xref: /linux/drivers/xen/privcmd.c (revision 06a130e42a5bfc84795464bff023bff4c16f58c5)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /******************************************************************************
3  * privcmd.c
4  *
5  * Interface to privileged domain-0 commands.
6  *
7  * Copyright (c) 2002-2004, K A Fraser, B Dragovic
8  */
9 
10 #define pr_fmt(fmt) "xen:" KBUILD_MODNAME ": " fmt
11 
12 #include <linux/eventfd.h>
13 #include <linux/file.h>
14 #include <linux/kernel.h>
15 #include <linux/module.h>
16 #include <linux/mutex.h>
17 #include <linux/poll.h>
18 #include <linux/sched.h>
19 #include <linux/slab.h>
20 #include <linux/srcu.h>
21 #include <linux/string.h>
22 #include <linux/workqueue.h>
23 #include <linux/errno.h>
24 #include <linux/mm.h>
25 #include <linux/mman.h>
26 #include <linux/uaccess.h>
27 #include <linux/swap.h>
28 #include <linux/highmem.h>
29 #include <linux/pagemap.h>
30 #include <linux/seq_file.h>
31 #include <linux/miscdevice.h>
32 #include <linux/moduleparam.h>
33 #include <linux/virtio_mmio.h>
34 
35 #include <asm/xen/hypervisor.h>
36 #include <asm/xen/hypercall.h>
37 
38 #include <xen/xen.h>
39 #include <xen/events.h>
40 #include <xen/privcmd.h>
41 #include <xen/interface/xen.h>
42 #include <xen/interface/memory.h>
43 #include <xen/interface/hvm/dm_op.h>
44 #include <xen/interface/hvm/ioreq.h>
45 #include <xen/features.h>
46 #include <xen/page.h>
47 #include <xen/xen-ops.h>
48 #include <xen/balloon.h>
49 #ifdef CONFIG_XEN_ACPI
50 #include <xen/acpi.h>
51 #endif
52 
53 #include "privcmd.h"
54 
55 MODULE_DESCRIPTION("Xen hypercall passthrough driver");
56 MODULE_LICENSE("GPL");
57 
58 #define PRIV_VMA_LOCKED ((void *)1)
59 
60 static unsigned int privcmd_dm_op_max_num = 16;
61 module_param_named(dm_op_max_nr_bufs, privcmd_dm_op_max_num, uint, 0644);
62 MODULE_PARM_DESC(dm_op_max_nr_bufs,
63 		 "Maximum number of buffers per dm_op hypercall");
64 
65 static unsigned int privcmd_dm_op_buf_max_size = 4096;
66 module_param_named(dm_op_buf_max_size, privcmd_dm_op_buf_max_size, uint,
67 		   0644);
68 MODULE_PARM_DESC(dm_op_buf_max_size,
69 		 "Maximum size of a dm_op hypercall buffer");
70 
71 struct privcmd_data {
72 	domid_t domid;
73 };
74 
75 static int privcmd_vma_range_is_mapped(
76                struct vm_area_struct *vma,
77                unsigned long addr,
78                unsigned long nr_pages);
79 
80 static long privcmd_ioctl_hypercall(struct file *file, void __user *udata)
81 {
82 	struct privcmd_data *data = file->private_data;
83 	struct privcmd_hypercall hypercall;
84 	long ret;
85 
86 	/* Disallow arbitrary hypercalls if restricted */
87 	if (data->domid != DOMID_INVALID)
88 		return -EPERM;
89 
90 	if (copy_from_user(&hypercall, udata, sizeof(hypercall)))
91 		return -EFAULT;
92 
93 	xen_preemptible_hcall_begin();
94 	ret = privcmd_call(hypercall.op,
95 			   hypercall.arg[0], hypercall.arg[1],
96 			   hypercall.arg[2], hypercall.arg[3],
97 			   hypercall.arg[4]);
98 	xen_preemptible_hcall_end();
99 
100 	return ret;
101 }
102 
103 static void free_page_list(struct list_head *pages)
104 {
105 	struct page *p, *n;
106 
107 	list_for_each_entry_safe(p, n, pages, lru)
108 		__free_page(p);
109 
110 	INIT_LIST_HEAD(pages);
111 }
112 
113 /*
114  * Given an array of items in userspace, return a list of pages
115  * containing the data.  If copying fails, either because of memory
116  * allocation failure or a problem reading user memory, return an
117  * error code; its up to the caller to dispose of any partial list.
118  */
119 static int gather_array(struct list_head *pagelist,
120 			unsigned nelem, size_t size,
121 			const void __user *data)
122 {
123 	unsigned pageidx;
124 	void *pagedata;
125 	int ret;
126 
127 	if (size > PAGE_SIZE)
128 		return 0;
129 
130 	pageidx = PAGE_SIZE;
131 	pagedata = NULL;	/* quiet, gcc */
132 	while (nelem--) {
133 		if (pageidx > PAGE_SIZE-size) {
134 			struct page *page = alloc_page(GFP_KERNEL);
135 
136 			ret = -ENOMEM;
137 			if (page == NULL)
138 				goto fail;
139 
140 			pagedata = page_address(page);
141 
142 			list_add_tail(&page->lru, pagelist);
143 			pageidx = 0;
144 		}
145 
146 		ret = -EFAULT;
147 		if (copy_from_user(pagedata + pageidx, data, size))
148 			goto fail;
149 
150 		data += size;
151 		pageidx += size;
152 	}
153 
154 	ret = 0;
155 
156 fail:
157 	return ret;
158 }
159 
160 /*
161  * Call function "fn" on each element of the array fragmented
162  * over a list of pages.
163  */
164 static int traverse_pages(unsigned nelem, size_t size,
165 			  struct list_head *pos,
166 			  int (*fn)(void *data, void *state),
167 			  void *state)
168 {
169 	void *pagedata;
170 	unsigned pageidx;
171 	int ret = 0;
172 
173 	BUG_ON(size > PAGE_SIZE);
174 
175 	pageidx = PAGE_SIZE;
176 	pagedata = NULL;	/* hush, gcc */
177 
178 	while (nelem--) {
179 		if (pageidx > PAGE_SIZE-size) {
180 			struct page *page;
181 			pos = pos->next;
182 			page = list_entry(pos, struct page, lru);
183 			pagedata = page_address(page);
184 			pageidx = 0;
185 		}
186 
187 		ret = (*fn)(pagedata + pageidx, state);
188 		if (ret)
189 			break;
190 		pageidx += size;
191 	}
192 
193 	return ret;
194 }
195 
196 /*
197  * Similar to traverse_pages, but use each page as a "block" of
198  * data to be processed as one unit.
199  */
200 static int traverse_pages_block(unsigned nelem, size_t size,
201 				struct list_head *pos,
202 				int (*fn)(void *data, int nr, void *state),
203 				void *state)
204 {
205 	void *pagedata;
206 	int ret = 0;
207 
208 	BUG_ON(size > PAGE_SIZE);
209 
210 	while (nelem) {
211 		int nr = (PAGE_SIZE/size);
212 		struct page *page;
213 		if (nr > nelem)
214 			nr = nelem;
215 		pos = pos->next;
216 		page = list_entry(pos, struct page, lru);
217 		pagedata = page_address(page);
218 		ret = (*fn)(pagedata, nr, state);
219 		if (ret)
220 			break;
221 		nelem -= nr;
222 	}
223 
224 	return ret;
225 }
226 
227 struct mmap_gfn_state {
228 	unsigned long va;
229 	struct vm_area_struct *vma;
230 	domid_t domain;
231 };
232 
233 static int mmap_gfn_range(void *data, void *state)
234 {
235 	struct privcmd_mmap_entry *msg = data;
236 	struct mmap_gfn_state *st = state;
237 	struct vm_area_struct *vma = st->vma;
238 	int rc;
239 
240 	/* Do not allow range to wrap the address space. */
241 	if ((msg->npages > (LONG_MAX >> PAGE_SHIFT)) ||
242 	    ((unsigned long)(msg->npages << PAGE_SHIFT) >= -st->va))
243 		return -EINVAL;
244 
245 	/* Range chunks must be contiguous in va space. */
246 	if ((msg->va != st->va) ||
247 	    ((msg->va+(msg->npages<<PAGE_SHIFT)) > vma->vm_end))
248 		return -EINVAL;
249 
250 	rc = xen_remap_domain_gfn_range(vma,
251 					msg->va & PAGE_MASK,
252 					msg->mfn, msg->npages,
253 					vma->vm_page_prot,
254 					st->domain, NULL);
255 	if (rc < 0)
256 		return rc;
257 
258 	st->va += msg->npages << PAGE_SHIFT;
259 
260 	return 0;
261 }
262 
263 static long privcmd_ioctl_mmap(struct file *file, void __user *udata)
264 {
265 	struct privcmd_data *data = file->private_data;
266 	struct privcmd_mmap mmapcmd;
267 	struct mm_struct *mm = current->mm;
268 	struct vm_area_struct *vma;
269 	int rc;
270 	LIST_HEAD(pagelist);
271 	struct mmap_gfn_state state;
272 
273 	/* We only support privcmd_ioctl_mmap_batch for non-auto-translated. */
274 	if (xen_feature(XENFEAT_auto_translated_physmap))
275 		return -ENOSYS;
276 
277 	if (copy_from_user(&mmapcmd, udata, sizeof(mmapcmd)))
278 		return -EFAULT;
279 
280 	/* If restriction is in place, check the domid matches */
281 	if (data->domid != DOMID_INVALID && data->domid != mmapcmd.dom)
282 		return -EPERM;
283 
284 	rc = gather_array(&pagelist,
285 			  mmapcmd.num, sizeof(struct privcmd_mmap_entry),
286 			  mmapcmd.entry);
287 
288 	if (rc || list_empty(&pagelist))
289 		goto out;
290 
291 	mmap_write_lock(mm);
292 
293 	{
294 		struct page *page = list_first_entry(&pagelist,
295 						     struct page, lru);
296 		struct privcmd_mmap_entry *msg = page_address(page);
297 
298 		vma = vma_lookup(mm, msg->va);
299 		rc = -EINVAL;
300 
301 		if (!vma || (msg->va != vma->vm_start) || vma->vm_private_data)
302 			goto out_up;
303 		vma->vm_private_data = PRIV_VMA_LOCKED;
304 	}
305 
306 	state.va = vma->vm_start;
307 	state.vma = vma;
308 	state.domain = mmapcmd.dom;
309 
310 	rc = traverse_pages(mmapcmd.num, sizeof(struct privcmd_mmap_entry),
311 			    &pagelist,
312 			    mmap_gfn_range, &state);
313 
314 
315 out_up:
316 	mmap_write_unlock(mm);
317 
318 out:
319 	free_page_list(&pagelist);
320 
321 	return rc;
322 }
323 
324 struct mmap_batch_state {
325 	domid_t domain;
326 	unsigned long va;
327 	struct vm_area_struct *vma;
328 	int index;
329 	/* A tristate:
330 	 *      0 for no errors
331 	 *      1 if at least one error has happened (and no
332 	 *          -ENOENT errors have happened)
333 	 *      -ENOENT if at least 1 -ENOENT has happened.
334 	 */
335 	int global_error;
336 	int version;
337 
338 	/* User-space gfn array to store errors in the second pass for V1. */
339 	xen_pfn_t __user *user_gfn;
340 	/* User-space int array to store errors in the second pass for V2. */
341 	int __user *user_err;
342 };
343 
344 /* auto translated dom0 note: if domU being created is PV, then gfn is
345  * mfn(addr on bus). If it's auto xlated, then gfn is pfn (input to HAP).
346  */
347 static int mmap_batch_fn(void *data, int nr, void *state)
348 {
349 	xen_pfn_t *gfnp = data;
350 	struct mmap_batch_state *st = state;
351 	struct vm_area_struct *vma = st->vma;
352 	struct page **pages = vma->vm_private_data;
353 	struct page **cur_pages = NULL;
354 	int ret;
355 
356 	if (xen_feature(XENFEAT_auto_translated_physmap))
357 		cur_pages = &pages[st->index];
358 
359 	BUG_ON(nr < 0);
360 	ret = xen_remap_domain_gfn_array(st->vma, st->va & PAGE_MASK, gfnp, nr,
361 					 (int *)gfnp, st->vma->vm_page_prot,
362 					 st->domain, cur_pages);
363 
364 	/* Adjust the global_error? */
365 	if (ret != nr) {
366 		if (ret == -ENOENT)
367 			st->global_error = -ENOENT;
368 		else {
369 			/* Record that at least one error has happened. */
370 			if (st->global_error == 0)
371 				st->global_error = 1;
372 		}
373 	}
374 	st->va += XEN_PAGE_SIZE * nr;
375 	st->index += nr / XEN_PFN_PER_PAGE;
376 
377 	return 0;
378 }
379 
380 static int mmap_return_error(int err, struct mmap_batch_state *st)
381 {
382 	int ret;
383 
384 	if (st->version == 1) {
385 		if (err) {
386 			xen_pfn_t gfn;
387 
388 			ret = get_user(gfn, st->user_gfn);
389 			if (ret < 0)
390 				return ret;
391 			/*
392 			 * V1 encodes the error codes in the 32bit top
393 			 * nibble of the gfn (with its known
394 			 * limitations vis-a-vis 64 bit callers).
395 			 */
396 			gfn |= (err == -ENOENT) ?
397 				PRIVCMD_MMAPBATCH_PAGED_ERROR :
398 				PRIVCMD_MMAPBATCH_MFN_ERROR;
399 			return __put_user(gfn, st->user_gfn++);
400 		} else
401 			st->user_gfn++;
402 	} else { /* st->version == 2 */
403 		if (err)
404 			return __put_user(err, st->user_err++);
405 		else
406 			st->user_err++;
407 	}
408 
409 	return 0;
410 }
411 
412 static int mmap_return_errors(void *data, int nr, void *state)
413 {
414 	struct mmap_batch_state *st = state;
415 	int *errs = data;
416 	int i;
417 	int ret;
418 
419 	for (i = 0; i < nr; i++) {
420 		ret = mmap_return_error(errs[i], st);
421 		if (ret < 0)
422 			return ret;
423 	}
424 	return 0;
425 }
426 
427 /* Allocate pfns that are then mapped with gfns from foreign domid. Update
428  * the vma with the page info to use later.
429  * Returns: 0 if success, otherwise -errno
430  */
431 static int alloc_empty_pages(struct vm_area_struct *vma, int numpgs)
432 {
433 	int rc;
434 	struct page **pages;
435 
436 	pages = kvcalloc(numpgs, sizeof(pages[0]), GFP_KERNEL);
437 	if (pages == NULL)
438 		return -ENOMEM;
439 
440 	rc = xen_alloc_unpopulated_pages(numpgs, pages);
441 	if (rc != 0) {
442 		pr_warn("%s Could not alloc %d pfns rc:%d\n", __func__,
443 			numpgs, rc);
444 		kvfree(pages);
445 		return -ENOMEM;
446 	}
447 	BUG_ON(vma->vm_private_data != NULL);
448 	vma->vm_private_data = pages;
449 
450 	return 0;
451 }
452 
453 static const struct vm_operations_struct privcmd_vm_ops;
454 
455 static long privcmd_ioctl_mmap_batch(
456 	struct file *file, void __user *udata, int version)
457 {
458 	struct privcmd_data *data = file->private_data;
459 	int ret;
460 	struct privcmd_mmapbatch_v2 m;
461 	struct mm_struct *mm = current->mm;
462 	struct vm_area_struct *vma;
463 	unsigned long nr_pages;
464 	LIST_HEAD(pagelist);
465 	struct mmap_batch_state state;
466 
467 	switch (version) {
468 	case 1:
469 		if (copy_from_user(&m, udata, sizeof(struct privcmd_mmapbatch)))
470 			return -EFAULT;
471 		/* Returns per-frame error in m.arr. */
472 		m.err = NULL;
473 		if (!access_ok(m.arr, m.num * sizeof(*m.arr)))
474 			return -EFAULT;
475 		break;
476 	case 2:
477 		if (copy_from_user(&m, udata, sizeof(struct privcmd_mmapbatch_v2)))
478 			return -EFAULT;
479 		/* Returns per-frame error code in m.err. */
480 		if (!access_ok(m.err, m.num * (sizeof(*m.err))))
481 			return -EFAULT;
482 		break;
483 	default:
484 		return -EINVAL;
485 	}
486 
487 	/* If restriction is in place, check the domid matches */
488 	if (data->domid != DOMID_INVALID && data->domid != m.dom)
489 		return -EPERM;
490 
491 	nr_pages = DIV_ROUND_UP(m.num, XEN_PFN_PER_PAGE);
492 	if ((m.num <= 0) || (nr_pages > (LONG_MAX >> PAGE_SHIFT)))
493 		return -EINVAL;
494 
495 	ret = gather_array(&pagelist, m.num, sizeof(xen_pfn_t), m.arr);
496 
497 	if (ret)
498 		goto out;
499 	if (list_empty(&pagelist)) {
500 		ret = -EINVAL;
501 		goto out;
502 	}
503 
504 	if (version == 2) {
505 		/* Zero error array now to only copy back actual errors. */
506 		if (clear_user(m.err, sizeof(int) * m.num)) {
507 			ret = -EFAULT;
508 			goto out;
509 		}
510 	}
511 
512 	mmap_write_lock(mm);
513 
514 	vma = find_vma(mm, m.addr);
515 	if (!vma ||
516 	    vma->vm_ops != &privcmd_vm_ops) {
517 		ret = -EINVAL;
518 		goto out_unlock;
519 	}
520 
521 	/*
522 	 * Caller must either:
523 	 *
524 	 * Map the whole VMA range, which will also allocate all the
525 	 * pages required for the auto_translated_physmap case.
526 	 *
527 	 * Or
528 	 *
529 	 * Map unmapped holes left from a previous map attempt (e.g.,
530 	 * because those foreign frames were previously paged out).
531 	 */
532 	if (vma->vm_private_data == NULL) {
533 		if (m.addr != vma->vm_start ||
534 		    m.addr + (nr_pages << PAGE_SHIFT) != vma->vm_end) {
535 			ret = -EINVAL;
536 			goto out_unlock;
537 		}
538 		if (xen_feature(XENFEAT_auto_translated_physmap)) {
539 			ret = alloc_empty_pages(vma, nr_pages);
540 			if (ret < 0)
541 				goto out_unlock;
542 		} else
543 			vma->vm_private_data = PRIV_VMA_LOCKED;
544 	} else {
545 		if (m.addr < vma->vm_start ||
546 		    m.addr + (nr_pages << PAGE_SHIFT) > vma->vm_end) {
547 			ret = -EINVAL;
548 			goto out_unlock;
549 		}
550 		if (privcmd_vma_range_is_mapped(vma, m.addr, nr_pages)) {
551 			ret = -EINVAL;
552 			goto out_unlock;
553 		}
554 	}
555 
556 	state.domain        = m.dom;
557 	state.vma           = vma;
558 	state.va            = m.addr;
559 	state.index         = 0;
560 	state.global_error  = 0;
561 	state.version       = version;
562 
563 	BUILD_BUG_ON(((PAGE_SIZE / sizeof(xen_pfn_t)) % XEN_PFN_PER_PAGE) != 0);
564 	/* mmap_batch_fn guarantees ret == 0 */
565 	BUG_ON(traverse_pages_block(m.num, sizeof(xen_pfn_t),
566 				    &pagelist, mmap_batch_fn, &state));
567 
568 	mmap_write_unlock(mm);
569 
570 	if (state.global_error) {
571 		/* Write back errors in second pass. */
572 		state.user_gfn = (xen_pfn_t *)m.arr;
573 		state.user_err = m.err;
574 		ret = traverse_pages_block(m.num, sizeof(xen_pfn_t),
575 					   &pagelist, mmap_return_errors, &state);
576 	} else
577 		ret = 0;
578 
579 	/* If we have not had any EFAULT-like global errors then set the global
580 	 * error to -ENOENT if necessary. */
581 	if ((ret == 0) && (state.global_error == -ENOENT))
582 		ret = -ENOENT;
583 
584 out:
585 	free_page_list(&pagelist);
586 	return ret;
587 
588 out_unlock:
589 	mmap_write_unlock(mm);
590 	goto out;
591 }
592 
593 static int lock_pages(
594 	struct privcmd_dm_op_buf kbufs[], unsigned int num,
595 	struct page *pages[], unsigned int nr_pages, unsigned int *pinned)
596 {
597 	unsigned int i, off = 0;
598 
599 	for (i = 0; i < num; ) {
600 		unsigned int requested;
601 		int page_count;
602 
603 		requested = DIV_ROUND_UP(
604 			offset_in_page(kbufs[i].uptr) + kbufs[i].size,
605 			PAGE_SIZE) - off;
606 		if (requested > nr_pages)
607 			return -ENOSPC;
608 
609 		page_count = pin_user_pages_fast(
610 			(unsigned long)kbufs[i].uptr + off * PAGE_SIZE,
611 			requested, FOLL_WRITE, pages);
612 		if (page_count <= 0)
613 			return page_count ? : -EFAULT;
614 
615 		*pinned += page_count;
616 		nr_pages -= page_count;
617 		pages += page_count;
618 
619 		off = (requested == page_count) ? 0 : off + page_count;
620 		i += !off;
621 	}
622 
623 	return 0;
624 }
625 
626 static void unlock_pages(struct page *pages[], unsigned int nr_pages)
627 {
628 	unpin_user_pages_dirty_lock(pages, nr_pages, true);
629 }
630 
631 static long privcmd_ioctl_dm_op(struct file *file, void __user *udata)
632 {
633 	struct privcmd_data *data = file->private_data;
634 	struct privcmd_dm_op kdata;
635 	struct privcmd_dm_op_buf *kbufs;
636 	unsigned int nr_pages = 0;
637 	struct page **pages = NULL;
638 	struct xen_dm_op_buf *xbufs = NULL;
639 	unsigned int i;
640 	long rc;
641 	unsigned int pinned = 0;
642 
643 	if (copy_from_user(&kdata, udata, sizeof(kdata)))
644 		return -EFAULT;
645 
646 	/* If restriction is in place, check the domid matches */
647 	if (data->domid != DOMID_INVALID && data->domid != kdata.dom)
648 		return -EPERM;
649 
650 	if (kdata.num == 0)
651 		return 0;
652 
653 	if (kdata.num > privcmd_dm_op_max_num)
654 		return -E2BIG;
655 
656 	kbufs = kcalloc(kdata.num, sizeof(*kbufs), GFP_KERNEL);
657 	if (!kbufs)
658 		return -ENOMEM;
659 
660 	if (copy_from_user(kbufs, kdata.ubufs,
661 			   sizeof(*kbufs) * kdata.num)) {
662 		rc = -EFAULT;
663 		goto out;
664 	}
665 
666 	for (i = 0; i < kdata.num; i++) {
667 		if (kbufs[i].size > privcmd_dm_op_buf_max_size) {
668 			rc = -E2BIG;
669 			goto out;
670 		}
671 
672 		if (!access_ok(kbufs[i].uptr,
673 			       kbufs[i].size)) {
674 			rc = -EFAULT;
675 			goto out;
676 		}
677 
678 		nr_pages += DIV_ROUND_UP(
679 			offset_in_page(kbufs[i].uptr) + kbufs[i].size,
680 			PAGE_SIZE);
681 	}
682 
683 	pages = kcalloc(nr_pages, sizeof(*pages), GFP_KERNEL);
684 	if (!pages) {
685 		rc = -ENOMEM;
686 		goto out;
687 	}
688 
689 	xbufs = kcalloc(kdata.num, sizeof(*xbufs), GFP_KERNEL);
690 	if (!xbufs) {
691 		rc = -ENOMEM;
692 		goto out;
693 	}
694 
695 	rc = lock_pages(kbufs, kdata.num, pages, nr_pages, &pinned);
696 	if (rc < 0)
697 		goto out;
698 
699 	for (i = 0; i < kdata.num; i++) {
700 		set_xen_guest_handle(xbufs[i].h, kbufs[i].uptr);
701 		xbufs[i].size = kbufs[i].size;
702 	}
703 
704 	xen_preemptible_hcall_begin();
705 	rc = HYPERVISOR_dm_op(kdata.dom, kdata.num, xbufs);
706 	xen_preemptible_hcall_end();
707 
708 out:
709 	unlock_pages(pages, pinned);
710 	kfree(xbufs);
711 	kfree(pages);
712 	kfree(kbufs);
713 
714 	return rc;
715 }
716 
717 static long privcmd_ioctl_restrict(struct file *file, void __user *udata)
718 {
719 	struct privcmd_data *data = file->private_data;
720 	domid_t dom;
721 
722 	if (copy_from_user(&dom, udata, sizeof(dom)))
723 		return -EFAULT;
724 
725 	/* Set restriction to the specified domain, or check it matches */
726 	if (data->domid == DOMID_INVALID)
727 		data->domid = dom;
728 	else if (data->domid != dom)
729 		return -EINVAL;
730 
731 	return 0;
732 }
733 
734 static long privcmd_ioctl_mmap_resource(struct file *file,
735 				struct privcmd_mmap_resource __user *udata)
736 {
737 	struct privcmd_data *data = file->private_data;
738 	struct mm_struct *mm = current->mm;
739 	struct vm_area_struct *vma;
740 	struct privcmd_mmap_resource kdata;
741 	xen_pfn_t *pfns = NULL;
742 	struct xen_mem_acquire_resource xdata = { };
743 	int rc;
744 
745 	if (copy_from_user(&kdata, udata, sizeof(kdata)))
746 		return -EFAULT;
747 
748 	/* If restriction is in place, check the domid matches */
749 	if (data->domid != DOMID_INVALID && data->domid != kdata.dom)
750 		return -EPERM;
751 
752 	/* Both fields must be set or unset */
753 	if (!!kdata.addr != !!kdata.num)
754 		return -EINVAL;
755 
756 	xdata.domid = kdata.dom;
757 	xdata.type = kdata.type;
758 	xdata.id = kdata.id;
759 
760 	if (!kdata.addr && !kdata.num) {
761 		/* Query the size of the resource. */
762 		rc = HYPERVISOR_memory_op(XENMEM_acquire_resource, &xdata);
763 		if (rc)
764 			return rc;
765 		return __put_user(xdata.nr_frames, &udata->num);
766 	}
767 
768 	mmap_write_lock(mm);
769 
770 	vma = find_vma(mm, kdata.addr);
771 	if (!vma || vma->vm_ops != &privcmd_vm_ops) {
772 		rc = -EINVAL;
773 		goto out;
774 	}
775 
776 	pfns = kcalloc(kdata.num, sizeof(*pfns), GFP_KERNEL | __GFP_NOWARN);
777 	if (!pfns) {
778 		rc = -ENOMEM;
779 		goto out;
780 	}
781 
782 	if (IS_ENABLED(CONFIG_XEN_AUTO_XLATE) &&
783 	    xen_feature(XENFEAT_auto_translated_physmap)) {
784 		unsigned int nr = DIV_ROUND_UP(kdata.num, XEN_PFN_PER_PAGE);
785 		struct page **pages;
786 		unsigned int i;
787 
788 		rc = alloc_empty_pages(vma, nr);
789 		if (rc < 0)
790 			goto out;
791 
792 		pages = vma->vm_private_data;
793 
794 		for (i = 0; i < kdata.num; i++) {
795 			xen_pfn_t pfn =
796 				page_to_xen_pfn(pages[i / XEN_PFN_PER_PAGE]);
797 
798 			pfns[i] = pfn + (i % XEN_PFN_PER_PAGE);
799 		}
800 	} else
801 		vma->vm_private_data = PRIV_VMA_LOCKED;
802 
803 	xdata.frame = kdata.idx;
804 	xdata.nr_frames = kdata.num;
805 	set_xen_guest_handle(xdata.frame_list, pfns);
806 
807 	xen_preemptible_hcall_begin();
808 	rc = HYPERVISOR_memory_op(XENMEM_acquire_resource, &xdata);
809 	xen_preemptible_hcall_end();
810 
811 	if (rc)
812 		goto out;
813 
814 	if (IS_ENABLED(CONFIG_XEN_AUTO_XLATE) &&
815 	    xen_feature(XENFEAT_auto_translated_physmap)) {
816 		rc = xen_remap_vma_range(vma, kdata.addr, kdata.num << PAGE_SHIFT);
817 	} else {
818 		unsigned int domid =
819 			(xdata.flags & XENMEM_rsrc_acq_caller_owned) ?
820 			DOMID_SELF : kdata.dom;
821 		int num, *errs = (int *)pfns;
822 
823 		BUILD_BUG_ON(sizeof(*errs) > sizeof(*pfns));
824 		num = xen_remap_domain_mfn_array(vma,
825 						 kdata.addr & PAGE_MASK,
826 						 pfns, kdata.num, errs,
827 						 vma->vm_page_prot,
828 						 domid);
829 		if (num < 0)
830 			rc = num;
831 		else if (num != kdata.num) {
832 			unsigned int i;
833 
834 			for (i = 0; i < num; i++) {
835 				rc = errs[i];
836 				if (rc < 0)
837 					break;
838 			}
839 		} else
840 			rc = 0;
841 	}
842 
843 out:
844 	mmap_write_unlock(mm);
845 	kfree(pfns);
846 
847 	return rc;
848 }
849 
850 static long privcmd_ioctl_pcidev_get_gsi(struct file *file, void __user *udata)
851 {
852 #if defined(CONFIG_XEN_ACPI)
853 	int rc = -EINVAL;
854 	struct privcmd_pcidev_get_gsi kdata;
855 
856 	if (copy_from_user(&kdata, udata, sizeof(kdata)))
857 		return -EFAULT;
858 
859 	if (IS_REACHABLE(CONFIG_XEN_PCIDEV_BACKEND))
860 		rc = pcistub_get_gsi_from_sbdf(kdata.sbdf);
861 
862 	if (rc < 0)
863 		return rc;
864 
865 	kdata.gsi = rc;
866 	if (copy_to_user(udata, &kdata, sizeof(kdata)))
867 		return -EFAULT;
868 
869 	return 0;
870 #else
871 	return -EINVAL;
872 #endif
873 }
874 
875 #ifdef CONFIG_XEN_PRIVCMD_EVENTFD
876 /* Irqfd support */
877 static struct workqueue_struct *irqfd_cleanup_wq;
878 static DEFINE_SPINLOCK(irqfds_lock);
879 DEFINE_STATIC_SRCU(irqfds_srcu);
880 static LIST_HEAD(irqfds_list);
881 
882 struct privcmd_kernel_irqfd {
883 	struct xen_dm_op_buf xbufs;
884 	domid_t dom;
885 	bool error;
886 	struct eventfd_ctx *eventfd;
887 	struct work_struct shutdown;
888 	wait_queue_entry_t wait;
889 	struct list_head list;
890 	poll_table pt;
891 };
892 
893 static void irqfd_deactivate(struct privcmd_kernel_irqfd *kirqfd)
894 {
895 	lockdep_assert_held(&irqfds_lock);
896 
897 	list_del_init(&kirqfd->list);
898 	queue_work(irqfd_cleanup_wq, &kirqfd->shutdown);
899 }
900 
901 static void irqfd_shutdown(struct work_struct *work)
902 {
903 	struct privcmd_kernel_irqfd *kirqfd =
904 		container_of(work, struct privcmd_kernel_irqfd, shutdown);
905 	u64 cnt;
906 
907 	/* Make sure irqfd has been initialized in assign path */
908 	synchronize_srcu(&irqfds_srcu);
909 
910 	eventfd_ctx_remove_wait_queue(kirqfd->eventfd, &kirqfd->wait, &cnt);
911 	eventfd_ctx_put(kirqfd->eventfd);
912 	kfree(kirqfd);
913 }
914 
915 static void irqfd_inject(struct privcmd_kernel_irqfd *kirqfd)
916 {
917 	u64 cnt;
918 	long rc;
919 
920 	eventfd_ctx_do_read(kirqfd->eventfd, &cnt);
921 
922 	xen_preemptible_hcall_begin();
923 	rc = HYPERVISOR_dm_op(kirqfd->dom, 1, &kirqfd->xbufs);
924 	xen_preemptible_hcall_end();
925 
926 	/* Don't repeat the error message for consecutive failures */
927 	if (rc && !kirqfd->error) {
928 		pr_err("Failed to configure irq for guest domain: %d\n",
929 		       kirqfd->dom);
930 	}
931 
932 	kirqfd->error = rc;
933 }
934 
935 static int
936 irqfd_wakeup(wait_queue_entry_t *wait, unsigned int mode, int sync, void *key)
937 {
938 	struct privcmd_kernel_irqfd *kirqfd =
939 		container_of(wait, struct privcmd_kernel_irqfd, wait);
940 	__poll_t flags = key_to_poll(key);
941 
942 	if (flags & EPOLLIN)
943 		irqfd_inject(kirqfd);
944 
945 	if (flags & EPOLLHUP) {
946 		unsigned long flags;
947 
948 		spin_lock_irqsave(&irqfds_lock, flags);
949 		irqfd_deactivate(kirqfd);
950 		spin_unlock_irqrestore(&irqfds_lock, flags);
951 	}
952 
953 	return 0;
954 }
955 
956 static void
957 irqfd_poll_func(struct file *file, wait_queue_head_t *wqh, poll_table *pt)
958 {
959 	struct privcmd_kernel_irqfd *kirqfd =
960 		container_of(pt, struct privcmd_kernel_irqfd, pt);
961 
962 	add_wait_queue_priority(wqh, &kirqfd->wait);
963 }
964 
965 static int privcmd_irqfd_assign(struct privcmd_irqfd *irqfd)
966 {
967 	struct privcmd_kernel_irqfd *kirqfd, *tmp;
968 	unsigned long flags;
969 	__poll_t events;
970 	struct fd f;
971 	void *dm_op;
972 	int ret, idx;
973 
974 	kirqfd = kzalloc(sizeof(*kirqfd) + irqfd->size, GFP_KERNEL);
975 	if (!kirqfd)
976 		return -ENOMEM;
977 	dm_op = kirqfd + 1;
978 
979 	if (copy_from_user(dm_op, u64_to_user_ptr(irqfd->dm_op), irqfd->size)) {
980 		ret = -EFAULT;
981 		goto error_kfree;
982 	}
983 
984 	kirqfd->xbufs.size = irqfd->size;
985 	set_xen_guest_handle(kirqfd->xbufs.h, dm_op);
986 	kirqfd->dom = irqfd->dom;
987 	INIT_WORK(&kirqfd->shutdown, irqfd_shutdown);
988 
989 	f = fdget(irqfd->fd);
990 	if (!fd_file(f)) {
991 		ret = -EBADF;
992 		goto error_kfree;
993 	}
994 
995 	kirqfd->eventfd = eventfd_ctx_fileget(fd_file(f));
996 	if (IS_ERR(kirqfd->eventfd)) {
997 		ret = PTR_ERR(kirqfd->eventfd);
998 		goto error_fd_put;
999 	}
1000 
1001 	/*
1002 	 * Install our own custom wake-up handling so we are notified via a
1003 	 * callback whenever someone signals the underlying eventfd.
1004 	 */
1005 	init_waitqueue_func_entry(&kirqfd->wait, irqfd_wakeup);
1006 	init_poll_funcptr(&kirqfd->pt, irqfd_poll_func);
1007 
1008 	spin_lock_irqsave(&irqfds_lock, flags);
1009 
1010 	list_for_each_entry(tmp, &irqfds_list, list) {
1011 		if (kirqfd->eventfd == tmp->eventfd) {
1012 			ret = -EBUSY;
1013 			spin_unlock_irqrestore(&irqfds_lock, flags);
1014 			goto error_eventfd;
1015 		}
1016 	}
1017 
1018 	idx = srcu_read_lock(&irqfds_srcu);
1019 	list_add_tail(&kirqfd->list, &irqfds_list);
1020 	spin_unlock_irqrestore(&irqfds_lock, flags);
1021 
1022 	/*
1023 	 * Check if there was an event already pending on the eventfd before we
1024 	 * registered, and trigger it as if we didn't miss it.
1025 	 */
1026 	events = vfs_poll(fd_file(f), &kirqfd->pt);
1027 	if (events & EPOLLIN)
1028 		irqfd_inject(kirqfd);
1029 
1030 	srcu_read_unlock(&irqfds_srcu, idx);
1031 
1032 	/*
1033 	 * Do not drop the file until the kirqfd is fully initialized, otherwise
1034 	 * we might race against the EPOLLHUP.
1035 	 */
1036 	fdput(f);
1037 	return 0;
1038 
1039 error_eventfd:
1040 	eventfd_ctx_put(kirqfd->eventfd);
1041 
1042 error_fd_put:
1043 	fdput(f);
1044 
1045 error_kfree:
1046 	kfree(kirqfd);
1047 	return ret;
1048 }
1049 
1050 static int privcmd_irqfd_deassign(struct privcmd_irqfd *irqfd)
1051 {
1052 	struct privcmd_kernel_irqfd *kirqfd;
1053 	struct eventfd_ctx *eventfd;
1054 	unsigned long flags;
1055 
1056 	eventfd = eventfd_ctx_fdget(irqfd->fd);
1057 	if (IS_ERR(eventfd))
1058 		return PTR_ERR(eventfd);
1059 
1060 	spin_lock_irqsave(&irqfds_lock, flags);
1061 
1062 	list_for_each_entry(kirqfd, &irqfds_list, list) {
1063 		if (kirqfd->eventfd == eventfd) {
1064 			irqfd_deactivate(kirqfd);
1065 			break;
1066 		}
1067 	}
1068 
1069 	spin_unlock_irqrestore(&irqfds_lock, flags);
1070 
1071 	eventfd_ctx_put(eventfd);
1072 
1073 	/*
1074 	 * Block until we know all outstanding shutdown jobs have completed so
1075 	 * that we guarantee there will not be any more interrupts once this
1076 	 * deassign function returns.
1077 	 */
1078 	flush_workqueue(irqfd_cleanup_wq);
1079 
1080 	return 0;
1081 }
1082 
1083 static long privcmd_ioctl_irqfd(struct file *file, void __user *udata)
1084 {
1085 	struct privcmd_data *data = file->private_data;
1086 	struct privcmd_irqfd irqfd;
1087 
1088 	if (copy_from_user(&irqfd, udata, sizeof(irqfd)))
1089 		return -EFAULT;
1090 
1091 	/* No other flags should be set */
1092 	if (irqfd.flags & ~PRIVCMD_IRQFD_FLAG_DEASSIGN)
1093 		return -EINVAL;
1094 
1095 	/* If restriction is in place, check the domid matches */
1096 	if (data->domid != DOMID_INVALID && data->domid != irqfd.dom)
1097 		return -EPERM;
1098 
1099 	if (irqfd.flags & PRIVCMD_IRQFD_FLAG_DEASSIGN)
1100 		return privcmd_irqfd_deassign(&irqfd);
1101 
1102 	return privcmd_irqfd_assign(&irqfd);
1103 }
1104 
1105 static int privcmd_irqfd_init(void)
1106 {
1107 	irqfd_cleanup_wq = alloc_workqueue("privcmd-irqfd-cleanup", 0, 0);
1108 	if (!irqfd_cleanup_wq)
1109 		return -ENOMEM;
1110 
1111 	return 0;
1112 }
1113 
1114 static void privcmd_irqfd_exit(void)
1115 {
1116 	struct privcmd_kernel_irqfd *kirqfd, *tmp;
1117 	unsigned long flags;
1118 
1119 	spin_lock_irqsave(&irqfds_lock, flags);
1120 
1121 	list_for_each_entry_safe(kirqfd, tmp, &irqfds_list, list)
1122 		irqfd_deactivate(kirqfd);
1123 
1124 	spin_unlock_irqrestore(&irqfds_lock, flags);
1125 
1126 	destroy_workqueue(irqfd_cleanup_wq);
1127 }
1128 
1129 /* Ioeventfd Support */
1130 #define QUEUE_NOTIFY_VQ_MASK 0xFFFF
1131 
1132 static DEFINE_MUTEX(ioreq_lock);
1133 static LIST_HEAD(ioreq_list);
1134 
1135 /* per-eventfd structure */
1136 struct privcmd_kernel_ioeventfd {
1137 	struct eventfd_ctx *eventfd;
1138 	struct list_head list;
1139 	u64 addr;
1140 	unsigned int addr_len;
1141 	unsigned int vq;
1142 };
1143 
1144 /* per-guest CPU / port structure */
1145 struct ioreq_port {
1146 	int vcpu;
1147 	unsigned int port;
1148 	struct privcmd_kernel_ioreq *kioreq;
1149 };
1150 
1151 /* per-guest structure */
1152 struct privcmd_kernel_ioreq {
1153 	domid_t dom;
1154 	unsigned int vcpus;
1155 	u64 uioreq;
1156 	struct ioreq *ioreq;
1157 	spinlock_t lock; /* Protects ioeventfds list */
1158 	struct list_head ioeventfds;
1159 	struct list_head list;
1160 	struct ioreq_port ports[] __counted_by(vcpus);
1161 };
1162 
1163 static irqreturn_t ioeventfd_interrupt(int irq, void *dev_id)
1164 {
1165 	struct ioreq_port *port = dev_id;
1166 	struct privcmd_kernel_ioreq *kioreq = port->kioreq;
1167 	struct ioreq *ioreq = &kioreq->ioreq[port->vcpu];
1168 	struct privcmd_kernel_ioeventfd *kioeventfd;
1169 	unsigned int state = STATE_IOREQ_READY;
1170 
1171 	if (ioreq->state != STATE_IOREQ_READY ||
1172 	    ioreq->type != IOREQ_TYPE_COPY || ioreq->dir != IOREQ_WRITE)
1173 		return IRQ_NONE;
1174 
1175 	/*
1176 	 * We need a barrier, smp_mb(), here to ensure reads are finished before
1177 	 * `state` is updated. Since the lock implementation ensures that
1178 	 * appropriate barrier will be added anyway, we can avoid adding
1179 	 * explicit barrier here.
1180 	 *
1181 	 * Ideally we don't need to update `state` within the locks, but we do
1182 	 * that here to avoid adding explicit barrier.
1183 	 */
1184 
1185 	spin_lock(&kioreq->lock);
1186 	ioreq->state = STATE_IOREQ_INPROCESS;
1187 
1188 	list_for_each_entry(kioeventfd, &kioreq->ioeventfds, list) {
1189 		if (ioreq->addr == kioeventfd->addr + VIRTIO_MMIO_QUEUE_NOTIFY &&
1190 		    ioreq->size == kioeventfd->addr_len &&
1191 		    (ioreq->data & QUEUE_NOTIFY_VQ_MASK) == kioeventfd->vq) {
1192 			eventfd_signal(kioeventfd->eventfd);
1193 			state = STATE_IORESP_READY;
1194 			break;
1195 		}
1196 	}
1197 	spin_unlock(&kioreq->lock);
1198 
1199 	/*
1200 	 * We need a barrier, smp_mb(), here to ensure writes are finished
1201 	 * before `state` is updated. Since the lock implementation ensures that
1202 	 * appropriate barrier will be added anyway, we can avoid adding
1203 	 * explicit barrier here.
1204 	 */
1205 
1206 	ioreq->state = state;
1207 
1208 	if (state == STATE_IORESP_READY) {
1209 		notify_remote_via_evtchn(port->port);
1210 		return IRQ_HANDLED;
1211 	}
1212 
1213 	return IRQ_NONE;
1214 }
1215 
1216 static void ioreq_free(struct privcmd_kernel_ioreq *kioreq)
1217 {
1218 	struct ioreq_port *ports = kioreq->ports;
1219 	int i;
1220 
1221 	lockdep_assert_held(&ioreq_lock);
1222 
1223 	list_del(&kioreq->list);
1224 
1225 	for (i = kioreq->vcpus - 1; i >= 0; i--)
1226 		unbind_from_irqhandler(irq_from_evtchn(ports[i].port), &ports[i]);
1227 
1228 	kfree(kioreq);
1229 }
1230 
1231 static
1232 struct privcmd_kernel_ioreq *alloc_ioreq(struct privcmd_ioeventfd *ioeventfd)
1233 {
1234 	struct privcmd_kernel_ioreq *kioreq;
1235 	struct mm_struct *mm = current->mm;
1236 	struct vm_area_struct *vma;
1237 	struct page **pages;
1238 	unsigned int *ports;
1239 	int ret, size, i;
1240 
1241 	lockdep_assert_held(&ioreq_lock);
1242 
1243 	size = struct_size(kioreq, ports, ioeventfd->vcpus);
1244 	kioreq = kzalloc(size, GFP_KERNEL);
1245 	if (!kioreq)
1246 		return ERR_PTR(-ENOMEM);
1247 
1248 	kioreq->dom = ioeventfd->dom;
1249 	kioreq->vcpus = ioeventfd->vcpus;
1250 	kioreq->uioreq = ioeventfd->ioreq;
1251 	spin_lock_init(&kioreq->lock);
1252 	INIT_LIST_HEAD(&kioreq->ioeventfds);
1253 
1254 	/* The memory for ioreq server must have been mapped earlier */
1255 	mmap_write_lock(mm);
1256 	vma = find_vma(mm, (unsigned long)ioeventfd->ioreq);
1257 	if (!vma) {
1258 		pr_err("Failed to find vma for ioreq page!\n");
1259 		mmap_write_unlock(mm);
1260 		ret = -EFAULT;
1261 		goto error_kfree;
1262 	}
1263 
1264 	pages = vma->vm_private_data;
1265 	kioreq->ioreq = (struct ioreq *)(page_to_virt(pages[0]));
1266 	mmap_write_unlock(mm);
1267 
1268 	ports = memdup_array_user(u64_to_user_ptr(ioeventfd->ports),
1269 				  kioreq->vcpus, sizeof(*ports));
1270 	if (IS_ERR(ports)) {
1271 		ret = PTR_ERR(ports);
1272 		goto error_kfree;
1273 	}
1274 
1275 	for (i = 0; i < kioreq->vcpus; i++) {
1276 		kioreq->ports[i].vcpu = i;
1277 		kioreq->ports[i].port = ports[i];
1278 		kioreq->ports[i].kioreq = kioreq;
1279 
1280 		ret = bind_evtchn_to_irqhandler_lateeoi(ports[i],
1281 				ioeventfd_interrupt, IRQF_SHARED, "ioeventfd",
1282 				&kioreq->ports[i]);
1283 		if (ret < 0)
1284 			goto error_unbind;
1285 	}
1286 
1287 	kfree(ports);
1288 
1289 	list_add_tail(&kioreq->list, &ioreq_list);
1290 
1291 	return kioreq;
1292 
1293 error_unbind:
1294 	while (--i >= 0)
1295 		unbind_from_irqhandler(irq_from_evtchn(ports[i]), &kioreq->ports[i]);
1296 
1297 	kfree(ports);
1298 error_kfree:
1299 	kfree(kioreq);
1300 	return ERR_PTR(ret);
1301 }
1302 
1303 static struct privcmd_kernel_ioreq *
1304 get_ioreq(struct privcmd_ioeventfd *ioeventfd, struct eventfd_ctx *eventfd)
1305 {
1306 	struct privcmd_kernel_ioreq *kioreq;
1307 	unsigned long flags;
1308 
1309 	list_for_each_entry(kioreq, &ioreq_list, list) {
1310 		struct privcmd_kernel_ioeventfd *kioeventfd;
1311 
1312 		/*
1313 		 * kioreq fields can be accessed here without a lock as they are
1314 		 * never updated after being added to the ioreq_list.
1315 		 */
1316 		if (kioreq->uioreq != ioeventfd->ioreq) {
1317 			continue;
1318 		} else if (kioreq->dom != ioeventfd->dom ||
1319 			   kioreq->vcpus != ioeventfd->vcpus) {
1320 			pr_err("Invalid ioeventfd configuration mismatch, dom (%u vs %u), vcpus (%u vs %u)\n",
1321 			       kioreq->dom, ioeventfd->dom, kioreq->vcpus,
1322 			       ioeventfd->vcpus);
1323 			return ERR_PTR(-EINVAL);
1324 		}
1325 
1326 		/* Look for a duplicate eventfd for the same guest */
1327 		spin_lock_irqsave(&kioreq->lock, flags);
1328 		list_for_each_entry(kioeventfd, &kioreq->ioeventfds, list) {
1329 			if (eventfd == kioeventfd->eventfd) {
1330 				spin_unlock_irqrestore(&kioreq->lock, flags);
1331 				return ERR_PTR(-EBUSY);
1332 			}
1333 		}
1334 		spin_unlock_irqrestore(&kioreq->lock, flags);
1335 
1336 		return kioreq;
1337 	}
1338 
1339 	/* Matching kioreq isn't found, allocate a new one */
1340 	return alloc_ioreq(ioeventfd);
1341 }
1342 
1343 static void ioeventfd_free(struct privcmd_kernel_ioeventfd *kioeventfd)
1344 {
1345 	list_del(&kioeventfd->list);
1346 	eventfd_ctx_put(kioeventfd->eventfd);
1347 	kfree(kioeventfd);
1348 }
1349 
1350 static int privcmd_ioeventfd_assign(struct privcmd_ioeventfd *ioeventfd)
1351 {
1352 	struct privcmd_kernel_ioeventfd *kioeventfd;
1353 	struct privcmd_kernel_ioreq *kioreq;
1354 	unsigned long flags;
1355 	struct fd f;
1356 	int ret;
1357 
1358 	/* Check for range overflow */
1359 	if (ioeventfd->addr + ioeventfd->addr_len < ioeventfd->addr)
1360 		return -EINVAL;
1361 
1362 	/* Vhost requires us to support length 1, 2, 4, and 8 */
1363 	if (!(ioeventfd->addr_len == 1 || ioeventfd->addr_len == 2 ||
1364 	      ioeventfd->addr_len == 4 || ioeventfd->addr_len == 8))
1365 		return -EINVAL;
1366 
1367 	/* 4096 vcpus limit enough ? */
1368 	if (!ioeventfd->vcpus || ioeventfd->vcpus > 4096)
1369 		return -EINVAL;
1370 
1371 	kioeventfd = kzalloc(sizeof(*kioeventfd), GFP_KERNEL);
1372 	if (!kioeventfd)
1373 		return -ENOMEM;
1374 
1375 	f = fdget(ioeventfd->event_fd);
1376 	if (!fd_file(f)) {
1377 		ret = -EBADF;
1378 		goto error_kfree;
1379 	}
1380 
1381 	kioeventfd->eventfd = eventfd_ctx_fileget(fd_file(f));
1382 	fdput(f);
1383 
1384 	if (IS_ERR(kioeventfd->eventfd)) {
1385 		ret = PTR_ERR(kioeventfd->eventfd);
1386 		goto error_kfree;
1387 	}
1388 
1389 	kioeventfd->addr = ioeventfd->addr;
1390 	kioeventfd->addr_len = ioeventfd->addr_len;
1391 	kioeventfd->vq = ioeventfd->vq;
1392 
1393 	mutex_lock(&ioreq_lock);
1394 	kioreq = get_ioreq(ioeventfd, kioeventfd->eventfd);
1395 	if (IS_ERR(kioreq)) {
1396 		mutex_unlock(&ioreq_lock);
1397 		ret = PTR_ERR(kioreq);
1398 		goto error_eventfd;
1399 	}
1400 
1401 	spin_lock_irqsave(&kioreq->lock, flags);
1402 	list_add_tail(&kioeventfd->list, &kioreq->ioeventfds);
1403 	spin_unlock_irqrestore(&kioreq->lock, flags);
1404 
1405 	mutex_unlock(&ioreq_lock);
1406 
1407 	return 0;
1408 
1409 error_eventfd:
1410 	eventfd_ctx_put(kioeventfd->eventfd);
1411 
1412 error_kfree:
1413 	kfree(kioeventfd);
1414 	return ret;
1415 }
1416 
1417 static int privcmd_ioeventfd_deassign(struct privcmd_ioeventfd *ioeventfd)
1418 {
1419 	struct privcmd_kernel_ioreq *kioreq, *tkioreq;
1420 	struct eventfd_ctx *eventfd;
1421 	unsigned long flags;
1422 	int ret = 0;
1423 
1424 	eventfd = eventfd_ctx_fdget(ioeventfd->event_fd);
1425 	if (IS_ERR(eventfd))
1426 		return PTR_ERR(eventfd);
1427 
1428 	mutex_lock(&ioreq_lock);
1429 	list_for_each_entry_safe(kioreq, tkioreq, &ioreq_list, list) {
1430 		struct privcmd_kernel_ioeventfd *kioeventfd, *tmp;
1431 		/*
1432 		 * kioreq fields can be accessed here without a lock as they are
1433 		 * never updated after being added to the ioreq_list.
1434 		 */
1435 		if (kioreq->dom != ioeventfd->dom ||
1436 		    kioreq->uioreq != ioeventfd->ioreq ||
1437 		    kioreq->vcpus != ioeventfd->vcpus)
1438 			continue;
1439 
1440 		spin_lock_irqsave(&kioreq->lock, flags);
1441 		list_for_each_entry_safe(kioeventfd, tmp, &kioreq->ioeventfds, list) {
1442 			if (eventfd == kioeventfd->eventfd) {
1443 				ioeventfd_free(kioeventfd);
1444 				spin_unlock_irqrestore(&kioreq->lock, flags);
1445 
1446 				if (list_empty(&kioreq->ioeventfds))
1447 					ioreq_free(kioreq);
1448 				goto unlock;
1449 			}
1450 		}
1451 		spin_unlock_irqrestore(&kioreq->lock, flags);
1452 		break;
1453 	}
1454 
1455 	pr_err("Ioeventfd isn't already assigned, dom: %u, addr: %llu\n",
1456 	       ioeventfd->dom, ioeventfd->addr);
1457 	ret = -ENODEV;
1458 
1459 unlock:
1460 	mutex_unlock(&ioreq_lock);
1461 	eventfd_ctx_put(eventfd);
1462 
1463 	return ret;
1464 }
1465 
1466 static long privcmd_ioctl_ioeventfd(struct file *file, void __user *udata)
1467 {
1468 	struct privcmd_data *data = file->private_data;
1469 	struct privcmd_ioeventfd ioeventfd;
1470 
1471 	if (copy_from_user(&ioeventfd, udata, sizeof(ioeventfd)))
1472 		return -EFAULT;
1473 
1474 	/* No other flags should be set */
1475 	if (ioeventfd.flags & ~PRIVCMD_IOEVENTFD_FLAG_DEASSIGN)
1476 		return -EINVAL;
1477 
1478 	/* If restriction is in place, check the domid matches */
1479 	if (data->domid != DOMID_INVALID && data->domid != ioeventfd.dom)
1480 		return -EPERM;
1481 
1482 	if (ioeventfd.flags & PRIVCMD_IOEVENTFD_FLAG_DEASSIGN)
1483 		return privcmd_ioeventfd_deassign(&ioeventfd);
1484 
1485 	return privcmd_ioeventfd_assign(&ioeventfd);
1486 }
1487 
1488 static void privcmd_ioeventfd_exit(void)
1489 {
1490 	struct privcmd_kernel_ioreq *kioreq, *tmp;
1491 	unsigned long flags;
1492 
1493 	mutex_lock(&ioreq_lock);
1494 	list_for_each_entry_safe(kioreq, tmp, &ioreq_list, list) {
1495 		struct privcmd_kernel_ioeventfd *kioeventfd, *tmp;
1496 
1497 		spin_lock_irqsave(&kioreq->lock, flags);
1498 		list_for_each_entry_safe(kioeventfd, tmp, &kioreq->ioeventfds, list)
1499 			ioeventfd_free(kioeventfd);
1500 		spin_unlock_irqrestore(&kioreq->lock, flags);
1501 
1502 		ioreq_free(kioreq);
1503 	}
1504 	mutex_unlock(&ioreq_lock);
1505 }
1506 #else
1507 static inline long privcmd_ioctl_irqfd(struct file *file, void __user *udata)
1508 {
1509 	return -EOPNOTSUPP;
1510 }
1511 
1512 static inline int privcmd_irqfd_init(void)
1513 {
1514 	return 0;
1515 }
1516 
1517 static inline void privcmd_irqfd_exit(void)
1518 {
1519 }
1520 
1521 static inline long privcmd_ioctl_ioeventfd(struct file *file, void __user *udata)
1522 {
1523 	return -EOPNOTSUPP;
1524 }
1525 
1526 static inline void privcmd_ioeventfd_exit(void)
1527 {
1528 }
1529 #endif /* CONFIG_XEN_PRIVCMD_EVENTFD */
1530 
1531 static long privcmd_ioctl(struct file *file,
1532 			  unsigned int cmd, unsigned long data)
1533 {
1534 	int ret = -ENOTTY;
1535 	void __user *udata = (void __user *) data;
1536 
1537 	switch (cmd) {
1538 	case IOCTL_PRIVCMD_HYPERCALL:
1539 		ret = privcmd_ioctl_hypercall(file, udata);
1540 		break;
1541 
1542 	case IOCTL_PRIVCMD_MMAP:
1543 		ret = privcmd_ioctl_mmap(file, udata);
1544 		break;
1545 
1546 	case IOCTL_PRIVCMD_MMAPBATCH:
1547 		ret = privcmd_ioctl_mmap_batch(file, udata, 1);
1548 		break;
1549 
1550 	case IOCTL_PRIVCMD_MMAPBATCH_V2:
1551 		ret = privcmd_ioctl_mmap_batch(file, udata, 2);
1552 		break;
1553 
1554 	case IOCTL_PRIVCMD_DM_OP:
1555 		ret = privcmd_ioctl_dm_op(file, udata);
1556 		break;
1557 
1558 	case IOCTL_PRIVCMD_RESTRICT:
1559 		ret = privcmd_ioctl_restrict(file, udata);
1560 		break;
1561 
1562 	case IOCTL_PRIVCMD_MMAP_RESOURCE:
1563 		ret = privcmd_ioctl_mmap_resource(file, udata);
1564 		break;
1565 
1566 	case IOCTL_PRIVCMD_IRQFD:
1567 		ret = privcmd_ioctl_irqfd(file, udata);
1568 		break;
1569 
1570 	case IOCTL_PRIVCMD_IOEVENTFD:
1571 		ret = privcmd_ioctl_ioeventfd(file, udata);
1572 		break;
1573 
1574 	case IOCTL_PRIVCMD_PCIDEV_GET_GSI:
1575 		ret = privcmd_ioctl_pcidev_get_gsi(file, udata);
1576 		break;
1577 
1578 	default:
1579 		break;
1580 	}
1581 
1582 	return ret;
1583 }
1584 
1585 static int privcmd_open(struct inode *ino, struct file *file)
1586 {
1587 	struct privcmd_data *data = kzalloc(sizeof(*data), GFP_KERNEL);
1588 
1589 	if (!data)
1590 		return -ENOMEM;
1591 
1592 	/* DOMID_INVALID implies no restriction */
1593 	data->domid = DOMID_INVALID;
1594 
1595 	file->private_data = data;
1596 	return 0;
1597 }
1598 
1599 static int privcmd_release(struct inode *ino, struct file *file)
1600 {
1601 	struct privcmd_data *data = file->private_data;
1602 
1603 	kfree(data);
1604 	return 0;
1605 }
1606 
1607 static void privcmd_close(struct vm_area_struct *vma)
1608 {
1609 	struct page **pages = vma->vm_private_data;
1610 	int numpgs = vma_pages(vma);
1611 	int numgfns = (vma->vm_end - vma->vm_start) >> XEN_PAGE_SHIFT;
1612 	int rc;
1613 
1614 	if (!xen_feature(XENFEAT_auto_translated_physmap) || !numpgs || !pages)
1615 		return;
1616 
1617 	rc = xen_unmap_domain_gfn_range(vma, numgfns, pages);
1618 	if (rc == 0)
1619 		xen_free_unpopulated_pages(numpgs, pages);
1620 	else
1621 		pr_crit("unable to unmap MFN range: leaking %d pages. rc=%d\n",
1622 			numpgs, rc);
1623 	kvfree(pages);
1624 }
1625 
1626 static vm_fault_t privcmd_fault(struct vm_fault *vmf)
1627 {
1628 	printk(KERN_DEBUG "privcmd_fault: vma=%p %lx-%lx, pgoff=%lx, uv=%p\n",
1629 	       vmf->vma, vmf->vma->vm_start, vmf->vma->vm_end,
1630 	       vmf->pgoff, (void *)vmf->address);
1631 
1632 	return VM_FAULT_SIGBUS;
1633 }
1634 
1635 static const struct vm_operations_struct privcmd_vm_ops = {
1636 	.close = privcmd_close,
1637 	.fault = privcmd_fault
1638 };
1639 
1640 static int privcmd_mmap(struct file *file, struct vm_area_struct *vma)
1641 {
1642 	/* DONTCOPY is essential for Xen because copy_page_range doesn't know
1643 	 * how to recreate these mappings */
1644 	vm_flags_set(vma, VM_IO | VM_PFNMAP | VM_DONTCOPY |
1645 			 VM_DONTEXPAND | VM_DONTDUMP);
1646 	vma->vm_ops = &privcmd_vm_ops;
1647 	vma->vm_private_data = NULL;
1648 
1649 	return 0;
1650 }
1651 
1652 /*
1653  * For MMAPBATCH*. This allows asserting the singleshot mapping
1654  * on a per pfn/pte basis. Mapping calls that fail with ENOENT
1655  * can be then retried until success.
1656  */
1657 static int is_mapped_fn(pte_t *pte, unsigned long addr, void *data)
1658 {
1659 	return pte_none(ptep_get(pte)) ? 0 : -EBUSY;
1660 }
1661 
1662 static int privcmd_vma_range_is_mapped(
1663 	           struct vm_area_struct *vma,
1664 	           unsigned long addr,
1665 	           unsigned long nr_pages)
1666 {
1667 	return apply_to_page_range(vma->vm_mm, addr, nr_pages << PAGE_SHIFT,
1668 				   is_mapped_fn, NULL) != 0;
1669 }
1670 
1671 const struct file_operations xen_privcmd_fops = {
1672 	.owner = THIS_MODULE,
1673 	.unlocked_ioctl = privcmd_ioctl,
1674 	.open = privcmd_open,
1675 	.release = privcmd_release,
1676 	.mmap = privcmd_mmap,
1677 };
1678 EXPORT_SYMBOL_GPL(xen_privcmd_fops);
1679 
1680 static struct miscdevice privcmd_dev = {
1681 	.minor = MISC_DYNAMIC_MINOR,
1682 	.name = "xen/privcmd",
1683 	.fops = &xen_privcmd_fops,
1684 };
1685 
1686 static int __init privcmd_init(void)
1687 {
1688 	int err;
1689 
1690 	if (!xen_domain())
1691 		return -ENODEV;
1692 
1693 	err = misc_register(&privcmd_dev);
1694 	if (err != 0) {
1695 		pr_err("Could not register Xen privcmd device\n");
1696 		return err;
1697 	}
1698 
1699 	err = misc_register(&xen_privcmdbuf_dev);
1700 	if (err != 0) {
1701 		pr_err("Could not register Xen hypercall-buf device\n");
1702 		goto err_privcmdbuf;
1703 	}
1704 
1705 	err = privcmd_irqfd_init();
1706 	if (err != 0) {
1707 		pr_err("irqfd init failed\n");
1708 		goto err_irqfd;
1709 	}
1710 
1711 	return 0;
1712 
1713 err_irqfd:
1714 	misc_deregister(&xen_privcmdbuf_dev);
1715 err_privcmdbuf:
1716 	misc_deregister(&privcmd_dev);
1717 	return err;
1718 }
1719 
1720 static void __exit privcmd_exit(void)
1721 {
1722 	privcmd_ioeventfd_exit();
1723 	privcmd_irqfd_exit();
1724 	misc_deregister(&privcmd_dev);
1725 	misc_deregister(&xen_privcmdbuf_dev);
1726 }
1727 
1728 module_init(privcmd_init);
1729 module_exit(privcmd_exit);
1730