projects/linux·blob master · refs

1205 lines · 30.5 KiB

1/******************************************************************************
2 * gntdev.c
3 *
4 * Device for accessing (in user-space) pages that have been granted by other
5 * domains.
6 *
7 * Copyright (c) 2006-2007, D G Murray.
8 * (c) 2009 Gerd Hoffmann <[email protected]>
9 * (c) 2018 Oleksandr Andrushchenko, EPAM Systems Inc.
10 *
11 * This program is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 * GNU General Public License for more details.
15 *
16 * You should have received a copy of the GNU General Public License
17 * along with this program; if not, write to the Free Software
18 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
19 */
20
21#undef DEBUG
22
23#define pr_fmt(fmt) "xen:" KBUILD_MODNAME ": " fmt
24
25#include <linux/dma-mapping.h>
26#include <linux/module.h>
27#include <linux/kernel.h>
28#include <linux/init.h>
29#include <linux/miscdevice.h>
30#include <linux/fs.h>
31#include <linux/uaccess.h>
32#include <linux/sched.h>
33#include <linux/sched/mm.h>
34#include <linux/spinlock.h>
35#include <linux/slab.h>
36#include <linux/highmem.h>
37#include <linux/refcount.h>
38#include <linux/workqueue.h>
39
40#include <xen/xen.h>
41#include <xen/grant_table.h>
42#include <xen/balloon.h>
43#include <xen/gntdev.h>
44#include <xen/events.h>
45#include <xen/page.h>
46#include <asm/xen/hypervisor.h>
47#include <asm/xen/hypercall.h>
48
49#include "gntdev-common.h"
50#ifdef CONFIG_XEN_GNTDEV_DMABUF
51#include "gntdev-dmabuf.h"
52#endif
53
54MODULE_LICENSE("GPL");
55MODULE_AUTHOR("Derek G. Murray <[email protected]>, "
56 "Gerd Hoffmann <[email protected]>");
57MODULE_DESCRIPTION("User-space granted page access driver");
58
59#define GNTDEV_COPY_BATCH 16
60
61struct gntdev_copy_batch {
62 struct gnttab_copy ops[GNTDEV_COPY_BATCH];
63 struct page *pages[GNTDEV_COPY_BATCH];
64 s16 __user *status[GNTDEV_COPY_BATCH];
65 unsigned int nr_ops;
66 unsigned int nr_pages;
67 bool writeable;
68 struct gntdev_copy_batch *next;
69};
70
71static unsigned int limit = 64*1024;
72module_param(limit, uint, 0644);
73MODULE_PARM_DESC(limit,
74 "Maximum number of grants that may be mapped by one mapping request");
75
76static void unmap_grant_pages(struct gntdev_grant_map *map,
77 int offset, int pages);
78
79static struct miscdevice gntdev_miscdev;
80
81/* ------------------------------------------------------------------ */
82
83bool gntdev_test_page_count(unsigned int count)
84{
85 return !count || count > limit;
86}
87
88static void gntdev_print_maps(struct gntdev_priv *priv,
89 char *text, int text_index)
90{
91#ifdef DEBUG
92 struct gntdev_grant_map *map;
93
94 pr_debug("%s: maps list (priv %p)\n", __func__, priv);
95 list_for_each_entry(map, &priv->maps, next)
96 pr_debug(" index %2d, count %2d %s\n",
97 map->index, map->count,
98 map->index == text_index && text ? text : "");
99#endif
100}
101
102static void gntdev_free_map(struct gntdev_grant_map *map)
103{
104 if (map == NULL)
105 return;
106
107#ifdef CONFIG_XEN_GRANT_DMA_ALLOC
108 if (map->dma_vaddr) {
109 struct gnttab_dma_alloc_args args;
110
111 args.dev = map->dma_dev;
112 args.coherent = !!(map->dma_flags & GNTDEV_DMA_FLAG_COHERENT);
113 args.nr_pages = map->count;
114 args.pages = map->pages;
115 args.frames = map->frames;
116 args.vaddr = map->dma_vaddr;
117 args.dev_bus_addr = map->dma_bus_addr;
118
119 gnttab_dma_free_pages(&args);
120 } else
121#endif
122 if (map->pages)
123 gnttab_free_pages(map->count, map->pages);
124
125#ifdef CONFIG_XEN_GRANT_DMA_ALLOC
126 kvfree(map->frames);
127#endif
128 kvfree(map->pages);
129 kvfree(map->grants);
130 kvfree(map->map_ops);
131 kvfree(map->unmap_ops);
132 kvfree(map->kmap_ops);
133 kvfree(map->kunmap_ops);
134 kvfree(map->being_removed);
135 kfree(map);
136}
137
138struct gntdev_grant_map *gntdev_alloc_map(struct gntdev_priv *priv, int count,
139 int dma_flags)
140{
141 struct gntdev_grant_map *add;
142 int i;
143
144 add = kzalloc_obj(*add);
145 if (NULL == add)
146 return NULL;
147
148 add->grants = kvmalloc_objs(add->grants[0], count);
149 add->map_ops = kvmalloc_objs(add->map_ops[0], count);
150 add->unmap_ops = kvmalloc_objs(add->unmap_ops[0], count);
151 add->pages = kvzalloc_objs(add->pages[0], count);
152 add->being_removed =
153 kvzalloc_objs(add->being_removed[0], count);
154 if (NULL == add->grants ||
155 NULL == add->map_ops ||
156 NULL == add->unmap_ops ||
157 NULL == add->pages ||
158 NULL == add->being_removed)
159 goto err;
160 if (xen_pv_domain()) {
161 add->kmap_ops = kvmalloc_objs(add->kmap_ops[0], count);
162 add->kunmap_ops = kvmalloc_objs(add->kunmap_ops[0], count);
163 if (NULL == add->kmap_ops || NULL == add->kunmap_ops)
164 goto err;
165 }
166
167#ifdef CONFIG_XEN_GRANT_DMA_ALLOC
168 add->dma_flags = dma_flags;
169
170 /*
171 * Check if this mapping is requested to be backed
172 * by a DMA buffer.
173 */
174 if (dma_flags & (GNTDEV_DMA_FLAG_WC | GNTDEV_DMA_FLAG_COHERENT)) {
175 struct gnttab_dma_alloc_args args;
176
177 add->frames = kvzalloc_objs(add->frames[0], count);
178 if (!add->frames)
179 goto err;
180
181 /* Remember the device, so we can free DMA memory. */
182 add->dma_dev = priv->dma_dev;
183
184 args.dev = priv->dma_dev;
185 args.coherent = !!(dma_flags & GNTDEV_DMA_FLAG_COHERENT);
186 args.nr_pages = count;
187 args.pages = add->pages;
188 args.frames = add->frames;
189
190 if (gnttab_dma_alloc_pages(&args))
191 goto err;
192
193 add->dma_vaddr = args.vaddr;
194 add->dma_bus_addr = args.dev_bus_addr;
195 } else
196#endif
197 if (gnttab_alloc_pages(count, add->pages))
198 goto err;
199
200 for (i = 0; i < count; i++) {
201 add->grants[i].domid = DOMID_INVALID;
202 add->grants[i].ref = INVALID_GRANT_REF;
203 add->map_ops[i].handle = INVALID_GRANT_HANDLE;
204 add->unmap_ops[i].handle = INVALID_GRANT_HANDLE;
205 if (xen_pv_domain()) {
206 add->kmap_ops[i].handle = INVALID_GRANT_HANDLE;
207 add->kunmap_ops[i].handle = INVALID_GRANT_HANDLE;
208 }
209 }
210
211 add->index = 0;
212 add->count = count;
213 refcount_set(&add->users, 1);
214
215 return add;
216
217err:
218 gntdev_free_map(add);
219 return NULL;
220}
221
222void gntdev_add_map(struct gntdev_priv *priv, struct gntdev_grant_map *add)
223{
224 struct gntdev_grant_map *map;
225
226 list_for_each_entry(map, &priv->maps, next) {
227 if (add->index + add->count < map->index) {
228 list_add_tail(&add->next, &map->next);
229 goto done;
230 }
231 add->index = map->index + map->count;
232 }
233 list_add_tail(&add->next, &priv->maps);
234
235done:
236 gntdev_print_maps(priv, "[new]", add->index);
237}
238
239static struct gntdev_grant_map *gntdev_find_map_index(struct gntdev_priv *priv,
240 int index, int count)
241{
242 struct gntdev_grant_map *map;
243
244 list_for_each_entry(map, &priv->maps, next) {
245 if (map->index != index)
246 continue;
247 if (count && map->count != count)
248 continue;
249 return map;
250 }
251 return NULL;
252}
253
254void gntdev_put_map(struct gntdev_priv *priv, struct gntdev_grant_map *map)
255{
256 if (!map)
257 return;
258
259 if (!refcount_dec_and_test(&map->users))
260 return;
261
262 if (map->pages && !xen_pv_domain()) {
263 /*
264 * Increment the reference count. This ensures that the
265 * subsequent call to unmap_grant_pages() will not wind up
266 * re-entering itself. It *can* wind up calling
267 * gntdev_put_map() recursively, but such calls will be with a
268 * reference count greater than 1, so they will return before
269 * this code is reached. The recursion depth is thus limited to
270 * 1. Do NOT use refcount_inc() here, as it will detect that
271 * the reference count is zero and WARN().
272 */
273 refcount_set(&map->users, 1);
274
275 /*
276 * Unmap the grants. This may or may not be asynchronous, so it
277 * is possible that the reference count is 1 on return, but it
278 * could also be greater than 1.
279 */
280 unmap_grant_pages(map, 0, map->count);
281
282 /* Check if the memory now needs to be freed */
283 if (!refcount_dec_and_test(&map->users))
284 return;
285
286 /*
287 * All pages have been returned to the hypervisor, so free the
288 * map.
289 */
290 }
291
292 if (xen_pv_domain() && map->notifier_init)
293 mmu_interval_notifier_remove(&map->notifier);
294
295 if (map->notify.flags & UNMAP_NOTIFY_SEND_EVENT) {
296 notify_remote_via_evtchn(map->notify.event);
297 evtchn_put(map->notify.event);
298 }
299 gntdev_free_map(map);
300}
301
302/* ------------------------------------------------------------------ */
303
304static int find_grant_ptes(pte_t *pte, unsigned long addr, void *data)
305{
306 struct gntdev_grant_map *map = data;
307 unsigned int pgnr = (addr - map->pages_vm_start) >> PAGE_SHIFT;
308 int flags = map->flags | GNTMAP_application_map | GNTMAP_contains_pte |
309 (1 << _GNTMAP_guest_avail0);
310 u64 pte_maddr;
311
312 BUG_ON(pgnr >= map->count);
313 pte_maddr = arbitrary_virt_to_machine(pte).maddr;
314
315 /* Note: this will perform a pte_mkspecial() through the hypercall. */
316 gnttab_set_map_op(&map->map_ops[pgnr], pte_maddr, flags,
317 map->grants[pgnr].ref,
318 map->grants[pgnr].domid);
319 gnttab_set_unmap_op(&map->unmap_ops[pgnr], pte_maddr, flags,
320 INVALID_GRANT_HANDLE);
321 return 0;
322}
323
324int gntdev_map_grant_pages(struct gntdev_grant_map *map)
325{
326 size_t alloced = 0;
327 int i, err = 0;
328
329 if (!xen_pv_domain()) {
330 /* Note: it could already be mapped */
331 if (map->map_ops[0].handle != INVALID_GRANT_HANDLE)
332 return 0;
333 for (i = 0; i < map->count; i++) {
334 unsigned long addr = (unsigned long)
335 pfn_to_kaddr(page_to_pfn(map->pages[i]));
336 gnttab_set_map_op(&map->map_ops[i], addr, map->flags,
337 map->grants[i].ref,
338 map->grants[i].domid);
339 gnttab_set_unmap_op(&map->unmap_ops[i], addr,
340 map->flags, INVALID_GRANT_HANDLE);
341 }
342 } else {
343 /*
344 * Setup the map_ops corresponding to the pte entries pointing
345 * to the kernel linear addresses of the struct pages.
346 * These ptes are completely different from the user ptes dealt
347 * with find_grant_ptes.
348 * Note that GNTMAP_device_map isn't needed here: The
349 * dev_bus_addr output field gets consumed only from ->map_ops,
350 * and by not requesting it when mapping we also avoid needing
351 * to mirror dev_bus_addr into ->unmap_ops (and holding an extra
352 * reference to the page in the hypervisor).
353 */
354 unsigned int flags = (map->flags & ~GNTMAP_device_map) |
355 GNTMAP_host_map;
356
357 for (i = 0; i < map->count; i++) {
358 unsigned long address = (unsigned long)
359 pfn_to_kaddr(page_to_pfn(map->pages[i]));
360 BUG_ON(PageHighMem(map->pages[i]));
361
362 gnttab_set_map_op(&map->kmap_ops[i], address, flags,
363 map->grants[i].ref,
364 map->grants[i].domid);
365 gnttab_set_unmap_op(&map->kunmap_ops[i], address,
366 flags, INVALID_GRANT_HANDLE);
367 }
368 }
369
370 pr_debug("map %d+%d\n", map->index, map->count);
371 err = gnttab_map_refs(map->map_ops, map->kmap_ops, map->pages,
372 map->count);
373
374 for (i = 0; i < map->count; i++) {
375 if (map->map_ops[i].status == GNTST_okay) {
376 map->unmap_ops[i].handle = map->map_ops[i].handle;
377 alloced++;
378 } else if (!err)
379 err = -EINVAL;
380
381 if (map->flags & GNTMAP_device_map)
382 map->unmap_ops[i].dev_bus_addr = map->map_ops[i].dev_bus_addr;
383
384 if (xen_pv_domain()) {
385 if (map->kmap_ops[i].status == GNTST_okay) {
386 alloced++;
387 map->kunmap_ops[i].handle = map->kmap_ops[i].handle;
388 } else if (!err)
389 err = -EINVAL;
390 }
391 }
392 atomic_add(alloced, &map->live_grants);
393 return err;
394}
395
396static void __unmap_grant_pages_done(int result,
397 struct gntab_unmap_queue_data *data)
398{
399 unsigned int i;
400 struct gntdev_grant_map *map = data->data;
401 unsigned int offset = data->unmap_ops - map->unmap_ops;
402 int successful_unmaps = 0;
403 int live_grants;
404
405 for (i = 0; i < data->count; i++) {
406 if (map->unmap_ops[offset + i].status == GNTST_okay &&
407 map->unmap_ops[offset + i].handle != INVALID_GRANT_HANDLE)
408 successful_unmaps++;
409
410 WARN_ON(map->unmap_ops[offset + i].status != GNTST_okay &&
411 map->unmap_ops[offset + i].handle != INVALID_GRANT_HANDLE);
412 pr_debug("unmap handle=%d st=%d\n",
413 map->unmap_ops[offset+i].handle,
414 map->unmap_ops[offset+i].status);
415 map->unmap_ops[offset+i].handle = INVALID_GRANT_HANDLE;
416 if (xen_pv_domain()) {
417 if (map->kunmap_ops[offset + i].status == GNTST_okay &&
418 map->kunmap_ops[offset + i].handle != INVALID_GRANT_HANDLE)
419 successful_unmaps++;
420
421 WARN_ON(map->kunmap_ops[offset + i].status != GNTST_okay &&
422 map->kunmap_ops[offset + i].handle != INVALID_GRANT_HANDLE);
423 pr_debug("kunmap handle=%u st=%d\n",
424 map->kunmap_ops[offset+i].handle,
425 map->kunmap_ops[offset+i].status);
426 map->kunmap_ops[offset+i].handle = INVALID_GRANT_HANDLE;
427 }
428 }
429
430 /*
431 * Decrease the live-grant counter. This must happen after the loop to
432 * prevent premature reuse of the grants by gnttab_mmap().
433 */
434 live_grants = atomic_sub_return(successful_unmaps, &map->live_grants);
435 if (WARN_ON(live_grants < 0))
436 pr_err("%s: live_grants became negative (%d) after unmapping %d pages!\n",
437 __func__, live_grants, successful_unmaps);
438
439 /* Release reference taken by __unmap_grant_pages */
440 gntdev_put_map(NULL, map);
441}
442
443static void __unmap_grant_pages(struct gntdev_grant_map *map, int offset,
444 int pages)
445{
446 if (map->notify.flags & UNMAP_NOTIFY_CLEAR_BYTE) {
447 int pgno = (map->notify.addr >> PAGE_SHIFT);
448
449 if (pgno >= offset && pgno < offset + pages) {
450 /* No need for kmap, pages are in lowmem */
451 uint8_t *tmp = pfn_to_kaddr(page_to_pfn(map->pages[pgno]));
452
453 tmp[map->notify.addr & (PAGE_SIZE-1)] = 0;
454 map->notify.flags &= ~UNMAP_NOTIFY_CLEAR_BYTE;
455 }
456 }
457
458 map->unmap_data.unmap_ops = map->unmap_ops + offset;
459 map->unmap_data.kunmap_ops = xen_pv_domain() ? map->kunmap_ops + offset : NULL;
460 map->unmap_data.pages = map->pages + offset;
461 map->unmap_data.count = pages;
462 map->unmap_data.done = __unmap_grant_pages_done;
463 map->unmap_data.data = map;
464 refcount_inc(&map->users); /* to keep map alive during async call below */
465
466 gnttab_unmap_refs_async(&map->unmap_data);
467}
468
469static void unmap_grant_pages(struct gntdev_grant_map *map, int offset,
470 int pages)
471{
472 int range;
473
474 if (atomic_read(&map->live_grants) == 0)
475 return; /* Nothing to do */
476
477 pr_debug("unmap %d+%d [%d+%d]\n", map->index, map->count, offset, pages);
478
479 /* It is possible the requested range will have a "hole" where we
480 * already unmapped some of the grants. Only unmap valid ranges.
481 */
482 while (pages) {
483 while (pages && map->being_removed[offset]) {
484 offset++;
485 pages--;
486 }
487 range = 0;
488 while (range < pages) {
489 if (map->being_removed[offset + range])
490 break;
491 map->being_removed[offset + range] = true;
492 range++;
493 }
494 if (range)
495 __unmap_grant_pages(map, offset, range);
496 offset += range;
497 pages -= range;
498 }
499}
500
501/* ------------------------------------------------------------------ */
502
503static void gntdev_vma_open(struct vm_area_struct *vma)
504{
505 struct gntdev_grant_map *map = vma->vm_private_data;
506
507 pr_debug("gntdev_vma_open %p\n", vma);
508 refcount_inc(&map->users);
509}
510
511static void gntdev_vma_close(struct vm_area_struct *vma)
512{
513 struct gntdev_grant_map *map = vma->vm_private_data;
514 struct file *file = vma->vm_file;
515 struct gntdev_priv *priv = file->private_data;
516
517 pr_debug("gntdev_vma_close %p\n", vma);
518
519 vma->vm_private_data = NULL;
520 gntdev_put_map(priv, map);
521}
522
523static struct page *gntdev_vma_find_normal_page(struct vm_area_struct *vma,
524 unsigned long addr)
525{
526 struct gntdev_grant_map *map = vma->vm_private_data;
527
528 return map->pages[(addr - map->pages_vm_start) >> PAGE_SHIFT];
529}
530
531static const struct vm_operations_struct gntdev_vmops = {
532 .open = gntdev_vma_open,
533 .close = gntdev_vma_close,
534 .find_normal_page = gntdev_vma_find_normal_page,
535};
536
537/* ------------------------------------------------------------------ */
538
539static bool gntdev_invalidate(struct mmu_interval_notifier *mn,
540 const struct mmu_notifier_range *range,
541 unsigned long cur_seq)
542{
543 struct gntdev_grant_map *map =
544 container_of(mn, struct gntdev_grant_map, notifier);
545 unsigned long mstart, mend;
546 unsigned long map_start, map_end;
547
548 if (!mmu_notifier_range_blockable(range))
549 return false;
550
551 map_start = map->pages_vm_start;
552 map_end = map->pages_vm_start + (map->count << PAGE_SHIFT);
553
554 /*
555 * If the VMA is split or otherwise changed the notifier is not
556 * updated, but we don't want to process VA's outside the modified
557 * VMA. FIXME: It would be much more understandable to just prevent
558 * modifying the VMA in the first place.
559 */
560 if (map_start >= range->end || map_end <= range->start)
561 return true;
562
563 mstart = max(range->start, map_start);
564 mend = min(range->end, map_end);
565 pr_debug("map %d+%d (%lx %lx), range %lx %lx, mrange %lx %lx\n",
566 map->index, map->count, map_start, map_end,
567 range->start, range->end, mstart, mend);
568 unmap_grant_pages(map, (mstart - map_start) >> PAGE_SHIFT,
569 (mend - mstart) >> PAGE_SHIFT);
570
571 return true;
572}
573
574static const struct mmu_interval_notifier_ops gntdev_mmu_ops = {
575 .invalidate = gntdev_invalidate,
576};
577
578/* ------------------------------------------------------------------ */
579
580static int gntdev_open(struct inode *inode, struct file *flip)
581{
582 struct gntdev_priv *priv;
583
584 priv = kzalloc_obj(*priv);
585 if (!priv)
586 return -ENOMEM;
587
588 INIT_LIST_HEAD(&priv->maps);
589 mutex_init(&priv->lock);
590
591 mutex_init(&priv->batch_lock);
592
593#ifdef CONFIG_XEN_GNTDEV_DMABUF
594 priv->dmabuf_priv = gntdev_dmabuf_init(flip);
595 if (IS_ERR(priv->dmabuf_priv)) {
596 int ret = PTR_ERR(priv->dmabuf_priv);
597
598 kfree(priv);
599 return ret;
600 }
601#endif
602
603 flip->private_data = priv;
604#ifdef CONFIG_XEN_GRANT_DMA_ALLOC
605 priv->dma_dev = gntdev_miscdev.this_device;
606 dma_coerce_mask_and_coherent(priv->dma_dev, DMA_BIT_MASK(64));
607#endif
608 pr_debug("priv %p\n", priv);
609
610 return 0;
611}
612
613static int gntdev_release(struct inode *inode, struct file *flip)
614{
615 struct gntdev_priv *priv = flip->private_data;
616 struct gntdev_grant_map *map;
617 struct gntdev_copy_batch *batch;
618
619 pr_debug("priv %p\n", priv);
620
621 mutex_lock(&priv->lock);
622 while (!list_empty(&priv->maps)) {
623 map = list_entry(priv->maps.next,
624 struct gntdev_grant_map, next);
625 list_del(&map->next);
626 gntdev_put_map(NULL /* already removed */, map);
627 }
628 mutex_unlock(&priv->lock);
629
630 mutex_lock(&priv->batch_lock);
631 while (priv->batch) {
632 batch = priv->batch;
633 priv->batch = batch->next;
634 kfree(batch);
635 }
636 mutex_unlock(&priv->batch_lock);
637
638#ifdef CONFIG_XEN_GNTDEV_DMABUF
639 gntdev_dmabuf_fini(priv->dmabuf_priv);
640#endif
641
642 kfree(priv);
643 return 0;
644}
645
646static long gntdev_ioctl_map_grant_ref(struct gntdev_priv *priv,
647 struct ioctl_gntdev_map_grant_ref __user *u)
648{
649 struct ioctl_gntdev_map_grant_ref op;
650 struct gntdev_grant_map *map;
651 int err;
652
653 if (copy_from_user(&op, u, sizeof(op)) != 0)
654 return -EFAULT;
655 pr_debug("priv %p, add %d\n", priv, op.count);
656 if (unlikely(gntdev_test_page_count(op.count)))
657 return -EINVAL;
658
659 err = -ENOMEM;
660 map = gntdev_alloc_map(priv, op.count, 0 /* This is not a dma-buf. */);
661 if (!map)
662 return err;
663
664 if (copy_from_user(map->grants, &u->refs,
665 sizeof(map->grants[0]) * op.count) != 0) {
666 gntdev_put_map(NULL, map);
667 return -EFAULT;
668 }
669
670 mutex_lock(&priv->lock);
671 gntdev_add_map(priv, map);
672 op.index = map->index << PAGE_SHIFT;
673
674 if (copy_to_user(u, &op, sizeof(op)) != 0) {
675 list_del(&map->next);
676 mutex_unlock(&priv->lock);
677 gntdev_put_map(priv, map);
678 return -EFAULT;
679 }
680
681 mutex_unlock(&priv->lock);
682 return 0;
683}
684
685static long gntdev_ioctl_unmap_grant_ref(struct gntdev_priv *priv,
686 struct ioctl_gntdev_unmap_grant_ref __user *u)
687{
688 struct ioctl_gntdev_unmap_grant_ref op;
689 struct gntdev_grant_map *map;
690 int err = -ENOENT;
691
692 if (copy_from_user(&op, u, sizeof(op)) != 0)
693 return -EFAULT;
694 pr_debug("priv %p, del %d+%d\n", priv, (int)op.index, (int)op.count);
695
696 mutex_lock(&priv->lock);
697 map = gntdev_find_map_index(priv, op.index >> PAGE_SHIFT, op.count);
698 if (map) {
699 list_del(&map->next);
700 err = 0;
701 }
702 mutex_unlock(&priv->lock);
703 if (map)
704 gntdev_put_map(priv, map);
705 return err;
706}
707
708static long gntdev_ioctl_get_offset_for_vaddr(struct gntdev_priv *priv,
709 struct ioctl_gntdev_get_offset_for_vaddr __user *u)
710{
711 struct ioctl_gntdev_get_offset_for_vaddr op;
712 struct vm_area_struct *vma;
713 struct gntdev_grant_map *map;
714 int rv = -EINVAL;
715
716 if (copy_from_user(&op, u, sizeof(op)) != 0)
717 return -EFAULT;
718 pr_debug("priv %p, offset for vaddr %lx\n", priv, (unsigned long)op.vaddr);
719
720 mmap_read_lock(current->mm);
721 vma = find_vma(current->mm, op.vaddr);
722 if (!vma || vma->vm_ops != &gntdev_vmops)
723 goto out_unlock;
724
725 map = vma->vm_private_data;
726 if (!map)
727 goto out_unlock;
728
729 op.offset = map->index << PAGE_SHIFT;
730 op.count = map->count;
731 rv = 0;
732
733 out_unlock:
734 mmap_read_unlock(current->mm);
735
736 if (rv == 0 && copy_to_user(u, &op, sizeof(op)) != 0)
737 return -EFAULT;
738 return rv;
739}
740
741static long gntdev_ioctl_notify(struct gntdev_priv *priv, void __user *u)
742{
743 struct ioctl_gntdev_unmap_notify op;
744 struct gntdev_grant_map *map;
745 int rc;
746 int out_flags;
747 evtchn_port_t out_event;
748
749 if (copy_from_user(&op, u, sizeof(op)))
750 return -EFAULT;
751
752 if (op.action & ~(UNMAP_NOTIFY_CLEAR_BYTE|UNMAP_NOTIFY_SEND_EVENT))
753 return -EINVAL;
754
755 /* We need to grab a reference to the event channel we are going to use
756 * to send the notify before releasing the reference we may already have
757 * (if someone has called this ioctl twice). This is required so that
758 * it is possible to change the clear_byte part of the notification
759 * without disturbing the event channel part, which may now be the last
760 * reference to that event channel.
761 */
762 if (op.action & UNMAP_NOTIFY_SEND_EVENT) {
763 if (evtchn_get(op.event_channel_port))
764 return -EINVAL;
765 }
766
767 out_flags = op.action;
768 out_event = op.event_channel_port;
769
770 mutex_lock(&priv->lock);
771
772 list_for_each_entry(map, &priv->maps, next) {
773 uint64_t begin = map->index << PAGE_SHIFT;
774 uint64_t end = (map->index + map->count) << PAGE_SHIFT;
775 if (op.index >= begin && op.index < end)
776 goto found;
777 }
778 rc = -ENOENT;
779 goto unlock_out;
780
781 found:
782 if ((op.action & UNMAP_NOTIFY_CLEAR_BYTE) &&
783 (map->flags & GNTMAP_readonly)) {
784 rc = -EINVAL;
785 goto unlock_out;
786 }
787
788 out_flags = map->notify.flags;
789 out_event = map->notify.event;
790
791 map->notify.flags = op.action;
792 map->notify.addr = op.index - (map->index << PAGE_SHIFT);
793 map->notify.event = op.event_channel_port;
794
795 rc = 0;
796
797 unlock_out:
798 mutex_unlock(&priv->lock);
799
800 /* Drop the reference to the event channel we did not save in the map */
801 if (out_flags & UNMAP_NOTIFY_SEND_EVENT)
802 evtchn_put(out_event);
803
804 return rc;
805}
806
807static int gntdev_get_page(struct gntdev_copy_batch *batch, void __user *virt,
808 unsigned long *gfn)
809{
810 unsigned long addr = (unsigned long)virt;
811 struct page *page;
812 unsigned long xen_pfn;
813 int ret;
814
815 ret = pin_user_pages_fast(addr, 1, batch->writeable ? FOLL_WRITE : 0, &page);
816 if (ret < 0)
817 return ret;
818
819 batch->pages[batch->nr_pages++] = page;
820
821 xen_pfn = page_to_xen_pfn(page) + XEN_PFN_DOWN(addr & ~PAGE_MASK);
822 *gfn = pfn_to_gfn(xen_pfn);
823
824 return 0;
825}
826
827static void gntdev_put_pages(struct gntdev_copy_batch *batch)
828{
829 unpin_user_pages_dirty_lock(batch->pages, batch->nr_pages, batch->writeable);
830 batch->nr_pages = 0;
831 batch->writeable = false;
832}
833
834static int gntdev_copy(struct gntdev_copy_batch *batch)
835{
836 unsigned int i;
837
838 gnttab_batch_copy(batch->ops, batch->nr_ops);
839 gntdev_put_pages(batch);
840
841 /*
842 * For each completed op, update the status if the op failed
843 * and all previous ops for the segment were successful.
844 */
845 for (i = 0; i < batch->nr_ops; i++) {
846 s16 status = batch->ops[i].status;
847 s16 old_status;
848
849 if (status == GNTST_okay)
850 continue;
851
852 if (__get_user(old_status, batch->status[i]))
853 return -EFAULT;
854
855 if (old_status != GNTST_okay)
856 continue;
857
858 if (__put_user(status, batch->status[i]))
859 return -EFAULT;
860 }
861
862 batch->nr_ops = 0;
863 return 0;
864}
865
866static int gntdev_grant_copy_seg(struct gntdev_copy_batch *batch,
867 struct gntdev_grant_copy_segment *seg,
868 s16 __user *status)
869{
870 uint16_t copied = 0;
871
872 /*
873 * Disallow local -> local copies since there is only space in
874 * batch->pages for one page per-op and this would be a very
875 * expensive memcpy().
876 */
877 if (!(seg->flags & (GNTCOPY_source_gref | GNTCOPY_dest_gref)))
878 return -EINVAL;
879
880 /* Can't cross page if source/dest is a grant ref. */
881 if (seg->flags & GNTCOPY_source_gref) {
882 if (seg->source.foreign.offset + seg->len > XEN_PAGE_SIZE)
883 return -EINVAL;
884 }
885 if (seg->flags & GNTCOPY_dest_gref) {
886 if (seg->dest.foreign.offset + seg->len > XEN_PAGE_SIZE)
887 return -EINVAL;
888 }
889
890 if (put_user(GNTST_okay, status))
891 return -EFAULT;
892
893 while (copied < seg->len) {
894 struct gnttab_copy *op;
895 void __user *virt;
896 size_t len, off;
897 unsigned long gfn;
898 int ret;
899
900 if (batch->nr_ops >= GNTDEV_COPY_BATCH) {
901 ret = gntdev_copy(batch);
902 if (ret < 0)
903 return ret;
904 }
905
906 len = seg->len - copied;
907
908 op = &batch->ops[batch->nr_ops];
909 op->flags = 0;
910
911 if (seg->flags & GNTCOPY_source_gref) {
912 op->source.u.ref = seg->source.foreign.ref;
913 op->source.domid = seg->source.foreign.domid;
914 op->source.offset = seg->source.foreign.offset + copied;
915 op->flags |= GNTCOPY_source_gref;
916 } else {
917 virt = seg->source.virt + copied;
918 off = (unsigned long)virt & ~XEN_PAGE_MASK;
919 len = min(len, (size_t)XEN_PAGE_SIZE - off);
920 batch->writeable = false;
921
922 ret = gntdev_get_page(batch, virt, &gfn);
923 if (ret < 0)
924 return ret;
925
926 op->source.u.gmfn = gfn;
927 op->source.domid = DOMID_SELF;
928 op->source.offset = off;
929 }
930
931 if (seg->flags & GNTCOPY_dest_gref) {
932 op->dest.u.ref = seg->dest.foreign.ref;
933 op->dest.domid = seg->dest.foreign.domid;
934 op->dest.offset = seg->dest.foreign.offset + copied;
935 op->flags |= GNTCOPY_dest_gref;
936 } else {
937 virt = seg->dest.virt + copied;
938 off = (unsigned long)virt & ~XEN_PAGE_MASK;
939 len = min(len, (size_t)XEN_PAGE_SIZE - off);
940 batch->writeable = true;
941
942 ret = gntdev_get_page(batch, virt, &gfn);
943 if (ret < 0)
944 return ret;
945
946 op->dest.u.gmfn = gfn;
947 op->dest.domid = DOMID_SELF;
948 op->dest.offset = off;
949 }
950
951 op->len = len;
952 copied += len;
953
954 batch->status[batch->nr_ops] = status;
955 batch->nr_ops++;
956 }
957
958 return 0;
959}
960
961static long gntdev_ioctl_grant_copy(struct gntdev_priv *priv, void __user *u)
962{
963 struct ioctl_gntdev_grant_copy copy;
964 struct gntdev_copy_batch *batch;
965 unsigned int i;
966 int ret = 0;
967
968 if (copy_from_user(&copy, u, sizeof(copy)))
969 return -EFAULT;
970
971 mutex_lock(&priv->batch_lock);
972 if (!priv->batch) {
973 batch = kmalloc_obj(*batch);
974 } else {
975 batch = priv->batch;
976 priv->batch = batch->next;
977 }
978 mutex_unlock(&priv->batch_lock);
979 if (!batch)
980 return -ENOMEM;
981
982 batch->nr_ops = 0;
983 batch->nr_pages = 0;
984
985 for (i = 0; i < copy.count; i++) {
986 struct gntdev_grant_copy_segment seg;
987
988 if (copy_from_user(&seg, &copy.segments[i], sizeof(seg))) {
989 ret = -EFAULT;
990 gntdev_put_pages(batch);
991 goto out;
992 }
993
994 ret = gntdev_grant_copy_seg(batch, &seg, &copy.segments[i].status);
995 if (ret < 0) {
996 gntdev_put_pages(batch);
997 goto out;
998 }
999
1000 cond_resched();
1001 }
1002 if (batch->nr_ops)
1003 ret = gntdev_copy(batch);
1004
1005 out:
1006 mutex_lock(&priv->batch_lock);
1007 batch->next = priv->batch;
1008 priv->batch = batch;
1009 mutex_unlock(&priv->batch_lock);
1010
1011 return ret;
1012}
1013
1014static long gntdev_ioctl(struct file *flip,
1015 unsigned int cmd, unsigned long arg)
1016{
1017 struct gntdev_priv *priv = flip->private_data;
1018 void __user *ptr = (void __user *)arg;
1019
1020 switch (cmd) {
1021 case IOCTL_GNTDEV_MAP_GRANT_REF:
1022 return gntdev_ioctl_map_grant_ref(priv, ptr);
1023
1024 case IOCTL_GNTDEV_UNMAP_GRANT_REF:
1025 return gntdev_ioctl_unmap_grant_ref(priv, ptr);
1026
1027 case IOCTL_GNTDEV_GET_OFFSET_FOR_VADDR:
1028 return gntdev_ioctl_get_offset_for_vaddr(priv, ptr);
1029
1030 case IOCTL_GNTDEV_SET_UNMAP_NOTIFY:
1031 return gntdev_ioctl_notify(priv, ptr);
1032
1033 case IOCTL_GNTDEV_GRANT_COPY:
1034 return gntdev_ioctl_grant_copy(priv, ptr);
1035
1036#ifdef CONFIG_XEN_GNTDEV_DMABUF
1037 case IOCTL_GNTDEV_DMABUF_EXP_FROM_REFS:
1038 return gntdev_ioctl_dmabuf_exp_from_refs(priv, ptr);
1039
1040 case IOCTL_GNTDEV_DMABUF_EXP_WAIT_RELEASED:
1041 return gntdev_ioctl_dmabuf_exp_wait_released(priv, ptr);
1042
1043 case IOCTL_GNTDEV_DMABUF_IMP_TO_REFS:
1044 return gntdev_ioctl_dmabuf_imp_to_refs(priv, ptr);
1045
1046 case IOCTL_GNTDEV_DMABUF_IMP_RELEASE:
1047 return gntdev_ioctl_dmabuf_imp_release(priv, ptr);
1048#endif
1049
1050 default:
1051 pr_debug("priv %p, unknown cmd %x\n", priv, cmd);
1052 return -ENOIOCTLCMD;
1053 }
1054
1055 return 0;
1056}
1057
1058static int gntdev_mmap(struct file *flip, struct vm_area_struct *vma)
1059{
1060 struct gntdev_priv *priv = flip->private_data;
1061 int index = vma->vm_pgoff;
1062 int count = vma_pages(vma);
1063 struct gntdev_grant_map *map;
1064 int err = -EINVAL;
1065
1066 if ((vma->vm_flags & VM_WRITE) && !(vma->vm_flags & VM_SHARED))
1067 return -EINVAL;
1068
1069 /*
1070 * Private mappings of foreign grant pages must never become
1071 * writable: they would take the COW path on granted pages.
1072 */
1073 if (!(vma->vm_flags & VM_SHARED))
1074 vm_flags_clear(vma, VM_MAYWRITE);
1075
1076 pr_debug("map %d+%d at %lx (pgoff %lx)\n",
1077 index, count, vma->vm_start, vma->vm_pgoff);
1078
1079 mutex_lock(&priv->lock);
1080 map = gntdev_find_map_index(priv, index, count);
1081 if (!map)
1082 goto unlock_out;
1083 if (!atomic_add_unless(&map->in_use, 1, 1))
1084 goto unlock_out;
1085
1086 refcount_inc(&map->users);
1087
1088 vma->vm_ops = &gntdev_vmops;
1089
1090 vm_flags_set(vma, VM_DONTEXPAND | VM_DONTDUMP | VM_MIXEDMAP);
1091
1092 if (xen_pv_domain())
1093 vm_flags_set(vma, VM_DONTCOPY);
1094
1095 vma->vm_private_data = map;
1096 if (map->flags) {
1097 if ((vma->vm_flags & VM_WRITE) &&
1098 (map->flags & GNTMAP_readonly))
1099 goto out_unlock_put;
1100 } else {
1101 map->flags = GNTMAP_host_map;
1102 if (!(vma->vm_flags & VM_WRITE))
1103 map->flags |= GNTMAP_readonly;
1104 }
1105
1106 map->pages_vm_start = vma->vm_start;
1107
1108 if (xen_pv_domain()) {
1109 err = mmu_interval_notifier_insert_locked(
1110 &map->notifier, vma->vm_mm, vma->vm_start,
1111 vma->vm_end - vma->vm_start, &gntdev_mmu_ops);
1112 if (err)
1113 goto out_unlock_put;
1114
1115 map->notifier_init = true;
1116 }
1117 mutex_unlock(&priv->lock);
1118
1119 if (xen_pv_domain()) {
1120 /*
1121 * gntdev takes the address of the PTE in find_grant_ptes() and
1122 * passes it to the hypervisor in gntdev_map_grant_pages(). The
1123 * purpose of the notifier is to prevent the hypervisor pointer
1124 * to the PTE from going stale.
1125 *
1126 * Since this vma's mappings can't be touched without the
1127 * mmap_lock, and we are holding it now, there is no need for
1128 * the notifier_range locking pattern.
1129 */
1130 mmu_interval_read_begin(&map->notifier);
1131
1132 err = apply_to_page_range(vma->vm_mm, vma->vm_start,
1133 vma->vm_end - vma->vm_start,
1134 find_grant_ptes, map);
1135 if (err) {
1136 pr_warn("find_grant_ptes() failure.\n");
1137 goto out_put_map;
1138 }
1139 }
1140
1141 err = gntdev_map_grant_pages(map);
1142 if (err)
1143 goto out_put_map;
1144
1145 if (!xen_pv_domain()) {
1146 err = vm_map_pages_zero(vma, map->pages, map->count);
1147 if (err)
1148 goto out_put_map;
1149 }
1150
1151 return 0;
1152
1153unlock_out:
1154 mutex_unlock(&priv->lock);
1155 return err;
1156
1157out_unlock_put:
1158 mutex_unlock(&priv->lock);
1159out_put_map:
1160 if (xen_pv_domain())
1161 unmap_grant_pages(map, 0, map->count);
1162 gntdev_put_map(priv, map);
1163 return err;
1164}
1165
1166static const struct file_operations gntdev_fops = {
1167 .owner = THIS_MODULE,
1168 .open = gntdev_open,
1169 .release = gntdev_release,
1170 .mmap = gntdev_mmap,
1171 .unlocked_ioctl = gntdev_ioctl
1172};
1173
1174static struct miscdevice gntdev_miscdev = {
1175 .minor = MISC_DYNAMIC_MINOR,
1176 .name = "xen/gntdev",
1177 .fops = &gntdev_fops,
1178};
1179
1180/* ------------------------------------------------------------------ */
1181
1182static int __init gntdev_init(void)
1183{
1184 int err;
1185
1186 if (!xen_domain())
1187 return -ENODEV;
1188
1189 err = misc_register(&gntdev_miscdev);
1190 if (err != 0) {
1191 pr_err("Could not register gntdev device\n");
1192 return err;
1193 }
1194 return 0;
1195}
1196
1197static void __exit gntdev_exit(void)
1198{
1199 misc_deregister(&gntdev_miscdev);
1200}
1201
1202module_init(gntdev_init);
1203module_exit(gntdev_exit);
1204
1205/* ------------------------------------------------------------------ */