xref: /openbmc/linux/drivers/xen/gntdev.c (revision f519cd13)
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    = kvcalloc(count, sizeof(add->grants[0]), GFP_KERNEL);
137 	add->map_ops   = kvcalloc(count, sizeof(add->map_ops[0]), GFP_KERNEL);
138 	add->unmap_ops = kvcalloc(count, sizeof(add->unmap_ops[0]), GFP_KERNEL);
139 	add->kmap_ops  = kvcalloc(count, sizeof(add->kmap_ops[0]), GFP_KERNEL);
140 	add->kunmap_ops = kvcalloc(count,
141 				   sizeof(add->kunmap_ops[0]), 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->kmap_ops  ||
147 	    NULL == add->kunmap_ops ||
148 	    NULL == add->pages)
149 		goto err;
150 
151 #ifdef CONFIG_XEN_GRANT_DMA_ALLOC
152 	add->dma_flags = dma_flags;
153 
154 	/*
155 	 * Check if this mapping is requested to be backed
156 	 * by a DMA buffer.
157 	 */
158 	if (dma_flags & (GNTDEV_DMA_FLAG_WC | GNTDEV_DMA_FLAG_COHERENT)) {
159 		struct gnttab_dma_alloc_args args;
160 
161 		add->frames = kvcalloc(count, sizeof(add->frames[0]),
162 				       GFP_KERNEL);
163 		if (!add->frames)
164 			goto err;
165 
166 		/* Remember the device, so we can free DMA memory. */
167 		add->dma_dev = priv->dma_dev;
168 
169 		args.dev = priv->dma_dev;
170 		args.coherent = !!(dma_flags & GNTDEV_DMA_FLAG_COHERENT);
171 		args.nr_pages = count;
172 		args.pages = add->pages;
173 		args.frames = add->frames;
174 
175 		if (gnttab_dma_alloc_pages(&args))
176 			goto err;
177 
178 		add->dma_vaddr = args.vaddr;
179 		add->dma_bus_addr = args.dev_bus_addr;
180 	} else
181 #endif
182 	if (gnttab_alloc_pages(count, add->pages))
183 		goto err;
184 
185 	for (i = 0; i < count; i++) {
186 		add->map_ops[i].handle = -1;
187 		add->unmap_ops[i].handle = -1;
188 		add->kmap_ops[i].handle = -1;
189 		add->kunmap_ops[i].handle = -1;
190 	}
191 
192 	add->index = 0;
193 	add->count = count;
194 	refcount_set(&add->users, 1);
195 
196 	return add;
197 
198 err:
199 	gntdev_free_map(add);
200 	return NULL;
201 }
202 
203 void gntdev_add_map(struct gntdev_priv *priv, struct gntdev_grant_map *add)
204 {
205 	struct gntdev_grant_map *map;
206 
207 	list_for_each_entry(map, &priv->maps, next) {
208 		if (add->index + add->count < map->index) {
209 			list_add_tail(&add->next, &map->next);
210 			goto done;
211 		}
212 		add->index = map->index + map->count;
213 	}
214 	list_add_tail(&add->next, &priv->maps);
215 
216 done:
217 	gntdev_print_maps(priv, "[new]", add->index);
218 }
219 
220 static struct gntdev_grant_map *gntdev_find_map_index(struct gntdev_priv *priv,
221 						      int index, int count)
222 {
223 	struct gntdev_grant_map *map;
224 
225 	list_for_each_entry(map, &priv->maps, next) {
226 		if (map->index != index)
227 			continue;
228 		if (count && map->count != count)
229 			continue;
230 		return map;
231 	}
232 	return NULL;
233 }
234 
235 void gntdev_put_map(struct gntdev_priv *priv, struct gntdev_grant_map *map)
236 {
237 	if (!map)
238 		return;
239 
240 	if (!refcount_dec_and_test(&map->users))
241 		return;
242 
243 	if (map->notify.flags & UNMAP_NOTIFY_SEND_EVENT) {
244 		notify_remote_via_evtchn(map->notify.event);
245 		evtchn_put(map->notify.event);
246 	}
247 
248 	if (map->pages && !use_ptemod)
249 		unmap_grant_pages(map, 0, map->count);
250 	gntdev_free_map(map);
251 }
252 
253 /* ------------------------------------------------------------------ */
254 
255 static int find_grant_ptes(pte_t *pte, unsigned long addr, void *data)
256 {
257 	struct gntdev_grant_map *map = data;
258 	unsigned int pgnr = (addr - map->vma->vm_start) >> PAGE_SHIFT;
259 	int flags = map->flags | GNTMAP_application_map | GNTMAP_contains_pte;
260 	u64 pte_maddr;
261 
262 	BUG_ON(pgnr >= map->count);
263 	pte_maddr = arbitrary_virt_to_machine(pte).maddr;
264 
265 	/*
266 	 * Set the PTE as special to force get_user_pages_fast() fall
267 	 * back to the slow path.  If this is not supported as part of
268 	 * the grant map, it will be done afterwards.
269 	 */
270 	if (xen_feature(XENFEAT_gnttab_map_avail_bits))
271 		flags |= (1 << _GNTMAP_guest_avail0);
272 
273 	gnttab_set_map_op(&map->map_ops[pgnr], pte_maddr, flags,
274 			  map->grants[pgnr].ref,
275 			  map->grants[pgnr].domid);
276 	gnttab_set_unmap_op(&map->unmap_ops[pgnr], pte_maddr, flags,
277 			    -1 /* handle */);
278 	return 0;
279 }
280 
281 #ifdef CONFIG_X86
282 static int set_grant_ptes_as_special(pte_t *pte, unsigned long addr, void *data)
283 {
284 	set_pte_at(current->mm, addr, pte, pte_mkspecial(*pte));
285 	return 0;
286 }
287 #endif
288 
289 int gntdev_map_grant_pages(struct gntdev_grant_map *map)
290 {
291 	int i, err = 0;
292 
293 	if (!use_ptemod) {
294 		/* Note: it could already be mapped */
295 		if (map->map_ops[0].handle != -1)
296 			return 0;
297 		for (i = 0; i < map->count; i++) {
298 			unsigned long addr = (unsigned long)
299 				pfn_to_kaddr(page_to_pfn(map->pages[i]));
300 			gnttab_set_map_op(&map->map_ops[i], addr, map->flags,
301 				map->grants[i].ref,
302 				map->grants[i].domid);
303 			gnttab_set_unmap_op(&map->unmap_ops[i], addr,
304 				map->flags, -1 /* handle */);
305 		}
306 	} else {
307 		/*
308 		 * Setup the map_ops corresponding to the pte entries pointing
309 		 * to the kernel linear addresses of the struct pages.
310 		 * These ptes are completely different from the user ptes dealt
311 		 * with find_grant_ptes.
312 		 */
313 		for (i = 0; i < map->count; i++) {
314 			unsigned long address = (unsigned long)
315 				pfn_to_kaddr(page_to_pfn(map->pages[i]));
316 			BUG_ON(PageHighMem(map->pages[i]));
317 
318 			gnttab_set_map_op(&map->kmap_ops[i], address,
319 				map->flags | GNTMAP_host_map,
320 				map->grants[i].ref,
321 				map->grants[i].domid);
322 			gnttab_set_unmap_op(&map->kunmap_ops[i], address,
323 				map->flags | GNTMAP_host_map, -1);
324 		}
325 	}
326 
327 	pr_debug("map %d+%d\n", map->index, map->count);
328 	err = gnttab_map_refs(map->map_ops, use_ptemod ? map->kmap_ops : NULL,
329 			map->pages, map->count);
330 	if (err)
331 		return err;
332 
333 	for (i = 0; i < map->count; i++) {
334 		if (map->map_ops[i].status) {
335 			err = -EINVAL;
336 			continue;
337 		}
338 
339 		map->unmap_ops[i].handle = map->map_ops[i].handle;
340 		if (use_ptemod)
341 			map->kunmap_ops[i].handle = map->kmap_ops[i].handle;
342 #ifdef CONFIG_XEN_GRANT_DMA_ALLOC
343 		else if (map->dma_vaddr) {
344 			unsigned long bfn;
345 
346 			bfn = pfn_to_bfn(page_to_pfn(map->pages[i]));
347 			map->unmap_ops[i].dev_bus_addr = __pfn_to_phys(bfn);
348 		}
349 #endif
350 	}
351 	return err;
352 }
353 
354 static int __unmap_grant_pages(struct gntdev_grant_map *map, int offset,
355 			       int pages)
356 {
357 	int i, err = 0;
358 	struct gntab_unmap_queue_data unmap_data;
359 
360 	if (map->notify.flags & UNMAP_NOTIFY_CLEAR_BYTE) {
361 		int pgno = (map->notify.addr >> PAGE_SHIFT);
362 		if (pgno >= offset && pgno < offset + pages) {
363 			/* No need for kmap, pages are in lowmem */
364 			uint8_t *tmp = pfn_to_kaddr(page_to_pfn(map->pages[pgno]));
365 			tmp[map->notify.addr & (PAGE_SIZE-1)] = 0;
366 			map->notify.flags &= ~UNMAP_NOTIFY_CLEAR_BYTE;
367 		}
368 	}
369 
370 	unmap_data.unmap_ops = map->unmap_ops + offset;
371 	unmap_data.kunmap_ops = use_ptemod ? map->kunmap_ops + offset : NULL;
372 	unmap_data.pages = map->pages + offset;
373 	unmap_data.count = pages;
374 
375 	err = gnttab_unmap_refs_sync(&unmap_data);
376 	if (err)
377 		return err;
378 
379 	for (i = 0; i < pages; i++) {
380 		if (map->unmap_ops[offset+i].status)
381 			err = -EINVAL;
382 		pr_debug("unmap handle=%d st=%d\n",
383 			map->unmap_ops[offset+i].handle,
384 			map->unmap_ops[offset+i].status);
385 		map->unmap_ops[offset+i].handle = -1;
386 	}
387 	return err;
388 }
389 
390 static int unmap_grant_pages(struct gntdev_grant_map *map, int offset,
391 			     int pages)
392 {
393 	int range, err = 0;
394 
395 	pr_debug("unmap %d+%d [%d+%d]\n", map->index, map->count, offset, pages);
396 
397 	/* It is possible the requested range will have a "hole" where we
398 	 * already unmapped some of the grants. Only unmap valid ranges.
399 	 */
400 	while (pages && !err) {
401 		while (pages && map->unmap_ops[offset].handle == -1) {
402 			offset++;
403 			pages--;
404 		}
405 		range = 0;
406 		while (range < pages) {
407 			if (map->unmap_ops[offset+range].handle == -1)
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 	down_read(&current->mm->mmap_sem);
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 	up_read(&current->mm->mmap_sem);
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 	unsigned int 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 };
724 
725 static int gntdev_get_page(struct gntdev_copy_batch *batch, void __user *virt,
726 			   bool writeable, unsigned long *gfn)
727 {
728 	unsigned long addr = (unsigned long)virt;
729 	struct page *page;
730 	unsigned long xen_pfn;
731 	int ret;
732 
733 	ret = get_user_pages_fast(addr, 1, writeable ? FOLL_WRITE : 0, &page);
734 	if (ret < 0)
735 		return ret;
736 
737 	batch->pages[batch->nr_pages++] = page;
738 
739 	xen_pfn = page_to_xen_pfn(page) + XEN_PFN_DOWN(addr & ~PAGE_MASK);
740 	*gfn = pfn_to_gfn(xen_pfn);
741 
742 	return 0;
743 }
744 
745 static void gntdev_put_pages(struct gntdev_copy_batch *batch)
746 {
747 	unsigned int i;
748 
749 	for (i = 0; i < batch->nr_pages; i++)
750 		put_page(batch->pages[i]);
751 	batch->nr_pages = 0;
752 }
753 
754 static int gntdev_copy(struct gntdev_copy_batch *batch)
755 {
756 	unsigned int i;
757 
758 	gnttab_batch_copy(batch->ops, batch->nr_ops);
759 	gntdev_put_pages(batch);
760 
761 	/*
762 	 * For each completed op, update the status if the op failed
763 	 * and all previous ops for the segment were successful.
764 	 */
765 	for (i = 0; i < batch->nr_ops; i++) {
766 		s16 status = batch->ops[i].status;
767 		s16 old_status;
768 
769 		if (status == GNTST_okay)
770 			continue;
771 
772 		if (__get_user(old_status, batch->status[i]))
773 			return -EFAULT;
774 
775 		if (old_status != GNTST_okay)
776 			continue;
777 
778 		if (__put_user(status, batch->status[i]))
779 			return -EFAULT;
780 	}
781 
782 	batch->nr_ops = 0;
783 	return 0;
784 }
785 
786 static int gntdev_grant_copy_seg(struct gntdev_copy_batch *batch,
787 				 struct gntdev_grant_copy_segment *seg,
788 				 s16 __user *status)
789 {
790 	uint16_t copied = 0;
791 
792 	/*
793 	 * Disallow local -> local copies since there is only space in
794 	 * batch->pages for one page per-op and this would be a very
795 	 * expensive memcpy().
796 	 */
797 	if (!(seg->flags & (GNTCOPY_source_gref | GNTCOPY_dest_gref)))
798 		return -EINVAL;
799 
800 	/* Can't cross page if source/dest is a grant ref. */
801 	if (seg->flags & GNTCOPY_source_gref) {
802 		if (seg->source.foreign.offset + seg->len > XEN_PAGE_SIZE)
803 			return -EINVAL;
804 	}
805 	if (seg->flags & GNTCOPY_dest_gref) {
806 		if (seg->dest.foreign.offset + seg->len > XEN_PAGE_SIZE)
807 			return -EINVAL;
808 	}
809 
810 	if (put_user(GNTST_okay, status))
811 		return -EFAULT;
812 
813 	while (copied < seg->len) {
814 		struct gnttab_copy *op;
815 		void __user *virt;
816 		size_t len, off;
817 		unsigned long gfn;
818 		int ret;
819 
820 		if (batch->nr_ops >= GNTDEV_COPY_BATCH) {
821 			ret = gntdev_copy(batch);
822 			if (ret < 0)
823 				return ret;
824 		}
825 
826 		len = seg->len - copied;
827 
828 		op = &batch->ops[batch->nr_ops];
829 		op->flags = 0;
830 
831 		if (seg->flags & GNTCOPY_source_gref) {
832 			op->source.u.ref = seg->source.foreign.ref;
833 			op->source.domid = seg->source.foreign.domid;
834 			op->source.offset = seg->source.foreign.offset + copied;
835 			op->flags |= GNTCOPY_source_gref;
836 		} else {
837 			virt = seg->source.virt + copied;
838 			off = (unsigned long)virt & ~XEN_PAGE_MASK;
839 			len = min(len, (size_t)XEN_PAGE_SIZE - off);
840 
841 			ret = gntdev_get_page(batch, virt, false, &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 
860 			ret = gntdev_get_page(batch, virt, true, &gfn);
861 			if (ret < 0)
862 				return ret;
863 
864 			op->dest.u.gmfn = gfn;
865 			op->dest.domid = DOMID_SELF;
866 			op->dest.offset = off;
867 		}
868 
869 		op->len = len;
870 		copied += len;
871 
872 		batch->status[batch->nr_ops] = status;
873 		batch->nr_ops++;
874 	}
875 
876 	return 0;
877 }
878 
879 static long gntdev_ioctl_grant_copy(struct gntdev_priv *priv, void __user *u)
880 {
881 	struct ioctl_gntdev_grant_copy copy;
882 	struct gntdev_copy_batch batch;
883 	unsigned int i;
884 	int ret = 0;
885 
886 	if (copy_from_user(&copy, u, sizeof(copy)))
887 		return -EFAULT;
888 
889 	batch.nr_ops = 0;
890 	batch.nr_pages = 0;
891 
892 	for (i = 0; i < copy.count; i++) {
893 		struct gntdev_grant_copy_segment seg;
894 
895 		if (copy_from_user(&seg, &copy.segments[i], sizeof(seg))) {
896 			ret = -EFAULT;
897 			goto out;
898 		}
899 
900 		ret = gntdev_grant_copy_seg(&batch, &seg, &copy.segments[i].status);
901 		if (ret < 0)
902 			goto out;
903 
904 		cond_resched();
905 	}
906 	if (batch.nr_ops)
907 		ret = gntdev_copy(&batch);
908 	return ret;
909 
910   out:
911 	gntdev_put_pages(&batch);
912 	return ret;
913 }
914 
915 static long gntdev_ioctl(struct file *flip,
916 			 unsigned int cmd, unsigned long arg)
917 {
918 	struct gntdev_priv *priv = flip->private_data;
919 	void __user *ptr = (void __user *)arg;
920 
921 	switch (cmd) {
922 	case IOCTL_GNTDEV_MAP_GRANT_REF:
923 		return gntdev_ioctl_map_grant_ref(priv, ptr);
924 
925 	case IOCTL_GNTDEV_UNMAP_GRANT_REF:
926 		return gntdev_ioctl_unmap_grant_ref(priv, ptr);
927 
928 	case IOCTL_GNTDEV_GET_OFFSET_FOR_VADDR:
929 		return gntdev_ioctl_get_offset_for_vaddr(priv, ptr);
930 
931 	case IOCTL_GNTDEV_SET_UNMAP_NOTIFY:
932 		return gntdev_ioctl_notify(priv, ptr);
933 
934 	case IOCTL_GNTDEV_GRANT_COPY:
935 		return gntdev_ioctl_grant_copy(priv, ptr);
936 
937 #ifdef CONFIG_XEN_GNTDEV_DMABUF
938 	case IOCTL_GNTDEV_DMABUF_EXP_FROM_REFS:
939 		return gntdev_ioctl_dmabuf_exp_from_refs(priv, use_ptemod, ptr);
940 
941 	case IOCTL_GNTDEV_DMABUF_EXP_WAIT_RELEASED:
942 		return gntdev_ioctl_dmabuf_exp_wait_released(priv, ptr);
943 
944 	case IOCTL_GNTDEV_DMABUF_IMP_TO_REFS:
945 		return gntdev_ioctl_dmabuf_imp_to_refs(priv, ptr);
946 
947 	case IOCTL_GNTDEV_DMABUF_IMP_RELEASE:
948 		return gntdev_ioctl_dmabuf_imp_release(priv, ptr);
949 #endif
950 
951 	default:
952 		pr_debug("priv %p, unknown cmd %x\n", priv, cmd);
953 		return -ENOIOCTLCMD;
954 	}
955 
956 	return 0;
957 }
958 
959 static int gntdev_mmap(struct file *flip, struct vm_area_struct *vma)
960 {
961 	struct gntdev_priv *priv = flip->private_data;
962 	int index = vma->vm_pgoff;
963 	int count = vma_pages(vma);
964 	struct gntdev_grant_map *map;
965 	int err = -EINVAL;
966 
967 	if ((vma->vm_flags & VM_WRITE) && !(vma->vm_flags & VM_SHARED))
968 		return -EINVAL;
969 
970 	pr_debug("map %d+%d at %lx (pgoff %lx)\n",
971 			index, count, vma->vm_start, vma->vm_pgoff);
972 
973 	mutex_lock(&priv->lock);
974 	map = gntdev_find_map_index(priv, index, count);
975 	if (!map)
976 		goto unlock_out;
977 	if (use_ptemod && map->vma)
978 		goto unlock_out;
979 	refcount_inc(&map->users);
980 
981 	vma->vm_ops = &gntdev_vmops;
982 
983 	vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP | VM_MIXEDMAP;
984 
985 	if (use_ptemod)
986 		vma->vm_flags |= VM_DONTCOPY;
987 
988 	vma->vm_private_data = map;
989 	if (map->flags) {
990 		if ((vma->vm_flags & VM_WRITE) &&
991 				(map->flags & GNTMAP_readonly))
992 			goto out_unlock_put;
993 	} else {
994 		map->flags = GNTMAP_host_map;
995 		if (!(vma->vm_flags & VM_WRITE))
996 			map->flags |= GNTMAP_readonly;
997 	}
998 
999 	if (use_ptemod) {
1000 		map->vma = vma;
1001 		err = mmu_interval_notifier_insert_locked(
1002 			&map->notifier, vma->vm_mm, vma->vm_start,
1003 			vma->vm_end - vma->vm_start, &gntdev_mmu_ops);
1004 		if (err)
1005 			goto out_unlock_put;
1006 	}
1007 	mutex_unlock(&priv->lock);
1008 
1009 	/*
1010 	 * gntdev takes the address of the PTE in find_grant_ptes() and passes
1011 	 * it to the hypervisor in gntdev_map_grant_pages(). The purpose of
1012 	 * the notifier is to prevent the hypervisor pointer to the PTE from
1013 	 * going stale.
1014 	 *
1015 	 * Since this vma's mappings can't be touched without the mmap_sem,
1016 	 * and we are holding it now, there is no need for the notifier_range
1017 	 * locking pattern.
1018 	 */
1019 	mmu_interval_read_begin(&map->notifier);
1020 
1021 	if (use_ptemod) {
1022 		map->pages_vm_start = vma->vm_start;
1023 		err = apply_to_page_range(vma->vm_mm, vma->vm_start,
1024 					  vma->vm_end - vma->vm_start,
1025 					  find_grant_ptes, map);
1026 		if (err) {
1027 			pr_warn("find_grant_ptes() failure.\n");
1028 			goto out_put_map;
1029 		}
1030 	}
1031 
1032 	err = gntdev_map_grant_pages(map);
1033 	if (err)
1034 		goto out_put_map;
1035 
1036 	if (!use_ptemod) {
1037 		err = vm_map_pages_zero(vma, map->pages, map->count);
1038 		if (err)
1039 			goto out_put_map;
1040 	} else {
1041 #ifdef CONFIG_X86
1042 		/*
1043 		 * If the PTEs were not made special by the grant map
1044 		 * hypercall, do so here.
1045 		 *
1046 		 * This is racy since the mapping is already visible
1047 		 * to userspace but userspace should be well-behaved
1048 		 * enough to not touch it until the mmap() call
1049 		 * returns.
1050 		 */
1051 		if (!xen_feature(XENFEAT_gnttab_map_avail_bits)) {
1052 			apply_to_page_range(vma->vm_mm, vma->vm_start,
1053 					    vma->vm_end - vma->vm_start,
1054 					    set_grant_ptes_as_special, NULL);
1055 		}
1056 #endif
1057 	}
1058 
1059 	return 0;
1060 
1061 unlock_out:
1062 	mutex_unlock(&priv->lock);
1063 	return err;
1064 
1065 out_unlock_put:
1066 	mutex_unlock(&priv->lock);
1067 out_put_map:
1068 	if (use_ptemod) {
1069 		unmap_grant_pages(map, 0, map->count);
1070 		if (map->vma) {
1071 			mmu_interval_notifier_remove(&map->notifier);
1072 			map->vma = NULL;
1073 		}
1074 	}
1075 	gntdev_put_map(priv, map);
1076 	return err;
1077 }
1078 
1079 static const struct file_operations gntdev_fops = {
1080 	.owner = THIS_MODULE,
1081 	.open = gntdev_open,
1082 	.release = gntdev_release,
1083 	.mmap = gntdev_mmap,
1084 	.unlocked_ioctl = gntdev_ioctl
1085 };
1086 
1087 static struct miscdevice gntdev_miscdev = {
1088 	.minor        = MISC_DYNAMIC_MINOR,
1089 	.name         = "xen/gntdev",
1090 	.fops         = &gntdev_fops,
1091 };
1092 
1093 /* ------------------------------------------------------------------ */
1094 
1095 static int __init gntdev_init(void)
1096 {
1097 	int err;
1098 
1099 	if (!xen_domain())
1100 		return -ENODEV;
1101 
1102 	use_ptemod = !xen_feature(XENFEAT_auto_translated_physmap);
1103 
1104 	err = misc_register(&gntdev_miscdev);
1105 	if (err != 0) {
1106 		pr_err("Could not register gntdev device\n");
1107 		return err;
1108 	}
1109 	return 0;
1110 }
1111 
1112 static void __exit gntdev_exit(void)
1113 {
1114 	misc_deregister(&gntdev_miscdev);
1115 }
1116 
1117 module_init(gntdev_init);
1118 module_exit(gntdev_exit);
1119 
1120 /* ------------------------------------------------------------------ */
1121