xref: /linux/drivers/xen/gntdev.c (revision e80a48bade619ec5a92230b3d4ae84bfc2746822)
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 <kraxel@redhat.com>
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 
54 MODULE_LICENSE("GPL");
55 MODULE_AUTHOR("Derek G. Murray <Derek.Murray@cl.cam.ac.uk>, "
56 	      "Gerd Hoffmann <kraxel@redhat.com>");
57 MODULE_DESCRIPTION("User-space granted page access driver");
58 
59 static unsigned int limit = 64*1024;
60 module_param(limit, uint, 0644);
61 MODULE_PARM_DESC(limit,
62 	"Maximum number of grants that may be mapped by one mapping request");
63 
64 /* True in PV mode, false otherwise */
65 static int use_ptemod;
66 
67 static void unmap_grant_pages(struct gntdev_grant_map *map,
68 			      int offset, int pages);
69 
70 static struct miscdevice gntdev_miscdev;
71 
72 /* ------------------------------------------------------------------ */
73 
74 bool gntdev_test_page_count(unsigned int count)
75 {
76 	return !count || count > limit;
77 }
78 
79 static void gntdev_print_maps(struct gntdev_priv *priv,
80 			      char *text, int text_index)
81 {
82 #ifdef DEBUG
83 	struct gntdev_grant_map *map;
84 
85 	pr_debug("%s: maps list (priv %p)\n", __func__, priv);
86 	list_for_each_entry(map, &priv->maps, next)
87 		pr_debug("  index %2d, count %2d %s\n",
88 		       map->index, map->count,
89 		       map->index == text_index && text ? text : "");
90 #endif
91 }
92 
93 static void gntdev_free_map(struct gntdev_grant_map *map)
94 {
95 	if (map == NULL)
96 		return;
97 
98 #ifdef CONFIG_XEN_GRANT_DMA_ALLOC
99 	if (map->dma_vaddr) {
100 		struct gnttab_dma_alloc_args args;
101 
102 		args.dev = map->dma_dev;
103 		args.coherent = !!(map->dma_flags & GNTDEV_DMA_FLAG_COHERENT);
104 		args.nr_pages = map->count;
105 		args.pages = map->pages;
106 		args.frames = map->frames;
107 		args.vaddr = map->dma_vaddr;
108 		args.dev_bus_addr = map->dma_bus_addr;
109 
110 		gnttab_dma_free_pages(&args);
111 	} else
112 #endif
113 	if (map->pages)
114 		gnttab_free_pages(map->count, map->pages);
115 
116 #ifdef CONFIG_XEN_GRANT_DMA_ALLOC
117 	kvfree(map->frames);
118 #endif
119 	kvfree(map->pages);
120 	kvfree(map->grants);
121 	kvfree(map->map_ops);
122 	kvfree(map->unmap_ops);
123 	kvfree(map->kmap_ops);
124 	kvfree(map->kunmap_ops);
125 	kvfree(map->being_removed);
126 	kfree(map);
127 }
128 
129 struct gntdev_grant_map *gntdev_alloc_map(struct gntdev_priv *priv, int count,
130 					  int dma_flags)
131 {
132 	struct gntdev_grant_map *add;
133 	int i;
134 
135 	add = kzalloc(sizeof(*add), GFP_KERNEL);
136 	if (NULL == add)
137 		return NULL;
138 
139 	add->grants    = kvmalloc_array(count, sizeof(add->grants[0]),
140 					GFP_KERNEL);
141 	add->map_ops   = kvmalloc_array(count, sizeof(add->map_ops[0]),
142 					GFP_KERNEL);
143 	add->unmap_ops = kvmalloc_array(count, sizeof(add->unmap_ops[0]),
144 					GFP_KERNEL);
145 	add->pages     = kvcalloc(count, sizeof(add->pages[0]), GFP_KERNEL);
146 	add->being_removed =
147 		kvcalloc(count, sizeof(add->being_removed[0]), GFP_KERNEL);
148 	if (NULL == add->grants    ||
149 	    NULL == add->map_ops   ||
150 	    NULL == add->unmap_ops ||
151 	    NULL == add->pages     ||
152 	    NULL == add->being_removed)
153 		goto err;
154 	if (use_ptemod) {
155 		add->kmap_ops   = kvmalloc_array(count, sizeof(add->kmap_ops[0]),
156 						 GFP_KERNEL);
157 		add->kunmap_ops = kvmalloc_array(count, sizeof(add->kunmap_ops[0]),
158 						 GFP_KERNEL);
159 		if (NULL == add->kmap_ops || NULL == add->kunmap_ops)
160 			goto err;
161 	}
162 
163 #ifdef CONFIG_XEN_GRANT_DMA_ALLOC
164 	add->dma_flags = dma_flags;
165 
166 	/*
167 	 * Check if this mapping is requested to be backed
168 	 * by a DMA buffer.
169 	 */
170 	if (dma_flags & (GNTDEV_DMA_FLAG_WC | GNTDEV_DMA_FLAG_COHERENT)) {
171 		struct gnttab_dma_alloc_args args;
172 
173 		add->frames = kvcalloc(count, sizeof(add->frames[0]),
174 				       GFP_KERNEL);
175 		if (!add->frames)
176 			goto err;
177 
178 		/* Remember the device, so we can free DMA memory. */
179 		add->dma_dev = priv->dma_dev;
180 
181 		args.dev = priv->dma_dev;
182 		args.coherent = !!(dma_flags & GNTDEV_DMA_FLAG_COHERENT);
183 		args.nr_pages = count;
184 		args.pages = add->pages;
185 		args.frames = add->frames;
186 
187 		if (gnttab_dma_alloc_pages(&args))
188 			goto err;
189 
190 		add->dma_vaddr = args.vaddr;
191 		add->dma_bus_addr = args.dev_bus_addr;
192 	} else
193 #endif
194 	if (gnttab_alloc_pages(count, add->pages))
195 		goto err;
196 
197 	for (i = 0; i < count; i++) {
198 		add->grants[i].domid = DOMID_INVALID;
199 		add->grants[i].ref = INVALID_GRANT_REF;
200 		add->map_ops[i].handle = INVALID_GRANT_HANDLE;
201 		add->unmap_ops[i].handle = INVALID_GRANT_HANDLE;
202 		if (use_ptemod) {
203 			add->kmap_ops[i].handle = INVALID_GRANT_HANDLE;
204 			add->kunmap_ops[i].handle = INVALID_GRANT_HANDLE;
205 		}
206 	}
207 
208 	add->index = 0;
209 	add->count = count;
210 	refcount_set(&add->users, 1);
211 
212 	return add;
213 
214 err:
215 	gntdev_free_map(add);
216 	return NULL;
217 }
218 
219 void gntdev_add_map(struct gntdev_priv *priv, struct gntdev_grant_map *add)
220 {
221 	struct gntdev_grant_map *map;
222 
223 	list_for_each_entry(map, &priv->maps, next) {
224 		if (add->index + add->count < map->index) {
225 			list_add_tail(&add->next, &map->next);
226 			goto done;
227 		}
228 		add->index = map->index + map->count;
229 	}
230 	list_add_tail(&add->next, &priv->maps);
231 
232 done:
233 	gntdev_print_maps(priv, "[new]", add->index);
234 }
235 
236 static struct gntdev_grant_map *gntdev_find_map_index(struct gntdev_priv *priv,
237 						      int index, int count)
238 {
239 	struct gntdev_grant_map *map;
240 
241 	list_for_each_entry(map, &priv->maps, next) {
242 		if (map->index != index)
243 			continue;
244 		if (count && map->count != count)
245 			continue;
246 		return map;
247 	}
248 	return NULL;
249 }
250 
251 void gntdev_put_map(struct gntdev_priv *priv, struct gntdev_grant_map *map)
252 {
253 	if (!map)
254 		return;
255 
256 	if (!refcount_dec_and_test(&map->users))
257 		return;
258 
259 	if (map->pages && !use_ptemod) {
260 		/*
261 		 * Increment the reference count.  This ensures that the
262 		 * subsequent call to unmap_grant_pages() will not wind up
263 		 * re-entering itself.  It *can* wind up calling
264 		 * gntdev_put_map() recursively, but such calls will be with a
265 		 * reference count greater than 1, so they will return before
266 		 * this code is reached.  The recursion depth is thus limited to
267 		 * 1.  Do NOT use refcount_inc() here, as it will detect that
268 		 * the reference count is zero and WARN().
269 		 */
270 		refcount_set(&map->users, 1);
271 
272 		/*
273 		 * Unmap the grants.  This may or may not be asynchronous, so it
274 		 * is possible that the reference count is 1 on return, but it
275 		 * could also be greater than 1.
276 		 */
277 		unmap_grant_pages(map, 0, map->count);
278 
279 		/* Check if the memory now needs to be freed */
280 		if (!refcount_dec_and_test(&map->users))
281 			return;
282 
283 		/*
284 		 * All pages have been returned to the hypervisor, so free the
285 		 * map.
286 		 */
287 	}
288 
289 	if (use_ptemod && map->notifier_init)
290 		mmu_interval_notifier_remove(&map->notifier);
291 
292 	if (map->notify.flags & UNMAP_NOTIFY_SEND_EVENT) {
293 		notify_remote_via_evtchn(map->notify.event);
294 		evtchn_put(map->notify.event);
295 	}
296 	gntdev_free_map(map);
297 }
298 
299 /* ------------------------------------------------------------------ */
300 
301 static int find_grant_ptes(pte_t *pte, unsigned long addr, void *data)
302 {
303 	struct gntdev_grant_map *map = data;
304 	unsigned int pgnr = (addr - map->pages_vm_start) >> PAGE_SHIFT;
305 	int flags = map->flags | GNTMAP_application_map | GNTMAP_contains_pte |
306 		    (1 << _GNTMAP_guest_avail0);
307 	u64 pte_maddr;
308 
309 	BUG_ON(pgnr >= map->count);
310 	pte_maddr = arbitrary_virt_to_machine(pte).maddr;
311 
312 	gnttab_set_map_op(&map->map_ops[pgnr], pte_maddr, flags,
313 			  map->grants[pgnr].ref,
314 			  map->grants[pgnr].domid);
315 	gnttab_set_unmap_op(&map->unmap_ops[pgnr], pte_maddr, flags,
316 			    INVALID_GRANT_HANDLE);
317 	return 0;
318 }
319 
320 int gntdev_map_grant_pages(struct gntdev_grant_map *map)
321 {
322 	size_t alloced = 0;
323 	int i, err = 0;
324 
325 	if (!use_ptemod) {
326 		/* Note: it could already be mapped */
327 		if (map->map_ops[0].handle != INVALID_GRANT_HANDLE)
328 			return 0;
329 		for (i = 0; i < map->count; i++) {
330 			unsigned long addr = (unsigned long)
331 				pfn_to_kaddr(page_to_pfn(map->pages[i]));
332 			gnttab_set_map_op(&map->map_ops[i], addr, map->flags,
333 				map->grants[i].ref,
334 				map->grants[i].domid);
335 			gnttab_set_unmap_op(&map->unmap_ops[i], addr,
336 				map->flags, INVALID_GRANT_HANDLE);
337 		}
338 	} else {
339 		/*
340 		 * Setup the map_ops corresponding to the pte entries pointing
341 		 * to the kernel linear addresses of the struct pages.
342 		 * These ptes are completely different from the user ptes dealt
343 		 * with find_grant_ptes.
344 		 * Note that GNTMAP_device_map isn't needed here: The
345 		 * dev_bus_addr output field gets consumed only from ->map_ops,
346 		 * and by not requesting it when mapping we also avoid needing
347 		 * to mirror dev_bus_addr into ->unmap_ops (and holding an extra
348 		 * reference to the page in the hypervisor).
349 		 */
350 		unsigned int flags = (map->flags & ~GNTMAP_device_map) |
351 				     GNTMAP_host_map;
352 
353 		for (i = 0; i < map->count; i++) {
354 			unsigned long address = (unsigned long)
355 				pfn_to_kaddr(page_to_pfn(map->pages[i]));
356 			BUG_ON(PageHighMem(map->pages[i]));
357 
358 			gnttab_set_map_op(&map->kmap_ops[i], address, flags,
359 				map->grants[i].ref,
360 				map->grants[i].domid);
361 			gnttab_set_unmap_op(&map->kunmap_ops[i], address,
362 				flags, INVALID_GRANT_HANDLE);
363 		}
364 	}
365 
366 	pr_debug("map %d+%d\n", map->index, map->count);
367 	err = gnttab_map_refs(map->map_ops, map->kmap_ops, map->pages,
368 			map->count);
369 
370 	for (i = 0; i < map->count; i++) {
371 		if (map->map_ops[i].status == GNTST_okay) {
372 			map->unmap_ops[i].handle = map->map_ops[i].handle;
373 			alloced++;
374 		} else if (!err)
375 			err = -EINVAL;
376 
377 		if (map->flags & GNTMAP_device_map)
378 			map->unmap_ops[i].dev_bus_addr = map->map_ops[i].dev_bus_addr;
379 
380 		if (use_ptemod) {
381 			if (map->kmap_ops[i].status == GNTST_okay) {
382 				alloced++;
383 				map->kunmap_ops[i].handle = map->kmap_ops[i].handle;
384 			} else if (!err)
385 				err = -EINVAL;
386 		}
387 	}
388 	atomic_add(alloced, &map->live_grants);
389 	return err;
390 }
391 
392 static void __unmap_grant_pages_done(int result,
393 		struct gntab_unmap_queue_data *data)
394 {
395 	unsigned int i;
396 	struct gntdev_grant_map *map = data->data;
397 	unsigned int offset = data->unmap_ops - map->unmap_ops;
398 	int successful_unmaps = 0;
399 	int live_grants;
400 
401 	for (i = 0; i < data->count; i++) {
402 		if (map->unmap_ops[offset + i].status == GNTST_okay &&
403 		    map->unmap_ops[offset + i].handle != INVALID_GRANT_HANDLE)
404 			successful_unmaps++;
405 
406 		WARN_ON(map->unmap_ops[offset + i].status != GNTST_okay &&
407 			map->unmap_ops[offset + i].handle != INVALID_GRANT_HANDLE);
408 		pr_debug("unmap handle=%d st=%d\n",
409 			map->unmap_ops[offset+i].handle,
410 			map->unmap_ops[offset+i].status);
411 		map->unmap_ops[offset+i].handle = INVALID_GRANT_HANDLE;
412 		if (use_ptemod) {
413 			if (map->kunmap_ops[offset + i].status == GNTST_okay &&
414 			    map->kunmap_ops[offset + i].handle != INVALID_GRANT_HANDLE)
415 				successful_unmaps++;
416 
417 			WARN_ON(map->kunmap_ops[offset + i].status != GNTST_okay &&
418 				map->kunmap_ops[offset + i].handle != INVALID_GRANT_HANDLE);
419 			pr_debug("kunmap handle=%u st=%d\n",
420 				 map->kunmap_ops[offset+i].handle,
421 				 map->kunmap_ops[offset+i].status);
422 			map->kunmap_ops[offset+i].handle = INVALID_GRANT_HANDLE;
423 		}
424 	}
425 
426 	/*
427 	 * Decrease the live-grant counter.  This must happen after the loop to
428 	 * prevent premature reuse of the grants by gnttab_mmap().
429 	 */
430 	live_grants = atomic_sub_return(successful_unmaps, &map->live_grants);
431 	if (WARN_ON(live_grants < 0))
432 		pr_err("%s: live_grants became negative (%d) after unmapping %d pages!\n",
433 		       __func__, live_grants, successful_unmaps);
434 
435 	/* Release reference taken by __unmap_grant_pages */
436 	gntdev_put_map(NULL, map);
437 }
438 
439 static void __unmap_grant_pages(struct gntdev_grant_map *map, int offset,
440 			       int pages)
441 {
442 	if (map->notify.flags & UNMAP_NOTIFY_CLEAR_BYTE) {
443 		int pgno = (map->notify.addr >> PAGE_SHIFT);
444 
445 		if (pgno >= offset && pgno < offset + pages) {
446 			/* No need for kmap, pages are in lowmem */
447 			uint8_t *tmp = pfn_to_kaddr(page_to_pfn(map->pages[pgno]));
448 
449 			tmp[map->notify.addr & (PAGE_SIZE-1)] = 0;
450 			map->notify.flags &= ~UNMAP_NOTIFY_CLEAR_BYTE;
451 		}
452 	}
453 
454 	map->unmap_data.unmap_ops = map->unmap_ops + offset;
455 	map->unmap_data.kunmap_ops = use_ptemod ? map->kunmap_ops + offset : NULL;
456 	map->unmap_data.pages = map->pages + offset;
457 	map->unmap_data.count = pages;
458 	map->unmap_data.done = __unmap_grant_pages_done;
459 	map->unmap_data.data = map;
460 	refcount_inc(&map->users); /* to keep map alive during async call below */
461 
462 	gnttab_unmap_refs_async(&map->unmap_data);
463 }
464 
465 static void unmap_grant_pages(struct gntdev_grant_map *map, int offset,
466 			      int pages)
467 {
468 	int range;
469 
470 	if (atomic_read(&map->live_grants) == 0)
471 		return; /* Nothing to do */
472 
473 	pr_debug("unmap %d+%d [%d+%d]\n", map->index, map->count, offset, pages);
474 
475 	/* It is possible the requested range will have a "hole" where we
476 	 * already unmapped some of the grants. Only unmap valid ranges.
477 	 */
478 	while (pages) {
479 		while (pages && map->being_removed[offset]) {
480 			offset++;
481 			pages--;
482 		}
483 		range = 0;
484 		while (range < pages) {
485 			if (map->being_removed[offset + range])
486 				break;
487 			map->being_removed[offset + range] = true;
488 			range++;
489 		}
490 		if (range)
491 			__unmap_grant_pages(map, offset, range);
492 		offset += range;
493 		pages -= range;
494 	}
495 }
496 
497 /* ------------------------------------------------------------------ */
498 
499 static void gntdev_vma_open(struct vm_area_struct *vma)
500 {
501 	struct gntdev_grant_map *map = vma->vm_private_data;
502 
503 	pr_debug("gntdev_vma_open %p\n", vma);
504 	refcount_inc(&map->users);
505 }
506 
507 static void gntdev_vma_close(struct vm_area_struct *vma)
508 {
509 	struct gntdev_grant_map *map = vma->vm_private_data;
510 	struct file *file = vma->vm_file;
511 	struct gntdev_priv *priv = file->private_data;
512 
513 	pr_debug("gntdev_vma_close %p\n", vma);
514 
515 	vma->vm_private_data = NULL;
516 	gntdev_put_map(priv, map);
517 }
518 
519 static struct page *gntdev_vma_find_special_page(struct vm_area_struct *vma,
520 						 unsigned long addr)
521 {
522 	struct gntdev_grant_map *map = vma->vm_private_data;
523 
524 	return map->pages[(addr - map->pages_vm_start) >> PAGE_SHIFT];
525 }
526 
527 static const struct vm_operations_struct gntdev_vmops = {
528 	.open = gntdev_vma_open,
529 	.close = gntdev_vma_close,
530 	.find_special_page = gntdev_vma_find_special_page,
531 };
532 
533 /* ------------------------------------------------------------------ */
534 
535 static bool gntdev_invalidate(struct mmu_interval_notifier *mn,
536 			      const struct mmu_notifier_range *range,
537 			      unsigned long cur_seq)
538 {
539 	struct gntdev_grant_map *map =
540 		container_of(mn, struct gntdev_grant_map, notifier);
541 	unsigned long mstart, mend;
542 	unsigned long map_start, map_end;
543 
544 	if (!mmu_notifier_range_blockable(range))
545 		return false;
546 
547 	map_start = map->pages_vm_start;
548 	map_end = map->pages_vm_start + (map->count << PAGE_SHIFT);
549 
550 	/*
551 	 * If the VMA is split or otherwise changed the notifier is not
552 	 * updated, but we don't want to process VA's outside the modified
553 	 * VMA. FIXME: It would be much more understandable to just prevent
554 	 * modifying the VMA in the first place.
555 	 */
556 	if (map_start >= range->end || map_end <= range->start)
557 		return true;
558 
559 	mstart = max(range->start, map_start);
560 	mend = min(range->end, map_end);
561 	pr_debug("map %d+%d (%lx %lx), range %lx %lx, mrange %lx %lx\n",
562 		 map->index, map->count, map_start, map_end,
563 		 range->start, range->end, mstart, mend);
564 	unmap_grant_pages(map, (mstart - map_start) >> PAGE_SHIFT,
565 			  (mend - mstart) >> PAGE_SHIFT);
566 
567 	return true;
568 }
569 
570 static const struct mmu_interval_notifier_ops gntdev_mmu_ops = {
571 	.invalidate = gntdev_invalidate,
572 };
573 
574 /* ------------------------------------------------------------------ */
575 
576 static int gntdev_open(struct inode *inode, struct file *flip)
577 {
578 	struct gntdev_priv *priv;
579 
580 	priv = kzalloc(sizeof(*priv), GFP_KERNEL);
581 	if (!priv)
582 		return -ENOMEM;
583 
584 	INIT_LIST_HEAD(&priv->maps);
585 	mutex_init(&priv->lock);
586 
587 #ifdef CONFIG_XEN_GNTDEV_DMABUF
588 	priv->dmabuf_priv = gntdev_dmabuf_init(flip);
589 	if (IS_ERR(priv->dmabuf_priv)) {
590 		int ret = PTR_ERR(priv->dmabuf_priv);
591 
592 		kfree(priv);
593 		return ret;
594 	}
595 #endif
596 
597 	flip->private_data = priv;
598 #ifdef CONFIG_XEN_GRANT_DMA_ALLOC
599 	priv->dma_dev = gntdev_miscdev.this_device;
600 	dma_coerce_mask_and_coherent(priv->dma_dev, DMA_BIT_MASK(64));
601 #endif
602 	pr_debug("priv %p\n", priv);
603 
604 	return 0;
605 }
606 
607 static int gntdev_release(struct inode *inode, struct file *flip)
608 {
609 	struct gntdev_priv *priv = flip->private_data;
610 	struct gntdev_grant_map *map;
611 
612 	pr_debug("priv %p\n", priv);
613 
614 	mutex_lock(&priv->lock);
615 	while (!list_empty(&priv->maps)) {
616 		map = list_entry(priv->maps.next,
617 				 struct gntdev_grant_map, next);
618 		list_del(&map->next);
619 		gntdev_put_map(NULL /* already removed */, map);
620 	}
621 	mutex_unlock(&priv->lock);
622 
623 #ifdef CONFIG_XEN_GNTDEV_DMABUF
624 	gntdev_dmabuf_fini(priv->dmabuf_priv);
625 #endif
626 
627 	kfree(priv);
628 	return 0;
629 }
630 
631 static long gntdev_ioctl_map_grant_ref(struct gntdev_priv *priv,
632 				       struct ioctl_gntdev_map_grant_ref __user *u)
633 {
634 	struct ioctl_gntdev_map_grant_ref op;
635 	struct gntdev_grant_map *map;
636 	int err;
637 
638 	if (copy_from_user(&op, u, sizeof(op)) != 0)
639 		return -EFAULT;
640 	pr_debug("priv %p, add %d\n", priv, op.count);
641 	if (unlikely(gntdev_test_page_count(op.count)))
642 		return -EINVAL;
643 
644 	err = -ENOMEM;
645 	map = gntdev_alloc_map(priv, op.count, 0 /* This is not a dma-buf. */);
646 	if (!map)
647 		return err;
648 
649 	if (copy_from_user(map->grants, &u->refs,
650 			   sizeof(map->grants[0]) * op.count) != 0) {
651 		gntdev_put_map(NULL, map);
652 		return -EFAULT;
653 	}
654 
655 	mutex_lock(&priv->lock);
656 	gntdev_add_map(priv, map);
657 	op.index = map->index << PAGE_SHIFT;
658 	mutex_unlock(&priv->lock);
659 
660 	if (copy_to_user(u, &op, sizeof(op)) != 0)
661 		return -EFAULT;
662 
663 	return 0;
664 }
665 
666 static long gntdev_ioctl_unmap_grant_ref(struct gntdev_priv *priv,
667 					 struct ioctl_gntdev_unmap_grant_ref __user *u)
668 {
669 	struct ioctl_gntdev_unmap_grant_ref op;
670 	struct gntdev_grant_map *map;
671 	int err = -ENOENT;
672 
673 	if (copy_from_user(&op, u, sizeof(op)) != 0)
674 		return -EFAULT;
675 	pr_debug("priv %p, del %d+%d\n", priv, (int)op.index, (int)op.count);
676 
677 	mutex_lock(&priv->lock);
678 	map = gntdev_find_map_index(priv, op.index >> PAGE_SHIFT, op.count);
679 	if (map) {
680 		list_del(&map->next);
681 		err = 0;
682 	}
683 	mutex_unlock(&priv->lock);
684 	if (map)
685 		gntdev_put_map(priv, map);
686 	return err;
687 }
688 
689 static long gntdev_ioctl_get_offset_for_vaddr(struct gntdev_priv *priv,
690 					      struct ioctl_gntdev_get_offset_for_vaddr __user *u)
691 {
692 	struct ioctl_gntdev_get_offset_for_vaddr op;
693 	struct vm_area_struct *vma;
694 	struct gntdev_grant_map *map;
695 	int rv = -EINVAL;
696 
697 	if (copy_from_user(&op, u, sizeof(op)) != 0)
698 		return -EFAULT;
699 	pr_debug("priv %p, offset for vaddr %lx\n", priv, (unsigned long)op.vaddr);
700 
701 	mmap_read_lock(current->mm);
702 	vma = find_vma(current->mm, op.vaddr);
703 	if (!vma || vma->vm_ops != &gntdev_vmops)
704 		goto out_unlock;
705 
706 	map = vma->vm_private_data;
707 	if (!map)
708 		goto out_unlock;
709 
710 	op.offset = map->index << PAGE_SHIFT;
711 	op.count = map->count;
712 	rv = 0;
713 
714  out_unlock:
715 	mmap_read_unlock(current->mm);
716 
717 	if (rv == 0 && copy_to_user(u, &op, sizeof(op)) != 0)
718 		return -EFAULT;
719 	return rv;
720 }
721 
722 static long gntdev_ioctl_notify(struct gntdev_priv *priv, void __user *u)
723 {
724 	struct ioctl_gntdev_unmap_notify op;
725 	struct gntdev_grant_map *map;
726 	int rc;
727 	int out_flags;
728 	evtchn_port_t out_event;
729 
730 	if (copy_from_user(&op, u, sizeof(op)))
731 		return -EFAULT;
732 
733 	if (op.action & ~(UNMAP_NOTIFY_CLEAR_BYTE|UNMAP_NOTIFY_SEND_EVENT))
734 		return -EINVAL;
735 
736 	/* We need to grab a reference to the event channel we are going to use
737 	 * to send the notify before releasing the reference we may already have
738 	 * (if someone has called this ioctl twice). This is required so that
739 	 * it is possible to change the clear_byte part of the notification
740 	 * without disturbing the event channel part, which may now be the last
741 	 * reference to that event channel.
742 	 */
743 	if (op.action & UNMAP_NOTIFY_SEND_EVENT) {
744 		if (evtchn_get(op.event_channel_port))
745 			return -EINVAL;
746 	}
747 
748 	out_flags = op.action;
749 	out_event = op.event_channel_port;
750 
751 	mutex_lock(&priv->lock);
752 
753 	list_for_each_entry(map, &priv->maps, next) {
754 		uint64_t begin = map->index << PAGE_SHIFT;
755 		uint64_t end = (map->index + map->count) << PAGE_SHIFT;
756 		if (op.index >= begin && op.index < end)
757 			goto found;
758 	}
759 	rc = -ENOENT;
760 	goto unlock_out;
761 
762  found:
763 	if ((op.action & UNMAP_NOTIFY_CLEAR_BYTE) &&
764 			(map->flags & GNTMAP_readonly)) {
765 		rc = -EINVAL;
766 		goto unlock_out;
767 	}
768 
769 	out_flags = map->notify.flags;
770 	out_event = map->notify.event;
771 
772 	map->notify.flags = op.action;
773 	map->notify.addr = op.index - (map->index << PAGE_SHIFT);
774 	map->notify.event = op.event_channel_port;
775 
776 	rc = 0;
777 
778  unlock_out:
779 	mutex_unlock(&priv->lock);
780 
781 	/* Drop the reference to the event channel we did not save in the map */
782 	if (out_flags & UNMAP_NOTIFY_SEND_EVENT)
783 		evtchn_put(out_event);
784 
785 	return rc;
786 }
787 
788 #define GNTDEV_COPY_BATCH 16
789 
790 struct gntdev_copy_batch {
791 	struct gnttab_copy ops[GNTDEV_COPY_BATCH];
792 	struct page *pages[GNTDEV_COPY_BATCH];
793 	s16 __user *status[GNTDEV_COPY_BATCH];
794 	unsigned int nr_ops;
795 	unsigned int nr_pages;
796 	bool writeable;
797 };
798 
799 static int gntdev_get_page(struct gntdev_copy_batch *batch, void __user *virt,
800 				unsigned long *gfn)
801 {
802 	unsigned long addr = (unsigned long)virt;
803 	struct page *page;
804 	unsigned long xen_pfn;
805 	int ret;
806 
807 	ret = pin_user_pages_fast(addr, 1, batch->writeable ? FOLL_WRITE : 0, &page);
808 	if (ret < 0)
809 		return ret;
810 
811 	batch->pages[batch->nr_pages++] = page;
812 
813 	xen_pfn = page_to_xen_pfn(page) + XEN_PFN_DOWN(addr & ~PAGE_MASK);
814 	*gfn = pfn_to_gfn(xen_pfn);
815 
816 	return 0;
817 }
818 
819 static void gntdev_put_pages(struct gntdev_copy_batch *batch)
820 {
821 	unpin_user_pages_dirty_lock(batch->pages, batch->nr_pages, batch->writeable);
822 	batch->nr_pages = 0;
823 	batch->writeable = false;
824 }
825 
826 static int gntdev_copy(struct gntdev_copy_batch *batch)
827 {
828 	unsigned int i;
829 
830 	gnttab_batch_copy(batch->ops, batch->nr_ops);
831 	gntdev_put_pages(batch);
832 
833 	/*
834 	 * For each completed op, update the status if the op failed
835 	 * and all previous ops for the segment were successful.
836 	 */
837 	for (i = 0; i < batch->nr_ops; i++) {
838 		s16 status = batch->ops[i].status;
839 		s16 old_status;
840 
841 		if (status == GNTST_okay)
842 			continue;
843 
844 		if (__get_user(old_status, batch->status[i]))
845 			return -EFAULT;
846 
847 		if (old_status != GNTST_okay)
848 			continue;
849 
850 		if (__put_user(status, batch->status[i]))
851 			return -EFAULT;
852 	}
853 
854 	batch->nr_ops = 0;
855 	return 0;
856 }
857 
858 static int gntdev_grant_copy_seg(struct gntdev_copy_batch *batch,
859 				 struct gntdev_grant_copy_segment *seg,
860 				 s16 __user *status)
861 {
862 	uint16_t copied = 0;
863 
864 	/*
865 	 * Disallow local -> local copies since there is only space in
866 	 * batch->pages for one page per-op and this would be a very
867 	 * expensive memcpy().
868 	 */
869 	if (!(seg->flags & (GNTCOPY_source_gref | GNTCOPY_dest_gref)))
870 		return -EINVAL;
871 
872 	/* Can't cross page if source/dest is a grant ref. */
873 	if (seg->flags & GNTCOPY_source_gref) {
874 		if (seg->source.foreign.offset + seg->len > XEN_PAGE_SIZE)
875 			return -EINVAL;
876 	}
877 	if (seg->flags & GNTCOPY_dest_gref) {
878 		if (seg->dest.foreign.offset + seg->len > XEN_PAGE_SIZE)
879 			return -EINVAL;
880 	}
881 
882 	if (put_user(GNTST_okay, status))
883 		return -EFAULT;
884 
885 	while (copied < seg->len) {
886 		struct gnttab_copy *op;
887 		void __user *virt;
888 		size_t len, off;
889 		unsigned long gfn;
890 		int ret;
891 
892 		if (batch->nr_ops >= GNTDEV_COPY_BATCH) {
893 			ret = gntdev_copy(batch);
894 			if (ret < 0)
895 				return ret;
896 		}
897 
898 		len = seg->len - copied;
899 
900 		op = &batch->ops[batch->nr_ops];
901 		op->flags = 0;
902 
903 		if (seg->flags & GNTCOPY_source_gref) {
904 			op->source.u.ref = seg->source.foreign.ref;
905 			op->source.domid = seg->source.foreign.domid;
906 			op->source.offset = seg->source.foreign.offset + copied;
907 			op->flags |= GNTCOPY_source_gref;
908 		} else {
909 			virt = seg->source.virt + copied;
910 			off = (unsigned long)virt & ~XEN_PAGE_MASK;
911 			len = min(len, (size_t)XEN_PAGE_SIZE - off);
912 			batch->writeable = false;
913 
914 			ret = gntdev_get_page(batch, virt, &gfn);
915 			if (ret < 0)
916 				return ret;
917 
918 			op->source.u.gmfn = gfn;
919 			op->source.domid = DOMID_SELF;
920 			op->source.offset = off;
921 		}
922 
923 		if (seg->flags & GNTCOPY_dest_gref) {
924 			op->dest.u.ref = seg->dest.foreign.ref;
925 			op->dest.domid = seg->dest.foreign.domid;
926 			op->dest.offset = seg->dest.foreign.offset + copied;
927 			op->flags |= GNTCOPY_dest_gref;
928 		} else {
929 			virt = seg->dest.virt + copied;
930 			off = (unsigned long)virt & ~XEN_PAGE_MASK;
931 			len = min(len, (size_t)XEN_PAGE_SIZE - off);
932 			batch->writeable = true;
933 
934 			ret = gntdev_get_page(batch, virt, &gfn);
935 			if (ret < 0)
936 				return ret;
937 
938 			op->dest.u.gmfn = gfn;
939 			op->dest.domid = DOMID_SELF;
940 			op->dest.offset = off;
941 		}
942 
943 		op->len = len;
944 		copied += len;
945 
946 		batch->status[batch->nr_ops] = status;
947 		batch->nr_ops++;
948 	}
949 
950 	return 0;
951 }
952 
953 static long gntdev_ioctl_grant_copy(struct gntdev_priv *priv, void __user *u)
954 {
955 	struct ioctl_gntdev_grant_copy copy;
956 	struct gntdev_copy_batch batch;
957 	unsigned int i;
958 	int ret = 0;
959 
960 	if (copy_from_user(&copy, u, sizeof(copy)))
961 		return -EFAULT;
962 
963 	batch.nr_ops = 0;
964 	batch.nr_pages = 0;
965 
966 	for (i = 0; i < copy.count; i++) {
967 		struct gntdev_grant_copy_segment seg;
968 
969 		if (copy_from_user(&seg, &copy.segments[i], sizeof(seg))) {
970 			ret = -EFAULT;
971 			goto out;
972 		}
973 
974 		ret = gntdev_grant_copy_seg(&batch, &seg, &copy.segments[i].status);
975 		if (ret < 0)
976 			goto out;
977 
978 		cond_resched();
979 	}
980 	if (batch.nr_ops)
981 		ret = gntdev_copy(&batch);
982 	return ret;
983 
984   out:
985 	gntdev_put_pages(&batch);
986 	return ret;
987 }
988 
989 static long gntdev_ioctl(struct file *flip,
990 			 unsigned int cmd, unsigned long arg)
991 {
992 	struct gntdev_priv *priv = flip->private_data;
993 	void __user *ptr = (void __user *)arg;
994 
995 	switch (cmd) {
996 	case IOCTL_GNTDEV_MAP_GRANT_REF:
997 		return gntdev_ioctl_map_grant_ref(priv, ptr);
998 
999 	case IOCTL_GNTDEV_UNMAP_GRANT_REF:
1000 		return gntdev_ioctl_unmap_grant_ref(priv, ptr);
1001 
1002 	case IOCTL_GNTDEV_GET_OFFSET_FOR_VADDR:
1003 		return gntdev_ioctl_get_offset_for_vaddr(priv, ptr);
1004 
1005 	case IOCTL_GNTDEV_SET_UNMAP_NOTIFY:
1006 		return gntdev_ioctl_notify(priv, ptr);
1007 
1008 	case IOCTL_GNTDEV_GRANT_COPY:
1009 		return gntdev_ioctl_grant_copy(priv, ptr);
1010 
1011 #ifdef CONFIG_XEN_GNTDEV_DMABUF
1012 	case IOCTL_GNTDEV_DMABUF_EXP_FROM_REFS:
1013 		return gntdev_ioctl_dmabuf_exp_from_refs(priv, use_ptemod, ptr);
1014 
1015 	case IOCTL_GNTDEV_DMABUF_EXP_WAIT_RELEASED:
1016 		return gntdev_ioctl_dmabuf_exp_wait_released(priv, ptr);
1017 
1018 	case IOCTL_GNTDEV_DMABUF_IMP_TO_REFS:
1019 		return gntdev_ioctl_dmabuf_imp_to_refs(priv, ptr);
1020 
1021 	case IOCTL_GNTDEV_DMABUF_IMP_RELEASE:
1022 		return gntdev_ioctl_dmabuf_imp_release(priv, ptr);
1023 #endif
1024 
1025 	default:
1026 		pr_debug("priv %p, unknown cmd %x\n", priv, cmd);
1027 		return -ENOIOCTLCMD;
1028 	}
1029 
1030 	return 0;
1031 }
1032 
1033 static int gntdev_mmap(struct file *flip, struct vm_area_struct *vma)
1034 {
1035 	struct gntdev_priv *priv = flip->private_data;
1036 	int index = vma->vm_pgoff;
1037 	int count = vma_pages(vma);
1038 	struct gntdev_grant_map *map;
1039 	int err = -EINVAL;
1040 
1041 	if ((vma->vm_flags & VM_WRITE) && !(vma->vm_flags & VM_SHARED))
1042 		return -EINVAL;
1043 
1044 	pr_debug("map %d+%d at %lx (pgoff %lx)\n",
1045 		 index, count, vma->vm_start, vma->vm_pgoff);
1046 
1047 	mutex_lock(&priv->lock);
1048 	map = gntdev_find_map_index(priv, index, count);
1049 	if (!map)
1050 		goto unlock_out;
1051 	if (!atomic_add_unless(&map->in_use, 1, 1))
1052 		goto unlock_out;
1053 
1054 	refcount_inc(&map->users);
1055 
1056 	vma->vm_ops = &gntdev_vmops;
1057 
1058 	vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP | VM_MIXEDMAP;
1059 
1060 	if (use_ptemod)
1061 		vma->vm_flags |= VM_DONTCOPY;
1062 
1063 	vma->vm_private_data = map;
1064 	if (map->flags) {
1065 		if ((vma->vm_flags & VM_WRITE) &&
1066 				(map->flags & GNTMAP_readonly))
1067 			goto out_unlock_put;
1068 	} else {
1069 		map->flags = GNTMAP_host_map;
1070 		if (!(vma->vm_flags & VM_WRITE))
1071 			map->flags |= GNTMAP_readonly;
1072 	}
1073 
1074 	map->pages_vm_start = vma->vm_start;
1075 
1076 	if (use_ptemod) {
1077 		err = mmu_interval_notifier_insert_locked(
1078 			&map->notifier, vma->vm_mm, vma->vm_start,
1079 			vma->vm_end - vma->vm_start, &gntdev_mmu_ops);
1080 		if (err)
1081 			goto out_unlock_put;
1082 
1083 		map->notifier_init = true;
1084 	}
1085 	mutex_unlock(&priv->lock);
1086 
1087 	if (use_ptemod) {
1088 		/*
1089 		 * gntdev takes the address of the PTE in find_grant_ptes() and
1090 		 * passes it to the hypervisor in gntdev_map_grant_pages(). The
1091 		 * purpose of the notifier is to prevent the hypervisor pointer
1092 		 * to the PTE from going stale.
1093 		 *
1094 		 * Since this vma's mappings can't be touched without the
1095 		 * mmap_lock, and we are holding it now, there is no need for
1096 		 * the notifier_range locking pattern.
1097 		 */
1098 		mmu_interval_read_begin(&map->notifier);
1099 
1100 		err = apply_to_page_range(vma->vm_mm, vma->vm_start,
1101 					  vma->vm_end - vma->vm_start,
1102 					  find_grant_ptes, map);
1103 		if (err) {
1104 			pr_warn("find_grant_ptes() failure.\n");
1105 			goto out_put_map;
1106 		}
1107 	}
1108 
1109 	err = gntdev_map_grant_pages(map);
1110 	if (err)
1111 		goto out_put_map;
1112 
1113 	if (!use_ptemod) {
1114 		err = vm_map_pages_zero(vma, map->pages, map->count);
1115 		if (err)
1116 			goto out_put_map;
1117 	}
1118 
1119 	return 0;
1120 
1121 unlock_out:
1122 	mutex_unlock(&priv->lock);
1123 	return err;
1124 
1125 out_unlock_put:
1126 	mutex_unlock(&priv->lock);
1127 out_put_map:
1128 	if (use_ptemod)
1129 		unmap_grant_pages(map, 0, map->count);
1130 	gntdev_put_map(priv, map);
1131 	return err;
1132 }
1133 
1134 static const struct file_operations gntdev_fops = {
1135 	.owner = THIS_MODULE,
1136 	.open = gntdev_open,
1137 	.release = gntdev_release,
1138 	.mmap = gntdev_mmap,
1139 	.unlocked_ioctl = gntdev_ioctl
1140 };
1141 
1142 static struct miscdevice gntdev_miscdev = {
1143 	.minor        = MISC_DYNAMIC_MINOR,
1144 	.name         = "xen/gntdev",
1145 	.fops         = &gntdev_fops,
1146 };
1147 
1148 /* ------------------------------------------------------------------ */
1149 
1150 static int __init gntdev_init(void)
1151 {
1152 	int err;
1153 
1154 	if (!xen_domain())
1155 		return -ENODEV;
1156 
1157 	use_ptemod = !xen_feature(XENFEAT_auto_translated_physmap);
1158 
1159 	err = misc_register(&gntdev_miscdev);
1160 	if (err != 0) {
1161 		pr_err("Could not register gntdev device\n");
1162 		return err;
1163 	}
1164 	return 0;
1165 }
1166 
1167 static void __exit gntdev_exit(void)
1168 {
1169 	misc_deregister(&gntdev_miscdev);
1170 }
1171 
1172 module_init(gntdev_init);
1173 module_exit(gntdev_exit);
1174 
1175 /* ------------------------------------------------------------------ */
1176