xref: /openbmc/linux/drivers/xen/gntdev.c (revision 0ad53fe3)
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 
72 bool gntdev_test_page_count(unsigned int count)
73 {
74 	return !count || count > limit;
75 }
76 
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 
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 
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 
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 
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 
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 
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 
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 
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 	}
385 	return err;
386 }
387 
388 static int unmap_grant_pages(struct gntdev_grant_map *map, int offset,
389 			     int pages)
390 {
391 	int range, err = 0;
392 
393 	pr_debug("unmap %d+%d [%d+%d]\n", map->index, map->count, offset, pages);
394 
395 	/* It is possible the requested range will have a "hole" where we
396 	 * already unmapped some of the grants. Only unmap valid ranges.
397 	 */
398 	while (pages && !err) {
399 		while (pages &&
400 		       map->unmap_ops[offset].handle == INVALID_GRANT_HANDLE) {
401 			offset++;
402 			pages--;
403 		}
404 		range = 0;
405 		while (range < pages) {
406 			if (map->unmap_ops[offset + range].handle ==
407 			    INVALID_GRANT_HANDLE)
408 				break;
409 			range++;
410 		}
411 		err = __unmap_grant_pages(map, offset, range);
412 		offset += range;
413 		pages -= range;
414 	}
415 
416 	return err;
417 }
418 
419 /* ------------------------------------------------------------------ */
420 
421 static void gntdev_vma_open(struct vm_area_struct *vma)
422 {
423 	struct gntdev_grant_map *map = vma->vm_private_data;
424 
425 	pr_debug("gntdev_vma_open %p\n", vma);
426 	refcount_inc(&map->users);
427 }
428 
429 static void gntdev_vma_close(struct vm_area_struct *vma)
430 {
431 	struct gntdev_grant_map *map = vma->vm_private_data;
432 	struct file *file = vma->vm_file;
433 	struct gntdev_priv *priv = file->private_data;
434 
435 	pr_debug("gntdev_vma_close %p\n", vma);
436 	if (use_ptemod) {
437 		WARN_ON(map->vma != vma);
438 		mmu_interval_notifier_remove(&map->notifier);
439 		map->vma = NULL;
440 	}
441 	vma->vm_private_data = NULL;
442 	gntdev_put_map(priv, map);
443 }
444 
445 static struct page *gntdev_vma_find_special_page(struct vm_area_struct *vma,
446 						 unsigned long addr)
447 {
448 	struct gntdev_grant_map *map = vma->vm_private_data;
449 
450 	return map->pages[(addr - map->pages_vm_start) >> PAGE_SHIFT];
451 }
452 
453 static const struct vm_operations_struct gntdev_vmops = {
454 	.open = gntdev_vma_open,
455 	.close = gntdev_vma_close,
456 	.find_special_page = gntdev_vma_find_special_page,
457 };
458 
459 /* ------------------------------------------------------------------ */
460 
461 static bool gntdev_invalidate(struct mmu_interval_notifier *mn,
462 			      const struct mmu_notifier_range *range,
463 			      unsigned long cur_seq)
464 {
465 	struct gntdev_grant_map *map =
466 		container_of(mn, struct gntdev_grant_map, notifier);
467 	unsigned long mstart, mend;
468 	int err;
469 
470 	if (!mmu_notifier_range_blockable(range))
471 		return false;
472 
473 	/*
474 	 * If the VMA is split or otherwise changed the notifier is not
475 	 * updated, but we don't want to process VA's outside the modified
476 	 * VMA. FIXME: It would be much more understandable to just prevent
477 	 * modifying the VMA in the first place.
478 	 */
479 	if (map->vma->vm_start >= range->end ||
480 	    map->vma->vm_end <= range->start)
481 		return true;
482 
483 	mstart = max(range->start, map->vma->vm_start);
484 	mend = min(range->end, map->vma->vm_end);
485 	pr_debug("map %d+%d (%lx %lx), range %lx %lx, mrange %lx %lx\n",
486 			map->index, map->count,
487 			map->vma->vm_start, map->vma->vm_end,
488 			range->start, range->end, mstart, mend);
489 	err = unmap_grant_pages(map,
490 				(mstart - map->vma->vm_start) >> PAGE_SHIFT,
491 				(mend - mstart) >> PAGE_SHIFT);
492 	WARN_ON(err);
493 
494 	return true;
495 }
496 
497 static const struct mmu_interval_notifier_ops gntdev_mmu_ops = {
498 	.invalidate = gntdev_invalidate,
499 };
500 
501 /* ------------------------------------------------------------------ */
502 
503 static int gntdev_open(struct inode *inode, struct file *flip)
504 {
505 	struct gntdev_priv *priv;
506 
507 	priv = kzalloc(sizeof(*priv), GFP_KERNEL);
508 	if (!priv)
509 		return -ENOMEM;
510 
511 	INIT_LIST_HEAD(&priv->maps);
512 	mutex_init(&priv->lock);
513 
514 #ifdef CONFIG_XEN_GNTDEV_DMABUF
515 	priv->dmabuf_priv = gntdev_dmabuf_init(flip);
516 	if (IS_ERR(priv->dmabuf_priv)) {
517 		int ret = PTR_ERR(priv->dmabuf_priv);
518 
519 		kfree(priv);
520 		return ret;
521 	}
522 #endif
523 
524 	flip->private_data = priv;
525 #ifdef CONFIG_XEN_GRANT_DMA_ALLOC
526 	priv->dma_dev = gntdev_miscdev.this_device;
527 	dma_coerce_mask_and_coherent(priv->dma_dev, DMA_BIT_MASK(64));
528 #endif
529 	pr_debug("priv %p\n", priv);
530 
531 	return 0;
532 }
533 
534 static int gntdev_release(struct inode *inode, struct file *flip)
535 {
536 	struct gntdev_priv *priv = flip->private_data;
537 	struct gntdev_grant_map *map;
538 
539 	pr_debug("priv %p\n", priv);
540 
541 	mutex_lock(&priv->lock);
542 	while (!list_empty(&priv->maps)) {
543 		map = list_entry(priv->maps.next,
544 				 struct gntdev_grant_map, next);
545 		list_del(&map->next);
546 		gntdev_put_map(NULL /* already removed */, map);
547 	}
548 	mutex_unlock(&priv->lock);
549 
550 #ifdef CONFIG_XEN_GNTDEV_DMABUF
551 	gntdev_dmabuf_fini(priv->dmabuf_priv);
552 #endif
553 
554 	kfree(priv);
555 	return 0;
556 }
557 
558 static long gntdev_ioctl_map_grant_ref(struct gntdev_priv *priv,
559 				       struct ioctl_gntdev_map_grant_ref __user *u)
560 {
561 	struct ioctl_gntdev_map_grant_ref op;
562 	struct gntdev_grant_map *map;
563 	int err;
564 
565 	if (copy_from_user(&op, u, sizeof(op)) != 0)
566 		return -EFAULT;
567 	pr_debug("priv %p, add %d\n", priv, op.count);
568 	if (unlikely(gntdev_test_page_count(op.count)))
569 		return -EINVAL;
570 
571 	err = -ENOMEM;
572 	map = gntdev_alloc_map(priv, op.count, 0 /* This is not a dma-buf. */);
573 	if (!map)
574 		return err;
575 
576 	if (copy_from_user(map->grants, &u->refs,
577 			   sizeof(map->grants[0]) * op.count) != 0) {
578 		gntdev_put_map(NULL, map);
579 		return -EFAULT;
580 	}
581 
582 	mutex_lock(&priv->lock);
583 	gntdev_add_map(priv, map);
584 	op.index = map->index << PAGE_SHIFT;
585 	mutex_unlock(&priv->lock);
586 
587 	if (copy_to_user(u, &op, sizeof(op)) != 0)
588 		return -EFAULT;
589 
590 	return 0;
591 }
592 
593 static long gntdev_ioctl_unmap_grant_ref(struct gntdev_priv *priv,
594 					 struct ioctl_gntdev_unmap_grant_ref __user *u)
595 {
596 	struct ioctl_gntdev_unmap_grant_ref op;
597 	struct gntdev_grant_map *map;
598 	int err = -ENOENT;
599 
600 	if (copy_from_user(&op, u, sizeof(op)) != 0)
601 		return -EFAULT;
602 	pr_debug("priv %p, del %d+%d\n", priv, (int)op.index, (int)op.count);
603 
604 	mutex_lock(&priv->lock);
605 	map = gntdev_find_map_index(priv, op.index >> PAGE_SHIFT, op.count);
606 	if (map) {
607 		list_del(&map->next);
608 		err = 0;
609 	}
610 	mutex_unlock(&priv->lock);
611 	if (map)
612 		gntdev_put_map(priv, map);
613 	return err;
614 }
615 
616 static long gntdev_ioctl_get_offset_for_vaddr(struct gntdev_priv *priv,
617 					      struct ioctl_gntdev_get_offset_for_vaddr __user *u)
618 {
619 	struct ioctl_gntdev_get_offset_for_vaddr op;
620 	struct vm_area_struct *vma;
621 	struct gntdev_grant_map *map;
622 	int rv = -EINVAL;
623 
624 	if (copy_from_user(&op, u, sizeof(op)) != 0)
625 		return -EFAULT;
626 	pr_debug("priv %p, offset for vaddr %lx\n", priv, (unsigned long)op.vaddr);
627 
628 	mmap_read_lock(current->mm);
629 	vma = find_vma(current->mm, op.vaddr);
630 	if (!vma || vma->vm_ops != &gntdev_vmops)
631 		goto out_unlock;
632 
633 	map = vma->vm_private_data;
634 	if (!map)
635 		goto out_unlock;
636 
637 	op.offset = map->index << PAGE_SHIFT;
638 	op.count = map->count;
639 	rv = 0;
640 
641  out_unlock:
642 	mmap_read_unlock(current->mm);
643 
644 	if (rv == 0 && copy_to_user(u, &op, sizeof(op)) != 0)
645 		return -EFAULT;
646 	return rv;
647 }
648 
649 static long gntdev_ioctl_notify(struct gntdev_priv *priv, void __user *u)
650 {
651 	struct ioctl_gntdev_unmap_notify op;
652 	struct gntdev_grant_map *map;
653 	int rc;
654 	int out_flags;
655 	evtchn_port_t out_event;
656 
657 	if (copy_from_user(&op, u, sizeof(op)))
658 		return -EFAULT;
659 
660 	if (op.action & ~(UNMAP_NOTIFY_CLEAR_BYTE|UNMAP_NOTIFY_SEND_EVENT))
661 		return -EINVAL;
662 
663 	/* We need to grab a reference to the event channel we are going to use
664 	 * to send the notify before releasing the reference we may already have
665 	 * (if someone has called this ioctl twice). This is required so that
666 	 * it is possible to change the clear_byte part of the notification
667 	 * without disturbing the event channel part, which may now be the last
668 	 * reference to that event channel.
669 	 */
670 	if (op.action & UNMAP_NOTIFY_SEND_EVENT) {
671 		if (evtchn_get(op.event_channel_port))
672 			return -EINVAL;
673 	}
674 
675 	out_flags = op.action;
676 	out_event = op.event_channel_port;
677 
678 	mutex_lock(&priv->lock);
679 
680 	list_for_each_entry(map, &priv->maps, next) {
681 		uint64_t begin = map->index << PAGE_SHIFT;
682 		uint64_t end = (map->index + map->count) << PAGE_SHIFT;
683 		if (op.index >= begin && op.index < end)
684 			goto found;
685 	}
686 	rc = -ENOENT;
687 	goto unlock_out;
688 
689  found:
690 	if ((op.action & UNMAP_NOTIFY_CLEAR_BYTE) &&
691 			(map->flags & GNTMAP_readonly)) {
692 		rc = -EINVAL;
693 		goto unlock_out;
694 	}
695 
696 	out_flags = map->notify.flags;
697 	out_event = map->notify.event;
698 
699 	map->notify.flags = op.action;
700 	map->notify.addr = op.index - (map->index << PAGE_SHIFT);
701 	map->notify.event = op.event_channel_port;
702 
703 	rc = 0;
704 
705  unlock_out:
706 	mutex_unlock(&priv->lock);
707 
708 	/* Drop the reference to the event channel we did not save in the map */
709 	if (out_flags & UNMAP_NOTIFY_SEND_EVENT)
710 		evtchn_put(out_event);
711 
712 	return rc;
713 }
714 
715 #define GNTDEV_COPY_BATCH 16
716 
717 struct gntdev_copy_batch {
718 	struct gnttab_copy ops[GNTDEV_COPY_BATCH];
719 	struct page *pages[GNTDEV_COPY_BATCH];
720 	s16 __user *status[GNTDEV_COPY_BATCH];
721 	unsigned int nr_ops;
722 	unsigned int nr_pages;
723 	bool writeable;
724 };
725 
726 static int gntdev_get_page(struct gntdev_copy_batch *batch, void __user *virt,
727 				unsigned long *gfn)
728 {
729 	unsigned long addr = (unsigned long)virt;
730 	struct page *page;
731 	unsigned long xen_pfn;
732 	int ret;
733 
734 	ret = pin_user_pages_fast(addr, 1, batch->writeable ? FOLL_WRITE : 0, &page);
735 	if (ret < 0)
736 		return ret;
737 
738 	batch->pages[batch->nr_pages++] = page;
739 
740 	xen_pfn = page_to_xen_pfn(page) + XEN_PFN_DOWN(addr & ~PAGE_MASK);
741 	*gfn = pfn_to_gfn(xen_pfn);
742 
743 	return 0;
744 }
745 
746 static void gntdev_put_pages(struct gntdev_copy_batch *batch)
747 {
748 	unpin_user_pages_dirty_lock(batch->pages, batch->nr_pages, batch->writeable);
749 	batch->nr_pages = 0;
750 	batch->writeable = false;
751 }
752 
753 static int gntdev_copy(struct gntdev_copy_batch *batch)
754 {
755 	unsigned int i;
756 
757 	gnttab_batch_copy(batch->ops, batch->nr_ops);
758 	gntdev_put_pages(batch);
759 
760 	/*
761 	 * For each completed op, update the status if the op failed
762 	 * and all previous ops for the segment were successful.
763 	 */
764 	for (i = 0; i < batch->nr_ops; i++) {
765 		s16 status = batch->ops[i].status;
766 		s16 old_status;
767 
768 		if (status == GNTST_okay)
769 			continue;
770 
771 		if (__get_user(old_status, batch->status[i]))
772 			return -EFAULT;
773 
774 		if (old_status != GNTST_okay)
775 			continue;
776 
777 		if (__put_user(status, batch->status[i]))
778 			return -EFAULT;
779 	}
780 
781 	batch->nr_ops = 0;
782 	return 0;
783 }
784 
785 static int gntdev_grant_copy_seg(struct gntdev_copy_batch *batch,
786 				 struct gntdev_grant_copy_segment *seg,
787 				 s16 __user *status)
788 {
789 	uint16_t copied = 0;
790 
791 	/*
792 	 * Disallow local -> local copies since there is only space in
793 	 * batch->pages for one page per-op and this would be a very
794 	 * expensive memcpy().
795 	 */
796 	if (!(seg->flags & (GNTCOPY_source_gref | GNTCOPY_dest_gref)))
797 		return -EINVAL;
798 
799 	/* Can't cross page if source/dest is a grant ref. */
800 	if (seg->flags & GNTCOPY_source_gref) {
801 		if (seg->source.foreign.offset + seg->len > XEN_PAGE_SIZE)
802 			return -EINVAL;
803 	}
804 	if (seg->flags & GNTCOPY_dest_gref) {
805 		if (seg->dest.foreign.offset + seg->len > XEN_PAGE_SIZE)
806 			return -EINVAL;
807 	}
808 
809 	if (put_user(GNTST_okay, status))
810 		return -EFAULT;
811 
812 	while (copied < seg->len) {
813 		struct gnttab_copy *op;
814 		void __user *virt;
815 		size_t len, off;
816 		unsigned long gfn;
817 		int ret;
818 
819 		if (batch->nr_ops >= GNTDEV_COPY_BATCH) {
820 			ret = gntdev_copy(batch);
821 			if (ret < 0)
822 				return ret;
823 		}
824 
825 		len = seg->len - copied;
826 
827 		op = &batch->ops[batch->nr_ops];
828 		op->flags = 0;
829 
830 		if (seg->flags & GNTCOPY_source_gref) {
831 			op->source.u.ref = seg->source.foreign.ref;
832 			op->source.domid = seg->source.foreign.domid;
833 			op->source.offset = seg->source.foreign.offset + copied;
834 			op->flags |= GNTCOPY_source_gref;
835 		} else {
836 			virt = seg->source.virt + copied;
837 			off = (unsigned long)virt & ~XEN_PAGE_MASK;
838 			len = min(len, (size_t)XEN_PAGE_SIZE - off);
839 			batch->writeable = false;
840 
841 			ret = gntdev_get_page(batch, virt, &gfn);
842 			if (ret < 0)
843 				return ret;
844 
845 			op->source.u.gmfn = gfn;
846 			op->source.domid = DOMID_SELF;
847 			op->source.offset = off;
848 		}
849 
850 		if (seg->flags & GNTCOPY_dest_gref) {
851 			op->dest.u.ref = seg->dest.foreign.ref;
852 			op->dest.domid = seg->dest.foreign.domid;
853 			op->dest.offset = seg->dest.foreign.offset + copied;
854 			op->flags |= GNTCOPY_dest_gref;
855 		} else {
856 			virt = seg->dest.virt + copied;
857 			off = (unsigned long)virt & ~XEN_PAGE_MASK;
858 			len = min(len, (size_t)XEN_PAGE_SIZE - off);
859 			batch->writeable = true;
860 
861 			ret = gntdev_get_page(batch, virt, &gfn);
862 			if (ret < 0)
863 				return ret;
864 
865 			op->dest.u.gmfn = gfn;
866 			op->dest.domid = DOMID_SELF;
867 			op->dest.offset = off;
868 		}
869 
870 		op->len = len;
871 		copied += len;
872 
873 		batch->status[batch->nr_ops] = status;
874 		batch->nr_ops++;
875 	}
876 
877 	return 0;
878 }
879 
880 static long gntdev_ioctl_grant_copy(struct gntdev_priv *priv, void __user *u)
881 {
882 	struct ioctl_gntdev_grant_copy copy;
883 	struct gntdev_copy_batch batch;
884 	unsigned int i;
885 	int ret = 0;
886 
887 	if (copy_from_user(&copy, u, sizeof(copy)))
888 		return -EFAULT;
889 
890 	batch.nr_ops = 0;
891 	batch.nr_pages = 0;
892 
893 	for (i = 0; i < copy.count; i++) {
894 		struct gntdev_grant_copy_segment seg;
895 
896 		if (copy_from_user(&seg, &copy.segments[i], sizeof(seg))) {
897 			ret = -EFAULT;
898 			goto out;
899 		}
900 
901 		ret = gntdev_grant_copy_seg(&batch, &seg, &copy.segments[i].status);
902 		if (ret < 0)
903 			goto out;
904 
905 		cond_resched();
906 	}
907 	if (batch.nr_ops)
908 		ret = gntdev_copy(&batch);
909 	return ret;
910 
911   out:
912 	gntdev_put_pages(&batch);
913 	return ret;
914 }
915 
916 static long gntdev_ioctl(struct file *flip,
917 			 unsigned int cmd, unsigned long arg)
918 {
919 	struct gntdev_priv *priv = flip->private_data;
920 	void __user *ptr = (void __user *)arg;
921 
922 	switch (cmd) {
923 	case IOCTL_GNTDEV_MAP_GRANT_REF:
924 		return gntdev_ioctl_map_grant_ref(priv, ptr);
925 
926 	case IOCTL_GNTDEV_UNMAP_GRANT_REF:
927 		return gntdev_ioctl_unmap_grant_ref(priv, ptr);
928 
929 	case IOCTL_GNTDEV_GET_OFFSET_FOR_VADDR:
930 		return gntdev_ioctl_get_offset_for_vaddr(priv, ptr);
931 
932 	case IOCTL_GNTDEV_SET_UNMAP_NOTIFY:
933 		return gntdev_ioctl_notify(priv, ptr);
934 
935 	case IOCTL_GNTDEV_GRANT_COPY:
936 		return gntdev_ioctl_grant_copy(priv, ptr);
937 
938 #ifdef CONFIG_XEN_GNTDEV_DMABUF
939 	case IOCTL_GNTDEV_DMABUF_EXP_FROM_REFS:
940 		return gntdev_ioctl_dmabuf_exp_from_refs(priv, use_ptemod, ptr);
941 
942 	case IOCTL_GNTDEV_DMABUF_EXP_WAIT_RELEASED:
943 		return gntdev_ioctl_dmabuf_exp_wait_released(priv, ptr);
944 
945 	case IOCTL_GNTDEV_DMABUF_IMP_TO_REFS:
946 		return gntdev_ioctl_dmabuf_imp_to_refs(priv, ptr);
947 
948 	case IOCTL_GNTDEV_DMABUF_IMP_RELEASE:
949 		return gntdev_ioctl_dmabuf_imp_release(priv, ptr);
950 #endif
951 
952 	default:
953 		pr_debug("priv %p, unknown cmd %x\n", priv, cmd);
954 		return -ENOIOCTLCMD;
955 	}
956 
957 	return 0;
958 }
959 
960 static int gntdev_mmap(struct file *flip, struct vm_area_struct *vma)
961 {
962 	struct gntdev_priv *priv = flip->private_data;
963 	int index = vma->vm_pgoff;
964 	int count = vma_pages(vma);
965 	struct gntdev_grant_map *map;
966 	int err = -EINVAL;
967 
968 	if ((vma->vm_flags & VM_WRITE) && !(vma->vm_flags & VM_SHARED))
969 		return -EINVAL;
970 
971 	pr_debug("map %d+%d at %lx (pgoff %lx)\n",
972 			index, count, vma->vm_start, vma->vm_pgoff);
973 
974 	mutex_lock(&priv->lock);
975 	map = gntdev_find_map_index(priv, index, count);
976 	if (!map)
977 		goto unlock_out;
978 	if (use_ptemod && map->vma)
979 		goto unlock_out;
980 	refcount_inc(&map->users);
981 
982 	vma->vm_ops = &gntdev_vmops;
983 
984 	vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP | VM_MIXEDMAP;
985 
986 	if (use_ptemod)
987 		vma->vm_flags |= VM_DONTCOPY;
988 
989 	vma->vm_private_data = map;
990 	if (map->flags) {
991 		if ((vma->vm_flags & VM_WRITE) &&
992 				(map->flags & GNTMAP_readonly))
993 			goto out_unlock_put;
994 	} else {
995 		map->flags = GNTMAP_host_map;
996 		if (!(vma->vm_flags & VM_WRITE))
997 			map->flags |= GNTMAP_readonly;
998 	}
999 
1000 	if (use_ptemod) {
1001 		map->vma = vma;
1002 		err = mmu_interval_notifier_insert_locked(
1003 			&map->notifier, vma->vm_mm, vma->vm_start,
1004 			vma->vm_end - vma->vm_start, &gntdev_mmu_ops);
1005 		if (err) {
1006 			map->vma = NULL;
1007 			goto out_unlock_put;
1008 		}
1009 	}
1010 	mutex_unlock(&priv->lock);
1011 
1012 	if (use_ptemod) {
1013 		/*
1014 		 * gntdev takes the address of the PTE in find_grant_ptes() and
1015 		 * passes it to the hypervisor in gntdev_map_grant_pages(). The
1016 		 * purpose of the notifier is to prevent the hypervisor pointer
1017 		 * to the PTE from going stale.
1018 		 *
1019 		 * Since this vma's mappings can't be touched without the
1020 		 * mmap_lock, and we are holding it now, there is no need for
1021 		 * the notifier_range locking pattern.
1022 		 */
1023 		mmu_interval_read_begin(&map->notifier);
1024 
1025 		map->pages_vm_start = vma->vm_start;
1026 		err = apply_to_page_range(vma->vm_mm, vma->vm_start,
1027 					  vma->vm_end - vma->vm_start,
1028 					  find_grant_ptes, map);
1029 		if (err) {
1030 			pr_warn("find_grant_ptes() failure.\n");
1031 			goto out_put_map;
1032 		}
1033 	}
1034 
1035 	err = gntdev_map_grant_pages(map);
1036 	if (err)
1037 		goto out_put_map;
1038 
1039 	if (!use_ptemod) {
1040 		err = vm_map_pages_zero(vma, map->pages, map->count);
1041 		if (err)
1042 			goto out_put_map;
1043 	}
1044 
1045 	return 0;
1046 
1047 unlock_out:
1048 	mutex_unlock(&priv->lock);
1049 	return err;
1050 
1051 out_unlock_put:
1052 	mutex_unlock(&priv->lock);
1053 out_put_map:
1054 	if (use_ptemod) {
1055 		unmap_grant_pages(map, 0, map->count);
1056 		if (map->vma) {
1057 			mmu_interval_notifier_remove(&map->notifier);
1058 			map->vma = NULL;
1059 		}
1060 	}
1061 	gntdev_put_map(priv, map);
1062 	return err;
1063 }
1064 
1065 static const struct file_operations gntdev_fops = {
1066 	.owner = THIS_MODULE,
1067 	.open = gntdev_open,
1068 	.release = gntdev_release,
1069 	.mmap = gntdev_mmap,
1070 	.unlocked_ioctl = gntdev_ioctl
1071 };
1072 
1073 static struct miscdevice gntdev_miscdev = {
1074 	.minor        = MISC_DYNAMIC_MINOR,
1075 	.name         = "xen/gntdev",
1076 	.fops         = &gntdev_fops,
1077 };
1078 
1079 /* ------------------------------------------------------------------ */
1080 
1081 static int __init gntdev_init(void)
1082 {
1083 	int err;
1084 
1085 	if (!xen_domain())
1086 		return -ENODEV;
1087 
1088 	use_ptemod = !xen_feature(XENFEAT_auto_translated_physmap);
1089 
1090 	err = misc_register(&gntdev_miscdev);
1091 	if (err != 0) {
1092 		pr_err("Could not register gntdev device\n");
1093 		return err;
1094 	}
1095 	return 0;
1096 }
1097 
1098 static void __exit gntdev_exit(void)
1099 {
1100 	misc_deregister(&gntdev_miscdev);
1101 }
1102 
1103 module_init(gntdev_init);
1104 module_exit(gntdev_exit);
1105 
1106 /* ------------------------------------------------------------------ */
1107