Loading...
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/kernel.h>
13#include <linux/module.h>
14#include <linux/sched.h>
15#include <linux/slab.h>
16#include <linux/string.h>
17#include <linux/errno.h>
18#include <linux/mm.h>
19#include <linux/mman.h>
20#include <linux/uaccess.h>
21#include <linux/swap.h>
22#include <linux/highmem.h>
23#include <linux/pagemap.h>
24#include <linux/seq_file.h>
25#include <linux/miscdevice.h>
26#include <linux/moduleparam.h>
27
28#include <asm/pgalloc.h>
29#include <asm/pgtable.h>
30#include <asm/tlb.h>
31#include <asm/xen/hypervisor.h>
32#include <asm/xen/hypercall.h>
33
34#include <xen/xen.h>
35#include <xen/privcmd.h>
36#include <xen/interface/xen.h>
37#include <xen/interface/memory.h>
38#include <xen/interface/hvm/dm_op.h>
39#include <xen/features.h>
40#include <xen/page.h>
41#include <xen/xen-ops.h>
42#include <xen/balloon.h>
43
44#include "privcmd.h"
45
46MODULE_LICENSE("GPL");
47
48#define PRIV_VMA_LOCKED ((void *)1)
49
50static unsigned int privcmd_dm_op_max_num = 16;
51module_param_named(dm_op_max_nr_bufs, privcmd_dm_op_max_num, uint, 0644);
52MODULE_PARM_DESC(dm_op_max_nr_bufs,
53 "Maximum number of buffers per dm_op hypercall");
54
55static unsigned int privcmd_dm_op_buf_max_size = 4096;
56module_param_named(dm_op_buf_max_size, privcmd_dm_op_buf_max_size, uint,
57 0644);
58MODULE_PARM_DESC(dm_op_buf_max_size,
59 "Maximum size of a dm_op hypercall buffer");
60
61struct privcmd_data {
62 domid_t domid;
63};
64
65static int privcmd_vma_range_is_mapped(
66 struct vm_area_struct *vma,
67 unsigned long addr,
68 unsigned long nr_pages);
69
70static long privcmd_ioctl_hypercall(struct file *file, void __user *udata)
71{
72 struct privcmd_data *data = file->private_data;
73 struct privcmd_hypercall hypercall;
74 long ret;
75
76 /* Disallow arbitrary hypercalls if restricted */
77 if (data->domid != DOMID_INVALID)
78 return -EPERM;
79
80 if (copy_from_user(&hypercall, udata, sizeof(hypercall)))
81 return -EFAULT;
82
83 xen_preemptible_hcall_begin();
84 ret = privcmd_call(hypercall.op,
85 hypercall.arg[0], hypercall.arg[1],
86 hypercall.arg[2], hypercall.arg[3],
87 hypercall.arg[4]);
88 xen_preemptible_hcall_end();
89
90 return ret;
91}
92
93static void free_page_list(struct list_head *pages)
94{
95 struct page *p, *n;
96
97 list_for_each_entry_safe(p, n, pages, lru)
98 __free_page(p);
99
100 INIT_LIST_HEAD(pages);
101}
102
103/*
104 * Given an array of items in userspace, return a list of pages
105 * containing the data. If copying fails, either because of memory
106 * allocation failure or a problem reading user memory, return an
107 * error code; its up to the caller to dispose of any partial list.
108 */
109static int gather_array(struct list_head *pagelist,
110 unsigned nelem, size_t size,
111 const void __user *data)
112{
113 unsigned pageidx;
114 void *pagedata;
115 int ret;
116
117 if (size > PAGE_SIZE)
118 return 0;
119
120 pageidx = PAGE_SIZE;
121 pagedata = NULL; /* quiet, gcc */
122 while (nelem--) {
123 if (pageidx > PAGE_SIZE-size) {
124 struct page *page = alloc_page(GFP_KERNEL);
125
126 ret = -ENOMEM;
127 if (page == NULL)
128 goto fail;
129
130 pagedata = page_address(page);
131
132 list_add_tail(&page->lru, pagelist);
133 pageidx = 0;
134 }
135
136 ret = -EFAULT;
137 if (copy_from_user(pagedata + pageidx, data, size))
138 goto fail;
139
140 data += size;
141 pageidx += size;
142 }
143
144 ret = 0;
145
146fail:
147 return ret;
148}
149
150/*
151 * Call function "fn" on each element of the array fragmented
152 * over a list of pages.
153 */
154static int traverse_pages(unsigned nelem, size_t size,
155 struct list_head *pos,
156 int (*fn)(void *data, void *state),
157 void *state)
158{
159 void *pagedata;
160 unsigned pageidx;
161 int ret = 0;
162
163 BUG_ON(size > PAGE_SIZE);
164
165 pageidx = PAGE_SIZE;
166 pagedata = NULL; /* hush, gcc */
167
168 while (nelem--) {
169 if (pageidx > PAGE_SIZE-size) {
170 struct page *page;
171 pos = pos->next;
172 page = list_entry(pos, struct page, lru);
173 pagedata = page_address(page);
174 pageidx = 0;
175 }
176
177 ret = (*fn)(pagedata + pageidx, state);
178 if (ret)
179 break;
180 pageidx += size;
181 }
182
183 return ret;
184}
185
186/*
187 * Similar to traverse_pages, but use each page as a "block" of
188 * data to be processed as one unit.
189 */
190static int traverse_pages_block(unsigned nelem, size_t size,
191 struct list_head *pos,
192 int (*fn)(void *data, int nr, void *state),
193 void *state)
194{
195 void *pagedata;
196 int ret = 0;
197
198 BUG_ON(size > PAGE_SIZE);
199
200 while (nelem) {
201 int nr = (PAGE_SIZE/size);
202 struct page *page;
203 if (nr > nelem)
204 nr = nelem;
205 pos = pos->next;
206 page = list_entry(pos, struct page, lru);
207 pagedata = page_address(page);
208 ret = (*fn)(pagedata, nr, state);
209 if (ret)
210 break;
211 nelem -= nr;
212 }
213
214 return ret;
215}
216
217struct mmap_gfn_state {
218 unsigned long va;
219 struct vm_area_struct *vma;
220 domid_t domain;
221};
222
223static int mmap_gfn_range(void *data, void *state)
224{
225 struct privcmd_mmap_entry *msg = data;
226 struct mmap_gfn_state *st = state;
227 struct vm_area_struct *vma = st->vma;
228 int rc;
229
230 /* Do not allow range to wrap the address space. */
231 if ((msg->npages > (LONG_MAX >> PAGE_SHIFT)) ||
232 ((unsigned long)(msg->npages << PAGE_SHIFT) >= -st->va))
233 return -EINVAL;
234
235 /* Range chunks must be contiguous in va space. */
236 if ((msg->va != st->va) ||
237 ((msg->va+(msg->npages<<PAGE_SHIFT)) > vma->vm_end))
238 return -EINVAL;
239
240 rc = xen_remap_domain_gfn_range(vma,
241 msg->va & PAGE_MASK,
242 msg->mfn, msg->npages,
243 vma->vm_page_prot,
244 st->domain, NULL);
245 if (rc < 0)
246 return rc;
247
248 st->va += msg->npages << PAGE_SHIFT;
249
250 return 0;
251}
252
253static long privcmd_ioctl_mmap(struct file *file, void __user *udata)
254{
255 struct privcmd_data *data = file->private_data;
256 struct privcmd_mmap mmapcmd;
257 struct mm_struct *mm = current->mm;
258 struct vm_area_struct *vma;
259 int rc;
260 LIST_HEAD(pagelist);
261 struct mmap_gfn_state state;
262
263 /* We only support privcmd_ioctl_mmap_batch for auto translated. */
264 if (xen_feature(XENFEAT_auto_translated_physmap))
265 return -ENOSYS;
266
267 if (copy_from_user(&mmapcmd, udata, sizeof(mmapcmd)))
268 return -EFAULT;
269
270 /* If restriction is in place, check the domid matches */
271 if (data->domid != DOMID_INVALID && data->domid != mmapcmd.dom)
272 return -EPERM;
273
274 rc = gather_array(&pagelist,
275 mmapcmd.num, sizeof(struct privcmd_mmap_entry),
276 mmapcmd.entry);
277
278 if (rc || list_empty(&pagelist))
279 goto out;
280
281 down_write(&mm->mmap_sem);
282
283 {
284 struct page *page = list_first_entry(&pagelist,
285 struct page, lru);
286 struct privcmd_mmap_entry *msg = page_address(page);
287
288 vma = find_vma(mm, msg->va);
289 rc = -EINVAL;
290
291 if (!vma || (msg->va != vma->vm_start) || vma->vm_private_data)
292 goto out_up;
293 vma->vm_private_data = PRIV_VMA_LOCKED;
294 }
295
296 state.va = vma->vm_start;
297 state.vma = vma;
298 state.domain = mmapcmd.dom;
299
300 rc = traverse_pages(mmapcmd.num, sizeof(struct privcmd_mmap_entry),
301 &pagelist,
302 mmap_gfn_range, &state);
303
304
305out_up:
306 up_write(&mm->mmap_sem);
307
308out:
309 free_page_list(&pagelist);
310
311 return rc;
312}
313
314struct mmap_batch_state {
315 domid_t domain;
316 unsigned long va;
317 struct vm_area_struct *vma;
318 int index;
319 /* A tristate:
320 * 0 for no errors
321 * 1 if at least one error has happened (and no
322 * -ENOENT errors have happened)
323 * -ENOENT if at least 1 -ENOENT has happened.
324 */
325 int global_error;
326 int version;
327
328 /* User-space gfn array to store errors in the second pass for V1. */
329 xen_pfn_t __user *user_gfn;
330 /* User-space int array to store errors in the second pass for V2. */
331 int __user *user_err;
332};
333
334/* auto translated dom0 note: if domU being created is PV, then gfn is
335 * mfn(addr on bus). If it's auto xlated, then gfn is pfn (input to HAP).
336 */
337static int mmap_batch_fn(void *data, int nr, void *state)
338{
339 xen_pfn_t *gfnp = data;
340 struct mmap_batch_state *st = state;
341 struct vm_area_struct *vma = st->vma;
342 struct page **pages = vma->vm_private_data;
343 struct page **cur_pages = NULL;
344 int ret;
345
346 if (xen_feature(XENFEAT_auto_translated_physmap))
347 cur_pages = &pages[st->index];
348
349 BUG_ON(nr < 0);
350 ret = xen_remap_domain_gfn_array(st->vma, st->va & PAGE_MASK, gfnp, nr,
351 (int *)gfnp, st->vma->vm_page_prot,
352 st->domain, cur_pages);
353
354 /* Adjust the global_error? */
355 if (ret != nr) {
356 if (ret == -ENOENT)
357 st->global_error = -ENOENT;
358 else {
359 /* Record that at least one error has happened. */
360 if (st->global_error == 0)
361 st->global_error = 1;
362 }
363 }
364 st->va += XEN_PAGE_SIZE * nr;
365 st->index += nr / XEN_PFN_PER_PAGE;
366
367 return 0;
368}
369
370static int mmap_return_error(int err, struct mmap_batch_state *st)
371{
372 int ret;
373
374 if (st->version == 1) {
375 if (err) {
376 xen_pfn_t gfn;
377
378 ret = get_user(gfn, st->user_gfn);
379 if (ret < 0)
380 return ret;
381 /*
382 * V1 encodes the error codes in the 32bit top
383 * nibble of the gfn (with its known
384 * limitations vis-a-vis 64 bit callers).
385 */
386 gfn |= (err == -ENOENT) ?
387 PRIVCMD_MMAPBATCH_PAGED_ERROR :
388 PRIVCMD_MMAPBATCH_MFN_ERROR;
389 return __put_user(gfn, st->user_gfn++);
390 } else
391 st->user_gfn++;
392 } else { /* st->version == 2 */
393 if (err)
394 return __put_user(err, st->user_err++);
395 else
396 st->user_err++;
397 }
398
399 return 0;
400}
401
402static int mmap_return_errors(void *data, int nr, void *state)
403{
404 struct mmap_batch_state *st = state;
405 int *errs = data;
406 int i;
407 int ret;
408
409 for (i = 0; i < nr; i++) {
410 ret = mmap_return_error(errs[i], st);
411 if (ret < 0)
412 return ret;
413 }
414 return 0;
415}
416
417/* Allocate pfns that are then mapped with gfns from foreign domid. Update
418 * the vma with the page info to use later.
419 * Returns: 0 if success, otherwise -errno
420 */
421static int alloc_empty_pages(struct vm_area_struct *vma, int numpgs)
422{
423 int rc;
424 struct page **pages;
425
426 pages = kcalloc(numpgs, sizeof(pages[0]), GFP_KERNEL);
427 if (pages == NULL)
428 return -ENOMEM;
429
430 rc = alloc_xenballooned_pages(numpgs, pages);
431 if (rc != 0) {
432 pr_warn("%s Could not alloc %d pfns rc:%d\n", __func__,
433 numpgs, rc);
434 kfree(pages);
435 return -ENOMEM;
436 }
437 BUG_ON(vma->vm_private_data != NULL);
438 vma->vm_private_data = pages;
439
440 return 0;
441}
442
443static const struct vm_operations_struct privcmd_vm_ops;
444
445static long privcmd_ioctl_mmap_batch(
446 struct file *file, void __user *udata, int version)
447{
448 struct privcmd_data *data = file->private_data;
449 int ret;
450 struct privcmd_mmapbatch_v2 m;
451 struct mm_struct *mm = current->mm;
452 struct vm_area_struct *vma;
453 unsigned long nr_pages;
454 LIST_HEAD(pagelist);
455 struct mmap_batch_state state;
456
457 switch (version) {
458 case 1:
459 if (copy_from_user(&m, udata, sizeof(struct privcmd_mmapbatch)))
460 return -EFAULT;
461 /* Returns per-frame error in m.arr. */
462 m.err = NULL;
463 if (!access_ok(m.arr, m.num * sizeof(*m.arr)))
464 return -EFAULT;
465 break;
466 case 2:
467 if (copy_from_user(&m, udata, sizeof(struct privcmd_mmapbatch_v2)))
468 return -EFAULT;
469 /* Returns per-frame error code in m.err. */
470 if (!access_ok(m.err, m.num * (sizeof(*m.err))))
471 return -EFAULT;
472 break;
473 default:
474 return -EINVAL;
475 }
476
477 /* If restriction is in place, check the domid matches */
478 if (data->domid != DOMID_INVALID && data->domid != m.dom)
479 return -EPERM;
480
481 nr_pages = DIV_ROUND_UP(m.num, XEN_PFN_PER_PAGE);
482 if ((m.num <= 0) || (nr_pages > (LONG_MAX >> PAGE_SHIFT)))
483 return -EINVAL;
484
485 ret = gather_array(&pagelist, m.num, sizeof(xen_pfn_t), m.arr);
486
487 if (ret)
488 goto out;
489 if (list_empty(&pagelist)) {
490 ret = -EINVAL;
491 goto out;
492 }
493
494 if (version == 2) {
495 /* Zero error array now to only copy back actual errors. */
496 if (clear_user(m.err, sizeof(int) * m.num)) {
497 ret = -EFAULT;
498 goto out;
499 }
500 }
501
502 down_write(&mm->mmap_sem);
503
504 vma = find_vma(mm, m.addr);
505 if (!vma ||
506 vma->vm_ops != &privcmd_vm_ops) {
507 ret = -EINVAL;
508 goto out_unlock;
509 }
510
511 /*
512 * Caller must either:
513 *
514 * Map the whole VMA range, which will also allocate all the
515 * pages required for the auto_translated_physmap case.
516 *
517 * Or
518 *
519 * Map unmapped holes left from a previous map attempt (e.g.,
520 * because those foreign frames were previously paged out).
521 */
522 if (vma->vm_private_data == NULL) {
523 if (m.addr != vma->vm_start ||
524 m.addr + (nr_pages << PAGE_SHIFT) != vma->vm_end) {
525 ret = -EINVAL;
526 goto out_unlock;
527 }
528 if (xen_feature(XENFEAT_auto_translated_physmap)) {
529 ret = alloc_empty_pages(vma, nr_pages);
530 if (ret < 0)
531 goto out_unlock;
532 } else
533 vma->vm_private_data = PRIV_VMA_LOCKED;
534 } else {
535 if (m.addr < vma->vm_start ||
536 m.addr + (nr_pages << PAGE_SHIFT) > vma->vm_end) {
537 ret = -EINVAL;
538 goto out_unlock;
539 }
540 if (privcmd_vma_range_is_mapped(vma, m.addr, nr_pages)) {
541 ret = -EINVAL;
542 goto out_unlock;
543 }
544 }
545
546 state.domain = m.dom;
547 state.vma = vma;
548 state.va = m.addr;
549 state.index = 0;
550 state.global_error = 0;
551 state.version = version;
552
553 BUILD_BUG_ON(((PAGE_SIZE / sizeof(xen_pfn_t)) % XEN_PFN_PER_PAGE) != 0);
554 /* mmap_batch_fn guarantees ret == 0 */
555 BUG_ON(traverse_pages_block(m.num, sizeof(xen_pfn_t),
556 &pagelist, mmap_batch_fn, &state));
557
558 up_write(&mm->mmap_sem);
559
560 if (state.global_error) {
561 /* Write back errors in second pass. */
562 state.user_gfn = (xen_pfn_t *)m.arr;
563 state.user_err = m.err;
564 ret = traverse_pages_block(m.num, sizeof(xen_pfn_t),
565 &pagelist, mmap_return_errors, &state);
566 } else
567 ret = 0;
568
569 /* If we have not had any EFAULT-like global errors then set the global
570 * error to -ENOENT if necessary. */
571 if ((ret == 0) && (state.global_error == -ENOENT))
572 ret = -ENOENT;
573
574out:
575 free_page_list(&pagelist);
576 return ret;
577
578out_unlock:
579 up_write(&mm->mmap_sem);
580 goto out;
581}
582
583static int lock_pages(
584 struct privcmd_dm_op_buf kbufs[], unsigned int num,
585 struct page *pages[], unsigned int nr_pages)
586{
587 unsigned int i;
588
589 for (i = 0; i < num; i++) {
590 unsigned int requested;
591 int pinned;
592
593 requested = DIV_ROUND_UP(
594 offset_in_page(kbufs[i].uptr) + kbufs[i].size,
595 PAGE_SIZE);
596 if (requested > nr_pages)
597 return -ENOSPC;
598
599 pinned = get_user_pages_fast(
600 (unsigned long) kbufs[i].uptr,
601 requested, FOLL_WRITE, pages);
602 if (pinned < 0)
603 return pinned;
604
605 nr_pages -= pinned;
606 pages += pinned;
607 }
608
609 return 0;
610}
611
612static void unlock_pages(struct page *pages[], unsigned int nr_pages)
613{
614 unsigned int i;
615
616 if (!pages)
617 return;
618
619 for (i = 0; i < nr_pages; i++) {
620 if (pages[i])
621 put_page(pages[i]);
622 }
623}
624
625static long privcmd_ioctl_dm_op(struct file *file, void __user *udata)
626{
627 struct privcmd_data *data = file->private_data;
628 struct privcmd_dm_op kdata;
629 struct privcmd_dm_op_buf *kbufs;
630 unsigned int nr_pages = 0;
631 struct page **pages = NULL;
632 struct xen_dm_op_buf *xbufs = NULL;
633 unsigned int i;
634 long rc;
635
636 if (copy_from_user(&kdata, udata, sizeof(kdata)))
637 return -EFAULT;
638
639 /* If restriction is in place, check the domid matches */
640 if (data->domid != DOMID_INVALID && data->domid != kdata.dom)
641 return -EPERM;
642
643 if (kdata.num == 0)
644 return 0;
645
646 if (kdata.num > privcmd_dm_op_max_num)
647 return -E2BIG;
648
649 kbufs = kcalloc(kdata.num, sizeof(*kbufs), GFP_KERNEL);
650 if (!kbufs)
651 return -ENOMEM;
652
653 if (copy_from_user(kbufs, kdata.ubufs,
654 sizeof(*kbufs) * kdata.num)) {
655 rc = -EFAULT;
656 goto out;
657 }
658
659 for (i = 0; i < kdata.num; i++) {
660 if (kbufs[i].size > privcmd_dm_op_buf_max_size) {
661 rc = -E2BIG;
662 goto out;
663 }
664
665 if (!access_ok(kbufs[i].uptr,
666 kbufs[i].size)) {
667 rc = -EFAULT;
668 goto out;
669 }
670
671 nr_pages += DIV_ROUND_UP(
672 offset_in_page(kbufs[i].uptr) + kbufs[i].size,
673 PAGE_SIZE);
674 }
675
676 pages = kcalloc(nr_pages, sizeof(*pages), GFP_KERNEL);
677 if (!pages) {
678 rc = -ENOMEM;
679 goto out;
680 }
681
682 xbufs = kcalloc(kdata.num, sizeof(*xbufs), GFP_KERNEL);
683 if (!xbufs) {
684 rc = -ENOMEM;
685 goto out;
686 }
687
688 rc = lock_pages(kbufs, kdata.num, pages, nr_pages);
689 if (rc)
690 goto out;
691
692 for (i = 0; i < kdata.num; i++) {
693 set_xen_guest_handle(xbufs[i].h, kbufs[i].uptr);
694 xbufs[i].size = kbufs[i].size;
695 }
696
697 xen_preemptible_hcall_begin();
698 rc = HYPERVISOR_dm_op(kdata.dom, kdata.num, xbufs);
699 xen_preemptible_hcall_end();
700
701out:
702 unlock_pages(pages, nr_pages);
703 kfree(xbufs);
704 kfree(pages);
705 kfree(kbufs);
706
707 return rc;
708}
709
710static long privcmd_ioctl_restrict(struct file *file, void __user *udata)
711{
712 struct privcmd_data *data = file->private_data;
713 domid_t dom;
714
715 if (copy_from_user(&dom, udata, sizeof(dom)))
716 return -EFAULT;
717
718 /* Set restriction to the specified domain, or check it matches */
719 if (data->domid == DOMID_INVALID)
720 data->domid = dom;
721 else if (data->domid != dom)
722 return -EINVAL;
723
724 return 0;
725}
726
727static long privcmd_ioctl_mmap_resource(struct file *file, void __user *udata)
728{
729 struct privcmd_data *data = file->private_data;
730 struct mm_struct *mm = current->mm;
731 struct vm_area_struct *vma;
732 struct privcmd_mmap_resource kdata;
733 xen_pfn_t *pfns = NULL;
734 struct xen_mem_acquire_resource xdata;
735 int rc;
736
737 if (copy_from_user(&kdata, udata, sizeof(kdata)))
738 return -EFAULT;
739
740 /* If restriction is in place, check the domid matches */
741 if (data->domid != DOMID_INVALID && data->domid != kdata.dom)
742 return -EPERM;
743
744 down_write(&mm->mmap_sem);
745
746 vma = find_vma(mm, kdata.addr);
747 if (!vma || vma->vm_ops != &privcmd_vm_ops) {
748 rc = -EINVAL;
749 goto out;
750 }
751
752 pfns = kcalloc(kdata.num, sizeof(*pfns), GFP_KERNEL);
753 if (!pfns) {
754 rc = -ENOMEM;
755 goto out;
756 }
757
758 if (IS_ENABLED(CONFIG_XEN_AUTO_XLATE) &&
759 xen_feature(XENFEAT_auto_translated_physmap)) {
760 unsigned int nr = DIV_ROUND_UP(kdata.num, XEN_PFN_PER_PAGE);
761 struct page **pages;
762 unsigned int i;
763
764 rc = alloc_empty_pages(vma, nr);
765 if (rc < 0)
766 goto out;
767
768 pages = vma->vm_private_data;
769 for (i = 0; i < kdata.num; i++) {
770 xen_pfn_t pfn =
771 page_to_xen_pfn(pages[i / XEN_PFN_PER_PAGE]);
772
773 pfns[i] = pfn + (i % XEN_PFN_PER_PAGE);
774 }
775 } else
776 vma->vm_private_data = PRIV_VMA_LOCKED;
777
778 memset(&xdata, 0, sizeof(xdata));
779 xdata.domid = kdata.dom;
780 xdata.type = kdata.type;
781 xdata.id = kdata.id;
782 xdata.frame = kdata.idx;
783 xdata.nr_frames = kdata.num;
784 set_xen_guest_handle(xdata.frame_list, pfns);
785
786 xen_preemptible_hcall_begin();
787 rc = HYPERVISOR_memory_op(XENMEM_acquire_resource, &xdata);
788 xen_preemptible_hcall_end();
789
790 if (rc)
791 goto out;
792
793 if (IS_ENABLED(CONFIG_XEN_AUTO_XLATE) &&
794 xen_feature(XENFEAT_auto_translated_physmap)) {
795 rc = xen_remap_vma_range(vma, kdata.addr, kdata.num << PAGE_SHIFT);
796 } else {
797 unsigned int domid =
798 (xdata.flags & XENMEM_rsrc_acq_caller_owned) ?
799 DOMID_SELF : kdata.dom;
800 int num;
801
802 num = xen_remap_domain_mfn_array(vma,
803 kdata.addr & PAGE_MASK,
804 pfns, kdata.num, (int *)pfns,
805 vma->vm_page_prot,
806 domid,
807 vma->vm_private_data);
808 if (num < 0)
809 rc = num;
810 else if (num != kdata.num) {
811 unsigned int i;
812
813 for (i = 0; i < num; i++) {
814 rc = pfns[i];
815 if (rc < 0)
816 break;
817 }
818 } else
819 rc = 0;
820 }
821
822out:
823 up_write(&mm->mmap_sem);
824 kfree(pfns);
825
826 return rc;
827}
828
829static long privcmd_ioctl(struct file *file,
830 unsigned int cmd, unsigned long data)
831{
832 int ret = -ENOTTY;
833 void __user *udata = (void __user *) data;
834
835 switch (cmd) {
836 case IOCTL_PRIVCMD_HYPERCALL:
837 ret = privcmd_ioctl_hypercall(file, udata);
838 break;
839
840 case IOCTL_PRIVCMD_MMAP:
841 ret = privcmd_ioctl_mmap(file, udata);
842 break;
843
844 case IOCTL_PRIVCMD_MMAPBATCH:
845 ret = privcmd_ioctl_mmap_batch(file, udata, 1);
846 break;
847
848 case IOCTL_PRIVCMD_MMAPBATCH_V2:
849 ret = privcmd_ioctl_mmap_batch(file, udata, 2);
850 break;
851
852 case IOCTL_PRIVCMD_DM_OP:
853 ret = privcmd_ioctl_dm_op(file, udata);
854 break;
855
856 case IOCTL_PRIVCMD_RESTRICT:
857 ret = privcmd_ioctl_restrict(file, udata);
858 break;
859
860 case IOCTL_PRIVCMD_MMAP_RESOURCE:
861 ret = privcmd_ioctl_mmap_resource(file, udata);
862 break;
863
864 default:
865 break;
866 }
867
868 return ret;
869}
870
871static int privcmd_open(struct inode *ino, struct file *file)
872{
873 struct privcmd_data *data = kzalloc(sizeof(*data), GFP_KERNEL);
874
875 if (!data)
876 return -ENOMEM;
877
878 /* DOMID_INVALID implies no restriction */
879 data->domid = DOMID_INVALID;
880
881 file->private_data = data;
882 return 0;
883}
884
885static int privcmd_release(struct inode *ino, struct file *file)
886{
887 struct privcmd_data *data = file->private_data;
888
889 kfree(data);
890 return 0;
891}
892
893static void privcmd_close(struct vm_area_struct *vma)
894{
895 struct page **pages = vma->vm_private_data;
896 int numpgs = vma_pages(vma);
897 int numgfns = (vma->vm_end - vma->vm_start) >> XEN_PAGE_SHIFT;
898 int rc;
899
900 if (!xen_feature(XENFEAT_auto_translated_physmap) || !numpgs || !pages)
901 return;
902
903 rc = xen_unmap_domain_gfn_range(vma, numgfns, pages);
904 if (rc == 0)
905 free_xenballooned_pages(numpgs, pages);
906 else
907 pr_crit("unable to unmap MFN range: leaking %d pages. rc=%d\n",
908 numpgs, rc);
909 kfree(pages);
910}
911
912static vm_fault_t privcmd_fault(struct vm_fault *vmf)
913{
914 printk(KERN_DEBUG "privcmd_fault: vma=%p %lx-%lx, pgoff=%lx, uv=%p\n",
915 vmf->vma, vmf->vma->vm_start, vmf->vma->vm_end,
916 vmf->pgoff, (void *)vmf->address);
917
918 return VM_FAULT_SIGBUS;
919}
920
921static const struct vm_operations_struct privcmd_vm_ops = {
922 .close = privcmd_close,
923 .fault = privcmd_fault
924};
925
926static int privcmd_mmap(struct file *file, struct vm_area_struct *vma)
927{
928 /* DONTCOPY is essential for Xen because copy_page_range doesn't know
929 * how to recreate these mappings */
930 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTCOPY |
931 VM_DONTEXPAND | VM_DONTDUMP;
932 vma->vm_ops = &privcmd_vm_ops;
933 vma->vm_private_data = NULL;
934
935 return 0;
936}
937
938/*
939 * For MMAPBATCH*. This allows asserting the singleshot mapping
940 * on a per pfn/pte basis. Mapping calls that fail with ENOENT
941 * can be then retried until success.
942 */
943static int is_mapped_fn(pte_t *pte, unsigned long addr, void *data)
944{
945 return pte_none(*pte) ? 0 : -EBUSY;
946}
947
948static int privcmd_vma_range_is_mapped(
949 struct vm_area_struct *vma,
950 unsigned long addr,
951 unsigned long nr_pages)
952{
953 return apply_to_page_range(vma->vm_mm, addr, nr_pages << PAGE_SHIFT,
954 is_mapped_fn, NULL) != 0;
955}
956
957const struct file_operations xen_privcmd_fops = {
958 .owner = THIS_MODULE,
959 .unlocked_ioctl = privcmd_ioctl,
960 .open = privcmd_open,
961 .release = privcmd_release,
962 .mmap = privcmd_mmap,
963};
964EXPORT_SYMBOL_GPL(xen_privcmd_fops);
965
966static struct miscdevice privcmd_dev = {
967 .minor = MISC_DYNAMIC_MINOR,
968 .name = "xen/privcmd",
969 .fops = &xen_privcmd_fops,
970};
971
972static int __init privcmd_init(void)
973{
974 int err;
975
976 if (!xen_domain())
977 return -ENODEV;
978
979 err = misc_register(&privcmd_dev);
980 if (err != 0) {
981 pr_err("Could not register Xen privcmd device\n");
982 return err;
983 }
984
985 err = misc_register(&xen_privcmdbuf_dev);
986 if (err != 0) {
987 pr_err("Could not register Xen hypercall-buf device\n");
988 misc_deregister(&privcmd_dev);
989 return err;
990 }
991
992 return 0;
993}
994
995static void __exit privcmd_exit(void)
996{
997 misc_deregister(&privcmd_dev);
998 misc_deregister(&xen_privcmdbuf_dev);
999}
1000
1001module_init(privcmd_init);
1002module_exit(privcmd_exit);
1/******************************************************************************
2 * privcmd.c
3 *
4 * Interface to privileged domain-0 commands.
5 *
6 * Copyright (c) 2002-2004, K A Fraser, B Dragovic
7 */
8
9#define pr_fmt(fmt) "xen:" KBUILD_MODNAME ": " fmt
10
11#include <linux/kernel.h>
12#include <linux/module.h>
13#include <linux/sched.h>
14#include <linux/slab.h>
15#include <linux/string.h>
16#include <linux/errno.h>
17#include <linux/mm.h>
18#include <linux/mman.h>
19#include <linux/uaccess.h>
20#include <linux/swap.h>
21#include <linux/highmem.h>
22#include <linux/pagemap.h>
23#include <linux/seq_file.h>
24#include <linux/miscdevice.h>
25#include <linux/moduleparam.h>
26
27#include <asm/pgalloc.h>
28#include <asm/pgtable.h>
29#include <asm/tlb.h>
30#include <asm/xen/hypervisor.h>
31#include <asm/xen/hypercall.h>
32
33#include <xen/xen.h>
34#include <xen/privcmd.h>
35#include <xen/interface/xen.h>
36#include <xen/interface/hvm/dm_op.h>
37#include <xen/features.h>
38#include <xen/page.h>
39#include <xen/xen-ops.h>
40#include <xen/balloon.h>
41
42#include "privcmd.h"
43
44MODULE_LICENSE("GPL");
45
46#define PRIV_VMA_LOCKED ((void *)1)
47
48static unsigned int privcmd_dm_op_max_num = 16;
49module_param_named(dm_op_max_nr_bufs, privcmd_dm_op_max_num, uint, 0644);
50MODULE_PARM_DESC(dm_op_max_nr_bufs,
51 "Maximum number of buffers per dm_op hypercall");
52
53static unsigned int privcmd_dm_op_buf_max_size = 4096;
54module_param_named(dm_op_buf_max_size, privcmd_dm_op_buf_max_size, uint,
55 0644);
56MODULE_PARM_DESC(dm_op_buf_max_size,
57 "Maximum size of a dm_op hypercall buffer");
58
59struct privcmd_data {
60 domid_t domid;
61};
62
63static int privcmd_vma_range_is_mapped(
64 struct vm_area_struct *vma,
65 unsigned long addr,
66 unsigned long nr_pages);
67
68static long privcmd_ioctl_hypercall(struct file *file, void __user *udata)
69{
70 struct privcmd_data *data = file->private_data;
71 struct privcmd_hypercall hypercall;
72 long ret;
73
74 /* Disallow arbitrary hypercalls if restricted */
75 if (data->domid != DOMID_INVALID)
76 return -EPERM;
77
78 if (copy_from_user(&hypercall, udata, sizeof(hypercall)))
79 return -EFAULT;
80
81 xen_preemptible_hcall_begin();
82 ret = privcmd_call(hypercall.op,
83 hypercall.arg[0], hypercall.arg[1],
84 hypercall.arg[2], hypercall.arg[3],
85 hypercall.arg[4]);
86 xen_preemptible_hcall_end();
87
88 return ret;
89}
90
91static void free_page_list(struct list_head *pages)
92{
93 struct page *p, *n;
94
95 list_for_each_entry_safe(p, n, pages, lru)
96 __free_page(p);
97
98 INIT_LIST_HEAD(pages);
99}
100
101/*
102 * Given an array of items in userspace, return a list of pages
103 * containing the data. If copying fails, either because of memory
104 * allocation failure or a problem reading user memory, return an
105 * error code; its up to the caller to dispose of any partial list.
106 */
107static int gather_array(struct list_head *pagelist,
108 unsigned nelem, size_t size,
109 const void __user *data)
110{
111 unsigned pageidx;
112 void *pagedata;
113 int ret;
114
115 if (size > PAGE_SIZE)
116 return 0;
117
118 pageidx = PAGE_SIZE;
119 pagedata = NULL; /* quiet, gcc */
120 while (nelem--) {
121 if (pageidx > PAGE_SIZE-size) {
122 struct page *page = alloc_page(GFP_KERNEL);
123
124 ret = -ENOMEM;
125 if (page == NULL)
126 goto fail;
127
128 pagedata = page_address(page);
129
130 list_add_tail(&page->lru, pagelist);
131 pageidx = 0;
132 }
133
134 ret = -EFAULT;
135 if (copy_from_user(pagedata + pageidx, data, size))
136 goto fail;
137
138 data += size;
139 pageidx += size;
140 }
141
142 ret = 0;
143
144fail:
145 return ret;
146}
147
148/*
149 * Call function "fn" on each element of the array fragmented
150 * over a list of pages.
151 */
152static int traverse_pages(unsigned nelem, size_t size,
153 struct list_head *pos,
154 int (*fn)(void *data, void *state),
155 void *state)
156{
157 void *pagedata;
158 unsigned pageidx;
159 int ret = 0;
160
161 BUG_ON(size > PAGE_SIZE);
162
163 pageidx = PAGE_SIZE;
164 pagedata = NULL; /* hush, gcc */
165
166 while (nelem--) {
167 if (pageidx > PAGE_SIZE-size) {
168 struct page *page;
169 pos = pos->next;
170 page = list_entry(pos, struct page, lru);
171 pagedata = page_address(page);
172 pageidx = 0;
173 }
174
175 ret = (*fn)(pagedata + pageidx, state);
176 if (ret)
177 break;
178 pageidx += size;
179 }
180
181 return ret;
182}
183
184/*
185 * Similar to traverse_pages, but use each page as a "block" of
186 * data to be processed as one unit.
187 */
188static int traverse_pages_block(unsigned nelem, size_t size,
189 struct list_head *pos,
190 int (*fn)(void *data, int nr, void *state),
191 void *state)
192{
193 void *pagedata;
194 int ret = 0;
195
196 BUG_ON(size > PAGE_SIZE);
197
198 while (nelem) {
199 int nr = (PAGE_SIZE/size);
200 struct page *page;
201 if (nr > nelem)
202 nr = nelem;
203 pos = pos->next;
204 page = list_entry(pos, struct page, lru);
205 pagedata = page_address(page);
206 ret = (*fn)(pagedata, nr, state);
207 if (ret)
208 break;
209 nelem -= nr;
210 }
211
212 return ret;
213}
214
215struct mmap_gfn_state {
216 unsigned long va;
217 struct vm_area_struct *vma;
218 domid_t domain;
219};
220
221static int mmap_gfn_range(void *data, void *state)
222{
223 struct privcmd_mmap_entry *msg = data;
224 struct mmap_gfn_state *st = state;
225 struct vm_area_struct *vma = st->vma;
226 int rc;
227
228 /* Do not allow range to wrap the address space. */
229 if ((msg->npages > (LONG_MAX >> PAGE_SHIFT)) ||
230 ((unsigned long)(msg->npages << PAGE_SHIFT) >= -st->va))
231 return -EINVAL;
232
233 /* Range chunks must be contiguous in va space. */
234 if ((msg->va != st->va) ||
235 ((msg->va+(msg->npages<<PAGE_SHIFT)) > vma->vm_end))
236 return -EINVAL;
237
238 rc = xen_remap_domain_gfn_range(vma,
239 msg->va & PAGE_MASK,
240 msg->mfn, msg->npages,
241 vma->vm_page_prot,
242 st->domain, NULL);
243 if (rc < 0)
244 return rc;
245
246 st->va += msg->npages << PAGE_SHIFT;
247
248 return 0;
249}
250
251static long privcmd_ioctl_mmap(struct file *file, void __user *udata)
252{
253 struct privcmd_data *data = file->private_data;
254 struct privcmd_mmap mmapcmd;
255 struct mm_struct *mm = current->mm;
256 struct vm_area_struct *vma;
257 int rc;
258 LIST_HEAD(pagelist);
259 struct mmap_gfn_state state;
260
261 /* We only support privcmd_ioctl_mmap_batch for auto translated. */
262 if (xen_feature(XENFEAT_auto_translated_physmap))
263 return -ENOSYS;
264
265 if (copy_from_user(&mmapcmd, udata, sizeof(mmapcmd)))
266 return -EFAULT;
267
268 /* If restriction is in place, check the domid matches */
269 if (data->domid != DOMID_INVALID && data->domid != mmapcmd.dom)
270 return -EPERM;
271
272 rc = gather_array(&pagelist,
273 mmapcmd.num, sizeof(struct privcmd_mmap_entry),
274 mmapcmd.entry);
275
276 if (rc || list_empty(&pagelist))
277 goto out;
278
279 down_write(&mm->mmap_sem);
280
281 {
282 struct page *page = list_first_entry(&pagelist,
283 struct page, lru);
284 struct privcmd_mmap_entry *msg = page_address(page);
285
286 vma = find_vma(mm, msg->va);
287 rc = -EINVAL;
288
289 if (!vma || (msg->va != vma->vm_start) || vma->vm_private_data)
290 goto out_up;
291 vma->vm_private_data = PRIV_VMA_LOCKED;
292 }
293
294 state.va = vma->vm_start;
295 state.vma = vma;
296 state.domain = mmapcmd.dom;
297
298 rc = traverse_pages(mmapcmd.num, sizeof(struct privcmd_mmap_entry),
299 &pagelist,
300 mmap_gfn_range, &state);
301
302
303out_up:
304 up_write(&mm->mmap_sem);
305
306out:
307 free_page_list(&pagelist);
308
309 return rc;
310}
311
312struct mmap_batch_state {
313 domid_t domain;
314 unsigned long va;
315 struct vm_area_struct *vma;
316 int index;
317 /* A tristate:
318 * 0 for no errors
319 * 1 if at least one error has happened (and no
320 * -ENOENT errors have happened)
321 * -ENOENT if at least 1 -ENOENT has happened.
322 */
323 int global_error;
324 int version;
325
326 /* User-space gfn array to store errors in the second pass for V1. */
327 xen_pfn_t __user *user_gfn;
328 /* User-space int array to store errors in the second pass for V2. */
329 int __user *user_err;
330};
331
332/* auto translated dom0 note: if domU being created is PV, then gfn is
333 * mfn(addr on bus). If it's auto xlated, then gfn is pfn (input to HAP).
334 */
335static int mmap_batch_fn(void *data, int nr, void *state)
336{
337 xen_pfn_t *gfnp = data;
338 struct mmap_batch_state *st = state;
339 struct vm_area_struct *vma = st->vma;
340 struct page **pages = vma->vm_private_data;
341 struct page **cur_pages = NULL;
342 int ret;
343
344 if (xen_feature(XENFEAT_auto_translated_physmap))
345 cur_pages = &pages[st->index];
346
347 BUG_ON(nr < 0);
348 ret = xen_remap_domain_gfn_array(st->vma, st->va & PAGE_MASK, gfnp, nr,
349 (int *)gfnp, st->vma->vm_page_prot,
350 st->domain, cur_pages);
351
352 /* Adjust the global_error? */
353 if (ret != nr) {
354 if (ret == -ENOENT)
355 st->global_error = -ENOENT;
356 else {
357 /* Record that at least one error has happened. */
358 if (st->global_error == 0)
359 st->global_error = 1;
360 }
361 }
362 st->va += XEN_PAGE_SIZE * nr;
363 st->index += nr / XEN_PFN_PER_PAGE;
364
365 return 0;
366}
367
368static int mmap_return_error(int err, struct mmap_batch_state *st)
369{
370 int ret;
371
372 if (st->version == 1) {
373 if (err) {
374 xen_pfn_t gfn;
375
376 ret = get_user(gfn, st->user_gfn);
377 if (ret < 0)
378 return ret;
379 /*
380 * V1 encodes the error codes in the 32bit top
381 * nibble of the gfn (with its known
382 * limitations vis-a-vis 64 bit callers).
383 */
384 gfn |= (err == -ENOENT) ?
385 PRIVCMD_MMAPBATCH_PAGED_ERROR :
386 PRIVCMD_MMAPBATCH_MFN_ERROR;
387 return __put_user(gfn, st->user_gfn++);
388 } else
389 st->user_gfn++;
390 } else { /* st->version == 2 */
391 if (err)
392 return __put_user(err, st->user_err++);
393 else
394 st->user_err++;
395 }
396
397 return 0;
398}
399
400static int mmap_return_errors(void *data, int nr, void *state)
401{
402 struct mmap_batch_state *st = state;
403 int *errs = data;
404 int i;
405 int ret;
406
407 for (i = 0; i < nr; i++) {
408 ret = mmap_return_error(errs[i], st);
409 if (ret < 0)
410 return ret;
411 }
412 return 0;
413}
414
415/* Allocate pfns that are then mapped with gfns from foreign domid. Update
416 * the vma with the page info to use later.
417 * Returns: 0 if success, otherwise -errno
418 */
419static int alloc_empty_pages(struct vm_area_struct *vma, int numpgs)
420{
421 int rc;
422 struct page **pages;
423
424 pages = kcalloc(numpgs, sizeof(pages[0]), GFP_KERNEL);
425 if (pages == NULL)
426 return -ENOMEM;
427
428 rc = alloc_xenballooned_pages(numpgs, pages);
429 if (rc != 0) {
430 pr_warn("%s Could not alloc %d pfns rc:%d\n", __func__,
431 numpgs, rc);
432 kfree(pages);
433 return -ENOMEM;
434 }
435 BUG_ON(vma->vm_private_data != NULL);
436 vma->vm_private_data = pages;
437
438 return 0;
439}
440
441static const struct vm_operations_struct privcmd_vm_ops;
442
443static long privcmd_ioctl_mmap_batch(
444 struct file *file, void __user *udata, int version)
445{
446 struct privcmd_data *data = file->private_data;
447 int ret;
448 struct privcmd_mmapbatch_v2 m;
449 struct mm_struct *mm = current->mm;
450 struct vm_area_struct *vma;
451 unsigned long nr_pages;
452 LIST_HEAD(pagelist);
453 struct mmap_batch_state state;
454
455 switch (version) {
456 case 1:
457 if (copy_from_user(&m, udata, sizeof(struct privcmd_mmapbatch)))
458 return -EFAULT;
459 /* Returns per-frame error in m.arr. */
460 m.err = NULL;
461 if (!access_ok(VERIFY_WRITE, m.arr, m.num * sizeof(*m.arr)))
462 return -EFAULT;
463 break;
464 case 2:
465 if (copy_from_user(&m, udata, sizeof(struct privcmd_mmapbatch_v2)))
466 return -EFAULT;
467 /* Returns per-frame error code in m.err. */
468 if (!access_ok(VERIFY_WRITE, m.err, m.num * (sizeof(*m.err))))
469 return -EFAULT;
470 break;
471 default:
472 return -EINVAL;
473 }
474
475 /* If restriction is in place, check the domid matches */
476 if (data->domid != DOMID_INVALID && data->domid != m.dom)
477 return -EPERM;
478
479 nr_pages = DIV_ROUND_UP(m.num, XEN_PFN_PER_PAGE);
480 if ((m.num <= 0) || (nr_pages > (LONG_MAX >> PAGE_SHIFT)))
481 return -EINVAL;
482
483 ret = gather_array(&pagelist, m.num, sizeof(xen_pfn_t), m.arr);
484
485 if (ret)
486 goto out;
487 if (list_empty(&pagelist)) {
488 ret = -EINVAL;
489 goto out;
490 }
491
492 if (version == 2) {
493 /* Zero error array now to only copy back actual errors. */
494 if (clear_user(m.err, sizeof(int) * m.num)) {
495 ret = -EFAULT;
496 goto out;
497 }
498 }
499
500 down_write(&mm->mmap_sem);
501
502 vma = find_vma(mm, m.addr);
503 if (!vma ||
504 vma->vm_ops != &privcmd_vm_ops) {
505 ret = -EINVAL;
506 goto out_unlock;
507 }
508
509 /*
510 * Caller must either:
511 *
512 * Map the whole VMA range, which will also allocate all the
513 * pages required for the auto_translated_physmap case.
514 *
515 * Or
516 *
517 * Map unmapped holes left from a previous map attempt (e.g.,
518 * because those foreign frames were previously paged out).
519 */
520 if (vma->vm_private_data == NULL) {
521 if (m.addr != vma->vm_start ||
522 m.addr + (nr_pages << PAGE_SHIFT) != vma->vm_end) {
523 ret = -EINVAL;
524 goto out_unlock;
525 }
526 if (xen_feature(XENFEAT_auto_translated_physmap)) {
527 ret = alloc_empty_pages(vma, nr_pages);
528 if (ret < 0)
529 goto out_unlock;
530 } else
531 vma->vm_private_data = PRIV_VMA_LOCKED;
532 } else {
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 (privcmd_vma_range_is_mapped(vma, m.addr, nr_pages)) {
539 ret = -EINVAL;
540 goto out_unlock;
541 }
542 }
543
544 state.domain = m.dom;
545 state.vma = vma;
546 state.va = m.addr;
547 state.index = 0;
548 state.global_error = 0;
549 state.version = version;
550
551 BUILD_BUG_ON(((PAGE_SIZE / sizeof(xen_pfn_t)) % XEN_PFN_PER_PAGE) != 0);
552 /* mmap_batch_fn guarantees ret == 0 */
553 BUG_ON(traverse_pages_block(m.num, sizeof(xen_pfn_t),
554 &pagelist, mmap_batch_fn, &state));
555
556 up_write(&mm->mmap_sem);
557
558 if (state.global_error) {
559 /* Write back errors in second pass. */
560 state.user_gfn = (xen_pfn_t *)m.arr;
561 state.user_err = m.err;
562 ret = traverse_pages_block(m.num, sizeof(xen_pfn_t),
563 &pagelist, mmap_return_errors, &state);
564 } else
565 ret = 0;
566
567 /* If we have not had any EFAULT-like global errors then set the global
568 * error to -ENOENT if necessary. */
569 if ((ret == 0) && (state.global_error == -ENOENT))
570 ret = -ENOENT;
571
572out:
573 free_page_list(&pagelist);
574 return ret;
575
576out_unlock:
577 up_write(&mm->mmap_sem);
578 goto out;
579}
580
581static int lock_pages(
582 struct privcmd_dm_op_buf kbufs[], unsigned int num,
583 struct page *pages[], unsigned int nr_pages)
584{
585 unsigned int i;
586
587 for (i = 0; i < num; i++) {
588 unsigned int requested;
589 int pinned;
590
591 requested = DIV_ROUND_UP(
592 offset_in_page(kbufs[i].uptr) + kbufs[i].size,
593 PAGE_SIZE);
594 if (requested > nr_pages)
595 return -ENOSPC;
596
597 pinned = get_user_pages_fast(
598 (unsigned long) kbufs[i].uptr,
599 requested, FOLL_WRITE, pages);
600 if (pinned < 0)
601 return pinned;
602
603 nr_pages -= pinned;
604 pages += pinned;
605 }
606
607 return 0;
608}
609
610static void unlock_pages(struct page *pages[], unsigned int nr_pages)
611{
612 unsigned int i;
613
614 if (!pages)
615 return;
616
617 for (i = 0; i < nr_pages; i++) {
618 if (pages[i])
619 put_page(pages[i]);
620 }
621}
622
623static long privcmd_ioctl_dm_op(struct file *file, void __user *udata)
624{
625 struct privcmd_data *data = file->private_data;
626 struct privcmd_dm_op kdata;
627 struct privcmd_dm_op_buf *kbufs;
628 unsigned int nr_pages = 0;
629 struct page **pages = NULL;
630 struct xen_dm_op_buf *xbufs = NULL;
631 unsigned int i;
632 long rc;
633
634 if (copy_from_user(&kdata, udata, sizeof(kdata)))
635 return -EFAULT;
636
637 /* If restriction is in place, check the domid matches */
638 if (data->domid != DOMID_INVALID && data->domid != kdata.dom)
639 return -EPERM;
640
641 if (kdata.num == 0)
642 return 0;
643
644 if (kdata.num > privcmd_dm_op_max_num)
645 return -E2BIG;
646
647 kbufs = kcalloc(kdata.num, sizeof(*kbufs), GFP_KERNEL);
648 if (!kbufs)
649 return -ENOMEM;
650
651 if (copy_from_user(kbufs, kdata.ubufs,
652 sizeof(*kbufs) * kdata.num)) {
653 rc = -EFAULT;
654 goto out;
655 }
656
657 for (i = 0; i < kdata.num; i++) {
658 if (kbufs[i].size > privcmd_dm_op_buf_max_size) {
659 rc = -E2BIG;
660 goto out;
661 }
662
663 if (!access_ok(VERIFY_WRITE, kbufs[i].uptr,
664 kbufs[i].size)) {
665 rc = -EFAULT;
666 goto out;
667 }
668
669 nr_pages += DIV_ROUND_UP(
670 offset_in_page(kbufs[i].uptr) + kbufs[i].size,
671 PAGE_SIZE);
672 }
673
674 pages = kcalloc(nr_pages, sizeof(*pages), GFP_KERNEL);
675 if (!pages) {
676 rc = -ENOMEM;
677 goto out;
678 }
679
680 xbufs = kcalloc(kdata.num, sizeof(*xbufs), GFP_KERNEL);
681 if (!xbufs) {
682 rc = -ENOMEM;
683 goto out;
684 }
685
686 rc = lock_pages(kbufs, kdata.num, pages, nr_pages);
687 if (rc)
688 goto out;
689
690 for (i = 0; i < kdata.num; i++) {
691 set_xen_guest_handle(xbufs[i].h, kbufs[i].uptr);
692 xbufs[i].size = kbufs[i].size;
693 }
694
695 xen_preemptible_hcall_begin();
696 rc = HYPERVISOR_dm_op(kdata.dom, kdata.num, xbufs);
697 xen_preemptible_hcall_end();
698
699out:
700 unlock_pages(pages, nr_pages);
701 kfree(xbufs);
702 kfree(pages);
703 kfree(kbufs);
704
705 return rc;
706}
707
708static long privcmd_ioctl_restrict(struct file *file, void __user *udata)
709{
710 struct privcmd_data *data = file->private_data;
711 domid_t dom;
712
713 if (copy_from_user(&dom, udata, sizeof(dom)))
714 return -EFAULT;
715
716 /* Set restriction to the specified domain, or check it matches */
717 if (data->domid == DOMID_INVALID)
718 data->domid = dom;
719 else if (data->domid != dom)
720 return -EINVAL;
721
722 return 0;
723}
724
725static long privcmd_ioctl(struct file *file,
726 unsigned int cmd, unsigned long data)
727{
728 int ret = -ENOTTY;
729 void __user *udata = (void __user *) data;
730
731 switch (cmd) {
732 case IOCTL_PRIVCMD_HYPERCALL:
733 ret = privcmd_ioctl_hypercall(file, udata);
734 break;
735
736 case IOCTL_PRIVCMD_MMAP:
737 ret = privcmd_ioctl_mmap(file, udata);
738 break;
739
740 case IOCTL_PRIVCMD_MMAPBATCH:
741 ret = privcmd_ioctl_mmap_batch(file, udata, 1);
742 break;
743
744 case IOCTL_PRIVCMD_MMAPBATCH_V2:
745 ret = privcmd_ioctl_mmap_batch(file, udata, 2);
746 break;
747
748 case IOCTL_PRIVCMD_DM_OP:
749 ret = privcmd_ioctl_dm_op(file, udata);
750 break;
751
752 case IOCTL_PRIVCMD_RESTRICT:
753 ret = privcmd_ioctl_restrict(file, udata);
754 break;
755
756 default:
757 break;
758 }
759
760 return ret;
761}
762
763static int privcmd_open(struct inode *ino, struct file *file)
764{
765 struct privcmd_data *data = kzalloc(sizeof(*data), GFP_KERNEL);
766
767 if (!data)
768 return -ENOMEM;
769
770 /* DOMID_INVALID implies no restriction */
771 data->domid = DOMID_INVALID;
772
773 file->private_data = data;
774 return 0;
775}
776
777static int privcmd_release(struct inode *ino, struct file *file)
778{
779 struct privcmd_data *data = file->private_data;
780
781 kfree(data);
782 return 0;
783}
784
785static void privcmd_close(struct vm_area_struct *vma)
786{
787 struct page **pages = vma->vm_private_data;
788 int numpgs = vma_pages(vma);
789 int numgfns = (vma->vm_end - vma->vm_start) >> XEN_PAGE_SHIFT;
790 int rc;
791
792 if (!xen_feature(XENFEAT_auto_translated_physmap) || !numpgs || !pages)
793 return;
794
795 rc = xen_unmap_domain_gfn_range(vma, numgfns, pages);
796 if (rc == 0)
797 free_xenballooned_pages(numpgs, pages);
798 else
799 pr_crit("unable to unmap MFN range: leaking %d pages. rc=%d\n",
800 numpgs, rc);
801 kfree(pages);
802}
803
804static int privcmd_fault(struct vm_fault *vmf)
805{
806 printk(KERN_DEBUG "privcmd_fault: vma=%p %lx-%lx, pgoff=%lx, uv=%p\n",
807 vmf->vma, vmf->vma->vm_start, vmf->vma->vm_end,
808 vmf->pgoff, (void *)vmf->address);
809
810 return VM_FAULT_SIGBUS;
811}
812
813static const struct vm_operations_struct privcmd_vm_ops = {
814 .close = privcmd_close,
815 .fault = privcmd_fault
816};
817
818static int privcmd_mmap(struct file *file, struct vm_area_struct *vma)
819{
820 /* DONTCOPY is essential for Xen because copy_page_range doesn't know
821 * how to recreate these mappings */
822 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTCOPY |
823 VM_DONTEXPAND | VM_DONTDUMP;
824 vma->vm_ops = &privcmd_vm_ops;
825 vma->vm_private_data = NULL;
826
827 return 0;
828}
829
830/*
831 * For MMAPBATCH*. This allows asserting the singleshot mapping
832 * on a per pfn/pte basis. Mapping calls that fail with ENOENT
833 * can be then retried until success.
834 */
835static int is_mapped_fn(pte_t *pte, struct page *pmd_page,
836 unsigned long addr, void *data)
837{
838 return pte_none(*pte) ? 0 : -EBUSY;
839}
840
841static int privcmd_vma_range_is_mapped(
842 struct vm_area_struct *vma,
843 unsigned long addr,
844 unsigned long nr_pages)
845{
846 return apply_to_page_range(vma->vm_mm, addr, nr_pages << PAGE_SHIFT,
847 is_mapped_fn, NULL) != 0;
848}
849
850const struct file_operations xen_privcmd_fops = {
851 .owner = THIS_MODULE,
852 .unlocked_ioctl = privcmd_ioctl,
853 .open = privcmd_open,
854 .release = privcmd_release,
855 .mmap = privcmd_mmap,
856};
857EXPORT_SYMBOL_GPL(xen_privcmd_fops);
858
859static struct miscdevice privcmd_dev = {
860 .minor = MISC_DYNAMIC_MINOR,
861 .name = "xen/privcmd",
862 .fops = &xen_privcmd_fops,
863};
864
865static int __init privcmd_init(void)
866{
867 int err;
868
869 if (!xen_domain())
870 return -ENODEV;
871
872 err = misc_register(&privcmd_dev);
873 if (err != 0) {
874 pr_err("Could not register Xen privcmd device\n");
875 return err;
876 }
877 return 0;
878}
879
880static void __exit privcmd_exit(void)
881{
882 misc_deregister(&privcmd_dev);
883}
884
885module_init(privcmd_init);
886module_exit(privcmd_exit);