xref: /openbmc/linux/drivers/xen/gntdev.c (revision 08283d30)
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 int limit = 1024*1024;
59 module_param(limit, int, 0644);
60 MODULE_PARM_DESC(limit, "Maximum number of grants that may be mapped by "
61 		"the gntdev device");
62 
63 static atomic_t pages_mapped = ATOMIC_INIT(0);
64 
65 static int use_ptemod;
66 #define populate_freeable_maps use_ptemod
67 
68 static int unmap_grant_pages(struct gntdev_grant_map *map,
69 			     int offset, int pages);
70 
71 static struct miscdevice gntdev_miscdev;
72 
73 /* ------------------------------------------------------------------ */
74 
75 bool gntdev_account_mapped_pages(int count)
76 {
77 	return atomic_add_return(count, &pages_mapped) > limit;
78 }
79 
80 static void gntdev_print_maps(struct gntdev_priv *priv,
81 			      char *text, int text_index)
82 {
83 #ifdef DEBUG
84 	struct gntdev_grant_map *map;
85 
86 	pr_debug("%s: maps list (priv %p)\n", __func__, priv);
87 	list_for_each_entry(map, &priv->maps, next)
88 		pr_debug("  index %2d, count %2d %s\n",
89 		       map->index, map->count,
90 		       map->index == text_index && text ? text : "");
91 #endif
92 }
93 
94 static void gntdev_free_map(struct gntdev_grant_map *map)
95 {
96 	if (map == NULL)
97 		return;
98 
99 #ifdef CONFIG_XEN_GRANT_DMA_ALLOC
100 	if (map->dma_vaddr) {
101 		struct gnttab_dma_alloc_args args;
102 
103 		args.dev = map->dma_dev;
104 		args.coherent = !!(map->dma_flags & GNTDEV_DMA_FLAG_COHERENT);
105 		args.nr_pages = map->count;
106 		args.pages = map->pages;
107 		args.frames = map->frames;
108 		args.vaddr = map->dma_vaddr;
109 		args.dev_bus_addr = map->dma_bus_addr;
110 
111 		gnttab_dma_free_pages(&args);
112 	} else
113 #endif
114 	if (map->pages)
115 		gnttab_free_pages(map->count, map->pages);
116 
117 #ifdef CONFIG_XEN_GRANT_DMA_ALLOC
118 	kfree(map->frames);
119 #endif
120 	kfree(map->pages);
121 	kfree(map->grants);
122 	kfree(map->map_ops);
123 	kfree(map->unmap_ops);
124 	kfree(map->kmap_ops);
125 	kfree(map->kunmap_ops);
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    = kcalloc(count, sizeof(add->grants[0]), GFP_KERNEL);
140 	add->map_ops   = kcalloc(count, sizeof(add->map_ops[0]), GFP_KERNEL);
141 	add->unmap_ops = kcalloc(count, sizeof(add->unmap_ops[0]), GFP_KERNEL);
142 	add->kmap_ops  = kcalloc(count, sizeof(add->kmap_ops[0]), GFP_KERNEL);
143 	add->kunmap_ops = kcalloc(count, sizeof(add->kunmap_ops[0]), GFP_KERNEL);
144 	add->pages     = kcalloc(count, sizeof(add->pages[0]), GFP_KERNEL);
145 	if (NULL == add->grants    ||
146 	    NULL == add->map_ops   ||
147 	    NULL == add->unmap_ops ||
148 	    NULL == add->kmap_ops  ||
149 	    NULL == add->kunmap_ops ||
150 	    NULL == add->pages)
151 		goto err;
152 
153 #ifdef CONFIG_XEN_GRANT_DMA_ALLOC
154 	add->dma_flags = dma_flags;
155 
156 	/*
157 	 * Check if this mapping is requested to be backed
158 	 * by a DMA buffer.
159 	 */
160 	if (dma_flags & (GNTDEV_DMA_FLAG_WC | GNTDEV_DMA_FLAG_COHERENT)) {
161 		struct gnttab_dma_alloc_args args;
162 
163 		add->frames = kcalloc(count, sizeof(add->frames[0]),
164 				      GFP_KERNEL);
165 		if (!add->frames)
166 			goto err;
167 
168 		/* Remember the device, so we can free DMA memory. */
169 		add->dma_dev = priv->dma_dev;
170 
171 		args.dev = priv->dma_dev;
172 		args.coherent = !!(dma_flags & GNTDEV_DMA_FLAG_COHERENT);
173 		args.nr_pages = count;
174 		args.pages = add->pages;
175 		args.frames = add->frames;
176 
177 		if (gnttab_dma_alloc_pages(&args))
178 			goto err;
179 
180 		add->dma_vaddr = args.vaddr;
181 		add->dma_bus_addr = args.dev_bus_addr;
182 	} else
183 #endif
184 	if (gnttab_alloc_pages(count, add->pages))
185 		goto err;
186 
187 	for (i = 0; i < count; i++) {
188 		add->map_ops[i].handle = -1;
189 		add->unmap_ops[i].handle = -1;
190 		add->kmap_ops[i].handle = -1;
191 		add->kunmap_ops[i].handle = -1;
192 	}
193 
194 	add->index = 0;
195 	add->count = count;
196 	refcount_set(&add->users, 1);
197 
198 	return add;
199 
200 err:
201 	gntdev_free_map(add);
202 	return NULL;
203 }
204 
205 void gntdev_add_map(struct gntdev_priv *priv, struct gntdev_grant_map *add)
206 {
207 	struct gntdev_grant_map *map;
208 
209 	list_for_each_entry(map, &priv->maps, next) {
210 		if (add->index + add->count < map->index) {
211 			list_add_tail(&add->next, &map->next);
212 			goto done;
213 		}
214 		add->index = map->index + map->count;
215 	}
216 	list_add_tail(&add->next, &priv->maps);
217 
218 done:
219 	gntdev_print_maps(priv, "[new]", add->index);
220 }
221 
222 static struct gntdev_grant_map *gntdev_find_map_index(struct gntdev_priv *priv,
223 						      int index, int count)
224 {
225 	struct gntdev_grant_map *map;
226 
227 	list_for_each_entry(map, &priv->maps, next) {
228 		if (map->index != index)
229 			continue;
230 		if (count && map->count != count)
231 			continue;
232 		return map;
233 	}
234 	return NULL;
235 }
236 
237 void gntdev_put_map(struct gntdev_priv *priv, struct gntdev_grant_map *map)
238 {
239 	if (!map)
240 		return;
241 
242 	if (!refcount_dec_and_test(&map->users))
243 		return;
244 
245 	atomic_sub(map->count, &pages_mapped);
246 
247 	if (map->notify.flags & UNMAP_NOTIFY_SEND_EVENT) {
248 		notify_remote_via_evtchn(map->notify.event);
249 		evtchn_put(map->notify.event);
250 	}
251 
252 	if (populate_freeable_maps && priv) {
253 		mutex_lock(&priv->lock);
254 		list_del(&map->next);
255 		mutex_unlock(&priv->lock);
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 	u64 pte_maddr;
271 
272 	BUG_ON(pgnr >= map->count);
273 	pte_maddr = arbitrary_virt_to_machine(pte).maddr;
274 
275 	/*
276 	 * Set the PTE as special to force get_user_pages_fast() fall
277 	 * back to the slow path.  If this is not supported as part of
278 	 * the grant map, it will be done afterwards.
279 	 */
280 	if (xen_feature(XENFEAT_gnttab_map_avail_bits))
281 		flags |= (1 << _GNTMAP_guest_avail0);
282 
283 	gnttab_set_map_op(&map->map_ops[pgnr], pte_maddr, flags,
284 			  map->grants[pgnr].ref,
285 			  map->grants[pgnr].domid);
286 	gnttab_set_unmap_op(&map->unmap_ops[pgnr], pte_maddr, flags,
287 			    -1 /* handle */);
288 	return 0;
289 }
290 
291 #ifdef CONFIG_X86
292 static int set_grant_ptes_as_special(pte_t *pte, unsigned long addr, void *data)
293 {
294 	set_pte_at(current->mm, addr, pte, pte_mkspecial(*pte));
295 	return 0;
296 }
297 #endif
298 
299 int gntdev_map_grant_pages(struct gntdev_grant_map *map)
300 {
301 	int i, err = 0;
302 
303 	if (!use_ptemod) {
304 		/* Note: it could already be mapped */
305 		if (map->map_ops[0].handle != -1)
306 			return 0;
307 		for (i = 0; i < map->count; i++) {
308 			unsigned long addr = (unsigned long)
309 				pfn_to_kaddr(page_to_pfn(map->pages[i]));
310 			gnttab_set_map_op(&map->map_ops[i], addr, map->flags,
311 				map->grants[i].ref,
312 				map->grants[i].domid);
313 			gnttab_set_unmap_op(&map->unmap_ops[i], addr,
314 				map->flags, -1 /* handle */);
315 		}
316 	} else {
317 		/*
318 		 * Setup the map_ops corresponding to the pte entries pointing
319 		 * to the kernel linear addresses of the struct pages.
320 		 * These ptes are completely different from the user ptes dealt
321 		 * with find_grant_ptes.
322 		 */
323 		for (i = 0; i < map->count; i++) {
324 			unsigned long address = (unsigned long)
325 				pfn_to_kaddr(page_to_pfn(map->pages[i]));
326 			BUG_ON(PageHighMem(map->pages[i]));
327 
328 			gnttab_set_map_op(&map->kmap_ops[i], address,
329 				map->flags | GNTMAP_host_map,
330 				map->grants[i].ref,
331 				map->grants[i].domid);
332 			gnttab_set_unmap_op(&map->kunmap_ops[i], address,
333 				map->flags | GNTMAP_host_map, -1);
334 		}
335 	}
336 
337 	pr_debug("map %d+%d\n", map->index, map->count);
338 	err = gnttab_map_refs(map->map_ops, use_ptemod ? map->kmap_ops : NULL,
339 			map->pages, map->count);
340 	if (err)
341 		return err;
342 
343 	for (i = 0; i < map->count; i++) {
344 		if (map->map_ops[i].status) {
345 			err = -EINVAL;
346 			continue;
347 		}
348 
349 		map->unmap_ops[i].handle = map->map_ops[i].handle;
350 		if (use_ptemod)
351 			map->kunmap_ops[i].handle = map->kmap_ops[i].handle;
352 #ifdef CONFIG_XEN_GRANT_DMA_ALLOC
353 		else if (map->dma_vaddr) {
354 			unsigned long bfn;
355 
356 			bfn = pfn_to_bfn(page_to_pfn(map->pages[i]));
357 			map->unmap_ops[i].dev_bus_addr = __pfn_to_phys(bfn);
358 		}
359 #endif
360 	}
361 	return err;
362 }
363 
364 static int __unmap_grant_pages(struct gntdev_grant_map *map, int offset,
365 			       int pages)
366 {
367 	int i, err = 0;
368 	struct gntab_unmap_queue_data unmap_data;
369 
370 	if (map->notify.flags & UNMAP_NOTIFY_CLEAR_BYTE) {
371 		int pgno = (map->notify.addr >> PAGE_SHIFT);
372 		if (pgno >= offset && pgno < offset + pages) {
373 			/* No need for kmap, pages are in lowmem */
374 			uint8_t *tmp = pfn_to_kaddr(page_to_pfn(map->pages[pgno]));
375 			tmp[map->notify.addr & (PAGE_SIZE-1)] = 0;
376 			map->notify.flags &= ~UNMAP_NOTIFY_CLEAR_BYTE;
377 		}
378 	}
379 
380 	unmap_data.unmap_ops = map->unmap_ops + offset;
381 	unmap_data.kunmap_ops = use_ptemod ? map->kunmap_ops + offset : NULL;
382 	unmap_data.pages = map->pages + offset;
383 	unmap_data.count = pages;
384 
385 	err = gnttab_unmap_refs_sync(&unmap_data);
386 	if (err)
387 		return err;
388 
389 	for (i = 0; i < pages; i++) {
390 		if (map->unmap_ops[offset+i].status)
391 			err = -EINVAL;
392 		pr_debug("unmap handle=%d st=%d\n",
393 			map->unmap_ops[offset+i].handle,
394 			map->unmap_ops[offset+i].status);
395 		map->unmap_ops[offset+i].handle = -1;
396 	}
397 	return err;
398 }
399 
400 static int unmap_grant_pages(struct gntdev_grant_map *map, int offset,
401 			     int pages)
402 {
403 	int range, err = 0;
404 
405 	pr_debug("unmap %d+%d [%d+%d]\n", map->index, map->count, offset, pages);
406 
407 	/* It is possible the requested range will have a "hole" where we
408 	 * already unmapped some of the grants. Only unmap valid ranges.
409 	 */
410 	while (pages && !err) {
411 		while (pages && map->unmap_ops[offset].handle == -1) {
412 			offset++;
413 			pages--;
414 		}
415 		range = 0;
416 		while (range < pages) {
417 			if (map->unmap_ops[offset+range].handle == -1)
418 				break;
419 			range++;
420 		}
421 		err = __unmap_grant_pages(map, offset, range);
422 		offset += range;
423 		pages -= range;
424 	}
425 
426 	return err;
427 }
428 
429 /* ------------------------------------------------------------------ */
430 
431 static void gntdev_vma_open(struct vm_area_struct *vma)
432 {
433 	struct gntdev_grant_map *map = vma->vm_private_data;
434 
435 	pr_debug("gntdev_vma_open %p\n", vma);
436 	refcount_inc(&map->users);
437 }
438 
439 static void gntdev_vma_close(struct vm_area_struct *vma)
440 {
441 	struct gntdev_grant_map *map = vma->vm_private_data;
442 	struct file *file = vma->vm_file;
443 	struct gntdev_priv *priv = file->private_data;
444 
445 	pr_debug("gntdev_vma_close %p\n", vma);
446 	if (use_ptemod) {
447 		/* It is possible that an mmu notifier could be running
448 		 * concurrently, so take priv->lock to ensure that the vma won't
449 		 * vanishing during the unmap_grant_pages call, since we will
450 		 * spin here until that completes. Such a concurrent call will
451 		 * not do any unmapping, since that has been done prior to
452 		 * closing the vma, but it may still iterate the unmap_ops list.
453 		 */
454 		mutex_lock(&priv->lock);
455 		map->vma = NULL;
456 		mutex_unlock(&priv->lock);
457 	}
458 	vma->vm_private_data = NULL;
459 	gntdev_put_map(priv, map);
460 }
461 
462 static struct page *gntdev_vma_find_special_page(struct vm_area_struct *vma,
463 						 unsigned long addr)
464 {
465 	struct gntdev_grant_map *map = vma->vm_private_data;
466 
467 	return map->pages[(addr - map->pages_vm_start) >> PAGE_SHIFT];
468 }
469 
470 static const struct vm_operations_struct gntdev_vmops = {
471 	.open = gntdev_vma_open,
472 	.close = gntdev_vma_close,
473 	.find_special_page = gntdev_vma_find_special_page,
474 };
475 
476 /* ------------------------------------------------------------------ */
477 
478 static bool in_range(struct gntdev_grant_map *map,
479 			      unsigned long start, unsigned long end)
480 {
481 	if (!map->vma)
482 		return false;
483 	if (map->vma->vm_start >= end)
484 		return false;
485 	if (map->vma->vm_end <= start)
486 		return false;
487 
488 	return true;
489 }
490 
491 static int unmap_if_in_range(struct gntdev_grant_map *map,
492 			      unsigned long start, unsigned long end,
493 			      bool blockable)
494 {
495 	unsigned long mstart, mend;
496 	int err;
497 
498 	if (!in_range(map, start, end))
499 		return 0;
500 
501 	if (!blockable)
502 		return -EAGAIN;
503 
504 	mstart = max(start, map->vma->vm_start);
505 	mend   = min(end,   map->vma->vm_end);
506 	pr_debug("map %d+%d (%lx %lx), range %lx %lx, mrange %lx %lx\n",
507 			map->index, map->count,
508 			map->vma->vm_start, map->vma->vm_end,
509 			start, end, mstart, mend);
510 	err = unmap_grant_pages(map,
511 				(mstart - map->vma->vm_start) >> PAGE_SHIFT,
512 				(mend - mstart) >> PAGE_SHIFT);
513 	WARN_ON(err);
514 
515 	return 0;
516 }
517 
518 static int mn_invl_range_start(struct mmu_notifier *mn,
519 			       const struct mmu_notifier_range *range)
520 {
521 	struct gntdev_priv *priv = container_of(mn, struct gntdev_priv, mn);
522 	struct gntdev_grant_map *map;
523 	int ret = 0;
524 
525 	if (mmu_notifier_range_blockable(range))
526 		mutex_lock(&priv->lock);
527 	else if (!mutex_trylock(&priv->lock))
528 		return -EAGAIN;
529 
530 	list_for_each_entry(map, &priv->maps, next) {
531 		ret = unmap_if_in_range(map, range->start, range->end,
532 					mmu_notifier_range_blockable(range));
533 		if (ret)
534 			goto out_unlock;
535 	}
536 	list_for_each_entry(map, &priv->freeable_maps, next) {
537 		ret = unmap_if_in_range(map, range->start, range->end,
538 					mmu_notifier_range_blockable(range));
539 		if (ret)
540 			goto out_unlock;
541 	}
542 
543 out_unlock:
544 	mutex_unlock(&priv->lock);
545 
546 	return ret;
547 }
548 
549 static void mn_release(struct mmu_notifier *mn,
550 		       struct mm_struct *mm)
551 {
552 	struct gntdev_priv *priv = container_of(mn, struct gntdev_priv, mn);
553 	struct gntdev_grant_map *map;
554 	int err;
555 
556 	mutex_lock(&priv->lock);
557 	list_for_each_entry(map, &priv->maps, next) {
558 		if (!map->vma)
559 			continue;
560 		pr_debug("map %d+%d (%lx %lx)\n",
561 				map->index, map->count,
562 				map->vma->vm_start, map->vma->vm_end);
563 		err = unmap_grant_pages(map, /* offset */ 0, map->count);
564 		WARN_ON(err);
565 	}
566 	list_for_each_entry(map, &priv->freeable_maps, next) {
567 		if (!map->vma)
568 			continue;
569 		pr_debug("map %d+%d (%lx %lx)\n",
570 				map->index, map->count,
571 				map->vma->vm_start, map->vma->vm_end);
572 		err = unmap_grant_pages(map, /* offset */ 0, map->count);
573 		WARN_ON(err);
574 	}
575 	mutex_unlock(&priv->lock);
576 }
577 
578 static const struct mmu_notifier_ops gntdev_mmu_ops = {
579 	.release                = mn_release,
580 	.invalidate_range_start = mn_invl_range_start,
581 };
582 
583 /* ------------------------------------------------------------------ */
584 
585 static int gntdev_open(struct inode *inode, struct file *flip)
586 {
587 	struct gntdev_priv *priv;
588 	int ret = 0;
589 
590 	priv = kzalloc(sizeof(*priv), GFP_KERNEL);
591 	if (!priv)
592 		return -ENOMEM;
593 
594 	INIT_LIST_HEAD(&priv->maps);
595 	INIT_LIST_HEAD(&priv->freeable_maps);
596 	mutex_init(&priv->lock);
597 
598 #ifdef CONFIG_XEN_GNTDEV_DMABUF
599 	priv->dmabuf_priv = gntdev_dmabuf_init(flip);
600 	if (IS_ERR(priv->dmabuf_priv)) {
601 		ret = PTR_ERR(priv->dmabuf_priv);
602 		kfree(priv);
603 		return ret;
604 	}
605 #endif
606 
607 	if (use_ptemod) {
608 		priv->mm = get_task_mm(current);
609 		if (!priv->mm) {
610 			kfree(priv);
611 			return -ENOMEM;
612 		}
613 		priv->mn.ops = &gntdev_mmu_ops;
614 		ret = mmu_notifier_register(&priv->mn, priv->mm);
615 		mmput(priv->mm);
616 	}
617 
618 	if (ret) {
619 		kfree(priv);
620 		return ret;
621 	}
622 
623 	flip->private_data = priv;
624 #ifdef CONFIG_XEN_GRANT_DMA_ALLOC
625 	priv->dma_dev = gntdev_miscdev.this_device;
626 	dma_coerce_mask_and_coherent(priv->dma_dev, DMA_BIT_MASK(64));
627 #endif
628 	pr_debug("priv %p\n", priv);
629 
630 	return 0;
631 }
632 
633 static int gntdev_release(struct inode *inode, struct file *flip)
634 {
635 	struct gntdev_priv *priv = flip->private_data;
636 	struct gntdev_grant_map *map;
637 
638 	pr_debug("priv %p\n", priv);
639 
640 	mutex_lock(&priv->lock);
641 	while (!list_empty(&priv->maps)) {
642 		map = list_entry(priv->maps.next,
643 				 struct gntdev_grant_map, next);
644 		list_del(&map->next);
645 		gntdev_put_map(NULL /* already removed */, map);
646 	}
647 	WARN_ON(!list_empty(&priv->freeable_maps));
648 	mutex_unlock(&priv->lock);
649 
650 #ifdef CONFIG_XEN_GNTDEV_DMABUF
651 	gntdev_dmabuf_fini(priv->dmabuf_priv);
652 #endif
653 
654 	if (use_ptemod)
655 		mmu_notifier_unregister(&priv->mn, priv->mm);
656 
657 	kfree(priv);
658 	return 0;
659 }
660 
661 static long gntdev_ioctl_map_grant_ref(struct gntdev_priv *priv,
662 				       struct ioctl_gntdev_map_grant_ref __user *u)
663 {
664 	struct ioctl_gntdev_map_grant_ref op;
665 	struct gntdev_grant_map *map;
666 	int err;
667 
668 	if (copy_from_user(&op, u, sizeof(op)) != 0)
669 		return -EFAULT;
670 	pr_debug("priv %p, add %d\n", priv, op.count);
671 	if (unlikely(op.count <= 0))
672 		return -EINVAL;
673 
674 	err = -ENOMEM;
675 	map = gntdev_alloc_map(priv, op.count, 0 /* This is not a dma-buf. */);
676 	if (!map)
677 		return err;
678 
679 	if (unlikely(gntdev_account_mapped_pages(op.count))) {
680 		pr_debug("can't map: over limit\n");
681 		gntdev_put_map(NULL, map);
682 		return err;
683 	}
684 
685 	if (copy_from_user(map->grants, &u->refs,
686 			   sizeof(map->grants[0]) * op.count) != 0) {
687 		gntdev_put_map(NULL, map);
688 		return -EFAULT;
689 	}
690 
691 	mutex_lock(&priv->lock);
692 	gntdev_add_map(priv, map);
693 	op.index = map->index << PAGE_SHIFT;
694 	mutex_unlock(&priv->lock);
695 
696 	if (copy_to_user(u, &op, sizeof(op)) != 0)
697 		return -EFAULT;
698 
699 	return 0;
700 }
701 
702 static long gntdev_ioctl_unmap_grant_ref(struct gntdev_priv *priv,
703 					 struct ioctl_gntdev_unmap_grant_ref __user *u)
704 {
705 	struct ioctl_gntdev_unmap_grant_ref op;
706 	struct gntdev_grant_map *map;
707 	int err = -ENOENT;
708 
709 	if (copy_from_user(&op, u, sizeof(op)) != 0)
710 		return -EFAULT;
711 	pr_debug("priv %p, del %d+%d\n", priv, (int)op.index, (int)op.count);
712 
713 	mutex_lock(&priv->lock);
714 	map = gntdev_find_map_index(priv, op.index >> PAGE_SHIFT, op.count);
715 	if (map) {
716 		list_del(&map->next);
717 		if (populate_freeable_maps)
718 			list_add_tail(&map->next, &priv->freeable_maps);
719 		err = 0;
720 	}
721 	mutex_unlock(&priv->lock);
722 	if (map)
723 		gntdev_put_map(priv, map);
724 	return err;
725 }
726 
727 static long gntdev_ioctl_get_offset_for_vaddr(struct gntdev_priv *priv,
728 					      struct ioctl_gntdev_get_offset_for_vaddr __user *u)
729 {
730 	struct ioctl_gntdev_get_offset_for_vaddr op;
731 	struct vm_area_struct *vma;
732 	struct gntdev_grant_map *map;
733 	int rv = -EINVAL;
734 
735 	if (copy_from_user(&op, u, sizeof(op)) != 0)
736 		return -EFAULT;
737 	pr_debug("priv %p, offset for vaddr %lx\n", priv, (unsigned long)op.vaddr);
738 
739 	down_read(&current->mm->mmap_sem);
740 	vma = find_vma(current->mm, op.vaddr);
741 	if (!vma || vma->vm_ops != &gntdev_vmops)
742 		goto out_unlock;
743 
744 	map = vma->vm_private_data;
745 	if (!map)
746 		goto out_unlock;
747 
748 	op.offset = map->index << PAGE_SHIFT;
749 	op.count = map->count;
750 	rv = 0;
751 
752  out_unlock:
753 	up_read(&current->mm->mmap_sem);
754 
755 	if (rv == 0 && copy_to_user(u, &op, sizeof(op)) != 0)
756 		return -EFAULT;
757 	return rv;
758 }
759 
760 static long gntdev_ioctl_notify(struct gntdev_priv *priv, void __user *u)
761 {
762 	struct ioctl_gntdev_unmap_notify op;
763 	struct gntdev_grant_map *map;
764 	int rc;
765 	int out_flags;
766 	unsigned int out_event;
767 
768 	if (copy_from_user(&op, u, sizeof(op)))
769 		return -EFAULT;
770 
771 	if (op.action & ~(UNMAP_NOTIFY_CLEAR_BYTE|UNMAP_NOTIFY_SEND_EVENT))
772 		return -EINVAL;
773 
774 	/* We need to grab a reference to the event channel we are going to use
775 	 * to send the notify before releasing the reference we may already have
776 	 * (if someone has called this ioctl twice). This is required so that
777 	 * it is possible to change the clear_byte part of the notification
778 	 * without disturbing the event channel part, which may now be the last
779 	 * reference to that event channel.
780 	 */
781 	if (op.action & UNMAP_NOTIFY_SEND_EVENT) {
782 		if (evtchn_get(op.event_channel_port))
783 			return -EINVAL;
784 	}
785 
786 	out_flags = op.action;
787 	out_event = op.event_channel_port;
788 
789 	mutex_lock(&priv->lock);
790 
791 	list_for_each_entry(map, &priv->maps, next) {
792 		uint64_t begin = map->index << PAGE_SHIFT;
793 		uint64_t end = (map->index + map->count) << PAGE_SHIFT;
794 		if (op.index >= begin && op.index < end)
795 			goto found;
796 	}
797 	rc = -ENOENT;
798 	goto unlock_out;
799 
800  found:
801 	if ((op.action & UNMAP_NOTIFY_CLEAR_BYTE) &&
802 			(map->flags & GNTMAP_readonly)) {
803 		rc = -EINVAL;
804 		goto unlock_out;
805 	}
806 
807 	out_flags = map->notify.flags;
808 	out_event = map->notify.event;
809 
810 	map->notify.flags = op.action;
811 	map->notify.addr = op.index - (map->index << PAGE_SHIFT);
812 	map->notify.event = op.event_channel_port;
813 
814 	rc = 0;
815 
816  unlock_out:
817 	mutex_unlock(&priv->lock);
818 
819 	/* Drop the reference to the event channel we did not save in the map */
820 	if (out_flags & UNMAP_NOTIFY_SEND_EVENT)
821 		evtchn_put(out_event);
822 
823 	return rc;
824 }
825 
826 #define GNTDEV_COPY_BATCH 16
827 
828 struct gntdev_copy_batch {
829 	struct gnttab_copy ops[GNTDEV_COPY_BATCH];
830 	struct page *pages[GNTDEV_COPY_BATCH];
831 	s16 __user *status[GNTDEV_COPY_BATCH];
832 	unsigned int nr_ops;
833 	unsigned int nr_pages;
834 };
835 
836 static int gntdev_get_page(struct gntdev_copy_batch *batch, void __user *virt,
837 			   bool writeable, unsigned long *gfn)
838 {
839 	unsigned long addr = (unsigned long)virt;
840 	struct page *page;
841 	unsigned long xen_pfn;
842 	int ret;
843 
844 	ret = get_user_pages_fast(addr, 1, writeable ? FOLL_WRITE : 0, &page);
845 	if (ret < 0)
846 		return ret;
847 
848 	batch->pages[batch->nr_pages++] = page;
849 
850 	xen_pfn = page_to_xen_pfn(page) + XEN_PFN_DOWN(addr & ~PAGE_MASK);
851 	*gfn = pfn_to_gfn(xen_pfn);
852 
853 	return 0;
854 }
855 
856 static void gntdev_put_pages(struct gntdev_copy_batch *batch)
857 {
858 	unsigned int i;
859 
860 	for (i = 0; i < batch->nr_pages; i++)
861 		put_page(batch->pages[i]);
862 	batch->nr_pages = 0;
863 }
864 
865 static int gntdev_copy(struct gntdev_copy_batch *batch)
866 {
867 	unsigned int i;
868 
869 	gnttab_batch_copy(batch->ops, batch->nr_ops);
870 	gntdev_put_pages(batch);
871 
872 	/*
873 	 * For each completed op, update the status if the op failed
874 	 * and all previous ops for the segment were successful.
875 	 */
876 	for (i = 0; i < batch->nr_ops; i++) {
877 		s16 status = batch->ops[i].status;
878 		s16 old_status;
879 
880 		if (status == GNTST_okay)
881 			continue;
882 
883 		if (__get_user(old_status, batch->status[i]))
884 			return -EFAULT;
885 
886 		if (old_status != GNTST_okay)
887 			continue;
888 
889 		if (__put_user(status, batch->status[i]))
890 			return -EFAULT;
891 	}
892 
893 	batch->nr_ops = 0;
894 	return 0;
895 }
896 
897 static int gntdev_grant_copy_seg(struct gntdev_copy_batch *batch,
898 				 struct gntdev_grant_copy_segment *seg,
899 				 s16 __user *status)
900 {
901 	uint16_t copied = 0;
902 
903 	/*
904 	 * Disallow local -> local copies since there is only space in
905 	 * batch->pages for one page per-op and this would be a very
906 	 * expensive memcpy().
907 	 */
908 	if (!(seg->flags & (GNTCOPY_source_gref | GNTCOPY_dest_gref)))
909 		return -EINVAL;
910 
911 	/* Can't cross page if source/dest is a grant ref. */
912 	if (seg->flags & GNTCOPY_source_gref) {
913 		if (seg->source.foreign.offset + seg->len > XEN_PAGE_SIZE)
914 			return -EINVAL;
915 	}
916 	if (seg->flags & GNTCOPY_dest_gref) {
917 		if (seg->dest.foreign.offset + seg->len > XEN_PAGE_SIZE)
918 			return -EINVAL;
919 	}
920 
921 	if (put_user(GNTST_okay, status))
922 		return -EFAULT;
923 
924 	while (copied < seg->len) {
925 		struct gnttab_copy *op;
926 		void __user *virt;
927 		size_t len, off;
928 		unsigned long gfn;
929 		int ret;
930 
931 		if (batch->nr_ops >= GNTDEV_COPY_BATCH) {
932 			ret = gntdev_copy(batch);
933 			if (ret < 0)
934 				return ret;
935 		}
936 
937 		len = seg->len - copied;
938 
939 		op = &batch->ops[batch->nr_ops];
940 		op->flags = 0;
941 
942 		if (seg->flags & GNTCOPY_source_gref) {
943 			op->source.u.ref = seg->source.foreign.ref;
944 			op->source.domid = seg->source.foreign.domid;
945 			op->source.offset = seg->source.foreign.offset + copied;
946 			op->flags |= GNTCOPY_source_gref;
947 		} else {
948 			virt = seg->source.virt + copied;
949 			off = (unsigned long)virt & ~XEN_PAGE_MASK;
950 			len = min(len, (size_t)XEN_PAGE_SIZE - off);
951 
952 			ret = gntdev_get_page(batch, virt, false, &gfn);
953 			if (ret < 0)
954 				return ret;
955 
956 			op->source.u.gmfn = gfn;
957 			op->source.domid = DOMID_SELF;
958 			op->source.offset = off;
959 		}
960 
961 		if (seg->flags & GNTCOPY_dest_gref) {
962 			op->dest.u.ref = seg->dest.foreign.ref;
963 			op->dest.domid = seg->dest.foreign.domid;
964 			op->dest.offset = seg->dest.foreign.offset + copied;
965 			op->flags |= GNTCOPY_dest_gref;
966 		} else {
967 			virt = seg->dest.virt + copied;
968 			off = (unsigned long)virt & ~XEN_PAGE_MASK;
969 			len = min(len, (size_t)XEN_PAGE_SIZE - off);
970 
971 			ret = gntdev_get_page(batch, virt, true, &gfn);
972 			if (ret < 0)
973 				return ret;
974 
975 			op->dest.u.gmfn = gfn;
976 			op->dest.domid = DOMID_SELF;
977 			op->dest.offset = off;
978 		}
979 
980 		op->len = len;
981 		copied += len;
982 
983 		batch->status[batch->nr_ops] = status;
984 		batch->nr_ops++;
985 	}
986 
987 	return 0;
988 }
989 
990 static long gntdev_ioctl_grant_copy(struct gntdev_priv *priv, void __user *u)
991 {
992 	struct ioctl_gntdev_grant_copy copy;
993 	struct gntdev_copy_batch batch;
994 	unsigned int i;
995 	int ret = 0;
996 
997 	if (copy_from_user(&copy, u, sizeof(copy)))
998 		return -EFAULT;
999 
1000 	batch.nr_ops = 0;
1001 	batch.nr_pages = 0;
1002 
1003 	for (i = 0; i < copy.count; i++) {
1004 		struct gntdev_grant_copy_segment seg;
1005 
1006 		if (copy_from_user(&seg, &copy.segments[i], sizeof(seg))) {
1007 			ret = -EFAULT;
1008 			goto out;
1009 		}
1010 
1011 		ret = gntdev_grant_copy_seg(&batch, &seg, &copy.segments[i].status);
1012 		if (ret < 0)
1013 			goto out;
1014 
1015 		cond_resched();
1016 	}
1017 	if (batch.nr_ops)
1018 		ret = gntdev_copy(&batch);
1019 	return ret;
1020 
1021   out:
1022 	gntdev_put_pages(&batch);
1023 	return ret;
1024 }
1025 
1026 static long gntdev_ioctl(struct file *flip,
1027 			 unsigned int cmd, unsigned long arg)
1028 {
1029 	struct gntdev_priv *priv = flip->private_data;
1030 	void __user *ptr = (void __user *)arg;
1031 
1032 	switch (cmd) {
1033 	case IOCTL_GNTDEV_MAP_GRANT_REF:
1034 		return gntdev_ioctl_map_grant_ref(priv, ptr);
1035 
1036 	case IOCTL_GNTDEV_UNMAP_GRANT_REF:
1037 		return gntdev_ioctl_unmap_grant_ref(priv, ptr);
1038 
1039 	case IOCTL_GNTDEV_GET_OFFSET_FOR_VADDR:
1040 		return gntdev_ioctl_get_offset_for_vaddr(priv, ptr);
1041 
1042 	case IOCTL_GNTDEV_SET_UNMAP_NOTIFY:
1043 		return gntdev_ioctl_notify(priv, ptr);
1044 
1045 	case IOCTL_GNTDEV_GRANT_COPY:
1046 		return gntdev_ioctl_grant_copy(priv, ptr);
1047 
1048 #ifdef CONFIG_XEN_GNTDEV_DMABUF
1049 	case IOCTL_GNTDEV_DMABUF_EXP_FROM_REFS:
1050 		return gntdev_ioctl_dmabuf_exp_from_refs(priv, use_ptemod, ptr);
1051 
1052 	case IOCTL_GNTDEV_DMABUF_EXP_WAIT_RELEASED:
1053 		return gntdev_ioctl_dmabuf_exp_wait_released(priv, ptr);
1054 
1055 	case IOCTL_GNTDEV_DMABUF_IMP_TO_REFS:
1056 		return gntdev_ioctl_dmabuf_imp_to_refs(priv, ptr);
1057 
1058 	case IOCTL_GNTDEV_DMABUF_IMP_RELEASE:
1059 		return gntdev_ioctl_dmabuf_imp_release(priv, ptr);
1060 #endif
1061 
1062 	default:
1063 		pr_debug("priv %p, unknown cmd %x\n", priv, cmd);
1064 		return -ENOIOCTLCMD;
1065 	}
1066 
1067 	return 0;
1068 }
1069 
1070 static int gntdev_mmap(struct file *flip, struct vm_area_struct *vma)
1071 {
1072 	struct gntdev_priv *priv = flip->private_data;
1073 	int index = vma->vm_pgoff;
1074 	int count = vma_pages(vma);
1075 	struct gntdev_grant_map *map;
1076 	int err = -EINVAL;
1077 
1078 	if ((vma->vm_flags & VM_WRITE) && !(vma->vm_flags & VM_SHARED))
1079 		return -EINVAL;
1080 
1081 	pr_debug("map %d+%d at %lx (pgoff %lx)\n",
1082 			index, count, vma->vm_start, vma->vm_pgoff);
1083 
1084 	mutex_lock(&priv->lock);
1085 	map = gntdev_find_map_index(priv, index, count);
1086 	if (!map)
1087 		goto unlock_out;
1088 	if (use_ptemod && map->vma)
1089 		goto unlock_out;
1090 	if (use_ptemod && priv->mm != vma->vm_mm) {
1091 		pr_warn("Huh? Other mm?\n");
1092 		goto unlock_out;
1093 	}
1094 
1095 	refcount_inc(&map->users);
1096 
1097 	vma->vm_ops = &gntdev_vmops;
1098 
1099 	vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP | VM_MIXEDMAP;
1100 
1101 	if (use_ptemod)
1102 		vma->vm_flags |= VM_DONTCOPY;
1103 
1104 	vma->vm_private_data = map;
1105 
1106 	if (use_ptemod)
1107 		map->vma = vma;
1108 
1109 	if (map->flags) {
1110 		if ((vma->vm_flags & VM_WRITE) &&
1111 				(map->flags & GNTMAP_readonly))
1112 			goto out_unlock_put;
1113 	} else {
1114 		map->flags = GNTMAP_host_map;
1115 		if (!(vma->vm_flags & VM_WRITE))
1116 			map->flags |= GNTMAP_readonly;
1117 	}
1118 
1119 	mutex_unlock(&priv->lock);
1120 
1121 	if (use_ptemod) {
1122 		map->pages_vm_start = vma->vm_start;
1123 		err = apply_to_page_range(vma->vm_mm, vma->vm_start,
1124 					  vma->vm_end - vma->vm_start,
1125 					  find_grant_ptes, map);
1126 		if (err) {
1127 			pr_warn("find_grant_ptes() failure.\n");
1128 			goto out_put_map;
1129 		}
1130 	}
1131 
1132 	err = gntdev_map_grant_pages(map);
1133 	if (err)
1134 		goto out_put_map;
1135 
1136 	if (!use_ptemod) {
1137 		err = vm_map_pages_zero(vma, map->pages, map->count);
1138 		if (err)
1139 			goto out_put_map;
1140 	} else {
1141 #ifdef CONFIG_X86
1142 		/*
1143 		 * If the PTEs were not made special by the grant map
1144 		 * hypercall, do so here.
1145 		 *
1146 		 * This is racy since the mapping is already visible
1147 		 * to userspace but userspace should be well-behaved
1148 		 * enough to not touch it until the mmap() call
1149 		 * returns.
1150 		 */
1151 		if (!xen_feature(XENFEAT_gnttab_map_avail_bits)) {
1152 			apply_to_page_range(vma->vm_mm, vma->vm_start,
1153 					    vma->vm_end - vma->vm_start,
1154 					    set_grant_ptes_as_special, NULL);
1155 		}
1156 #endif
1157 	}
1158 
1159 	return 0;
1160 
1161 unlock_out:
1162 	mutex_unlock(&priv->lock);
1163 	return err;
1164 
1165 out_unlock_put:
1166 	mutex_unlock(&priv->lock);
1167 out_put_map:
1168 	if (use_ptemod) {
1169 		map->vma = NULL;
1170 		unmap_grant_pages(map, 0, map->count);
1171 	}
1172 	gntdev_put_map(priv, map);
1173 	return err;
1174 }
1175 
1176 static const struct file_operations gntdev_fops = {
1177 	.owner = THIS_MODULE,
1178 	.open = gntdev_open,
1179 	.release = gntdev_release,
1180 	.mmap = gntdev_mmap,
1181 	.unlocked_ioctl = gntdev_ioctl
1182 };
1183 
1184 static struct miscdevice gntdev_miscdev = {
1185 	.minor        = MISC_DYNAMIC_MINOR,
1186 	.name         = "xen/gntdev",
1187 	.fops         = &gntdev_fops,
1188 };
1189 
1190 /* ------------------------------------------------------------------ */
1191 
1192 static int __init gntdev_init(void)
1193 {
1194 	int err;
1195 
1196 	if (!xen_domain())
1197 		return -ENODEV;
1198 
1199 	use_ptemod = !xen_feature(XENFEAT_auto_translated_physmap);
1200 
1201 	err = misc_register(&gntdev_miscdev);
1202 	if (err != 0) {
1203 		pr_err("Could not register gntdev device\n");
1204 		return err;
1205 	}
1206 	return 0;
1207 }
1208 
1209 static void __exit gntdev_exit(void)
1210 {
1211 	misc_deregister(&gntdev_miscdev);
1212 }
1213 
1214 module_init(gntdev_init);
1215 module_exit(gntdev_exit);
1216 
1217 /* ------------------------------------------------------------------ */
1218