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