1	// SPDX-License-Identifier: GPL-2.0-only
2	/*
3	* VFIO: IOMMU DMA mapping support for Type1 IOMMU
4	*
5	* Copyright (C) 2012 Red Hat, Inc. All rights reserved.
6	* Author: Alex Williamson <alex.williamson@redhat.com>
7	*
8	* Derived from original vfio:
9	* Copyright 2010 Cisco Systems, Inc. All rights reserved.
10	* Author: Tom Lyon, pugs@cisco.com
11	*
12	* We arbitrarily define a Type1 IOMMU as one matching the below code.
13	* It could be called the x86 IOMMU as it's designed for AMD-Vi & Intel
14	* VT-d, but that makes it harder to re-use as theoretically anyone
15	* implementing a similar IOMMU could make use of this. We expect the
16	* IOMMU to support the IOMMU API and have few to no restrictions around
17	* the IOVA range that can be mapped. The Type1 IOMMU is currently
18	* optimized for relatively static mappings of a userspace process with
19	* userspace pages pinned into memory. We also assume devices and IOMMU
20	* domains are PCI based as the IOMMU API is still centered around a
21	* device/bus interface rather than a group interface.
22	*/
23
24	#include <linux/compat.h>
25	#include <linux/device.h>
26	#include <linux/fs.h>
27	#include <linux/highmem.h>
28	#include <linux/iommu.h>
29	#include <linux/module.h>
30	#include <linux/mm.h>
31	#include <linux/kthread.h>
32	#include <linux/rbtree.h>
33	#include <linux/sched/signal.h>
34	#include <linux/sched/mm.h>
35	#include <linux/slab.h>
36	#include <linux/uaccess.h>
37	#include <linux/vfio.h>
38	#include <linux/workqueue.h>
39	#include <linux/notifier.h>
40	#include "vfio.h"
41
42	#define DRIVER_VERSION "0.2"
43	#define DRIVER_AUTHOR "Alex Williamson <alex.williamson@redhat.com>"
44	#define DRIVER_DESC "Type1 IOMMU driver for VFIO"
45
46	static bool allow_unsafe_interrupts;
47	module_param_named(allow_unsafe_interrupts,
48	allow_unsafe_interrupts, bool, S_IRUGO | S_IWUSR);
49	MODULE_PARM_DESC(allow_unsafe_interrupts,
50	"Enable VFIO IOMMU support for on platforms without interrupt remapping support.");
51
52	static bool disable_hugepages;
53	module_param_named(disable_hugepages,
54	disable_hugepages, bool, S_IRUGO | S_IWUSR);
55	MODULE_PARM_DESC(disable_hugepages,
56	"Disable VFIO IOMMU support for IOMMU hugepages.");
57
58	static unsigned int dma_entry_limit __read_mostly = U16_MAX;
59	module_param_named(dma_entry_limit, dma_entry_limit, uint, 0644);
60	MODULE_PARM_DESC(dma_entry_limit,
61	"Maximum number of user DMA mappings per container (65535).");
62
63	struct vfio_iommu {
64	struct list_head domain_list;
65	struct list_head iova_list;
66	struct mutex lock;
67	struct rb_root dma_list;
68	struct list_head device_list;
69	struct mutex device_list_lock;
70	unsigned int dma_avail;
71	unsigned int vaddr_invalid_count;
72	uint64_t pgsize_bitmap;
73	uint64_t num_non_pinned_groups;
74	bool v2;
75	bool nesting;
76	bool dirty_page_tracking;
77	struct list_head emulated_iommu_groups;
78	};
79
80	struct vfio_domain {
81	struct iommu_domain *domain;
82	struct list_head next;
83	struct list_head group_list;
84	bool fgsp : 1; /* Fine-grained super pages */
85	bool enforce_cache_coherency : 1;
86	};
87
88	struct vfio_dma {
89	struct rb_node node;
90	dma_addr_t iova; /* Device address */
91	unsigned long vaddr; /* Process virtual addr */
92	size_t size; /* Map size (bytes) */
93	int prot; /* IOMMU_READ/WRITE */
94	bool iommu_mapped;
95	bool lock_cap; /* capable(CAP_IPC_LOCK) */
96	bool vaddr_invalid;
97	struct task_struct *task;
98	struct rb_root pfn_list; /* Ex-user pinned pfn list */
99	unsigned long *bitmap;
100	struct mm_struct *mm;
101	size_t locked_vm;
102	};
103
104	struct vfio_batch {
105	struct page **pages; /* for pin_user_pages_remote */
106	struct page *fallback_page; /* if pages alloc fails */
107	int capacity; /* length of pages array */
108	int size; /* of batch currently */
109	int offset; /* of next entry in pages */
110	};
111
112	struct vfio_iommu_group {
113	struct iommu_group *iommu_group;
114	struct list_head next;
115	bool pinned_page_dirty_scope;
116	};
117
118	struct vfio_iova {
119	struct list_head list;
120	dma_addr_t start;
121	dma_addr_t end;
122	};
123
124	/*
125	* Guest RAM pinning working set or DMA target
126	*/
127	struct vfio_pfn {
128	struct rb_node node;
129	dma_addr_t iova; /* Device address */
130	unsigned long pfn; /* Host pfn */
131	unsigned int ref_count;
132	};
133
134	struct vfio_regions {
135	struct list_head list;
136	dma_addr_t iova;
137	phys_addr_t phys;
138	size_t len;
139	};
140
141	#define DIRTY_BITMAP_BYTES(n) (ALIGN(n, BITS_PER_TYPE(u64)) / BITS_PER_BYTE)
142
143	/*
144	* Input argument of number of bits to bitmap_set() is unsigned integer, which
145	* further casts to signed integer for unaligned multi-bit operation,
146	* __bitmap_set().
147	* Then maximum bitmap size supported is 2^31 bits divided by 2^3 bits/byte,
148	* that is 2^28 (256 MB) which maps to 2^31 * 2^12 = 2^43 (8TB) on 4K page
149	* system.
150	*/
151	#define DIRTY_BITMAP_PAGES_MAX ((u64)INT_MAX)
152	#define DIRTY_BITMAP_SIZE_MAX DIRTY_BITMAP_BYTES(DIRTY_BITMAP_PAGES_MAX)
153
154	static int put_pfn(unsigned long pfn, int prot);
155
156	static struct vfio_iommu_group*
157	vfio_iommu_find_iommu_group(struct vfio_iommu *iommu,
158	struct iommu_group *iommu_group);
159
160	/*
161	* This code handles mapping and unmapping of user data buffers
162	* into DMA'ble space using the IOMMU
163	*/
164
165	static struct vfio_dma *vfio_find_dma(struct vfio_iommu *iommu,
166	dma_addr_t start, size_t size)
167	{
168	struct rb_node *node = iommu->dma_list.rb_node;
169
170	while (node) {
171	struct vfio_dma *dma = rb_entry(node, struct vfio_dma, node);
172
173	if (start + size <= dma->iova)
174	node = node->rb_left;
175	else if (start >= dma->iova + dma->size)
176	node = node->rb_right;
177	else
178	return dma;
179	}
180
181	return NULL;
182	}
183
184	static struct rb_node *vfio_find_dma_first_node(struct vfio_iommu *iommu,
185	dma_addr_t start, u64 size)
186	{
187	struct rb_node *res = NULL;
188	struct rb_node *node = iommu->dma_list.rb_node;
189	struct vfio_dma *dma_res = NULL;
190
191	while (node) {
192	struct vfio_dma *dma = rb_entry(node, struct vfio_dma, node);
193
194	if (start < dma->iova + dma->size) {
195	res = node;
196	dma_res = dma;
197	if (start >= dma->iova)
198	break;
199	node = node->rb_left;
200	} else {
201	node = node->rb_right;
202	}
203	}
204	if (res && size && dma_res->iova >= start + size)
205	res = NULL;
206	return res;
207	}
208
209	static void vfio_link_dma(struct vfio_iommu *iommu, struct vfio_dma *new)
210	{
211	struct rb_node **link = &iommu->dma_list.rb_node, *parent = NULL;
212	struct vfio_dma *dma;
213
214	while (*link) {
215	parent = *link;
216	dma = rb_entry(parent, struct vfio_dma, node);
217
218	if (new->iova + new->size <= dma->iova)
219	link = &(*link)->rb_left;
220	else
221	link = &(*link)->rb_right;
222	}
223
224	rb_link_node(node: &new->node, parent, rb_link: link);
225	rb_insert_color(&new->node, &iommu->dma_list);
226	}
227
228	static void vfio_unlink_dma(struct vfio_iommu *iommu, struct vfio_dma *old)
229	{
230	rb_erase(&old->node, &iommu->dma_list);
231	}
232
233
234	static int vfio_dma_bitmap_alloc(struct vfio_dma *dma, size_t pgsize)
235	{
236	uint64_t npages = dma->size / pgsize;
237
238	if (npages > DIRTY_BITMAP_PAGES_MAX)
239	return -EINVAL;
240
241	/*
242	* Allocate extra 64 bits that are used to calculate shift required for
243	* bitmap_shift_left() to manipulate and club unaligned number of pages
244	* in adjacent vfio_dma ranges.
245	*/
246	dma->bitmap = kvzalloc(DIRTY_BITMAP_BYTES(npages) + sizeof(u64),
247	GFP_KERNEL);
248	if (!dma->bitmap)
249	return -ENOMEM;
250
251	return 0;
252	}
253
254	static void vfio_dma_bitmap_free(struct vfio_dma *dma)
255	{
256	kvfree(addr: dma->bitmap);
257	dma->bitmap = NULL;
258	}
259
260	static void vfio_dma_populate_bitmap(struct vfio_dma *dma, size_t pgsize)
261	{
262	struct rb_node *p;
263	unsigned long pgshift = __ffs(pgsize);
264
265	for (p = rb_first(&dma->pfn_list); p; p = rb_next(p)) {
266	struct vfio_pfn *vpfn = rb_entry(p, struct vfio_pfn, node);
267
268	bitmap_set(map: dma->bitmap, start: (vpfn->iova - dma->iova) >> pgshift, nbits: 1);
269	}
270	}
271
272	static void vfio_iommu_populate_bitmap_full(struct vfio_iommu *iommu)
273	{
274	struct rb_node *n;
275	unsigned long pgshift = __ffs(iommu->pgsize_bitmap);
276
277	for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
278	struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
279
280	bitmap_set(map: dma->bitmap, start: 0, nbits: dma->size >> pgshift);
281	}
282	}
283
284	static int vfio_dma_bitmap_alloc_all(struct vfio_iommu *iommu, size_t pgsize)
285	{
286	struct rb_node *n;
287
288	for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
289	struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
290	int ret;
291
292	ret = vfio_dma_bitmap_alloc(dma, pgsize);
293	if (ret) {
294	struct rb_node *p;
295
296	for (p = rb_prev(n); p; p = rb_prev(p)) {
297	struct vfio_dma *dma = rb_entry(n,
298	struct vfio_dma, node);
299
300	vfio_dma_bitmap_free(dma);
301	}
302	return ret;
303	}
304	vfio_dma_populate_bitmap(dma, pgsize);
305	}
306	return 0;
307	}
308
309	static void vfio_dma_bitmap_free_all(struct vfio_iommu *iommu)
310	{
311	struct rb_node *n;
312
313	for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
314	struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
315
316	vfio_dma_bitmap_free(dma);
317	}
318	}
319
320	/*
321	* Helper Functions for host iova-pfn list
322	*/
323	static struct vfio_pfn *vfio_find_vpfn(struct vfio_dma *dma, dma_addr_t iova)
324	{
325	struct vfio_pfn *vpfn;
326	struct rb_node *node = dma->pfn_list.rb_node;
327
328	while (node) {
329	vpfn = rb_entry(node, struct vfio_pfn, node);
330
331	if (iova < vpfn->iova)
332	node = node->rb_left;
333	else if (iova > vpfn->iova)
334	node = node->rb_right;
335	else
336	return vpfn;
337	}
338	return NULL;
339	}
340
341	static void vfio_link_pfn(struct vfio_dma *dma,
342	struct vfio_pfn *new)
343	{
344	struct rb_node **link, *parent = NULL;
345	struct vfio_pfn *vpfn;
346
347	link = &dma->pfn_list.rb_node;
348	while (*link) {
349	parent = *link;
350	vpfn = rb_entry(parent, struct vfio_pfn, node);
351
352	if (new->iova < vpfn->iova)
353	link = &(*link)->rb_left;
354	else
355	link = &(*link)->rb_right;
356	}
357
358	rb_link_node(node: &new->node, parent, rb_link: link);
359	rb_insert_color(&new->node, &dma->pfn_list);
360	}
361
362	static void vfio_unlink_pfn(struct vfio_dma *dma, struct vfio_pfn *old)
363	{
364	rb_erase(&old->node, &dma->pfn_list);
365	}
366
367	static int vfio_add_to_pfn_list(struct vfio_dma *dma, dma_addr_t iova,
368	unsigned long pfn)
369	{
370	struct vfio_pfn *vpfn;
371
372	vpfn = kzalloc(size: sizeof(*vpfn), GFP_KERNEL);
373	if (!vpfn)
374	return -ENOMEM;
375
376	vpfn->iova = iova;
377	vpfn->pfn = pfn;
378	vpfn->ref_count = 1;
379	vfio_link_pfn(dma, new: vpfn);
380	return 0;
381	}
382
383	static void vfio_remove_from_pfn_list(struct vfio_dma *dma,
384	struct vfio_pfn *vpfn)
385	{
386	vfio_unlink_pfn(dma, old: vpfn);
387	kfree(objp: vpfn);
388	}
389
390	static struct vfio_pfn *vfio_iova_get_vfio_pfn(struct vfio_dma *dma,
391	unsigned long iova)
392	{
393	struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova);
394
395	if (vpfn)
396	vpfn->ref_count++;
397	return vpfn;
398	}
399
400	static int vfio_iova_put_vfio_pfn(struct vfio_dma *dma, struct vfio_pfn *vpfn)
401	{
402	int ret = 0;
403
404	vpfn->ref_count--;
405	if (!vpfn->ref_count) {
406	ret = put_pfn(pfn: vpfn->pfn, prot: dma->prot);
407	vfio_remove_from_pfn_list(dma, vpfn);
408	}
409	return ret;
410	}
411
412	static int mm_lock_acct(struct task_struct *task, struct mm_struct *mm,
413	bool lock_cap, long npage)
414	{
415	int ret = mmap_write_lock_killable(mm);
416
417	if (ret)
418	return ret;
419
420	ret = __account_locked_vm(mm, abs(npage), inc: npage > 0, task, bypass_rlim: lock_cap);
421	mmap_write_unlock(mm);
422	return ret;
423	}
424
425	static int vfio_lock_acct(struct vfio_dma *dma, long npage, bool async)
426	{
427	struct mm_struct *mm;
428	int ret;
429
430	if (!npage)
431	return 0;
432
433	mm = dma->mm;
434	if (async && !mmget_not_zero(mm))
435	return -ESRCH; /* process exited */
436
437	ret = mm_lock_acct(task: dma->task, mm, lock_cap: dma->lock_cap, npage);
438	if (!ret)
439	dma->locked_vm += npage;
440
441	if (async)
442	mmput(mm);
443
444	return ret;
445	}
446
447	/*
448	* Some mappings aren't backed by a struct page, for example an mmap'd
449	* MMIO range for our own or another device. These use a different
450	* pfn conversion and shouldn't be tracked as locked pages.
451	* For compound pages, any driver that sets the reserved bit in head
452	* page needs to set the reserved bit in all subpages to be safe.
453	*/
454	static bool is_invalid_reserved_pfn(unsigned long pfn)
455	{
456	if (pfn_valid(pfn))
457	return PageReserved(pfn_to_page(pfn));
458
459	return true;
460	}
461
462	static int put_pfn(unsigned long pfn, int prot)
463	{
464	if (!is_invalid_reserved_pfn(pfn)) {
465	struct page *page = pfn_to_page(pfn);
466
467	unpin_user_pages_dirty_lock(pages: &page, npages: 1, make_dirty: prot & IOMMU_WRITE);
468	return 1;
469	}
470	return 0;
471	}
472
473	#define VFIO_BATCH_MAX_CAPACITY (PAGE_SIZE / sizeof(struct page *))
474
475	static void vfio_batch_init(struct vfio_batch *batch)
476	{
477	batch->size = 0;
478	batch->offset = 0;
479
480	if (unlikely(disable_hugepages))
481	goto fallback;
482
483	batch->pages = (struct page **) __get_free_page(GFP_KERNEL);
484	if (!batch->pages)
485	goto fallback;
486
487	batch->capacity = VFIO_BATCH_MAX_CAPACITY;
488	return;
489
490	fallback:
491	batch->pages = &batch->fallback_page;
492	batch->capacity = 1;
493	}
494
495	static void vfio_batch_unpin(struct vfio_batch *batch, struct vfio_dma *dma)
496	{
497	while (batch->size) {
498	unsigned long pfn = page_to_pfn(batch->pages[batch->offset]);
499
500	put_pfn(pfn, prot: dma->prot);
501	batch->offset++;
502	batch->size--;
503	}
504	}
505
506	static void vfio_batch_fini(struct vfio_batch *batch)
507	{
508	if (batch->capacity == VFIO_BATCH_MAX_CAPACITY)
509	free_page((unsigned long)batch->pages);
510	}
511
512	static int follow_fault_pfn(struct vm_area_struct *vma, struct mm_struct *mm,
513	unsigned long vaddr, unsigned long *pfn,
514	bool write_fault)
515	{
516	pte_t *ptep;
517	pte_t pte;
518	spinlock_t *ptl;
519	int ret;
520
521	ret = follow_pte(mm: vma->vm_mm, address: vaddr, ptepp: &ptep, ptlp: &ptl);
522	if (ret) {
523	bool unlocked = false;
524
525	ret = fixup_user_fault(mm, address: vaddr,
526	fault_flags: FAULT_FLAG_REMOTE |
527	(write_fault ? FAULT_FLAG_WRITE : 0),
528	unlocked: &unlocked);
529	if (unlocked)
530	return -EAGAIN;
531
532	if (ret)
533	return ret;
534
535	ret = follow_pte(mm: vma->vm_mm, address: vaddr, ptepp: &ptep, ptlp: &ptl);
536	if (ret)
537	return ret;
538	}
539
540	pte = ptep_get(ptep);
541
542	if (write_fault && !pte_write(pte))
543	ret = -EFAULT;
544	else
545	*pfn = pte_pfn(pte);
546
547	pte_unmap_unlock(ptep, ptl);
548	return ret;
549	}
550
551	/*
552	* Returns the positive number of pfns successfully obtained or a negative
553	* error code.
554	*/
555	static int vaddr_get_pfns(struct mm_struct *mm, unsigned long vaddr,
556	long npages, int prot, unsigned long *pfn,
557	struct page **pages)
558	{
559	struct vm_area_struct *vma;
560	unsigned int flags = 0;
561	int ret;
562
563	if (prot & IOMMU_WRITE)
564	flags |= FOLL_WRITE;
565
566	mmap_read_lock(mm);
567	ret = pin_user_pages_remote(mm, start: vaddr, nr_pages: npages, gup_flags: flags | FOLL_LONGTERM,
568	pages, NULL);
569	if (ret > 0) {
570	*pfn = page_to_pfn(pages[0]);
571	goto done;
572	}
573
574	vaddr = untagged_addr_remote(mm, vaddr);
575
576	retry:
577	vma = vma_lookup(mm, addr: vaddr);
578
579	if (vma && vma->vm_flags & VM_PFNMAP) {
580	ret = follow_fault_pfn(vma, mm, vaddr, pfn, write_fault: prot & IOMMU_WRITE);
581	if (ret == -EAGAIN)
582	goto retry;
583
584	if (!ret) {
585	if (is_invalid_reserved_pfn(pfn: *pfn))
586	ret = 1;
587	else
588	ret = -EFAULT;
589	}
590	}
591	done:
592	mmap_read_unlock(mm);
593	return ret;
594	}
595
596	/*
597	* Attempt to pin pages. We really don't want to track all the pfns and
598	* the iommu can only map chunks of consecutive pfns anyway, so get the
599	* first page and all consecutive pages with the same locking.
600	*/
601	static long vfio_pin_pages_remote(struct vfio_dma *dma, unsigned long vaddr,
602	long npage, unsigned long *pfn_base,
603	unsigned long limit, struct vfio_batch *batch)
604	{
605	unsigned long pfn;
606	struct mm_struct *mm = current->mm;
607	long ret, pinned = 0, lock_acct = 0;
608	bool rsvd;
609	dma_addr_t iova = vaddr - dma->vaddr + dma->iova;
610
611	/* This code path is only user initiated */
612	if (!mm)
613	return -ENODEV;
614
615	if (batch->size) {
616	/* Leftover pages in batch from an earlier call. */
617	*pfn_base = page_to_pfn(batch->pages[batch->offset]);
618	pfn = *pfn_base;
619	rsvd = is_invalid_reserved_pfn(pfn: *pfn_base);
620	} else {
621	*pfn_base = 0;
622	}
623
624	while (npage) {
625	if (!batch->size) {
626	/* Empty batch, so refill it. */
627	long req_pages = min_t(long, npage, batch->capacity);
628
629	ret = vaddr_get_pfns(mm, vaddr, npages: req_pages, prot: dma->prot,
630	pfn: &pfn, pages: batch->pages);
631	if (ret < 0)
632	goto unpin_out;
633
634	batch->size = ret;
635	batch->offset = 0;
636
637	if (!*pfn_base) {
638	*pfn_base = pfn;
639	rsvd = is_invalid_reserved_pfn(pfn: *pfn_base);
640	}
641	}
642
643	/*
644	* pfn is preset for the first iteration of this inner loop and
645	* updated at the end to handle a VM_PFNMAP pfn. In that case,
646	* batch->pages isn't valid (there's no struct page), so allow
647	* batch->pages to be touched only when there's more than one
648	* pfn to check, which guarantees the pfns are from a
649	* !VM_PFNMAP vma.
650	*/
651	while (true) {
652	if (pfn != *pfn_base + pinned ||
653	rsvd != is_invalid_reserved_pfn(pfn))
654	goto out;
655
656	/*
657	* Reserved pages aren't counted against the user,
658	* externally pinned pages are already counted against
659	* the user.
660	*/
661	if (!rsvd && !vfio_find_vpfn(dma, iova)) {
662	if (!dma->lock_cap &&
663	mm->locked_vm + lock_acct + 1 > limit) {
664	pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n",
665	__func__, limit << PAGE_SHIFT);
666	ret = -ENOMEM;
667	goto unpin_out;
668	}
669	lock_acct++;
670	}
671
672	pinned++;
673	npage--;
674	vaddr += PAGE_SIZE;
675	iova += PAGE_SIZE;
676	batch->offset++;
677	batch->size--;
678
679	if (!batch->size)
680	break;
681
682	pfn = page_to_pfn(batch->pages[batch->offset]);
683	}
684
685	if (unlikely(disable_hugepages))
686	break;
687	}
688
689	out:
690	ret = vfio_lock_acct(dma, npage: lock_acct, async: false);
691
692	unpin_out:
693	if (batch->size == 1 && !batch->offset) {
694	/* May be a VM_PFNMAP pfn, which the batch can't remember. */
695	put_pfn(pfn, prot: dma->prot);
696	batch->size = 0;
697	}
698
699	if (ret < 0) {
700	if (pinned && !rsvd) {
701	for (pfn = *pfn_base ; pinned ; pfn++, pinned--)
702	put_pfn(pfn, prot: dma->prot);
703	}
704	vfio_batch_unpin(batch, dma);
705
706	return ret;
707	}
708
709	return pinned;
710	}
711
712	static long vfio_unpin_pages_remote(struct vfio_dma *dma, dma_addr_t iova,
713	unsigned long pfn, long npage,
714	bool do_accounting)
715	{
716	long unlocked = 0, locked = 0;
717	long i;
718
719	for (i = 0; i < npage; i++, iova += PAGE_SIZE) {
720	if (put_pfn(pfn: pfn++, prot: dma->prot)) {
721	unlocked++;
722	if (vfio_find_vpfn(dma, iova))
723	locked++;
724	}
725	}
726
727	if (do_accounting)
728	vfio_lock_acct(dma, npage: locked - unlocked, async: true);
729
730	return unlocked;
731	}
732
733	static int vfio_pin_page_external(struct vfio_dma *dma, unsigned long vaddr,
734	unsigned long *pfn_base, bool do_accounting)
735	{
736	struct page *pages[1];
737	struct mm_struct *mm;
738	int ret;
739
740	mm = dma->mm;
741	if (!mmget_not_zero(mm))
742	return -ENODEV;
743
744	ret = vaddr_get_pfns(mm, vaddr, npages: 1, prot: dma->prot, pfn: pfn_base, pages);
745	if (ret != 1)
746	goto out;
747
748	ret = 0;
749
750	if (do_accounting && !is_invalid_reserved_pfn(pfn: *pfn_base)) {
751	ret = vfio_lock_acct(dma, npage: 1, async: false);
752	if (ret) {
753	put_pfn(pfn: *pfn_base, prot: dma->prot);
754	if (ret == -ENOMEM)
755	pr_warn("%s: Task %s (%d) RLIMIT_MEMLOCK "
756	"(%ld) exceeded\n", __func__,
757	dma->task->comm, task_pid_nr(dma->task),
758	task_rlimit(dma->task, RLIMIT_MEMLOCK));
759	}
760	}
761
762	out:
763	mmput(mm);
764	return ret;
765	}
766
767	static int vfio_unpin_page_external(struct vfio_dma *dma, dma_addr_t iova,
768	bool do_accounting)
769	{
770	int unlocked;
771	struct vfio_pfn *vpfn = vfio_find_vpfn(dma, iova);
772
773	if (!vpfn)
774	return 0;
775
776	unlocked = vfio_iova_put_vfio_pfn(dma, vpfn);
777
778	if (do_accounting)
779	vfio_lock_acct(dma, npage: -unlocked, async: true);
780
781	return unlocked;
782	}
783
784	static int vfio_iommu_type1_pin_pages(void *iommu_data,
785	struct iommu_group *iommu_group,
786	dma_addr_t user_iova,
787	int npage, int prot,
788	struct page **pages)
789	{
790	struct vfio_iommu *iommu = iommu_data;
791	struct vfio_iommu_group *group;
792	int i, j, ret;
793	unsigned long remote_vaddr;
794	struct vfio_dma *dma;
795	bool do_accounting;
796
797	if (!iommu || !pages)
798	return -EINVAL;
799
800	/* Supported for v2 version only */
801	if (!iommu->v2)
802	return -EACCES;
803
804	mutex_lock(&iommu->lock);
805
806	if (WARN_ONCE(iommu->vaddr_invalid_count,
807	"vfio_pin_pages not allowed with VFIO_UPDATE_VADDR\n")) {
808	ret = -EBUSY;
809	goto pin_done;
810	}
811
812	/* Fail if no dma_umap notifier is registered */
813	if (list_empty(head: &iommu->device_list)) {
814	ret = -EINVAL;
815	goto pin_done;
816	}
817
818	/*
819	* If iommu capable domain exist in the container then all pages are
820	* already pinned and accounted. Accounting should be done if there is no
821	* iommu capable domain in the container.
822	*/
823	do_accounting = list_empty(head: &iommu->domain_list);
824
825	for (i = 0; i < npage; i++) {
826	unsigned long phys_pfn;
827	dma_addr_t iova;
828	struct vfio_pfn *vpfn;
829
830	iova = user_iova + PAGE_SIZE * i;
831	dma = vfio_find_dma(iommu, start: iova, PAGE_SIZE);
832	if (!dma) {
833	ret = -EINVAL;
834	goto pin_unwind;
835	}
836
837	if ((dma->prot & prot) != prot) {
838	ret = -EPERM;
839	goto pin_unwind;
840	}
841
842	vpfn = vfio_iova_get_vfio_pfn(dma, iova);
843	if (vpfn) {
844	pages[i] = pfn_to_page(vpfn->pfn);
845	continue;
846	}
847
848	remote_vaddr = dma->vaddr + (iova - dma->iova);
849	ret = vfio_pin_page_external(dma, vaddr: remote_vaddr, pfn_base: &phys_pfn,
850	do_accounting);
851	if (ret)
852	goto pin_unwind;
853
854	if (!pfn_valid(pfn: phys_pfn)) {
855	ret = -EINVAL;
856	goto pin_unwind;
857	}
858
859	ret = vfio_add_to_pfn_list(dma, iova, pfn: phys_pfn);
860	if (ret) {
861	if (put_pfn(pfn: phys_pfn, prot: dma->prot) && do_accounting)
862	vfio_lock_acct(dma, npage: -1, async: true);
863	goto pin_unwind;
864	}
865
866	pages[i] = pfn_to_page(phys_pfn);
867
868	if (iommu->dirty_page_tracking) {
869	unsigned long pgshift = __ffs(iommu->pgsize_bitmap);
870
871	/*
872	* Bitmap populated with the smallest supported page
873	* size
874	*/
875	bitmap_set(map: dma->bitmap,
876	start: (iova - dma->iova) >> pgshift, nbits: 1);
877	}
878	}
879	ret = i;
880
881	group = vfio_iommu_find_iommu_group(iommu, iommu_group);
882	if (!group->pinned_page_dirty_scope) {
883	group->pinned_page_dirty_scope = true;
884	iommu->num_non_pinned_groups--;
885	}
886
887	goto pin_done;
888
889	pin_unwind:
890	pages[i] = NULL;
891	for (j = 0; j < i; j++) {
892	dma_addr_t iova;
893
894	iova = user_iova + PAGE_SIZE * j;
895	dma = vfio_find_dma(iommu, start: iova, PAGE_SIZE);
896	vfio_unpin_page_external(dma, iova, do_accounting);
897	pages[j] = NULL;
898	}
899	pin_done:
900	mutex_unlock(lock: &iommu->lock);
901	return ret;
902	}
903
904	static void vfio_iommu_type1_unpin_pages(void *iommu_data,
905	dma_addr_t user_iova, int npage)
906	{
907	struct vfio_iommu *iommu = iommu_data;
908	bool do_accounting;
909	int i;
910
911	/* Supported for v2 version only */
912	if (WARN_ON(!iommu->v2))
913	return;
914
915	mutex_lock(&iommu->lock);
916
917	do_accounting = list_empty(head: &iommu->domain_list);
918	for (i = 0; i < npage; i++) {
919	dma_addr_t iova = user_iova + PAGE_SIZE * i;
920	struct vfio_dma *dma;
921
922	dma = vfio_find_dma(iommu, start: iova, PAGE_SIZE);
923	if (!dma)
924	break;
925
926	vfio_unpin_page_external(dma, iova, do_accounting);
927	}
928
929	mutex_unlock(lock: &iommu->lock);
930
931	WARN_ON(i != npage);
932	}
933
934	static long vfio_sync_unpin(struct vfio_dma *dma, struct vfio_domain *domain,
935	struct list_head *regions,
936	struct iommu_iotlb_gather *iotlb_gather)
937	{
938	long unlocked = 0;
939	struct vfio_regions *entry, *next;
940
941	iommu_iotlb_sync(domain: domain->domain, iotlb_gather);
942
943	list_for_each_entry_safe(entry, next, regions, list) {
944	unlocked += vfio_unpin_pages_remote(dma,
945	iova: entry->iova,
946	pfn: entry->phys >> PAGE_SHIFT,
947	npage: entry->len >> PAGE_SHIFT,
948	do_accounting: false);
949	list_del(entry: &entry->list);
950	kfree(objp: entry);
951	}
952
953	cond_resched();
954
955	return unlocked;
956	}
957
958	/*
959	* Generally, VFIO needs to unpin remote pages after each IOTLB flush.
960	* Therefore, when using IOTLB flush sync interface, VFIO need to keep track
961	* of these regions (currently using a list).
962	*
963	* This value specifies maximum number of regions for each IOTLB flush sync.
964	*/
965	#define VFIO_IOMMU_TLB_SYNC_MAX 512
966
967	static size_t unmap_unpin_fast(struct vfio_domain *domain,
968	struct vfio_dma *dma, dma_addr_t *iova,
969	size_t len, phys_addr_t phys, long *unlocked,
970	struct list_head *unmapped_list,
971	int *unmapped_cnt,
972	struct iommu_iotlb_gather *iotlb_gather)
973	{
974	size_t unmapped = 0;
975	struct vfio_regions *entry = kzalloc(size: sizeof(*entry), GFP_KERNEL);
976
977	if (entry) {
978	unmapped = iommu_unmap_fast(domain: domain->domain, iova: *iova, size: len,
979	iotlb_gather);
980
981	if (!unmapped) {
982	kfree(objp: entry);
983	} else {
984	entry->iova = *iova;
985	entry->phys = phys;
986	entry->len = unmapped;
987	list_add_tail(new: &entry->list, head: unmapped_list);
988
989	*iova += unmapped;
990	(*unmapped_cnt)++;
991	}
992	}
993
994	/*
995	* Sync if the number of fast-unmap regions hits the limit
996	* or in case of errors.
997	*/
998	if (*unmapped_cnt >= VFIO_IOMMU_TLB_SYNC_MAX || !unmapped) {
999	*unlocked += vfio_sync_unpin(dma, domain, regions: unmapped_list,
1000	iotlb_gather);
1001	*unmapped_cnt = 0;
1002	}
1003
1004	return unmapped;
1005	}
1006
1007	static size_t unmap_unpin_slow(struct vfio_domain *domain,
1008	struct vfio_dma *dma, dma_addr_t *iova,
1009	size_t len, phys_addr_t phys,
1010	long *unlocked)
1011	{
1012	size_t unmapped = iommu_unmap(domain: domain->domain, iova: *iova, size: len);
1013
1014	if (unmapped) {
1015	*unlocked += vfio_unpin_pages_remote(dma, iova: *iova,
1016	pfn: phys >> PAGE_SHIFT,
1017	npage: unmapped >> PAGE_SHIFT,
1018	do_accounting: false);
1019	*iova += unmapped;
1020	cond_resched();
1021	}
1022	return unmapped;
1023	}
1024
1025	static long vfio_unmap_unpin(struct vfio_iommu *iommu, struct vfio_dma *dma,
1026	bool do_accounting)
1027	{
1028	dma_addr_t iova = dma->iova, end = dma->iova + dma->size;
1029	struct vfio_domain *domain, *d;
1030	LIST_HEAD(unmapped_region_list);
1031	struct iommu_iotlb_gather iotlb_gather;
1032	int unmapped_region_cnt = 0;
1033	long unlocked = 0;
1034
1035	if (!dma->size)
1036	return 0;
1037
1038	if (list_empty(head: &iommu->domain_list))
1039	return 0;
1040
1041	/*
1042	* We use the IOMMU to track the physical addresses, otherwise we'd
1043	* need a much more complicated tracking system. Unfortunately that
1044	* means we need to use one of the iommu domains to figure out the
1045	* pfns to unpin. The rest need to be unmapped in advance so we have
1046	* no iommu translations remaining when the pages are unpinned.
1047	*/
1048	domain = d = list_first_entry(&iommu->domain_list,
1049	struct vfio_domain, next);
1050
1051	list_for_each_entry_continue(d, &iommu->domain_list, next) {
1052	iommu_unmap(domain: d->domain, iova: dma->iova, size: dma->size);
1053	cond_resched();
1054	}
1055
1056	iommu_iotlb_gather_init(gather: &iotlb_gather);
1057	while (iova < end) {
1058	size_t unmapped, len;
1059	phys_addr_t phys, next;
1060
1061	phys = iommu_iova_to_phys(domain: domain->domain, iova);
1062	if (WARN_ON(!phys)) {
1063	iova += PAGE_SIZE;
1064	continue;
1065	}
1066
1067	/*
1068	* To optimize for fewer iommu_unmap() calls, each of which
1069	* may require hardware cache flushing, try to find the
1070	* largest contiguous physical memory chunk to unmap.
1071	*/
1072	for (len = PAGE_SIZE;
1073	!domain->fgsp && iova + len < end; len += PAGE_SIZE) {
1074	next = iommu_iova_to_phys(domain: domain->domain, iova: iova + len);
1075	if (next != phys + len)
1076	break;
1077	}
1078
1079	/*
1080	* First, try to use fast unmap/unpin. In case of failure,
1081	* switch to slow unmap/unpin path.
1082	*/
1083	unmapped = unmap_unpin_fast(domain, dma, iova: &iova, len, phys,
1084	unlocked: &unlocked, unmapped_list: &unmapped_region_list,
1085	unmapped_cnt: &unmapped_region_cnt,
1086	iotlb_gather: &iotlb_gather);
1087	if (!unmapped) {
1088	unmapped = unmap_unpin_slow(domain, dma, iova: &iova, len,
1089	phys, unlocked: &unlocked);
1090	if (WARN_ON(!unmapped))
1091	break;
1092	}
1093	}
1094
1095	dma->iommu_mapped = false;
1096
1097	if (unmapped_region_cnt) {
1098	unlocked += vfio_sync_unpin(dma, domain, regions: &unmapped_region_list,
1099	iotlb_gather: &iotlb_gather);
1100	}
1101
1102	if (do_accounting) {
1103	vfio_lock_acct(dma, npage: -unlocked, async: true);
1104	return 0;
1105	}
1106	return unlocked;
1107	}
1108
1109	static void vfio_remove_dma(struct vfio_iommu *iommu, struct vfio_dma *dma)
1110	{
1111	WARN_ON(!RB_EMPTY_ROOT(&dma->pfn_list));
1112	vfio_unmap_unpin(iommu, dma, do_accounting: true);
1113	vfio_unlink_dma(iommu, old: dma);
1114	put_task_struct(t: dma->task);
1115	mmdrop(mm: dma->mm);
1116	vfio_dma_bitmap_free(dma);
1117	if (dma->vaddr_invalid)
1118	iommu->vaddr_invalid_count--;
1119	kfree(objp: dma);
1120	iommu->dma_avail++;
1121	}
1122
1123	static void vfio_update_pgsize_bitmap(struct vfio_iommu *iommu)
1124	{
1125	struct vfio_domain *domain;
1126
1127	iommu->pgsize_bitmap = ULONG_MAX;
1128
1129	list_for_each_entry(domain, &iommu->domain_list, next)
1130	iommu->pgsize_bitmap &= domain->domain->pgsize_bitmap;
1131
1132	/*
1133	* In case the IOMMU supports page sizes smaller than PAGE_SIZE
1134	* we pretend PAGE_SIZE is supported and hide sub-PAGE_SIZE sizes.
1135	* That way the user will be able to map/unmap buffers whose size/
1136	* start address is aligned with PAGE_SIZE. Pinning code uses that
1137	* granularity while iommu driver can use the sub-PAGE_SIZE size
1138	* to map the buffer.
1139	*/
1140	if (iommu->pgsize_bitmap & ~PAGE_MASK) {
1141	iommu->pgsize_bitmap &= PAGE_MASK;
1142	iommu->pgsize_bitmap |= PAGE_SIZE;
1143	}
1144	}
1145
1146	static int update_user_bitmap(u64 __user *bitmap, struct vfio_iommu *iommu,
1147	struct vfio_dma *dma, dma_addr_t base_iova,
1148	size_t pgsize)
1149	{
1150	unsigned long pgshift = __ffs(pgsize);
1151	unsigned long nbits = dma->size >> pgshift;
1152	unsigned long bit_offset = (dma->iova - base_iova) >> pgshift;
1153	unsigned long copy_offset = bit_offset / BITS_PER_LONG;
1154	unsigned long shift = bit_offset % BITS_PER_LONG;
1155	unsigned long leftover;
1156
1157	/*
1158	* mark all pages dirty if any IOMMU capable device is not able
1159	* to report dirty pages and all pages are pinned and mapped.
1160	*/
1161	if (iommu->num_non_pinned_groups && dma->iommu_mapped)
1162	bitmap_set(map: dma->bitmap, start: 0, nbits);
1163
1164	if (shift) {
1165	bitmap_shift_left(dst: dma->bitmap, src: dma->bitmap, shift,
1166	nbits: nbits + shift);
1167
1168	if (copy_from_user(to: &leftover,
1169	from: (void __user *)(bitmap + copy_offset),
1170	n: sizeof(leftover)))
1171	return -EFAULT;
1172
1173	bitmap_or(dst: dma->bitmap, src1: dma->bitmap, src2: &leftover, nbits: shift);
1174	}
1175
1176	if (copy_to_user(to: (void __user *)(bitmap + copy_offset), from: dma->bitmap,
1177	DIRTY_BITMAP_BYTES(nbits + shift)))
1178	return -EFAULT;
1179
1180	return 0;
1181	}
1182
1183	static int vfio_iova_dirty_bitmap(u64 __user *bitmap, struct vfio_iommu *iommu,
1184	dma_addr_t iova, size_t size, size_t pgsize)
1185	{
1186	struct vfio_dma *dma;
1187	struct rb_node *n;
1188	unsigned long pgshift = __ffs(pgsize);
1189	int ret;
1190
1191	/*
1192	* GET_BITMAP request must fully cover vfio_dma mappings. Multiple
1193	* vfio_dma mappings may be clubbed by specifying large ranges, but
1194	* there must not be any previous mappings bisected by the range.
1195	* An error will be returned if these conditions are not met.
1196	*/
1197	dma = vfio_find_dma(iommu, start: iova, size: 1);
1198	if (dma && dma->iova != iova)
1199	return -EINVAL;
1200
1201	dma = vfio_find_dma(iommu, start: iova + size - 1, size: 0);
1202	if (dma && dma->iova + dma->size != iova + size)
1203	return -EINVAL;
1204
1205	for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
1206	struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
1207
1208	if (dma->iova < iova)
1209	continue;
1210
1211	if (dma->iova > iova + size - 1)
1212	break;
1213
1214	ret = update_user_bitmap(bitmap, iommu, dma, base_iova: iova, pgsize);
1215	if (ret)
1216	return ret;
1217
1218	/*
1219	* Re-populate bitmap to include all pinned pages which are
1220	* considered as dirty but exclude pages which are unpinned and
1221	* pages which are marked dirty by vfio_dma_rw()
1222	*/
1223	bitmap_clear(map: dma->bitmap, start: 0, nbits: dma->size >> pgshift);
1224	vfio_dma_populate_bitmap(dma, pgsize);
1225	}
1226	return 0;
1227	}
1228
1229	static int verify_bitmap_size(uint64_t npages, uint64_t bitmap_size)
1230	{
1231	if (!npages || !bitmap_size || (bitmap_size > DIRTY_BITMAP_SIZE_MAX) ||
1232	(bitmap_size < DIRTY_BITMAP_BYTES(npages)))
1233	return -EINVAL;
1234
1235	return 0;
1236	}
1237
1238	/*
1239	* Notify VFIO drivers using vfio_register_emulated_iommu_dev() to invalidate
1240	* and unmap iovas within the range we're about to unmap. Drivers MUST unpin
1241	* pages in response to an invalidation.
1242	*/
1243	static void vfio_notify_dma_unmap(struct vfio_iommu *iommu,
1244	struct vfio_dma *dma)
1245	{
1246	struct vfio_device *device;
1247
1248	if (list_empty(head: &iommu->device_list))
1249	return;
1250
1251	/*
1252	* The device is expected to call vfio_unpin_pages() for any IOVA it has
1253	* pinned within the range. Since vfio_unpin_pages() will eventually
1254	* call back down to this code and try to obtain the iommu->lock we must
1255	* drop it.
1256	*/
1257	mutex_lock(&iommu->device_list_lock);
1258	mutex_unlock(lock: &iommu->lock);
1259
1260	list_for_each_entry(device, &iommu->device_list, iommu_entry)
1261	device->ops->dma_unmap(device, dma->iova, dma->size);
1262
1263	mutex_unlock(lock: &iommu->device_list_lock);
1264	mutex_lock(&iommu->lock);
1265	}
1266
1267	static int vfio_dma_do_unmap(struct vfio_iommu *iommu,
1268	struct vfio_iommu_type1_dma_unmap *unmap,
1269	struct vfio_bitmap *bitmap)
1270	{
1271	struct vfio_dma *dma, *dma_last = NULL;
1272	size_t unmapped = 0, pgsize;
1273	int ret = -EINVAL, retries = 0;
1274	unsigned long pgshift;
1275	dma_addr_t iova = unmap->iova;
1276	u64 size = unmap->size;
1277	bool unmap_all = unmap->flags & VFIO_DMA_UNMAP_FLAG_ALL;
1278	bool invalidate_vaddr = unmap->flags & VFIO_DMA_UNMAP_FLAG_VADDR;
1279	struct rb_node *n, *first_n;
1280
1281	mutex_lock(&iommu->lock);
1282
1283	/* Cannot update vaddr if mdev is present. */
1284	if (invalidate_vaddr && !list_empty(head: &iommu->emulated_iommu_groups)) {
1285	ret = -EBUSY;
1286	goto unlock;
1287	}
1288
1289	pgshift = __ffs(iommu->pgsize_bitmap);
1290	pgsize = (size_t)1 << pgshift;
1291
1292	if (iova & (pgsize - 1))
1293	goto unlock;
1294
1295	if (unmap_all) {
1296	if (iova || size)
1297	goto unlock;
1298	size = U64_MAX;
1299	} else if (!size || size & (pgsize - 1) ||
1300	iova + size - 1 < iova || size > SIZE_MAX) {
1301	goto unlock;
1302	}
1303
1304	/* When dirty tracking is enabled, allow only min supported pgsize */
1305	if ((unmap->flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) &&
1306	(!iommu->dirty_page_tracking || (bitmap->pgsize != pgsize))) {
1307	goto unlock;
1308	}
1309
1310	WARN_ON((pgsize - 1) & PAGE_MASK);
1311	again:
1312	/*
1313	* vfio-iommu-type1 (v1) - User mappings were coalesced together to
1314	* avoid tracking individual mappings. This means that the granularity
1315	* of the original mapping was lost and the user was allowed to attempt
1316	* to unmap any range. Depending on the contiguousness of physical
1317	* memory and page sizes supported by the IOMMU, arbitrary unmaps may
1318	* or may not have worked. We only guaranteed unmap granularity
1319	* matching the original mapping; even though it was untracked here,
1320	* the original mappings are reflected in IOMMU mappings. This
1321	* resulted in a couple unusual behaviors. First, if a range is not
1322	* able to be unmapped, ex. a set of 4k pages that was mapped as a
1323	* 2M hugepage into the IOMMU, the unmap ioctl returns success but with
1324	* a zero sized unmap. Also, if an unmap request overlaps the first
1325	* address of a hugepage, the IOMMU will unmap the entire hugepage.
1326	* This also returns success and the returned unmap size reflects the
1327	* actual size unmapped.
1328	*
1329	* We attempt to maintain compatibility with this "v1" interface, but
1330	* we take control out of the hands of the IOMMU. Therefore, an unmap
1331	* request offset from the beginning of the original mapping will
1332	* return success with zero sized unmap. And an unmap request covering
1333	* the first iova of mapping will unmap the entire range.
1334	*
1335	* The v2 version of this interface intends to be more deterministic.
1336	* Unmap requests must fully cover previous mappings. Multiple
1337	* mappings may still be unmaped by specifying large ranges, but there
1338	* must not be any previous mappings bisected by the range. An error
1339	* will be returned if these conditions are not met. The v2 interface
1340	* will only return success and a size of zero if there were no
1341	* mappings within the range.
1342	*/
1343	if (iommu->v2 && !unmap_all) {
1344	dma = vfio_find_dma(iommu, start: iova, size: 1);
1345	if (dma && dma->iova != iova)
1346	goto unlock;
1347
1348	dma = vfio_find_dma(iommu, start: iova + size - 1, size: 0);
1349	if (dma && dma->iova + dma->size != iova + size)
1350	goto unlock;
1351	}
1352
1353	ret = 0;
1354	n = first_n = vfio_find_dma_first_node(iommu, start: iova, size);
1355
1356	while (n) {
1357	dma = rb_entry(n, struct vfio_dma, node);
1358	if (dma->iova >= iova + size)
1359	break;
1360
1361	if (!iommu->v2 && iova > dma->iova)
1362	break;
1363
1364	if (invalidate_vaddr) {
1365	if (dma->vaddr_invalid) {
1366	struct rb_node *last_n = n;
1367
1368	for (n = first_n; n != last_n; n = rb_next(n)) {
1369	dma = rb_entry(n,
1370	struct vfio_dma, node);
1371	dma->vaddr_invalid = false;
1372	iommu->vaddr_invalid_count--;
1373	}
1374	ret = -EINVAL;
1375	unmapped = 0;
1376	break;
1377	}
1378	dma->vaddr_invalid = true;
1379	iommu->vaddr_invalid_count++;
1380	unmapped += dma->size;
1381	n = rb_next(n);
1382	continue;
1383	}
1384
1385	if (!RB_EMPTY_ROOT(&dma->pfn_list)) {
1386	if (dma_last == dma) {
1387	BUG_ON(++retries > 10);
1388	} else {
1389	dma_last = dma;
1390	retries = 0;
1391	}
1392
1393	vfio_notify_dma_unmap(iommu, dma);
1394	goto again;
1395	}
1396
1397	if (unmap->flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) {
1398	ret = update_user_bitmap(bitmap: bitmap->data, iommu, dma,
1399	base_iova: iova, pgsize);
1400	if (ret)
1401	break;
1402	}
1403
1404	unmapped += dma->size;
1405	n = rb_next(n);
1406	vfio_remove_dma(iommu, dma);
1407	}
1408
1409	unlock:
1410	mutex_unlock(lock: &iommu->lock);
1411
1412	/* Report how much was unmapped */
1413	unmap->size = unmapped;
1414
1415	return ret;
1416	}
1417
1418	static int vfio_iommu_map(struct vfio_iommu *iommu, dma_addr_t iova,
1419	unsigned long pfn, long npage, int prot)
1420	{
1421	struct vfio_domain *d;
1422	int ret;
1423
1424	list_for_each_entry(d, &iommu->domain_list, next) {
1425	ret = iommu_map(domain: d->domain, iova, paddr: (phys_addr_t)pfn << PAGE_SHIFT,
1426	size: npage << PAGE_SHIFT, prot: prot | IOMMU_CACHE,
1427	GFP_KERNEL_ACCOUNT);
1428	if (ret)
1429	goto unwind;
1430
1431	cond_resched();
1432	}
1433
1434	return 0;
1435
1436	unwind:
1437	list_for_each_entry_continue_reverse(d, &iommu->domain_list, next) {
1438	iommu_unmap(domain: d->domain, iova, size: npage << PAGE_SHIFT);
1439	cond_resched();
1440	}
1441
1442	return ret;
1443	}
1444
1445	static int vfio_pin_map_dma(struct vfio_iommu *iommu, struct vfio_dma *dma,
1446	size_t map_size)
1447	{
1448	dma_addr_t iova = dma->iova;
1449	unsigned long vaddr = dma->vaddr;
1450	struct vfio_batch batch;
1451	size_t size = map_size;
1452	long npage;
1453	unsigned long pfn, limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
1454	int ret = 0;
1455
1456	vfio_batch_init(batch: &batch);
1457
1458	while (size) {
1459	/* Pin a contiguous chunk of memory */
1460	npage = vfio_pin_pages_remote(dma, vaddr: vaddr + dma->size,
1461	npage: size >> PAGE_SHIFT, pfn_base: &pfn, limit,
1462	batch: &batch);
1463	if (npage <= 0) {
1464	WARN_ON(!npage);
1465	ret = (int)npage;
1466	break;
1467	}
1468
1469	/* Map it! */
1470	ret = vfio_iommu_map(iommu, iova: iova + dma->size, pfn, npage,
1471	prot: dma->prot);
1472	if (ret) {
1473	vfio_unpin_pages_remote(dma, iova: iova + dma->size, pfn,
1474	npage, do_accounting: true);
1475	vfio_batch_unpin(batch: &batch, dma);
1476	break;
1477	}
1478
1479	size -= npage << PAGE_SHIFT;
1480	dma->size += npage << PAGE_SHIFT;
1481	}
1482
1483	vfio_batch_fini(batch: &batch);
1484	dma->iommu_mapped = true;
1485
1486	if (ret)
1487	vfio_remove_dma(iommu, dma);
1488
1489	return ret;
1490	}
1491
1492	/*
1493	* Check dma map request is within a valid iova range
1494	*/
1495	static bool vfio_iommu_iova_dma_valid(struct vfio_iommu *iommu,
1496	dma_addr_t start, dma_addr_t end)
1497	{
1498	struct list_head *iova = &iommu->iova_list;
1499	struct vfio_iova *node;
1500
1501	list_for_each_entry(node, iova, list) {
1502	if (start >= node->start && end <= node->end)
1503	return true;
1504	}
1505
1506	/*
1507	* Check for list_empty() as well since a container with
1508	* a single mdev device will have an empty list.
1509	*/
1510	return list_empty(head: iova);
1511	}
1512
1513	static int vfio_change_dma_owner(struct vfio_dma *dma)
1514	{
1515	struct task_struct *task = current->group_leader;
1516	struct mm_struct *mm = current->mm;
1517	long npage = dma->locked_vm;
1518	bool lock_cap;
1519	int ret;
1520
1521	if (mm == dma->mm)
1522	return 0;
1523
1524	lock_cap = capable(CAP_IPC_LOCK);
1525	ret = mm_lock_acct(task, mm, lock_cap, npage);
1526	if (ret)
1527	return ret;
1528
1529	if (mmget_not_zero(mm: dma->mm)) {
1530	mm_lock_acct(task: dma->task, mm: dma->mm, lock_cap: dma->lock_cap, npage: -npage);
1531	mmput(dma->mm);
1532	}
1533
1534	if (dma->task != task) {
1535	put_task_struct(t: dma->task);
1536	dma->task = get_task_struct(t: task);
1537	}
1538	mmdrop(mm: dma->mm);
1539	dma->mm = mm;
1540	mmgrab(mm: dma->mm);
1541	dma->lock_cap = lock_cap;
1542	return 0;
1543	}
1544
1545	static int vfio_dma_do_map(struct vfio_iommu *iommu,
1546	struct vfio_iommu_type1_dma_map *map)
1547	{
1548	bool set_vaddr = map->flags & VFIO_DMA_MAP_FLAG_VADDR;
1549	dma_addr_t iova = map->iova;
1550	unsigned long vaddr = map->vaddr;
1551	size_t size = map->size;
1552	int ret = 0, prot = 0;
1553	size_t pgsize;
1554	struct vfio_dma *dma;
1555
1556	/* Verify that none of our __u64 fields overflow */
1557	if (map->size != size || map->vaddr != vaddr || map->iova != iova)
1558	return -EINVAL;
1559
1560	/* READ/WRITE from device perspective */
1561	if (map->flags & VFIO_DMA_MAP_FLAG_WRITE)
1562	prot |= IOMMU_WRITE;
1563	if (map->flags & VFIO_DMA_MAP_FLAG_READ)
1564	prot |= IOMMU_READ;
1565
1566	if ((prot && set_vaddr) || (!prot && !set_vaddr))
1567	return -EINVAL;
1568
1569	mutex_lock(&iommu->lock);
1570
1571	pgsize = (size_t)1 << __ffs(iommu->pgsize_bitmap);
1572
1573	WARN_ON((pgsize - 1) & PAGE_MASK);
1574
1575	if (!size || (size | iova | vaddr) & (pgsize - 1)) {
1576	ret = -EINVAL;
1577	goto out_unlock;
1578	}
1579
1580	/* Don't allow IOVA or virtual address wrap */
1581	if (iova + size - 1 < iova || vaddr + size - 1 < vaddr) {
1582	ret = -EINVAL;
1583	goto out_unlock;
1584	}
1585
1586	dma = vfio_find_dma(iommu, start: iova, size);
1587	if (set_vaddr) {
1588	if (!dma) {
1589	ret = -ENOENT;
1590	} else if (!dma->vaddr_invalid || dma->iova != iova ||
1591	dma->size != size) {
1592	ret = -EINVAL;
1593	} else {
1594	ret = vfio_change_dma_owner(dma);
1595	if (ret)
1596	goto out_unlock;
1597	dma->vaddr = vaddr;
1598	dma->vaddr_invalid = false;
1599	iommu->vaddr_invalid_count--;
1600	}
1601	goto out_unlock;
1602	} else if (dma) {
1603	ret = -EEXIST;
1604	goto out_unlock;
1605	}
1606
1607	if (!iommu->dma_avail) {
1608	ret = -ENOSPC;
1609	goto out_unlock;
1610	}
1611
1612	if (!vfio_iommu_iova_dma_valid(iommu, start: iova, end: iova + size - 1)) {
1613	ret = -EINVAL;
1614	goto out_unlock;
1615	}
1616
1617	dma = kzalloc(size: sizeof(*dma), GFP_KERNEL);
1618	if (!dma) {
1619	ret = -ENOMEM;
1620	goto out_unlock;
1621	}
1622
1623	iommu->dma_avail--;
1624	dma->iova = iova;
1625	dma->vaddr = vaddr;
1626	dma->prot = prot;
1627
1628	/*
1629	* We need to be able to both add to a task's locked memory and test
1630	* against the locked memory limit and we need to be able to do both
1631	* outside of this call path as pinning can be asynchronous via the
1632	* external interfaces for mdev devices. RLIMIT_MEMLOCK requires a
1633	* task_struct. Save the group_leader so that all DMA tracking uses
1634	* the same task, to make debugging easier. VM locked pages requires
1635	* an mm_struct, so grab the mm in case the task dies.
1636	*/
1637	get_task_struct(current->group_leader);
1638	dma->task = current->group_leader;
1639	dma->lock_cap = capable(CAP_IPC_LOCK);
1640	dma->mm = current->mm;
1641	mmgrab(mm: dma->mm);
1642
1643	dma->pfn_list = RB_ROOT;
1644
1645	/* Insert zero-sized and grow as we map chunks of it */
1646	vfio_link_dma(iommu, new: dma);
1647
1648	/* Don't pin and map if container doesn't contain IOMMU capable domain*/
1649	if (list_empty(head: &iommu->domain_list))
1650	dma->size = size;
1651	else
1652	ret = vfio_pin_map_dma(iommu, dma, map_size: size);
1653
1654	if (!ret && iommu->dirty_page_tracking) {
1655	ret = vfio_dma_bitmap_alloc(dma, pgsize);
1656	if (ret)
1657	vfio_remove_dma(iommu, dma);
1658	}
1659
1660	out_unlock:
1661	mutex_unlock(lock: &iommu->lock);
1662	return ret;
1663	}
1664
1665	static int vfio_iommu_replay(struct vfio_iommu *iommu,
1666	struct vfio_domain *domain)
1667	{
1668	struct vfio_batch batch;
1669	struct vfio_domain *d = NULL;
1670	struct rb_node *n;
1671	unsigned long limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
1672	int ret;
1673
1674	/* Arbitrarily pick the first domain in the list for lookups */
1675	if (!list_empty(head: &iommu->domain_list))
1676	d = list_first_entry(&iommu->domain_list,
1677	struct vfio_domain, next);
1678
1679	vfio_batch_init(batch: &batch);
1680
1681	n = rb_first(&iommu->dma_list);
1682
1683	for (; n; n = rb_next(n)) {
1684	struct vfio_dma *dma;
1685	dma_addr_t iova;
1686
1687	dma = rb_entry(n, struct vfio_dma, node);
1688	iova = dma->iova;
1689
1690	while (iova < dma->iova + dma->size) {
1691	phys_addr_t phys;
1692	size_t size;
1693
1694	if (dma->iommu_mapped) {
1695	phys_addr_t p;
1696	dma_addr_t i;
1697
1698	if (WARN_ON(!d)) { /* mapped w/o a domain?! */
1699	ret = -EINVAL;
1700	goto unwind;
1701	}
1702
1703	phys = iommu_iova_to_phys(domain: d->domain, iova);
1704
1705	if (WARN_ON(!phys)) {
1706	iova += PAGE_SIZE;
1707	continue;
1708	}
1709
1710	size = PAGE_SIZE;
1711	p = phys + size;
1712	i = iova + size;
1713	while (i < dma->iova + dma->size &&
1714	p == iommu_iova_to_phys(domain: d->domain, iova: i)) {
1715	size += PAGE_SIZE;
1716	p += PAGE_SIZE;
1717	i += PAGE_SIZE;
1718	}
1719	} else {
1720	unsigned long pfn;
1721	unsigned long vaddr = dma->vaddr +
1722	(iova - dma->iova);
1723	size_t n = dma->iova + dma->size - iova;
1724	long npage;
1725
1726	npage = vfio_pin_pages_remote(dma, vaddr,
1727	npage: n >> PAGE_SHIFT,
1728	pfn_base: &pfn, limit,
1729	batch: &batch);
1730	if (npage <= 0) {
1731	WARN_ON(!npage);
1732	ret = (int)npage;
1733	goto unwind;
1734	}
1735
1736	phys = pfn << PAGE_SHIFT;
1737	size = npage << PAGE_SHIFT;
1738	}
1739
1740	ret = iommu_map(domain: domain->domain, iova, paddr: phys, size,
1741	prot: dma->prot | IOMMU_CACHE,
1742	GFP_KERNEL_ACCOUNT);
1743	if (ret) {
1744	if (!dma->iommu_mapped) {
1745	vfio_unpin_pages_remote(dma, iova,
1746	pfn: phys >> PAGE_SHIFT,
1747	npage: size >> PAGE_SHIFT,
1748	do_accounting: true);
1749	vfio_batch_unpin(batch: &batch, dma);
1750	}
1751	goto unwind;
1752	}
1753
1754	iova += size;
1755	}
1756	}
1757
1758	/* All dmas are now mapped, defer to second tree walk for unwind */
1759	for (n = rb_first(&iommu->dma_list); n; n = rb_next(n)) {
1760	struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
1761
1762	dma->iommu_mapped = true;
1763	}
1764
1765	vfio_batch_fini(batch: &batch);
1766	return 0;
1767
1768	unwind:
1769	for (; n; n = rb_prev(n)) {
1770	struct vfio_dma *dma = rb_entry(n, struct vfio_dma, node);
1771	dma_addr_t iova;
1772
1773	if (dma->iommu_mapped) {
1774	iommu_unmap(domain: domain->domain, iova: dma->iova, size: dma->size);
1775	continue;
1776	}
1777
1778	iova = dma->iova;
1779	while (iova < dma->iova + dma->size) {
1780	phys_addr_t phys, p;
1781	size_t size;
1782	dma_addr_t i;
1783
1784	phys = iommu_iova_to_phys(domain: domain->domain, iova);
1785	if (!phys) {
1786	iova += PAGE_SIZE;
1787	continue;
1788	}
1789
1790	size = PAGE_SIZE;
1791	p = phys + size;
1792	i = iova + size;
1793	while (i < dma->iova + dma->size &&
1794	p == iommu_iova_to_phys(domain: domain->domain, iova: i)) {
1795	size += PAGE_SIZE;
1796	p += PAGE_SIZE;
1797	i += PAGE_SIZE;
1798	}
1799
1800	iommu_unmap(domain: domain->domain, iova, size);
1801	vfio_unpin_pages_remote(dma, iova, pfn: phys >> PAGE_SHIFT,
1802	npage: size >> PAGE_SHIFT, do_accounting: true);
1803	}
1804	}
1805
1806	vfio_batch_fini(batch: &batch);
1807	return ret;
1808	}
1809
1810	/*
1811	* We change our unmap behavior slightly depending on whether the IOMMU
1812	* supports fine-grained superpages. IOMMUs like AMD-Vi will use a superpage
1813	* for practically any contiguous power-of-two mapping we give it. This means
1814	* we don't need to look for contiguous chunks ourselves to make unmapping
1815	* more efficient. On IOMMUs with coarse-grained super pages, like Intel VT-d
1816	* with discrete 2M/1G/512G/1T superpages, identifying contiguous chunks
1817	* significantly boosts non-hugetlbfs mappings and doesn't seem to hurt when
1818	* hugetlbfs is in use.
1819	*/
1820	static void vfio_test_domain_fgsp(struct vfio_domain *domain, struct list_head *regions)
1821	{
1822	int ret, order = get_order(PAGE_SIZE * 2);
1823	struct vfio_iova *region;
1824	struct page *pages;
1825	dma_addr_t start;
1826
1827	pages = alloc_pages(GFP_KERNEL | __GFP_ZERO, order);
1828	if (!pages)
1829	return;
1830
1831	list_for_each_entry(region, regions, list) {
1832	start = ALIGN(region->start, PAGE_SIZE * 2);
1833	if (start >= region->end || (region->end - start < PAGE_SIZE * 2))
1834	continue;
1835
1836	ret = iommu_map(domain: domain->domain, iova: start, page_to_phys(pages), PAGE_SIZE * 2,
1837	IOMMU_READ | IOMMU_WRITE | IOMMU_CACHE,
1838	GFP_KERNEL_ACCOUNT);
1839	if (!ret) {
1840	size_t unmapped = iommu_unmap(domain: domain->domain, iova: start, PAGE_SIZE);
1841
1842	if (unmapped == PAGE_SIZE)
1843	iommu_unmap(domain: domain->domain, iova: start + PAGE_SIZE, PAGE_SIZE);
1844	else
1845	domain->fgsp = true;
1846	}
1847	break;
1848	}
1849
1850	__free_pages(page: pages, order);
1851	}
1852
1853	static struct vfio_iommu_group *find_iommu_group(struct vfio_domain *domain,
1854	struct iommu_group *iommu_group)
1855	{
1856	struct vfio_iommu_group *g;
1857
1858	list_for_each_entry(g, &domain->group_list, next) {
1859	if (g->iommu_group == iommu_group)
1860	return g;
1861	}
1862
1863	return NULL;
1864	}
1865
1866	static struct vfio_iommu_group*
1867	vfio_iommu_find_iommu_group(struct vfio_iommu *iommu,
1868	struct iommu_group *iommu_group)
1869	{
1870	struct vfio_iommu_group *group;
1871	struct vfio_domain *domain;
1872
1873	list_for_each_entry(domain, &iommu->domain_list, next) {
1874	group = find_iommu_group(domain, iommu_group);
1875	if (group)
1876	return group;
1877	}
1878
1879	list_for_each_entry(group, &iommu->emulated_iommu_groups, next)
1880	if (group->iommu_group == iommu_group)
1881	return group;
1882	return NULL;
1883	}
1884
1885	static bool vfio_iommu_has_sw_msi(struct list_head *group_resv_regions,
1886	phys_addr_t *base)
1887	{
1888	struct iommu_resv_region *region;
1889	bool ret = false;
1890
1891	list_for_each_entry(region, group_resv_regions, list) {
1892	/*
1893	* The presence of any 'real' MSI regions should take
1894	* precedence over the software-managed one if the
1895	* IOMMU driver happens to advertise both types.
1896	*/
1897	if (region->type == IOMMU_RESV_MSI) {
1898	ret = false;
1899	break;
1900	}
1901
1902	if (region->type == IOMMU_RESV_SW_MSI) {
1903	*base = region->start;
1904	ret = true;
1905	}
1906	}
1907
1908	return ret;
1909	}
1910
1911	/*
1912	* This is a helper function to insert an address range to iova list.
1913	* The list is initially created with a single entry corresponding to
1914	* the IOMMU domain geometry to which the device group is attached.
1915	* The list aperture gets modified when a new domain is added to the
1916	* container if the new aperture doesn't conflict with the current one
1917	* or with any existing dma mappings. The list is also modified to
1918	* exclude any reserved regions associated with the device group.
1919	*/
1920	static int vfio_iommu_iova_insert(struct list_head *head,
1921	dma_addr_t start, dma_addr_t end)
1922	{
1923	struct vfio_iova *region;
1924
1925	region = kmalloc(size: sizeof(*region), GFP_KERNEL);
1926	if (!region)
1927	return -ENOMEM;
1928
1929	INIT_LIST_HEAD(list: &region->list);
1930	region->start = start;
1931	region->end = end;
1932
1933	list_add_tail(new: &region->list, head);
1934	return 0;
1935	}
1936
1937	/*
1938	* Check the new iommu aperture conflicts with existing aper or with any
1939	* existing dma mappings.
1940	*/
1941	static bool vfio_iommu_aper_conflict(struct vfio_iommu *iommu,
1942	dma_addr_t start, dma_addr_t end)
1943	{
1944	struct vfio_iova *first, *last;
1945	struct list_head *iova = &iommu->iova_list;
1946
1947	if (list_empty(head: iova))
1948	return false;
1949
1950	/* Disjoint sets, return conflict */
1951	first = list_first_entry(iova, struct vfio_iova, list);
1952	last = list_last_entry(iova, struct vfio_iova, list);
1953	if (start > last->end || end < first->start)
1954	return true;
1955
1956	/* Check for any existing dma mappings below the new start */
1957	if (start > first->start) {
1958	if (vfio_find_dma(iommu, start: first->start, size: start - first->start))
1959	return true;
1960	}
1961
1962	/* Check for any existing dma mappings beyond the new end */
1963	if (end < last->end) {
1964	if (vfio_find_dma(iommu, start: end + 1, size: last->end - end))
1965	return true;
1966	}
1967
1968	return false;
1969	}
1970
1971	/*
1972	* Resize iommu iova aperture window. This is called only if the new
1973	* aperture has no conflict with existing aperture and dma mappings.
1974	*/
1975	static int vfio_iommu_aper_resize(struct list_head *iova,
1976	dma_addr_t start, dma_addr_t end)
1977	{
1978	struct vfio_iova *node, *next;
1979
1980	if (list_empty(head: iova))
1981	return vfio_iommu_iova_insert(head: iova, start, end);
1982
1983	/* Adjust iova list start */
1984	list_for_each_entry_safe(node, next, iova, list) {
1985	if (start < node->start)
1986	break;
1987	if (start >= node->start && start < node->end) {
1988	node->start = start;
1989	break;
1990	}
1991	/* Delete nodes before new start */
1992	list_del(entry: &node->list);
1993	kfree(objp: node);
1994	}
1995
1996	/* Adjust iova list end */
1997	list_for_each_entry_safe(node, next, iova, list) {
1998	if (end > node->end)
1999	continue;
2000	if (end > node->start && end <= node->end) {
2001	node->end = end;
2002	continue;
2003	}
2004	/* Delete nodes after new end */
2005	list_del(entry: &node->list);
2006	kfree(objp: node);
2007	}
2008
2009	return 0;
2010	}
2011
2012	/*
2013	* Check reserved region conflicts with existing dma mappings
2014	*/
2015	static bool vfio_iommu_resv_conflict(struct vfio_iommu *iommu,
2016	struct list_head *resv_regions)
2017	{
2018	struct iommu_resv_region *region;
2019
2020	/* Check for conflict with existing dma mappings */
2021	list_for_each_entry(region, resv_regions, list) {
2022	if (region->type == IOMMU_RESV_DIRECT_RELAXABLE)
2023	continue;
2024
2025	if (vfio_find_dma(iommu, start: region->start, size: region->length))
2026	return true;
2027	}
2028
2029	return false;
2030	}
2031
2032	/*
2033	* Check iova region overlap with reserved regions and
2034	* exclude them from the iommu iova range
2035	*/
2036	static int vfio_iommu_resv_exclude(struct list_head *iova,
2037	struct list_head *resv_regions)
2038	{
2039	struct iommu_resv_region *resv;
2040	struct vfio_iova *n, *next;
2041
2042	list_for_each_entry(resv, resv_regions, list) {
2043	phys_addr_t start, end;
2044
2045	if (resv->type == IOMMU_RESV_DIRECT_RELAXABLE)
2046	continue;
2047
2048	start = resv->start;
2049	end = resv->start + resv->length - 1;
2050
2051	list_for_each_entry_safe(n, next, iova, list) {
2052	int ret = 0;
2053
2054	/* No overlap */
2055	if (start > n->end || end < n->start)
2056	continue;
2057	/*
2058	* Insert a new node if current node overlaps with the
2059	* reserve region to exclude that from valid iova range.
2060	* Note that, new node is inserted before the current
2061	* node and finally the current node is deleted keeping
2062	* the list updated and sorted.
2063	*/
2064	if (start > n->start)
2065	ret = vfio_iommu_iova_insert(head: &n->list, start: n->start,
2066	end: start - 1);
2067	if (!ret && end < n->end)
2068	ret = vfio_iommu_iova_insert(head: &n->list, start: end + 1,
2069	end: n->end);
2070	if (ret)
2071	return ret;
2072
2073	list_del(entry: &n->list);
2074	kfree(objp: n);
2075	}
2076	}
2077
2078	if (list_empty(head: iova))
2079	return -EINVAL;
2080
2081	return 0;
2082	}
2083
2084	static void vfio_iommu_resv_free(struct list_head *resv_regions)
2085	{
2086	struct iommu_resv_region *n, *next;
2087
2088	list_for_each_entry_safe(n, next, resv_regions, list) {
2089	list_del(entry: &n->list);
2090	kfree(objp: n);
2091	}
2092	}
2093
2094	static void vfio_iommu_iova_free(struct list_head *iova)
2095	{
2096	struct vfio_iova *n, *next;
2097
2098	list_for_each_entry_safe(n, next, iova, list) {
2099	list_del(entry: &n->list);
2100	kfree(objp: n);
2101	}
2102	}
2103
2104	static int vfio_iommu_iova_get_copy(struct vfio_iommu *iommu,
2105	struct list_head *iova_copy)
2106	{
2107	struct list_head *iova = &iommu->iova_list;
2108	struct vfio_iova *n;
2109	int ret;
2110
2111	list_for_each_entry(n, iova, list) {
2112	ret = vfio_iommu_iova_insert(head: iova_copy, start: n->start, end: n->end);
2113	if (ret)
2114	goto out_free;
2115	}
2116
2117	return 0;
2118
2119	out_free:
2120	vfio_iommu_iova_free(iova: iova_copy);
2121	return ret;
2122	}
2123
2124	static void vfio_iommu_iova_insert_copy(struct vfio_iommu *iommu,
2125	struct list_head *iova_copy)
2126	{
2127	struct list_head *iova = &iommu->iova_list;
2128
2129	vfio_iommu_iova_free(iova);
2130
2131	list_splice_tail(list: iova_copy, head: iova);
2132	}
2133
2134	static int vfio_iommu_domain_alloc(struct device *dev, void *data)
2135	{
2136	struct iommu_domain **domain = data;
2137
2138	*domain = iommu_domain_alloc(bus: dev->bus);
2139	return 1; /* Don't iterate */
2140	}
2141
2142	static int vfio_iommu_type1_attach_group(void *iommu_data,
2143	struct iommu_group *iommu_group, enum vfio_group_type type)
2144	{
2145	struct vfio_iommu *iommu = iommu_data;
2146	struct vfio_iommu_group *group;
2147	struct vfio_domain *domain, *d;
2148	bool resv_msi;
2149	phys_addr_t resv_msi_base = 0;
2150	struct iommu_domain_geometry *geo;
2151	LIST_HEAD(iova_copy);
2152	LIST_HEAD(group_resv_regions);
2153	int ret = -EBUSY;
2154
2155	mutex_lock(&iommu->lock);
2156
2157	/* Attach could require pinning, so disallow while vaddr is invalid. */
2158	if (iommu->vaddr_invalid_count)
2159	goto out_unlock;
2160
2161	/* Check for duplicates */
2162	ret = -EINVAL;
2163	if (vfio_iommu_find_iommu_group(iommu, iommu_group))
2164	goto out_unlock;
2165
2166	ret = -ENOMEM;
2167	group = kzalloc(size: sizeof(*group), GFP_KERNEL);
2168	if (!group)
2169	goto out_unlock;
2170	group->iommu_group = iommu_group;
2171
2172	if (type == VFIO_EMULATED_IOMMU) {
2173	list_add(new: &group->next, head: &iommu->emulated_iommu_groups);
2174	/*
2175	* An emulated IOMMU group cannot dirty memory directly, it can
2176	* only use interfaces that provide dirty tracking.
2177	* The iommu scope can only be promoted with the addition of a
2178	* dirty tracking group.
2179	*/
2180	group->pinned_page_dirty_scope = true;
2181	ret = 0;
2182	goto out_unlock;
2183	}
2184
2185	ret = -ENOMEM;
2186	domain = kzalloc(size: sizeof(*domain), GFP_KERNEL);
2187	if (!domain)
2188	goto out_free_group;
2189
2190	/*
2191	* Going via the iommu_group iterator avoids races, and trivially gives
2192	* us a representative device for the IOMMU API call. We don't actually
2193	* want to iterate beyond the first device (if any).
2194	*/
2195	ret = -EIO;
2196	iommu_group_for_each_dev(group: iommu_group, data: &domain->domain,
2197	fn: vfio_iommu_domain_alloc);
2198	if (!domain->domain)
2199	goto out_free_domain;
2200
2201	if (iommu->nesting) {
2202	ret = iommu_enable_nesting(domain: domain->domain);
2203	if (ret)
2204	goto out_domain;
2205	}
2206
2207	ret = iommu_attach_group(domain: domain->domain, group: group->iommu_group);
2208	if (ret)
2209	goto out_domain;
2210
2211	/* Get aperture info */
2212	geo = &domain->domain->geometry;
2213	if (vfio_iommu_aper_conflict(iommu, start: geo->aperture_start,
2214	end: geo->aperture_end)) {
2215	ret = -EINVAL;
2216	goto out_detach;
2217	}
2218
2219	ret = iommu_get_group_resv_regions(group: iommu_group, head: &group_resv_regions);
2220	if (ret)
2221	goto out_detach;
2222
2223	if (vfio_iommu_resv_conflict(iommu, resv_regions: &group_resv_regions)) {
2224	ret = -EINVAL;
2225	goto out_detach;
2226	}
2227
2228	/*
2229	* We don't want to work on the original iova list as the list
2230	* gets modified and in case of failure we have to retain the
2231	* original list. Get a copy here.
2232	*/
2233	ret = vfio_iommu_iova_get_copy(iommu, iova_copy: &iova_copy);
2234	if (ret)
2235	goto out_detach;
2236
2237	ret = vfio_iommu_aper_resize(iova: &iova_copy, start: geo->aperture_start,
2238	end: geo->aperture_end);
2239	if (ret)
2240	goto out_detach;
2241
2242	ret = vfio_iommu_resv_exclude(iova: &iova_copy, resv_regions: &group_resv_regions);
2243	if (ret)
2244	goto out_detach;
2245
2246	resv_msi = vfio_iommu_has_sw_msi(group_resv_regions: &group_resv_regions, base: &resv_msi_base);
2247
2248	INIT_LIST_HEAD(list: &domain->group_list);
2249	list_add(new: &group->next, head: &domain->group_list);
2250
2251	if (!allow_unsafe_interrupts &&
2252	!iommu_group_has_isolated_msi(group: iommu_group)) {
2253	pr_warn("%s: No interrupt remapping support. Use the module param \"allow_unsafe_interrupts\" to enable VFIO IOMMU support on this platform\n",
2254	__func__);
2255	ret = -EPERM;
2256	goto out_detach;
2257	}
2258
2259	/*
2260	* If the IOMMU can block non-coherent operations (ie PCIe TLPs with
2261	* no-snoop set) then VFIO always turns this feature on because on Intel
2262	* platforms it optimizes KVM to disable wbinvd emulation.
2263	*/
2264	if (domain->domain->ops->enforce_cache_coherency)
2265	domain->enforce_cache_coherency =
2266	domain->domain->ops->enforce_cache_coherency(
2267	domain->domain);
2268
2269	/*
2270	* Try to match an existing compatible domain. We don't want to
2271	* preclude an IOMMU driver supporting multiple bus_types and being
2272	* able to include different bus_types in the same IOMMU domain, so
2273	* we test whether the domains use the same iommu_ops rather than
2274	* testing if they're on the same bus_type.
2275	*/
2276	list_for_each_entry(d, &iommu->domain_list, next) {
2277	if (d->domain->ops == domain->domain->ops &&
2278	d->enforce_cache_coherency ==
2279	domain->enforce_cache_coherency) {
2280	iommu_detach_group(domain: domain->domain, group: group->iommu_group);
2281	if (!iommu_attach_group(domain: d->domain,
2282	group: group->iommu_group)) {
2283	list_add(new: &group->next, head: &d->group_list);
2284	iommu_domain_free(domain: domain->domain);
2285	kfree(objp: domain);
2286	goto done;
2287	}
2288
2289	ret = iommu_attach_group(domain: domain->domain,
2290	group: group->iommu_group);
2291	if (ret)
2292	goto out_domain;
2293	}
2294	}
2295
2296	vfio_test_domain_fgsp(domain, regions: &iova_copy);
2297
2298	/* replay mappings on new domains */
2299	ret = vfio_iommu_replay(iommu, domain);
2300	if (ret)
2301	goto out_detach;
2302
2303	if (resv_msi) {
2304	ret = iommu_get_msi_cookie(domain: domain->domain, base: resv_msi_base);
2305	if (ret && ret != -ENODEV)
2306	goto out_detach;
2307	}
2308
2309	list_add(new: &domain->next, head: &iommu->domain_list);
2310	vfio_update_pgsize_bitmap(iommu);
2311	done:
2312	/* Delete the old one and insert new iova list */
2313	vfio_iommu_iova_insert_copy(iommu, iova_copy: &iova_copy);
2314
2315	/*
2316	* An iommu backed group can dirty memory directly and therefore
2317	* demotes the iommu scope until it declares itself dirty tracking
2318	* capable via the page pinning interface.
2319	*/
2320	iommu->num_non_pinned_groups++;
2321	mutex_unlock(lock: &iommu->lock);
2322	vfio_iommu_resv_free(resv_regions: &group_resv_regions);
2323
2324	return 0;
2325
2326	out_detach:
2327	iommu_detach_group(domain: domain->domain, group: group->iommu_group);
2328	out_domain:
2329	iommu_domain_free(domain: domain->domain);
2330	vfio_iommu_iova_free(iova: &iova_copy);
2331	vfio_iommu_resv_free(resv_regions: &group_resv_regions);
2332	out_free_domain:
2333	kfree(objp: domain);
2334	out_free_group:
2335	kfree(objp: group);
2336	out_unlock:
2337	mutex_unlock(lock: &iommu->lock);
2338	return ret;
2339	}
2340
2341	static void vfio_iommu_unmap_unpin_all(struct vfio_iommu *iommu)
2342	{
2343	struct rb_node *node;
2344
2345	while ((node = rb_first(&iommu->dma_list)))
2346	vfio_remove_dma(iommu, rb_entry(node, struct vfio_dma, node));
2347	}
2348
2349	static void vfio_iommu_unmap_unpin_reaccount(struct vfio_iommu *iommu)
2350	{
2351	struct rb_node *n, *p;
2352
2353	n = rb_first(&iommu->dma_list);
2354	for (; n; n = rb_next(n)) {
2355	struct vfio_dma *dma;
2356	long locked = 0, unlocked = 0;
2357
2358	dma = rb_entry(n, struct vfio_dma, node);
2359	unlocked += vfio_unmap_unpin(iommu, dma, do_accounting: false);
2360	p = rb_first(&dma->pfn_list);
2361	for (; p; p = rb_next(p)) {
2362	struct vfio_pfn *vpfn = rb_entry(p, struct vfio_pfn,
2363	node);
2364
2365	if (!is_invalid_reserved_pfn(pfn: vpfn->pfn))
2366	locked++;
2367	}
2368	vfio_lock_acct(dma, npage: locked - unlocked, async: true);
2369	}
2370	}
2371
2372	/*
2373	* Called when a domain is removed in detach. It is possible that
2374	* the removed domain decided the iova aperture window. Modify the
2375	* iova aperture with the smallest window among existing domains.
2376	*/
2377	static void vfio_iommu_aper_expand(struct vfio_iommu *iommu,
2378	struct list_head *iova_copy)
2379	{
2380	struct vfio_domain *domain;
2381	struct vfio_iova *node;
2382	dma_addr_t start = 0;
2383	dma_addr_t end = (dma_addr_t)~0;
2384
2385	if (list_empty(head: iova_copy))
2386	return;
2387
2388	list_for_each_entry(domain, &iommu->domain_list, next) {
2389	struct iommu_domain_geometry *geo = &domain->domain->geometry;
2390
2391	if (geo->aperture_start > start)
2392	start = geo->aperture_start;
2393	if (geo->aperture_end < end)
2394	end = geo->aperture_end;
2395	}
2396
2397	/* Modify aperture limits. The new aper is either same or bigger */
2398	node = list_first_entry(iova_copy, struct vfio_iova, list);
2399	node->start = start;
2400	node = list_last_entry(iova_copy, struct vfio_iova, list);
2401	node->end = end;
2402	}
2403
2404	/*
2405	* Called when a group is detached. The reserved regions for that
2406	* group can be part of valid iova now. But since reserved regions
2407	* may be duplicated among groups, populate the iova valid regions
2408	* list again.
2409	*/
2410	static int vfio_iommu_resv_refresh(struct vfio_iommu *iommu,
2411	struct list_head *iova_copy)
2412	{
2413	struct vfio_domain *d;
2414	struct vfio_iommu_group *g;
2415	struct vfio_iova *node;
2416	dma_addr_t start, end;
2417	LIST_HEAD(resv_regions);
2418	int ret;
2419
2420	if (list_empty(head: iova_copy))
2421	return -EINVAL;
2422
2423	list_for_each_entry(d, &iommu->domain_list, next) {
2424	list_for_each_entry(g, &d->group_list, next) {
2425	ret = iommu_get_group_resv_regions(group: g->iommu_group,
2426	head: &resv_regions);
2427	if (ret)
2428	goto done;
2429	}
2430	}
2431
2432	node = list_first_entry(iova_copy, struct vfio_iova, list);
2433	start = node->start;
2434	node = list_last_entry(iova_copy, struct vfio_iova, list);
2435	end = node->end;
2436
2437	/* purge the iova list and create new one */
2438	vfio_iommu_iova_free(iova: iova_copy);
2439
2440	ret = vfio_iommu_aper_resize(iova: iova_copy, start, end);
2441	if (ret)
2442	goto done;
2443
2444	/* Exclude current reserved regions from iova ranges */
2445	ret = vfio_iommu_resv_exclude(iova: iova_copy, resv_regions: &resv_regions);
2446	done:
2447	vfio_iommu_resv_free(resv_regions: &resv_regions);
2448	return ret;
2449	}
2450
2451	static void vfio_iommu_type1_detach_group(void *iommu_data,
2452	struct iommu_group *iommu_group)
2453	{
2454	struct vfio_iommu *iommu = iommu_data;
2455	struct vfio_domain *domain;
2456	struct vfio_iommu_group *group;
2457	bool update_dirty_scope = false;
2458	LIST_HEAD(iova_copy);
2459
2460	mutex_lock(&iommu->lock);
2461	list_for_each_entry(group, &iommu->emulated_iommu_groups, next) {
2462	if (group->iommu_group != iommu_group)
2463	continue;
2464	update_dirty_scope = !group->pinned_page_dirty_scope;
2465	list_del(entry: &group->next);
2466	kfree(objp: group);
2467
2468	if (list_empty(head: &iommu->emulated_iommu_groups) &&
2469	list_empty(head: &iommu->domain_list)) {
2470	WARN_ON(!list_empty(&iommu->device_list));
2471	vfio_iommu_unmap_unpin_all(iommu);
2472	}
2473	goto detach_group_done;
2474	}
2475
2476	/*
2477	* Get a copy of iova list. This will be used to update
2478	* and to replace the current one later. Please note that
2479	* we will leave the original list as it is if update fails.
2480	*/
2481	vfio_iommu_iova_get_copy(iommu, iova_copy: &iova_copy);
2482
2483	list_for_each_entry(domain, &iommu->domain_list, next) {
2484	group = find_iommu_group(domain, iommu_group);
2485	if (!group)
2486	continue;
2487
2488	iommu_detach_group(domain: domain->domain, group: group->iommu_group);
2489	update_dirty_scope = !group->pinned_page_dirty_scope;
2490	list_del(entry: &group->next);
2491	kfree(objp: group);
2492	/*
2493	* Group ownership provides privilege, if the group list is
2494	* empty, the domain goes away. If it's the last domain with
2495	* iommu and external domain doesn't exist, then all the
2496	* mappings go away too. If it's the last domain with iommu and
2497	* external domain exist, update accounting
2498	*/
2499	if (list_empty(head: &domain->group_list)) {
2500	if (list_is_singular(head: &iommu->domain_list)) {
2501	if (list_empty(head: &iommu->emulated_iommu_groups)) {
2502	WARN_ON(!list_empty(
2503	&iommu->device_list));
2504	vfio_iommu_unmap_unpin_all(iommu);
2505	} else {
2506	vfio_iommu_unmap_unpin_reaccount(iommu);
2507	}
2508	}
2509	iommu_domain_free(domain: domain->domain);
2510	list_del(entry: &domain->next);
2511	kfree(objp: domain);
2512	vfio_iommu_aper_expand(iommu, iova_copy: &iova_copy);
2513	vfio_update_pgsize_bitmap(iommu);
2514	}
2515	break;
2516	}
2517
2518	if (!vfio_iommu_resv_refresh(iommu, iova_copy: &iova_copy))
2519	vfio_iommu_iova_insert_copy(iommu, iova_copy: &iova_copy);
2520	else
2521	vfio_iommu_iova_free(iova: &iova_copy);
2522
2523	detach_group_done:
2524	/*
2525	* Removal of a group without dirty tracking may allow the iommu scope
2526	* to be promoted.
2527	*/
2528	if (update_dirty_scope) {
2529	iommu->num_non_pinned_groups--;
2530	if (iommu->dirty_page_tracking)
2531	vfio_iommu_populate_bitmap_full(iommu);
2532	}
2533	mutex_unlock(lock: &iommu->lock);
2534	}
2535
2536	static void *vfio_iommu_type1_open(unsigned long arg)
2537	{
2538	struct vfio_iommu *iommu;
2539
2540	iommu = kzalloc(size: sizeof(*iommu), GFP_KERNEL);
2541	if (!iommu)
2542	return ERR_PTR(error: -ENOMEM);
2543
2544	switch (arg) {
2545	case VFIO_TYPE1_IOMMU:
2546	break;
2547	case VFIO_TYPE1_NESTING_IOMMU:
2548	iommu->nesting = true;
2549	fallthrough;
2550	case VFIO_TYPE1v2_IOMMU:
2551	iommu->v2 = true;
2552	break;
2553	default:
2554	kfree(objp: iommu);
2555	return ERR_PTR(error: -EINVAL);
2556	}
2557
2558	INIT_LIST_HEAD(list: &iommu->domain_list);
2559	INIT_LIST_HEAD(list: &iommu->iova_list);
2560	iommu->dma_list = RB_ROOT;
2561	iommu->dma_avail = dma_entry_limit;
2562	mutex_init(&iommu->lock);
2563	mutex_init(&iommu->device_list_lock);
2564	INIT_LIST_HEAD(list: &iommu->device_list);
2565	iommu->pgsize_bitmap = PAGE_MASK;
2566	INIT_LIST_HEAD(list: &iommu->emulated_iommu_groups);
2567
2568	return iommu;
2569	}
2570
2571	static void vfio_release_domain(struct vfio_domain *domain)
2572	{
2573	struct vfio_iommu_group *group, *group_tmp;
2574
2575	list_for_each_entry_safe(group, group_tmp,
2576	&domain->group_list, next) {
2577	iommu_detach_group(domain: domain->domain, group: group->iommu_group);
2578	list_del(entry: &group->next);
2579	kfree(objp: group);
2580	}
2581
2582	iommu_domain_free(domain: domain->domain);
2583	}
2584
2585	static void vfio_iommu_type1_release(void *iommu_data)
2586	{
2587	struct vfio_iommu *iommu = iommu_data;
2588	struct vfio_domain *domain, *domain_tmp;
2589	struct vfio_iommu_group *group, *next_group;
2590
2591	list_for_each_entry_safe(group, next_group,
2592	&iommu->emulated_iommu_groups, next) {
2593	list_del(entry: &group->next);
2594	kfree(objp: group);
2595	}
2596
2597	vfio_iommu_unmap_unpin_all(iommu);
2598
2599	list_for_each_entry_safe(domain, domain_tmp,
2600	&iommu->domain_list, next) {
2601	vfio_release_domain(domain);
2602	list_del(entry: &domain->next);
2603	kfree(objp: domain);
2604	}
2605
2606	vfio_iommu_iova_free(iova: &iommu->iova_list);
2607
2608	kfree(objp: iommu);
2609	}
2610
2611	static int vfio_domains_have_enforce_cache_coherency(struct vfio_iommu *iommu)
2612	{
2613	struct vfio_domain *domain;
2614	int ret = 1;
2615
2616	mutex_lock(&iommu->lock);
2617	list_for_each_entry(domain, &iommu->domain_list, next) {
2618	if (!(domain->enforce_cache_coherency)) {
2619	ret = 0;
2620	break;
2621	}
2622	}
2623	mutex_unlock(lock: &iommu->lock);
2624
2625	return ret;
2626	}
2627
2628	static bool vfio_iommu_has_emulated(struct vfio_iommu *iommu)
2629	{
2630	bool ret;
2631
2632	mutex_lock(&iommu->lock);
2633	ret = !list_empty(head: &iommu->emulated_iommu_groups);
2634	mutex_unlock(lock: &iommu->lock);
2635	return ret;
2636	}
2637
2638	static int vfio_iommu_type1_check_extension(struct vfio_iommu *iommu,
2639	unsigned long arg)
2640	{
2641	switch (arg) {
2642	case VFIO_TYPE1_IOMMU:
2643	case VFIO_TYPE1v2_IOMMU:
2644	case VFIO_TYPE1_NESTING_IOMMU:
2645	case VFIO_UNMAP_ALL:
2646	return 1;
2647	case VFIO_UPDATE_VADDR:
2648	/*
2649	* Disable this feature if mdevs are present. They cannot
2650	* safely pin/unpin/rw while vaddrs are being updated.
2651	*/
2652	return iommu && !vfio_iommu_has_emulated(iommu);
2653	case VFIO_DMA_CC_IOMMU:
2654	if (!iommu)
2655	return 0;
2656	return vfio_domains_have_enforce_cache_coherency(iommu);
2657	default:
2658	return 0;
2659	}
2660	}
2661
2662	static int vfio_iommu_iova_add_cap(struct vfio_info_cap *caps,
2663	struct vfio_iommu_type1_info_cap_iova_range *cap_iovas,
2664	size_t size)
2665	{
2666	struct vfio_info_cap_header *header;
2667	struct vfio_iommu_type1_info_cap_iova_range *iova_cap;
2668
2669	header = vfio_info_cap_add(caps, size,
2670	VFIO_IOMMU_TYPE1_INFO_CAP_IOVA_RANGE, version: 1);
2671	if (IS_ERR(ptr: header))
2672	return PTR_ERR(ptr: header);
2673
2674	iova_cap = container_of(header,
2675	struct vfio_iommu_type1_info_cap_iova_range,
2676	header);
2677	iova_cap->nr_iovas = cap_iovas->nr_iovas;
2678	memcpy(iova_cap->iova_ranges, cap_iovas->iova_ranges,
2679	cap_iovas->nr_iovas * sizeof(*cap_iovas->iova_ranges));
2680	return 0;
2681	}
2682
2683	static int vfio_iommu_iova_build_caps(struct vfio_iommu *iommu,
2684	struct vfio_info_cap *caps)
2685	{
2686	struct vfio_iommu_type1_info_cap_iova_range *cap_iovas;
2687	struct vfio_iova *iova;
2688	size_t size;
2689	int iovas = 0, i = 0, ret;
2690
2691	list_for_each_entry(iova, &iommu->iova_list, list)
2692	iovas++;
2693
2694	if (!iovas) {
2695	/*
2696	* Return 0 as a container with a single mdev device
2697	* will have an empty list
2698	*/
2699	return 0;
2700	}
2701
2702	size = struct_size(cap_iovas, iova_ranges, iovas);
2703
2704	cap_iovas = kzalloc(size, GFP_KERNEL);
2705	if (!cap_iovas)
2706	return -ENOMEM;
2707
2708	cap_iovas->nr_iovas = iovas;
2709
2710	list_for_each_entry(iova, &iommu->iova_list, list) {
2711	cap_iovas->iova_ranges[i].start = iova->start;
2712	cap_iovas->iova_ranges[i].end = iova->end;
2713	i++;
2714	}
2715
2716	ret = vfio_iommu_iova_add_cap(caps, cap_iovas, size);
2717
2718	kfree(objp: cap_iovas);
2719	return ret;
2720	}
2721
2722	static int vfio_iommu_migration_build_caps(struct vfio_iommu *iommu,
2723	struct vfio_info_cap *caps)
2724	{
2725	struct vfio_iommu_type1_info_cap_migration cap_mig = {};
2726
2727	cap_mig.header.id = VFIO_IOMMU_TYPE1_INFO_CAP_MIGRATION;
2728	cap_mig.header.version = 1;
2729
2730	cap_mig.flags = 0;
2731	/* support minimum pgsize */
2732	cap_mig.pgsize_bitmap = (size_t)1 << __ffs(iommu->pgsize_bitmap);
2733	cap_mig.max_dirty_bitmap_size = DIRTY_BITMAP_SIZE_MAX;
2734
2735	return vfio_info_add_capability(caps, cap: &cap_mig.header, size: sizeof(cap_mig));
2736	}
2737
2738	static int vfio_iommu_dma_avail_build_caps(struct vfio_iommu *iommu,
2739	struct vfio_info_cap *caps)
2740	{
2741	struct vfio_iommu_type1_info_dma_avail cap_dma_avail;
2742
2743	cap_dma_avail.header.id = VFIO_IOMMU_TYPE1_INFO_DMA_AVAIL;
2744	cap_dma_avail.header.version = 1;
2745
2746	cap_dma_avail.avail = iommu->dma_avail;
2747
2748	return vfio_info_add_capability(caps, cap: &cap_dma_avail.header,
2749	size: sizeof(cap_dma_avail));
2750	}
2751
2752	static int vfio_iommu_type1_get_info(struct vfio_iommu *iommu,
2753	unsigned long arg)
2754	{
2755	struct vfio_iommu_type1_info info = {};
2756	unsigned long minsz;
2757	struct vfio_info_cap caps = { .buf = NULL, .size = 0 };
2758	int ret;
2759
2760	minsz = offsetofend(struct vfio_iommu_type1_info, iova_pgsizes);
2761
2762	if (copy_from_user(to: &info, from: (void __user *)arg, n: minsz))
2763	return -EFAULT;
2764
2765	if (info.argsz < minsz)
2766	return -EINVAL;
2767
2768	minsz = min_t(size_t, info.argsz, sizeof(info));
2769
2770	mutex_lock(&iommu->lock);
2771	info.flags = VFIO_IOMMU_INFO_PGSIZES;
2772
2773	info.iova_pgsizes = iommu->pgsize_bitmap;
2774
2775	ret = vfio_iommu_migration_build_caps(iommu, caps: &caps);
2776
2777	if (!ret)
2778	ret = vfio_iommu_dma_avail_build_caps(iommu, caps: &caps);
2779
2780	if (!ret)
2781	ret = vfio_iommu_iova_build_caps(iommu, caps: &caps);
2782
2783	mutex_unlock(lock: &iommu->lock);
2784
2785	if (ret)
2786	return ret;
2787
2788	if (caps.size) {
2789	info.flags |= VFIO_IOMMU_INFO_CAPS;
2790
2791	if (info.argsz < sizeof(info) + caps.size) {
2792	info.argsz = sizeof(info) + caps.size;
2793	} else {
2794	vfio_info_cap_shift(caps: &caps, offset: sizeof(info));
2795	if (copy_to_user(to: (void __user *)arg +
2796	sizeof(info), from: caps.buf,
2797	n: caps.size)) {
2798	kfree(objp: caps.buf);
2799	return -EFAULT;
2800	}
2801	info.cap_offset = sizeof(info);
2802	}
2803
2804	kfree(objp: caps.buf);
2805	}
2806
2807	return copy_to_user(to: (void __user *)arg, from: &info, n: minsz) ?
2808	-EFAULT : 0;
2809	}
2810
2811	static int vfio_iommu_type1_map_dma(struct vfio_iommu *iommu,
2812	unsigned long arg)
2813	{
2814	struct vfio_iommu_type1_dma_map map;
2815	unsigned long minsz;
2816	uint32_t mask = VFIO_DMA_MAP_FLAG_READ | VFIO_DMA_MAP_FLAG_WRITE |
2817	VFIO_DMA_MAP_FLAG_VADDR;
2818
2819	minsz = offsetofend(struct vfio_iommu_type1_dma_map, size);
2820
2821	if (copy_from_user(to: &map, from: (void __user *)arg, n: minsz))
2822	return -EFAULT;
2823
2824	if (map.argsz < minsz || map.flags & ~mask)
2825	return -EINVAL;
2826
2827	return vfio_dma_do_map(iommu, map: &map);
2828	}
2829
2830	static int vfio_iommu_type1_unmap_dma(struct vfio_iommu *iommu,
2831	unsigned long arg)
2832	{
2833	struct vfio_iommu_type1_dma_unmap unmap;
2834	struct vfio_bitmap bitmap = { 0 };
2835	uint32_t mask = VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP |
2836	VFIO_DMA_UNMAP_FLAG_VADDR |
2837	VFIO_DMA_UNMAP_FLAG_ALL;
2838	unsigned long minsz;
2839	int ret;
2840
2841	minsz = offsetofend(struct vfio_iommu_type1_dma_unmap, size);
2842
2843	if (copy_from_user(to: &unmap, from: (void __user *)arg, n: minsz))
2844	return -EFAULT;
2845
2846	if (unmap.argsz < minsz || unmap.flags & ~mask)
2847	return -EINVAL;
2848
2849	if ((unmap.flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) &&
2850	(unmap.flags & (VFIO_DMA_UNMAP_FLAG_ALL |
2851	VFIO_DMA_UNMAP_FLAG_VADDR)))
2852	return -EINVAL;
2853
2854	if (unmap.flags & VFIO_DMA_UNMAP_FLAG_GET_DIRTY_BITMAP) {
2855	unsigned long pgshift;
2856
2857	if (unmap.argsz < (minsz + sizeof(bitmap)))
2858	return -EINVAL;
2859
2860	if (copy_from_user(to: &bitmap,
2861	from: (void __user *)(arg + minsz),
2862	n: sizeof(bitmap)))
2863	return -EFAULT;
2864
2865	if (!access_ok((void __user *)bitmap.data, bitmap.size))
2866	return -EINVAL;
2867
2868	pgshift = __ffs(bitmap.pgsize);
2869	ret = verify_bitmap_size(npages: unmap.size >> pgshift,
2870	bitmap_size: bitmap.size);
2871	if (ret)
2872	return ret;
2873	}
2874
2875	ret = vfio_dma_do_unmap(iommu, unmap: &unmap, bitmap: &bitmap);
2876	if (ret)
2877	return ret;
2878
2879	return copy_to_user(to: (void __user *)arg, from: &unmap, n: minsz) ?
2880	-EFAULT : 0;
2881	}
2882
2883	static int vfio_iommu_type1_dirty_pages(struct vfio_iommu *iommu,
2884	unsigned long arg)
2885	{
2886	struct vfio_iommu_type1_dirty_bitmap dirty;
2887	uint32_t mask = VFIO_IOMMU_DIRTY_PAGES_FLAG_START |
2888	VFIO_IOMMU_DIRTY_PAGES_FLAG_STOP |
2889	VFIO_IOMMU_DIRTY_PAGES_FLAG_GET_BITMAP;
2890	unsigned long minsz;
2891	int ret = 0;
2892
2893	if (!iommu->v2)
2894	return -EACCES;
2895
2896	minsz = offsetofend(struct vfio_iommu_type1_dirty_bitmap, flags);
2897
2898	if (copy_from_user(to: &dirty, from: (void __user *)arg, n: minsz))
2899	return -EFAULT;
2900
2901	if (dirty.argsz < minsz || dirty.flags & ~mask)
2902	return -EINVAL;
2903
2904	/* only one flag should be set at a time */
2905	if (__ffs(dirty.flags) != __fls(word: dirty.flags))
2906	return -EINVAL;
2907
2908	if (dirty.flags & VFIO_IOMMU_DIRTY_PAGES_FLAG_START) {
2909	size_t pgsize;
2910
2911	mutex_lock(&iommu->lock);
2912	pgsize = 1 << __ffs(iommu->pgsize_bitmap);
2913	if (!iommu->dirty_page_tracking) {
2914	ret = vfio_dma_bitmap_alloc_all(iommu, pgsize);
2915	if (!ret)
2916	iommu->dirty_page_tracking = true;
2917	}
2918	mutex_unlock(lock: &iommu->lock);
2919	return ret;
2920	} else if (dirty.flags & VFIO_IOMMU_DIRTY_PAGES_FLAG_STOP) {
2921	mutex_lock(&iommu->lock);
2922	if (iommu->dirty_page_tracking) {
2923	iommu->dirty_page_tracking = false;
2924	vfio_dma_bitmap_free_all(iommu);
2925	}
2926	mutex_unlock(lock: &iommu->lock);
2927	return 0;
2928	} else if (dirty.flags & VFIO_IOMMU_DIRTY_PAGES_FLAG_GET_BITMAP) {
2929	struct vfio_iommu_type1_dirty_bitmap_get range;
2930	unsigned long pgshift;
2931	size_t data_size = dirty.argsz - minsz;
2932	size_t iommu_pgsize;
2933
2934	if (!data_size || data_size < sizeof(range))
2935	return -EINVAL;
2936
2937	if (copy_from_user(to: &range, from: (void __user *)(arg + minsz),
2938	n: sizeof(range)))
2939	return -EFAULT;
2940
2941	if (range.iova + range.size < range.iova)
2942	return -EINVAL;
2943	if (!access_ok((void __user *)range.bitmap.data,
2944	range.bitmap.size))
2945	return -EINVAL;
2946
2947	pgshift = __ffs(range.bitmap.pgsize);
2948	ret = verify_bitmap_size(npages: range.size >> pgshift,
2949	bitmap_size: range.bitmap.size);
2950	if (ret)
2951	return ret;
2952
2953	mutex_lock(&iommu->lock);
2954
2955	iommu_pgsize = (size_t)1 << __ffs(iommu->pgsize_bitmap);
2956
2957	/* allow only smallest supported pgsize */
2958	if (range.bitmap.pgsize != iommu_pgsize) {
2959	ret = -EINVAL;
2960	goto out_unlock;
2961	}
2962	if (range.iova & (iommu_pgsize - 1)) {
2963	ret = -EINVAL;
2964	goto out_unlock;
2965	}
2966	if (!range.size || range.size & (iommu_pgsize - 1)) {
2967	ret = -EINVAL;
2968	goto out_unlock;
2969	}
2970
2971	if (iommu->dirty_page_tracking)
2972	ret = vfio_iova_dirty_bitmap(bitmap: range.bitmap.data,
2973	iommu, iova: range.iova,
2974	size: range.size,
2975	pgsize: range.bitmap.pgsize);
2976	else
2977	ret = -EINVAL;
2978	out_unlock:
2979	mutex_unlock(lock: &iommu->lock);
2980
2981	return ret;
2982	}
2983
2984	return -EINVAL;
2985	}
2986
2987	static long vfio_iommu_type1_ioctl(void *iommu_data,
2988	unsigned int cmd, unsigned long arg)
2989	{
2990	struct vfio_iommu *iommu = iommu_data;
2991
2992	switch (cmd) {
2993	case VFIO_CHECK_EXTENSION:
2994	return vfio_iommu_type1_check_extension(iommu, arg);
2995	case VFIO_IOMMU_GET_INFO:
2996	return vfio_iommu_type1_get_info(iommu, arg);
2997	case VFIO_IOMMU_MAP_DMA:
2998	return vfio_iommu_type1_map_dma(iommu, arg);
2999	case VFIO_IOMMU_UNMAP_DMA:
3000	return vfio_iommu_type1_unmap_dma(iommu, arg);
3001	case VFIO_IOMMU_DIRTY_PAGES:
3002	return vfio_iommu_type1_dirty_pages(iommu, arg);
3003	default:
3004	return -ENOTTY;
3005	}
3006	}
3007
3008	static void vfio_iommu_type1_register_device(void *iommu_data,
3009	struct vfio_device *vdev)
3010	{
3011	struct vfio_iommu *iommu = iommu_data;
3012
3013	if (!vdev->ops->dma_unmap)
3014	return;
3015
3016	/*
3017	* list_empty(&iommu->device_list) is tested under the iommu->lock while
3018	* iteration for dma_unmap must be done under the device_list_lock.
3019	* Holding both locks here allows avoiding the device_list_lock in
3020	* several fast paths. See vfio_notify_dma_unmap()
3021	*/
3022	mutex_lock(&iommu->lock);
3023	mutex_lock(&iommu->device_list_lock);
3024	list_add(new: &vdev->iommu_entry, head: &iommu->device_list);
3025	mutex_unlock(lock: &iommu->device_list_lock);
3026	mutex_unlock(lock: &iommu->lock);
3027	}
3028
3029	static void vfio_iommu_type1_unregister_device(void *iommu_data,
3030	struct vfio_device *vdev)
3031	{
3032	struct vfio_iommu *iommu = iommu_data;
3033
3034	if (!vdev->ops->dma_unmap)
3035	return;
3036
3037	mutex_lock(&iommu->lock);
3038	mutex_lock(&iommu->device_list_lock);
3039	list_del(entry: &vdev->iommu_entry);
3040	mutex_unlock(lock: &iommu->device_list_lock);
3041	mutex_unlock(lock: &iommu->lock);
3042	}
3043
3044	static int vfio_iommu_type1_dma_rw_chunk(struct vfio_iommu *iommu,
3045	dma_addr_t user_iova, void *data,
3046	size_t count, bool write,
3047	size_t *copied)
3048	{
3049	struct mm_struct *mm;
3050	unsigned long vaddr;
3051	struct vfio_dma *dma;
3052	bool kthread = current->mm == NULL;
3053	size_t offset;
3054
3055	*copied = 0;
3056
3057	dma = vfio_find_dma(iommu, start: user_iova, size: 1);
3058	if (!dma)
3059	return -EINVAL;
3060
3061	if ((write && !(dma->prot & IOMMU_WRITE)) ||
3062	!(dma->prot & IOMMU_READ))
3063	return -EPERM;
3064
3065	mm = dma->mm;
3066	if (!mmget_not_zero(mm))
3067	return -EPERM;
3068
3069	if (kthread)
3070	kthread_use_mm(mm);
3071	else if (current->mm != mm)
3072	goto out;
3073
3074	offset = user_iova - dma->iova;
3075
3076	if (count > dma->size - offset)
3077	count = dma->size - offset;
3078
3079	vaddr = dma->vaddr + offset;
3080
3081	if (write) {
3082	*copied = copy_to_user(to: (void __user *)vaddr, from: data,
3083	n: count) ? 0 : count;
3084	if (*copied && iommu->dirty_page_tracking) {
3085	unsigned long pgshift = __ffs(iommu->pgsize_bitmap);
3086	/*
3087	* Bitmap populated with the smallest supported page
3088	* size
3089	*/
3090	bitmap_set(map: dma->bitmap, start: offset >> pgshift,
3091	nbits: ((offset + *copied - 1) >> pgshift) -
3092	(offset >> pgshift) + 1);
3093	}
3094	} else
3095	*copied = copy_from_user(to: data, from: (void __user *)vaddr,
3096	n: count) ? 0 : count;
3097	if (kthread)
3098	kthread_unuse_mm(mm);
3099	out:
3100	mmput(mm);
3101	return *copied ? 0 : -EFAULT;
3102	}
3103
3104	static int vfio_iommu_type1_dma_rw(void *iommu_data, dma_addr_t user_iova,
3105	void *data, size_t count, bool write)
3106	{
3107	struct vfio_iommu *iommu = iommu_data;
3108	int ret = 0;
3109	size_t done;
3110
3111	mutex_lock(&iommu->lock);
3112
3113	if (WARN_ONCE(iommu->vaddr_invalid_count,
3114	"vfio_dma_rw not allowed with VFIO_UPDATE_VADDR\n")) {
3115	ret = -EBUSY;
3116	goto out;
3117	}
3118
3119	while (count > 0) {
3120	ret = vfio_iommu_type1_dma_rw_chunk(iommu, user_iova, data,
3121	count, write, copied: &done);
3122	if (ret)
3123	break;
3124
3125	count -= done;
3126	data += done;
3127	user_iova += done;
3128	}
3129
3130	out:
3131	mutex_unlock(lock: &iommu->lock);
3132	return ret;
3133	}
3134
3135	static struct iommu_domain *
3136	vfio_iommu_type1_group_iommu_domain(void *iommu_data,
3137	struct iommu_group *iommu_group)
3138	{
3139	struct iommu_domain *domain = ERR_PTR(error: -ENODEV);
3140	struct vfio_iommu *iommu = iommu_data;
3141	struct vfio_domain *d;
3142
3143	if (!iommu || !iommu_group)
3144	return ERR_PTR(error: -EINVAL);
3145
3146	mutex_lock(&iommu->lock);
3147	list_for_each_entry(d, &iommu->domain_list, next) {
3148	if (find_iommu_group(domain: d, iommu_group)) {
3149	domain = d->domain;
3150	break;
3151	}
3152	}
3153	mutex_unlock(lock: &iommu->lock);
3154
3155	return domain;
3156	}
3157
3158	static const struct vfio_iommu_driver_ops vfio_iommu_driver_ops_type1 = {
3159	.name = "vfio-iommu-type1",
3160	.owner = THIS_MODULE,
3161	.open = vfio_iommu_type1_open,
3162	.release = vfio_iommu_type1_release,
3163	.ioctl = vfio_iommu_type1_ioctl,
3164	.attach_group = vfio_iommu_type1_attach_group,
3165	.detach_group = vfio_iommu_type1_detach_group,
3166	.pin_pages = vfio_iommu_type1_pin_pages,
3167	.unpin_pages = vfio_iommu_type1_unpin_pages,
3168	.register_device = vfio_iommu_type1_register_device,
3169	.unregister_device = vfio_iommu_type1_unregister_device,
3170	.dma_rw = vfio_iommu_type1_dma_rw,
3171	.group_iommu_domain = vfio_iommu_type1_group_iommu_domain,
3172	};
3173
3174	static int __init vfio_iommu_type1_init(void)
3175	{
3176	return vfio_register_iommu_driver(ops: &vfio_iommu_driver_ops_type1);
3177	}
3178
3179	static void __exit vfio_iommu_type1_cleanup(void)
3180	{
3181	vfio_unregister_iommu_driver(ops: &vfio_iommu_driver_ops_type1);
3182	}
3183
3184	module_init(vfio_iommu_type1_init);
3185	module_exit(vfio_iommu_type1_cleanup);
3186
3187	MODULE_VERSION(DRIVER_VERSION);
3188	MODULE_LICENSE("GPL v2");
3189	MODULE_AUTHOR(DRIVER_AUTHOR);
3190	MODULE_DESCRIPTION(DRIVER_DESC);
3191