the table to every group.
646 	 * Groups have been tested for compatibility at the attach time.
647 	 */
648 	list_for_each_entry(tcegrp, &container->group_list, next) {
649 		table_group = iommu_group_get_iommudata(tcegrp->grp);
650 
651 		ret = table_group->ops->set_window(table_group, num, tbl);
652 		if (ret)
653 			goto unset_exit;
654 	}
655 
656 	container->tables[num] = tbl;
657 
658 	/* Return start address assigned by platform in create_table() */
659 	*start_addr = tbl->it_offset << tbl->it_page_shift;
660 
661 	return 0;
662 
663 unset_exit:
664 	list_for_each_entry(tcegrp, &container->group_list, next) {
665 		table_group = iommu_group_get_iommudata(tcegrp->grp);
666 		table_group->ops->unset_window(table_group, num);
667 	}
668 	tce_iommu_free_table(tbl);
669 
670 	return ret;
671 }
672 
673 static long tce_iommu_remove_window(struct tce_container *container,
674 		__u64 start_addr)
675 {
676 	struct iommu_table_group *table_group = NULL;
677 	struct iommu_table *tbl;
678 	struct tce_iommu_group *tcegrp;
679 	int num;
680 
681 	num = tce_iommu_find_table(container, start_addr, &tbl);
682 	if (num < 0)
683 		return -EINVAL;
684 
685 	BUG_ON(!tbl->it_size);
686 
687 	/* Detach groups from IOMMUs */
688 	list_for_each_entry(tcegrp, &container->group_list, next) {
689 		table_group = iommu_group_get_iommudata(tcegrp->grp);
690 
691 		/*
692 		 * SPAPR TCE IOMMU exposes the default DMA window to
693 		 * the guest via dma32_window_start/size of
694 		 * VFIO_IOMMU_SPAPR_TCE_GET_INFO. Some platforms allow
695 		 * the userspace to remove this window, some do not so
696 		 * here we check for the platform capability.
697 		 */
698 		if (!table_group->ops || !table_group->ops->unset_window)
699 			return -EPERM;
700 
701 		table_group->ops->unset_window(table_group, num);
702 	}
703 
704 	/* Free table */
705 	tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
706 	tce_iommu_free_table(tbl);
707 	container->tables[num] = NULL;
708 
709 	return 0;
710 }
711 
712 static long tce_iommu_ioctl(void *iommu_data,
713 				 unsigned int cmd, unsigned long arg)
714 {
715 	struct tce_container *container = iommu_data;
716 	unsigned long minsz, ddwsz;
717 	long ret;
718 
719 	switch (cmd) {
720 	case VFIO_CHECK_EXTENSION:
721 		switch (arg) {
722 		case VFIO_SPAPR_TCE_IOMMU:
723 		case VFIO_SPAPR_TCE_v2_IOMMU:
724 			ret = 1;
725 			break;
726 		default:
727 			ret = vfio_spapr_iommu_eeh_ioctl(NULL, cmd, arg);
728 			break;
729 		}
730 
731 		return (ret < 0) ? 0 : ret;
732 
733 	case VFIO_IOMMU_SPAPR_TCE_GET_INFO: {
734 		struct vfio_iommu_spapr_tce_info info;
735 		struct tce_iommu_group *tcegrp;
736 		struct iommu_table_group *table_group;
737 
738 		if (!tce_groups_attached(container))
739 			return -ENXIO;
740 
741 		tcegrp = list_first_entry(&container->group_list,
742 				struct tce_iommu_group, next);
743 		table_group = iommu_group_get_iommudata(tcegrp->grp);
744 
745 		if (!table_group)
746 			return -ENXIO;
747 
748 		minsz = offsetofend(struct vfio_iommu_spapr_tce_info,
749 				dma32_window_size);
750 
751 		if (copy_from_user(&info, (void __user *)arg, minsz))
752 			return -EFAULT;
753 
754 		if (info.argsz < minsz)
755 			return -EINVAL;
756 
757 		info.dma32_window_start = table_group->tce32_start;
758 		info.dma32_window_size = table_group->tce32_size;
759 		info.flags = 0;
760 		memset(&info.ddw, 0, sizeof(info.ddw));
761 
762 		if (table_group->max_dynamic_windows_supported &&
763 				container->v2) {
764 			info.flags |= VFIO_IOMMU_SPAPR_INFO_DDW;
765 			info.ddw.pgsizes = table_group->pgsizes;
766 			info.ddw.max_dynamic_windows_supported =
767 				table_group->max_dynamic_windows_supported;
768 			info.ddw.levels = table_group->max_levels;
769 		}
770 
771 		ddwsz = offsetofend(struct vfio_iommu_spapr_tce_info, ddw);
772 
773 		if (info.argsz >= ddwsz)
774 			minsz = ddwsz;
775 
776 		if (copy_to_user((void __user *)arg, &info, minsz))
777 			return -EFAULT;
778 
779 		return 0;
780 	}
781 	case VFIO_IOMMU_MAP_DMA: {
782 		struct vfio_iommu_type1_dma_map param;
783 		struct iommu_table *tbl = NULL;
784 		long num;
785 		enum dma_data_direction direction;
786 
787 		if (!container->enabled)
788 			return -EPERM;
789 
790 		minsz = offsetofend(struct vfio_iommu_type1_dma_map, size);
791 
792 		if (copy_from_user(&param, (void __user *)arg, minsz))
793 			return -EFAULT;
794 
795 		if (param.argsz < minsz)
796 			return -EINVAL;
797 
798 		if (param.flags & ~(VFIO_DMA_MAP_FLAG_READ |
799 				VFIO_DMA_MAP_FLAG_WRITE))
800 			return -EINVAL;
801 
802 		num = tce_iommu_find_table(container, param.iova, &tbl);
803 		if (num < 0)
804 			return -ENXIO;
805 
806 		if ((param.size & ~IOMMU_PAGE_MASK(tbl)) ||
807 				(param.vaddr & ~IOMMU_PAGE_MASK(tbl)))
808 			return -EINVAL;
809 
810 		/* iova is checked by the IOMMU API */
811 		if (param.flags & VFIO_DMA_MAP_FLAG_READ) {
812 			if (param.flags & VFIO_DMA_MAP_FLAG_WRITE)
813 				direction = DMA_BIDIRECTIONAL;
814 			else
815 				direction = DMA_TO_DEVICE;
816 		} else {
817 			if (param.flags & VFIO_DMA_MAP_FLAG_WRITE)
818 				direction = DMA_FROM_DEVICE;
819 			else
820 				return -EINVAL;
821 		}
822 
823 		ret = iommu_tce_put_param_check(tbl, param.iova, param.vaddr);
824 		if (ret)
825 			return ret;
826 
827 		if (container->v2)
828 			ret = tce_iommu_build_v2(container, tbl,
829 					param.iova >> tbl->it_page_shift,
830 					param.vaddr,
831 					param.size >> tbl->it_page_shift,
832 					direction);
833 		else
834 			ret = tce_iommu_build(container, tbl,
835 					param.iova >> tbl->it_page_shift,
836 					param.vaddr,
837 					param.size >> tbl->it_page_shift,
838 					direction);
839 
840 		iommu_flush_tce(tbl);
841 
842 		return ret;
843 	}
844 	case VFIO_IOMMU_UNMAP_DMA: {
845 		struct vfio_iommu_type1_dma_unmap param;
846 		struct iommu_table *tbl = NULL;
847 		long num;
848 
849 		if (!container->enabled)
850 			return -EPERM;
851 
852 		minsz = offsetofend(struct vfio_iommu_type1_dma_unmap,
853 				size);
854 
855 		if (copy_from_user(&param, (void __user *)arg, minsz))
856 			return -EFAULT;
857 
858 		if (param.argsz < minsz)
859 			return -EINVAL;
860 
861 		/* No flag is supported now */
862 		if (param.flags)
863 			return -EINVAL;
864 
865 		num = tce_iommu_find_table(container, param.iova, &tbl);
866 		if (num < 0)
867 			return -ENXIO;
868 
869 		if (param.size & ~IOMMU_PAGE_MASK(tbl))
870 			return -EINVAL;
871 
872 		ret = iommu_tce_clear_param_check(tbl, param.iova, 0,
873 				param.size >> tbl->it_page_shift);
874 		if (ret)
875 			return ret;
876 
877 		ret = tce_iommu_clear(container, tbl,
878 				param.iova >> tbl->it_page_shift,
879 				param.size >> tbl->it_page_shift);
880 		iommu_flush_tce(tbl);
881 
882 		return ret;
883 	}
884 	case VFIO_IOMMU_SPAPR_REGISTER_MEMORY: {
885 		struct vfio_iommu_spapr_register_memory param;
886 
887 		if (!container->v2)
888 			break;
889 
890 		minsz = offsetofend(struct vfio_iommu_spapr_register_memory,
891 				size);
892 
893 		if (copy_from_user(&param, (void __user *)arg, minsz))
894 			return -EFAULT;
895 
896 		if (param.argsz < minsz)
897 			return -EINVAL;
898 
899 		/* No flag is supported now */
900 		if (param.flags)
901 			return -EINVAL;
902 
903 		mutex_lock(&container->lock);
904 		ret = tce_iommu_register_pages(container, param.vaddr,
905 				param.size);
906 		mutex_unlock(&container->lock);
907 
908 		return ret;
909 	}
910 	case VFIO_IOMMU_SPAPR_UNREGISTER_MEMORY: {
911 		struct vfio_iommu_spapr_register_memory param;
912 
913 		if (!container->v2)
914 			break;
915 
916 		minsz = offsetofend(struct vfio_iommu_spapr_register_memory,
917 				size);
918 
919 		if (copy_from_user(&param, (void __user *)arg, minsz))
920 			return -EFAULT;
921 
922 		if (param.argsz < minsz)
923 			return -EINVAL;
924 
925 		/* No flag is supported now */
926 		if (param.flags)
927 			return -EINVAL;
928 
929 		mutex_lock(&container->lock);
930 		ret = tce_iommu_unregister_pages(container, param.vaddr,
931 				param.size);
932 		mutex_unlock(&container->lock);
933 
934 		return ret;
935 	}
936 	case VFIO_IOMMU_ENABLE:
937 		if (container->v2)
938 			break;
939 
940 		mutex_lock(&container->lock);
941 		ret = tce_iommu_enable(container);
942 		mutex_unlock(&container->lock);
943 		return ret;
944 
945 
946 	case VFIO_IOMMU_DISABLE:
947 		if (container->v2)
948 			break;
949 
950 		mutex_lock(&container->lock);
951 		tce_iommu_disable(container);
952 		mutex_unlock(&container->lock);
953 		return 0;
954 
955 	case VFIO_EEH_PE_OP: {
956 		struct tce_iommu_group *tcegrp;
957 
958 		ret = 0;
959 		list_for_each_entry(tcegrp, &container->group_list, next) {
960 			ret = vfio_spapr_iommu_eeh_ioctl(tcegrp->grp,
961 					cmd, arg);
962 			if (ret)
963 				return ret;
964 		}
965 		return ret;
966 	}
967 
968 	case VFIO_IOMMU_SPAPR_TCE_CREATE: {
969 		struct vfio_iommu_spapr_tce_create create;
970 
971 		if (!container->v2)
972 			break;
973 
974 		if (!tce_groups_attached(container))
975 			return -ENXIO;
976 
977 		minsz = offsetofend(struct vfio_iommu_spapr_tce_create,
978 				start_addr);
979 
980 		if (copy_from_user(&create, (void __user *)arg, minsz))
981 			return -EFAULT;
982 
983 		if (create.argsz < minsz)
984 			return -EINVAL;
985 
986 		if (create.flags)
987 			return -EINVAL;
988 
989 		mutex_lock(&container->lock);
990 
991 		ret = tce_iommu_create_window(container, create.page_shift,
992 				create.window_size, create.levels,
993 				&create.start_addr);
994 
995 		mutex_unlock(&container->lock);
996 
997 		if (!ret && copy_to_user((void __user *)arg, &create, minsz))
998 			ret = -EFAULT;
999 
1000 		return ret;
1001 	}
1002 	case VFIO_IOMMU_SPAPR_TCE_REMOVE: {
1003 		struct vfio_iommu_spapr_tce_remove remove;
1004 
1005 		if (!container->v2)
1006 			break;
1007 
1008 		if (!tce_groups_attached(container))
1009 			return -ENXIO;
1010 
1011 		minsz = offsetofend(struct vfio_iommu_spapr_tce_remove,
1012 				start_addr);
1013 
1014 		if (copy_from_user(&remove, (void __user *)arg, minsz))
1015 			return -EFAULT;
1016 
1017 		if (remove.argsz < minsz)
1018 			return -EINVAL;
1019 
1020 		if (remove.flags)
1021 			return -EINVAL;
1022 
1023 		mutex_lock(&container->lock);
1024 
1025 		ret = tce_iommu_remove_window(container, remove.start_addr);
1026 
1027 		mutex_unlock(&container->lock);
1028 
1029 		return ret;
1030 	}
1031 	}
1032 
1033 	return -ENOTTY;
1034 }
1035 
1036 static void tce_iommu_release_ownership(struct tce_container *container,
1037 		struct iommu_table_group *table_group)
1038 {
1039 	int i;
1040 
1041 	for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1042 		struct iommu_table *tbl = container->tables[i];
1043 
1044 		if (!tbl)
1045 			continue;
1046 
1047 		tce_iommu_clear(container, tbl, tbl->it_offset, tbl->it_size);
1048 		tce_iommu_userspace_view_free(tbl);
1049 		if (tbl->it_map)
1050 			iommu_release_ownership(tbl);
1051 
1052 		container->tables[i] = NULL;
1053 	}
1054 }
1055 
1056 static int tce_iommu_take_ownership(struct tce_container *container,
1057 		struct iommu_table_group *table_group)
1058 {
1059 	int i, j, rc = 0;
1060 
1061 	for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1062 		struct iommu_table *tbl = table_group->tables[i];
1063 
1064 		if (!tbl || !tbl->it_map)
1065 			continue;
1066 
1067 		rc = tce_iommu_userspace_view_alloc(tbl);
1068 		if (!rc)
1069 			rc = iommu_take_ownership(tbl);
1070 
1071 		if (rc) {
1072 			for (j = 0; j < i; ++j)
1073 				iommu_release_ownership(
1074 						table_group->tables[j]);
1075 
1076 			return rc;
1077 		}
1078 	}
1079 
1080 	for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1081 		container->tables[i] = table_group->tables[i];
1082 
1083 	return 0;
1084 }
1085 
1086 static void tce_iommu_release_ownership_ddw(struct tce_container *container,
1087 		struct iommu_table_group *table_group)
1088 {
1089 	long i;
1090 
1091 	if (!table_group->ops->unset_window) {
1092 		WARN_ON_ONCE(1);
1093 		return;
1094 	}
1095 
1096 	for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1097 		table_group->ops->unset_window(table_group, i);
1098 
1099 	table_group->ops->release_ownership(table_group);
1100 }
1101 
1102 static long tce_iommu_take_ownership_ddw(struct tce_container *container,
1103 		struct iommu_table_group *table_group)
1104 {
1105 	long i, ret = 0;
1106 	struct iommu_table *tbl = NULL;
1107 
1108 	if (!table_group->ops->create_table || !table_group->ops->set_window ||
1109 			!table_group->ops->release_ownership) {
1110 		WARN_ON_ONCE(1);
1111 		return -EFAULT;
1112 	}
1113 
1114 	table_group->ops->take_ownership(table_group);
1115 
1116 	/*
1117 	 * If it the first group attached, check if there is
1118 	 * a default DMA window and create one if none as
1119 	 * the userspace expects it to exist.
1120 	 */
1121 	if (!tce_groups_attached(container) && !container->tables[0]) {
1122 		ret = tce_iommu_create_table(container,
1123 				table_group,
1124 				0, /* window number */
1125 				IOMMU_PAGE_SHIFT_4K,
1126 				table_group->tce32_size,
1127 				1, /* default levels */
1128 				&tbl);
1129 		if (ret)
1130 			goto release_exit;
1131 		else
1132 			container->tables[0] = tbl;
1133 	}
1134 
1135 	/* Set all windows to the new group */
1136 	for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i) {
1137 		tbl = container->tables[i];
1138 
1139 		if (!tbl)
1140 			continue;
1141 
1142 		/* Set the default window to a new group */
1143 		ret = table_group->ops->set_window(table_group, i, tbl);
1144 		if (ret)
1145 			goto release_exit;
1146 	}
1147 
1148 	return 0;
1149 
1150 release_exit:
1151 	for (i = 0; i < IOMMU_TABLE_GROUP_MAX_TABLES; ++i)
1152 		table_group->ops->unset_window(table_group, i);
1153 
1154 	table_group->ops->release_ownership(table_group);
1155 
1156 	return ret;
1157 }
1158 
1159 static int tce_iommu_attach_group(void *iommu_data,
1160 		struct iommu_group *iommu_group)
1161 {
1162 	int ret;
1163 	struct tce_container *container = iommu_data;
1164 	struct iommu_table_group *table_group;
1165 	struct tce_iommu_group *tcegrp = NULL;
1166 
1167 	mutex_lock(&container->lock);
1168 
1169 	/* pr_debug("tce_vfio: Attaching group #%u to iommu %p\n",
1170 			iommu_group_id(iommu_group), iommu_group); */
1171 	table_group = iommu_group_get_iommudata(iommu_group);
1172 
1173 	if (tce_groups_attached(container) && (!table_group->ops ||
1174 			!table_group->ops->take_ownership ||
1175 			!table_group->ops->release_ownership)) {
1176 		ret = -EBUSY;
1177 		goto unlock_exit;
1178 	}
1179 
1180 	/* Check if new group has the same iommu_ops (i.e. compatible) */
1181 	list_for_each_entry(tcegrp, &container->group_list, next) {
1182 		struct iommu_table_group *table_group_tmp;
1183 
1184 		if (tcegrp->grp == iommu_group) {
1185 			pr_warn("tce_vfio: Group %d is already attached\n",
1186 					iommu_group_id(iommu_group));
1187 			ret = -EBUSY;
1188 			goto unlock_exit;
1189 		}
1190 		table_group_tmp = iommu_group_get_iommudata(tcegrp->grp);
1191 		if (table_group_tmp->ops != table_group->ops) {
1192 			pr_warn("tce_vfio: Group %d is incompatible with group %d\n",
1193 					iommu_group_id(iommu_group),
1194 					iommu_group_id(tcegrp->grp));
1195 			ret = -EPERM;
1196 			goto unlock_exit;
1197 		}
1198 	}
1199 
1200 	tcegrp = kzalloc(sizeof(*tcegrp), GFP_KERNEL);
1201 	if (!tcegrp) {
1202 		ret = -ENOMEM;
1203 		goto unlock_exit;
1204 	}
1205 
1206 	if (!table_group->ops || !table_group->ops->take_ownership ||
1207 			!table_group->ops->release_ownership)
1208 		ret = tce_iommu_take_ownership(container, table_group);
1209 	else
1210 		ret = tce_iommu_take_ownership_ddw(container, table_group);
1211 
1212 	if (!ret) {
1213 		tcegrp->grp = iommu_group;
1214 		list_add(&tcegrp->next, &container->group_list);
1215 	}
1216 
1217 unlock_exit:
1218 	if (ret && tcegrp)
1219 		kfree(tcegrp);
1220 
1221 	mutex_unlock(&container->lock);
1222 
1223 	return ret;
1224 }
1225 
1226 static void tce_iommu_detach_group(void *iommu_data,
1227 		struct iommu_group *iommu_group)
1228 {
1229 	struct tce_container *container = iommu_data;
1230 	struct iommu_table_group *table_group;
1231 	bool found = false;
1232 	struct tce_iommu_group *tcegrp;
1233 
1234 	mutex_lock(&container->lock);
1235 
1236 	list_for_each_entry(tcegrp, &container->group_list, next) {
1237 		if (tcegrp->grp == iommu_group) {
1238 			found = true;
1239 			break;
1240 		}
1241 	}
1242 
1243 	if (!found) {
1244 		pr_warn("tce_vfio: detaching unattached group #%u\n",
1245 				iommu_group_id(iommu_group));
1246 		goto unlock_exit;
1247 	}
1248 
1249 	list_del(&tcegrp->next);
1250 	kfree(tcegrp);
1251 
1252 	table_group = iommu_group_get_iommudata(iommu_group);
1253 	BUG_ON(!table_group);
1254 
1255 	if (!table_group->ops || !table_group->ops->release_ownership)
1256 		tce_iommu_release_ownership(container, table_group);
1257 	else
1258 		tce_iommu_release_ownership_ddw(container, table_group);
1259 
1260 unlock_exit:
1261 	mutex_unlock(&container->lock);
1262 }
1263 
1264 const struct vfio_iommu_driver_ops tce_iommu_driver_ops = {
1265 	.name		= "iommu-vfio-powerpc",
1266 	.owner		= THIS_MODULE,
1267 	.open		= tce_iommu_open,
1268 	.release	= tce_iommu_release,
1269 	.ioctl		= tce_iommu_ioctl,
1270 	.attach_group	= tce_iommu_attach_group,
1271 	.detach_group	= tce_iommu_detach_group,
1272 };
1273 
1274 static int __init tce_iommu_init(void)
1275 {
1276 	return vfio_register_iommu_driver(&tce_iommu_driver_ops);
1277 }
1278 
1279 static void __exit tce_iommu_cleanup(void)
1280 {
1281 	vfio_unregister_iommu_driver(&tce_iommu_driver_ops);
1282 }
1283 
1284 module_init(tce_iommu_init);
1285 module_exit(tce_iommu_cleanup);
1286 
1287 MODULE_VERSION(DRIVER_VERSION);
1288 MODULE_LICENSE("GPL v2");
1289 MODULE_AUTHOR(DRIVER_AUTHOR);
1290 MODULE_DESCRIPTION(DRIVER_DESC);
1291 
1292